BLOG POST
AI 日报
Reddit Mods Sued by YouTuber Ethan Klein Fight Efforts to Unmask Them
Klein has attempted to subpoena Discord and Reddit for information that would reveal the identity of moderators of a subreddit critical of him. The moderators' lawyers fear their clients will be physically attacked if the subpoenas go through.
Landlords Demand Tenants’ Workplace Logins to Scrape Their Paystubs
Screenshots shared with 404 Media show tenant screening services ApproveShield and Argyle taking much more data than they need. “Opt-out means no housing.”
How Ruby Went Off the Rails
What happened to RubyGems, Bundler, and the Open Source drama that controls the internet infrastructure.
Anthropic Launches Claude Sonnet 4.5
The model was released under the vendor's AI Safety Level 3 protection, which helps prevent dangerous inputs and outputs.
Industrial Robot Demand Doubled in 10 Years
China is dominating robotic installations, while European countries are beginning to lag behind.
Huawei details open-source AI development roadmap at Huawei Connect 2025
Open-source AI development took centre stage at Huawei Connect 2025 last week, with Huawei laying out implementation timelines and the technical specifics around making its entire AI software stack publicly available by year-end. The announcements came with context that matters to developers: frank acknowledgement of past friction, specific commitments about what components will be released, […]
The post Huawei details open-source AI development roadmap at Huawei Connect 2025 appeared first on AI News.
Users Can Shop From Etsy and Shopify in ChatGPT as OpenAI Launches New Agentic Commerce Protocol
Agentic Commerce Protocol (ACP), an open standard co-developed with Stripe, allows programmatic commerce flows between buyers, AI agents, and businesses.
The post Users Can Shop From Etsy and Shopify in ChatGPT as OpenAI Launches New Agentic Commerce Protocol appeared first on Analytics India Magazine.
Anthropic Launches Claude Sonnet 4.5, Touts It as ‘Best Coding Model in the World’
Claude Sonnet 4.5 achieved top scores on the SWE-bench Verified evaluation, which tests real-world software coding skills.
The post Anthropic Launches Claude Sonnet 4.5, Touts It as ‘Best Coding Model in the World’ appeared first on Analytics India Magazine.
Microsoft Brings ‘Vibe Working’ to 365 Copilot With Agent Mode and Office Agent
Office Agent in Copilot chat, powered by Anthropic models, brings presentation and document creation into a chat-first interface.
The post Microsoft Brings ‘Vibe Working’ to 365 Copilot With Agent Mode and Office Agent appeared first on Analytics India Magazine.
Lovable Introduces Cloud and AI Features to Build Full-Stack Apps
Lovable AI, powered by Google’s Gemini models, allows users to add artificial intelligence features to their apps without setup, API keys, or separate billing.
The post Lovable Introduces Cloud and AI Features to Build Full-Stack Apps appeared first on Analytics India Magazine.
NIT Rourkela Patents AI Model to Boost Road Safety Through Vehicle-to-Vehicle Communication
This could pave way for safer roads, smarter traffic management and the future of autonomous mobility in India.
The post NIT Rourkela Patents AI Model to Boost Road Safety Through Vehicle-to-Vehicle Communication appeared first on Analytics India Magazine.
Chennai Startup Agnikul Reveals Plans for its Fully Reusable Rockets
The company said it will ensure no part of its rockets is left behind or expended.
The post Chennai Startup Agnikul Reveals Plans for its Fully Reusable Rockets appeared first on Analytics India Magazine.
WhatsApp Vs Arattai – Can Made in India Messenger Win?
India’s swadeshi tech push faces its biggest challenge: can homegrown apps like Aratai rival WhatsApp? This debate explores patriotic sentiment, privacy concerns, and network effects. Featuring ministerial endorsements and Zoho’s swadeshi success, it examines whether India can build globally competitive digital products while riding the Make in India wave.
The post WhatsApp Vs Arattai – Can Made in India Messenger Win? appeared first on Analytics India Magazine.
Sahamati Labs, Google Cloud to Bring AI into India’s Account Aggregator Network
The partnership will use generative AI to build predictive insights, personalised recommendations, and fraud detection systems.
The post Sahamati Labs, Google Cloud to Bring AI into India’s Account Aggregator Network appeared first on Analytics India Magazine.
How Pure Storage India R&D Centre Built Pure KVA
Pure KVA, one of Pure Storage’s flagship innovations from India, helps enterprises reduce AI infrastructure costs by optimising GPU usage.
The post How Pure Storage India R&D Centre Built Pure KVA appeared first on Analytics India Magazine.
Spencer Kimball on How CockroachDB Will Power Efficient AI Agents
Spencer Kimball sees AI agents outnumber humans and CockroachDB is building the database backbone to handle that.
The post Spencer Kimball on How CockroachDB Will Power Efficient AI Agents appeared first on Analytics India Magazine.
Anthropic says its new AI model “maintained focus” for 30 hours on multistep tasks
Latest Claude model beats OpenAI and Google on coding tests.
Sports piracy site Streameast returns after US government let domain expire
ICE's HSI branch shut down Streameast in 2024.
EA will be a very different company under private ownership
Analysts see trouble for studios like BioWare, potential for Saudi meddling.
F-Droid says Google’s new sideloading restrictions will kill the project
F-Droid sees Google's plans as a threat to FOSS apps.
ESA will pay an Italian company nearly $50 million to design a mini-Starship
This is a "step forward in the transition of European launch systems toward full reusability."
Senators try to halt shuttle move, saying “little evidence” of public demand
"Removing an item from the National Collection is not a viable solution..."
Fortnite disables Peacemaker emote that might resemble a swastika
Epic will "inquire into our partner's creative intentions" before making a final decision.
Taiwan pressured to move 50% of chip production to US or lose protection
Trump official plots “impossible” deal moving Taiwan’s chip supply chain into US.
It’s official: EA is selling to private equity in $55 billion deal
Investors will receive $210 per share, well above the stock's all-time high.
Big AI firms pump money into world models as LLM advances slow
World models aim to navigate the physical world by learning from videos, robotic data.
ZR1, GTD, and America’s new Nürburgring war
Ford and Chevy set near-identical lap times with very different cars; we drove both.
Will AI make cheaper personalized insurance premiums possible? Here’s why it’s a slippery slope
In this age of algorithms, personalized insurance is encountering legal limits, especially the ban on discrimination. Can we balance accurate pricing, solidarity, and fairness?
Nvidia Is Quaking in Its Boots
"We're up against a formidable, innovative, hungry, fast-moving, underregulated [competitor]."
The post Nvidia Is Quaking in Its Boots appeared first on Futurism.
Trump’s Grip on Reality Questioned After He Shares and Then Deletes Bizarre AI-Generated Video
Is the president confused?
The post Trump’s Grip on Reality Questioned After He Shares and Then Deletes Bizarre AI-Generated Video appeared first on Futurism.
OpenAI Is Preparing to Launch a Social App for AI-Generated Videos
The platform appears to closely resemble TikTok and is powered by Sora 2, OpenAI's latest video generation model.
Marissa Mayer Is Dissolving Her Sunshine Startup Lab
After seven rocky years, the company’s assets will be sold to Dazzle, a new AI firm that Mayer founded.
Building a Streamlit Python UI for LLaVA with OpenAI API Integration
Table of Contents Building a Streamlit Python UI for LLaVA with OpenAI API Integration Why Streamlit Python for Multimodal Apps? What Is Streamlit Python? The Streamlit Python-Based UI We Will Build in This Lesson Why Not FastAPI or Django? Configuring…
The post Building a Streamlit Python UI for LLaVA with OpenAI API Integration appeared first on PyImageSearch.
Stripe Joins With OpenAI to Allow Etsy Purchases Via ChatGPT
The payment-processing company Stripe Inc. is working with OpenAI to allow US shoppers to buy products through ChatGPT, the popular artificial-intelligence tool.
Google to Pay $24.5 Million to End Trump Jan. 6 Ban Lawsuit
Alphabet Inc.’s Google agreed to pay $24.5 million to resolve Donald Trump’s claims that being banished from his YouTube channel after the Jan. 6, 2021, riot at the US Capitol was illegal censorship, according to a court filing.
Verizon Is in Talks to Purchase EchoStar Wireless Spectrum
Verizon Communications Inc. is in discussions with EchoStar Corp. about purchasing some of its wireless spectrum, according to people familiar with the matter.
AI Data Centers Are Sending Power Bills Soaring
Wholesale electricity costs as much as 267% more than it did five years ago in areas near data centers. That’s being passed on to customers.
S&P Says It Expects To Cut Electronic Arts to Junk Status
S&P Global Ratings said it expects to cut Electronic Arts Inc.’s credit grades to junk status after the video-game maker’s buyout deal is closed, citing the expected addition of about $20 billion of debt from the transaction.
Pressure Grows on SingTel as Optus Emergency Call Blunders Mount
As Optus reels from another emergency call outage, the crisis has spread beyond Australia’s borders. The escalating reputational damage to the nation’s second-largest phone company is now bearing down on its parent Singapore Telecommunications Ltd.
Robinhood Shares Climb 12% on Growth in Prediction-Market Trades
Robinhood Markets Inc. shares rose over 12% after Chief Executive Officer Vladimir Tenev said the online-stock brokerage has made strong inroads into prediction markets, where people wager on elections, sporting matches and other widely followed events.
EA Agrees to Largest Leveraged Buyout in History | Bloomberg Tech 9/29/2025
Bloomberg’s Caroline Hyde and Ed Ludlow discuss Electronic Arts’ agreement to be sold in a record-breaking leveraged buyout. Plus, Huawei plans to ramp up production of its most advanced AI chips in a bid to take market share from Nvidia. And, Peloton is preparing to launch new AI software in its latest bid to revive the struggling fitness firm. (Source: Bloomberg)
Saudi Plans for Video-Game Hub Grow With $55 Billion EA Deal
Saudi Arabia is accelerating plans to transform itself into a hub for gamers with its blockbuster deal to take Electronic Arts Inc. private.
EA Agrees to $55 Billion Sale in Largest Leveraged Buyout on Record
Electronic Arts Inc. has agreed to sell itself in the largest leveraged buyout on record to a group of investors that includes a firm managed by President Donald Trump’s son-in-law Jared Kushner and Saudi Arabia’s sovereign wealth fund.
US Hits Sanctioned Firms’ Subsidiaries, Drawing Chinese Rebuke
The Trump administration is dramatically expanding US sanctions to capture subsidiaries of blacklisted companies — a crackdown that drew a swift rebuke from China, where key tech giants are already subject to stringent American trade curbs.
Anthropic Says New Model Can Code on Its Own for 30 Hours Straight
Anthropic is releasing a new artificial intelligence model that is designed to code longer and more effectively than prior versions, its latest attempt to stay ahead of rivals like OpenAI in offering tools for software developers. The new model, called Claude Sonnet 4.5, is better at following instructions and can code on its own for up to 30 hours straight, the company said on Monday. By comparison, a previous model called Claude Opus 4 is said to be able to field coding tasks for up to seven h
Walmsley’s Dream Hire Badly Needs to Find Some Blockbuster Drugs
Nearly nine years ago, when Emma Walmsley was preparing to take over as CEO of GSK Plc, she wrote in her diary that Luke Miels — then an executive at rival AstraZeneca Plc — would be “a dream appointment for this adventure.”
Black Forest Labs, AI Image Partner for Meta, in Talks for Funding at $4 Billion Valuation
German artificial intelligence startup Black Forest Labs is in discussions to raise funding that could boost its valuation to around $4 billion, according to people familiar with the matter.
Blue Owl, xAI Bets Boost Qatar Fund’s $500 Billion US Pledge
The $524 billion Qatar Investment Authority is stepping up its investments in the US with high-profile commitments to artificial intelligence and digital infrastructure, underscoring how Doha is keeping its economic priorities separate from geopolitical tensions in the Middle East.
Asahi Suspends Some Beverage Shipments After Hit by Cyberattack
Asahi Group Holdings Ltd. suspended some beverage shipments on Monday after a cyberattack hobbled some of the drinks company’s operations in Japan.
ArcelorMittal May Get $491 Million IDC South Africa Unit Bid
South Africa’s top development finance institution is working with financial advisers on a possible bid of about 8.5 billion rand ($491 million) for control of ArcelorMittal SA’s business in the country, people familiar with the situation said.
Comcast Promotes Michael Cavanagh to Co-Chief Executive Officer
Comcast Corp. named Michael Cavanagh as co-chief executive officer alongside Brian Roberts on Monday and added him to the company’s board of directors.
Pimco CEO Roman on AI Financing, Private Markets, Fixed Income
Manny Roman, CEO at Pacific Investment Management Co. (Pimco), discusses investing in AI infrastructure, global opportunities in fixed income, and why tariff uncertainties pose a risk to US stocks. He spoke Monday on “Bloomberg Surveillance.” (Source: Bloomberg)
Electric Aircraft Firm Beta Technologies Files for US IPO
Beta Technologies Inc., an electric aircraft producer, filed for a US initial public offering, disclosing growing revenue and widening losses.
Netanyahu Discusses AI in Private Meeting With Tech Investors
Israeli Prime Minister Benjamin Netanyahu met with US tech executives and investors in New York on Sunday to discuss the prospects for using artificial intelligence to boost his country’s economy and military capabilities, according to people familiar with the matter.
Electronic Arts Said to Near Buyout Deal at About $210 Per Share
Video game maker Electronic Arts Inc. is in advanced talks to go private at about $210 a share, according to people familiar with the matter.
DeepSeek Debuts ‘Sparse Attention’ Method in Next-Gen AI Model
DeepSeek updated an experimental AI model Monday in what it called a step toward next-generation artificial intelligence.
Fed’s Waller Says Technology Can Deliver More Choice in Payments
US regulators ought to welcome the ways new technologies could benefit financial firms as well as consumers, while imposing guardrails that promote safety, Federal Reserve Governor Christopher Waller said.
OpenAI Floats the Promise of a More Useful Personal AI
Welcome to Tech In Depth, our daily newsletter about the business of tech from Bloomberg’s journalists around the world. Today, Austin Carr looks at OpenAI’s effort to offer ChatGPT users a personal daily briefing — and all-knowing assistant.
US to See $350 Billion Nuclear Boom to Power AI, Report Says
Soaring demand for electricity will drive a $350 billion nuclear spending boom in the US, boosting output from reactors by 63% by 2050, according to Bloomberg Intelligence.
Peloton Is Set to Kick Off Turnaround with Hardware, AI Product Launch
Peloton Interactive Inc. is preparing to kick off its second comeback attempt in three years, following a previous effort that fell short of expectations and lacked any major consumer-facing upgrades.
Xiaomi Says New IPhone Rival Outpacing Previous Model at Launch
Xiaomi Corp. said it’s confident its new iPhone rival will exceed the sales of its previous model, saying the phone saw record demand immediately after its debut on Saturday.
OpenAI Launches Parental Controls for ChatGPT After Teen’s Death
OpenAI is launching parental controls for ChatGPT, which it announced following a lawsuit alleging a teenager who died by suicide this spring relied on the popular chatbot as a coach.
YouTube to pay $22 million to Trump to settle post-January 6 ban case
In total, YouTube will pay $24.5 million to settle the suit brought by President Donald Trump.
EA's boss won't have to answer to Wall Street. That doesn't mean the pressure's over.
Andrew Wilson has helmed the video-game company since 2013 and will remain CEO after the deal closes.
US Navy techs are stripping parts from fighter jets and subs to keep other ones working
Long delays to get parts for F/A-18 fighters and Virginia-class submarines have forced maintainers to strip from other aircraft and subs.
I made Ina Garten's 'outrageous brownies,' the most popular dessert at her famous store. They tasted incredible.
Ina Garten said these "outrageous brownies" flew off the shelves at her famous Barefoot Contessa store. Her easy recipe is the perfect dessert.
Move over, doodles. There's a new 'It' dog in town.
Dachshunds could break into the American Kennel Club's top five most popular breeds this year — and weiner-dog influencers might help them do it.
How to get Ariana Grande tickets: Chicago, Atlanta, Los Angeles, and more
Though it won't kick off until the summer of 2026, Ariana Grande's coming tour is already sold out. Here's where to find tickets from resellers.
I paid $300 to spend 2 nights in an Airstream trailer for the first time. I'll never do it again in a city.
I stayed in a luxury Airstream trailer in Vienna. It had a bathtub and a comfy bed, but it was also noisy and lacked storage hacks.
Charlie Javice sentenced to 7 years in prison for JPMorgan Chase fraud
Charlie Javice got a 7-year prison sentence for defrauding JPMorgan Chase in the $175M sale of her fintech startup Frank.
The Army and Marine Corps are paying more to repair vehicles that are less ready to fight, watchdog says
Costs for vehicle repairs have soared, but the services are getting less availability and mission readiness in return.
Vintage photos show how dangerous railways, mills, and other workplaces used to be
Working in railways, mines, and mills caused thousands of deaths in the early 20th century and before. Photos show the dangerous workplaces.
Microsoft is buying millions of carbon credits. Meet the investor working to make them real.
Big power users like Microsoft and the Mercedes F1 team are buying carbon credits despite the questions surrounding the practice.
The Tiny Team era is here
AI-powered startups are proving the Tiny Team era is here. Five founders and employees share the pros and cons of working alongside AI agents.
I expected the Ryder Cup to be one big frat party. The level of corporate polish caught me by surprise.
There was light rowdiness on Friday — boos, drinks, porta potties — but the contrast to corporate America struck me more than the chaos itself.
What Kamala Harris' style has looked like throughout her career, from suits to sparkling gowns
Kamala Harris usually wears pantsuits, Converse sneakers, and pearl jewelry. Over the years, though, she's incorporated gowns and jeans.
I was sick of spending my travel budget on weddings, so now I turn every invitation into a vacation
Over the past few years, I've been invited to lots of weddings that required me to travel. Now, I like to turn every invite into a mini vacation.
Marissa Mayer's shuttered photo app was just too beautiful for this cruel world
Former Yahoo CEO Marissa Mayer's photo-sharing app never really caught on. So she's reportedly moving on — to a new AI personal assistant.
My family of 5 went on our first cruise. We had fun, but there are 4 things I wish we'd known before getting on the ship.
There are a few things I wish I had known before my first cruise with my family. Next time, I'll turn my phone on airplane mode and pack formal wear.
I ate dinner at one of Ina Garten's favorite restaurants. It was worth every penny.
Nick & Toni's, a favorite of Ina Garten, is a cozy restaurant that offers delicious dishes. You might even spot some A-listers while dining there.
22 musicians we'd love to see perform the Super Bowl halftime show
Bad Bunny is officially the 2026 Super Bowl halftime show headliner. Here are other musicians who deserve a shot, too.
Prologis exec explains why the company has an edge in the red-hot data center space
During Business Insider's "The Resiliency Playbook" event, a Prologis executive explained why the company is primed to develop data centers.
OpenAI Is Preparing to Launch a Social App for AI-Generated Videos
The platform appears to closely resemble TikTok and is powered by Sora 2, OpenAI's latest video generation model.
Marissa Mayer Is Dissolving Her Sunshine Startup Lab
After seven rocky years, the company’s assets will be sold to Dazzle, a new AI firm that Mayer founded.
Are H-1B Changes ‘A Strategic Opening’ For Startup Creation Or An ‘Insurmountable Founder Tax’?
Will the new $100,000 fee for H-1B applications help or hurt the U.S. startup ecosystem? Investors and an immigration attorney who spoke with Crunchbase News were divided over the impact of the changes.
Exclusive: Trucker’s Son Bucks Logistics Funding Decline with $40M Raise For Startup Alvys
Alvys, an AI-powered logistics software provider, has raised $40 million in an RTP Global-led Series B funding, the company tells Crunchbase News exclusively.
Biotech Share Of US Funding Hits Lowest Point In Crunchbase History
The share of U.S. startup investment going to biotech companies has hit the lowest level in more than 20 years, as an ever-growing proportion of funding goes to AI upstarts in other sectors. We take a look at what else we found in a data dive.
Timing The Market Is A Myth: Here’s How To Stay Ahead Of It
Mistimed market entry is one of the top five reasons startups fail, writes guest author Ashish Kakran of Sierra Ventures. While the perfect time to go to market is hard to calculate, he explains how just a slight edge ahead of the pack can make all the difference.
Claude Sonnet 4.5 Is Here
Customers process exabytes of data daily on Databricks, and generative AI is already...
How AI Agents Transform the Healthcare Sector
TL;DR AI agents are autonomous, goal-driven systems that perceive, reason, and act – handling multi-step clinical workflows with human-like adaptability (e.g., diagnosing and offering mental health support) Unlike rule-based automation, […]
The post How AI Agents Transform the Healthcare Sector appeared first on Datafloq.
Disconnected Demographic Data? Here’s How Data Appending Can Fix It
B2B marketing strategies suffer from blind spots due to incomplete demographic data. Decision-makers cannot tailor their approaches effectively without vital client demographic information. When B2B enterprise leaders use poor-quality demographic […]
The post Disconnected Demographic Data? Here’s How Data Appending Can Fix It appeared first on Datafloq.
The Role of Synthetic Data in AI Model Training
Artificial Intelligence (AI) and Machine Learning (ML) are revolutionizing industries with smart automation, predictive analytics, and larger scale AI-enabled decisions, but any AI production deployment is only as good as […]
The post The Role of Synthetic Data in AI Model Training appeared first on Datafloq.
Data is at the Center of Scientific Discovery Inside MIT’s New AI-Powered Platform
AI-powered tools have become more common in scientific research and development, especially for predicting outcomes or suggesting possible experiments using datasets. However, most of these systems only work with limited Read more…
The post Data is at the Center of Scientific Discovery Inside MIT’s New AI-Powered Platform appeared first on BigDATAwire.
Practical ways to use MCP at work
🧠 Market analysis, create design-grade prototypes, test your product’s APIs, build your own Chief of Staff and more…
✨ New on Code-Tree.dev — Discover Suggested Profiles!
We’ve just rolled out a small but mighty new feature on code-tree.dev
: profile suggestions! 🎉
Now, when you visit someone’s Code Tree profile, you’ll see a couple of recommended profiles just below their links — complete with avatars, names, and clickable handles.
At its core, Code Tree is about helping developers and makers showcase what they build, what they love, and what they stand for — in one clean, beautiful profile.
But we’re also about discovery.
- Find other devs building in the same stack.
- Explore profiles with similar tools or interests.
- Get inspired by how others present their work.
Build your profile in minutes, link your projects, tools, and stack — and join a growing community of makers.
Got feedback or feature ideas? We’d love to hear from you! Drop a comment or reach out on GitHub 💬
Happy building!
— The Code Tree team 🌱
Automating AWS Lambda Deployment using GitHub Actions
Why Teams Are Automating Lambda Deployments
With the growing adoption of serverless architectures, AWS Lambda has become a core compute solution for running event-driven workloads. However, manually deploying Lambda functions introduces the risk of inconsistency, downtime, and human error. DevOps teams are increasingly automating these processes using CI/CD pipelines.
GitHub Actions provides a powerful platform to integrate automation directly into the version control system. This empowers developers to trigger deployments automatically on code pushes, PR merges, or manually through workflow dispatches.
Core Components of the Automation Workflow
Preparing Your Lambda Function Code
Organize your function code in a directory structure that’s easy to zip and upload. Make sure to include only necessary dependencies. If your code relies on external Python packages, use a requirements.txt and deploy with dependencies zipped in a package directory.GitHub Actions Workflow File (.github/workflows/deploy.yml)
Create a GitHub Actions workflow YAML file to define your deployment pipeline. A basic Python example looks like this:
name: Deploy Lambda
on:
push:
branches:
- main
jobs:
deploy:
runs-on: ubuntu-latest
steps:
- name: Checkout Code
uses: actions/checkout@v3
- name: Set up Python
uses: actions/setup-python@v4
with:
python-version: '3.9'
- name: Install dependencies
run: |
pip install -r requirements.txt -t package
cd package
zip -r ../function.zip .
cd ..
zip -g function.zip lambda_function.py
- name: Deploy to Lambda
uses: aws-actions/aws-lambda-deploy@v1
with:
aws-access-key-id: ${{ secrets.AWS_ACCESS_KEY_ID }}
aws-secret-access-key: ${{ secrets.AWS_SECRET_ACCESS_KEY }}
aws-region: us-east-1
function-name: myLambdaFunction
zip-file: function.zip
- AWS IAM Permissions and Secrets Configuration Create an IAM user with permissions to update Lambda functions via lambda:UpdateFunctionCode and store its credentials in GitHub Secrets. Make sure to avoid environment variable leakage by reviewing workflow logs with temporary output disabled.
Secret: AWS_ACCESS_KEY_ID
Secret: AWS_SECRET_ACCESS_KEY
Advanced Features for Production Pipelines
Branch Filtering: Deploy only on specific branches like main or release/*.
Workflow Dispatch: Trigger manual deploys using workflow_dispatch:.
Environment Promotion: Deploy to dev, staging, and prod using environment protection rules and matrix builds.
Monitoring: Integrate Slack, Datadog, or Amazon CloudWatch for post-deployment notifications.
Conclusion
Automating AWS Lambda deployments using GitHub Actions leads to faster delivery cycles, reproducible builds, and minimized manual tasks. By defining a clear release workflow, setting up the right permissions, and using environment configurations, engineering teams can streamline serverless development at scale.
CTA: Explore GitHub Actions Marketplace for more Lambda integration tools.
Building a Full Stack Web App from scratch: First Steps
I am Constanza Díaz, Web developer, industrial designer, UX/UI enthusiast, aspiring entrepreneur, and self-taught digital product creator — currently building my own web application and documenting the process.
I've been working as a front end developer for the last 6 years and now I'm planning to develop my own web application named HandyFEM, a digital platform designed to connect women professionals in skilled trades with clients.
I am doing this as a way to learn by doing, as artificial intelligence is evolving so quickly and becoming a must for every developer and digital product creator and owner. At the same time, I will build my own personal project, which I am developing as a business project at Innoemprèn Sostenible, a program for entrepreneurs at Tecnocampus University in Barcelona, Spain.
Building this app will be a good way to create a very complete application where multiple features can be added — this is a very big challenge as a designer-developer-entrepreneur and a true one-woman band, but also a great way to show all my multiple skills in one place :D.
After a few attempts of "vibe coding" without a plan using Vercel v0 and Lovable, I realized my app is too ambitious to be done in only one prompt :D. So I decided to create a professional approach and document it, so that I can learn everything during the process and get feedback from other devs and designers.
I'm documenting every step — from mapping the app structure to experimenting with AI-driven prototyping.
For this project, I will start with Vercel v0 as an AI assistant for generating frontend components, with Supabase as the backend layer. This blog series will serve both as a technical journal and a record of decisions that shape the MVP.
AI tools like Vercel v0 can generate impressive amounts of code quickly. The risk? Losing track of fundamentals if the product lacks a clear structure.
To avoid gaps, I started by outlining the navigation map of HandyFEM: a high-level diagram of screens, user flows, and connections. This ensures that all the important features won’t be overlooked later.
Here’s the first structural map created in FigJam. It defines the scope of the MVP and keeps development aligned:
Caption: HandyFEM Navigation Map (MVP) — Yellow: buttons | Green: components (will develop more detail later).
Key sections:
- Landing Page: entry point with signup, login, and directory access.
- Auth: client and professional signup, plus login.
- Role-Based Dashboards: tailored panels for clients and professionals.
- Public Directory: searchable list of professionals with filters.
- Professional Profiles: detailed pages with contact options.
- Messaging System: list of conversations and direct chat.
Dashboard ≠ Profile
Dashboards serve as control centers. Profiles are just editable data. Keeping them separate opens room for metrics, history, or verification in the future.Roles must be explicit
Defining client vs. professional flows early makes the product scalable without major refactors.Planning is leverage
One diagram clarified priorities and prevented the common trap of coding first and restructuring later.
The next step is translating this map into Jira epics and tasks, preparing a Sprint 0 focused on:
- Setting up Supabase tables and relationships.
- Configuring authentication (email + password).
- Establishing the initial Next.js + Tailwind architecture.
This structured approach allows me to integrate AI tools without sacrificing control or long-term scalability.
👉 This post opens the series on building HandyFEM from scratch. The upcoming entries will cover Jira planning, Sprint 0 setup, and the integration of Vercel v0 into the workflow. Follow the hashtag #HandyFEMapp if you don’t want to miss the progress!
📎 For those who want to dive deeper or leave comments on my board, the full navigation map is available here: Open FigJam Board
🔗 Previous: none (this is the first post)
🔗 Next: Coming soon: From Diagrams to Tickets (Organizing HandyFEM with Jira)
Zero to Cloud: 12 Weeks of AWS re/Start + Free AWS CCP Exam
I graduated on the AWS re/start program under Ajira Digital as part of Cohort 12, where we had an intensive 12-week program that gave the basics to those interested in joining cloud and nailing the practitioner examinations. Here is all you need to know about the program.
AWS re/Start is a free, full-time training program created by Amazon Web Services to help individuals from diverse backgrounds launch careers in the cloud.
The program runs across the African continent through regional partners (with Ajira serving as the delivery partner in Kenya). In just three months, the learners would have gone through an intensive mix of:
1.Cloud Fundamentals — AWS services, security, databases, networking.
2.Technical Skills — Linux, Python scripting, version control.
3.Professional Skills — communication, teamwork, resume coaching, and interview preparation.
- Certification Prep — structured guidance for the AWS Certified Cloud Practitioner exam.
The program is designed for anyone curious about cloud technology, whether you’re a recent graduate, career switcher, or someone looking to upskill.
The 12 weeks are highly structured and immersive. A typical week includes:
- Morning sessions for theory and new concepts.
- Hands-on labs where you practice AWS services in real time.
- Afternoon sessions for deep dives and collaborative problem-solving.
- Career labs dedicated to employability training — CV building, interview prep, and workplace skills.
The schedule is full-time, Monday to Friday, and demands complete commitment. After the 12 weeks, the aws practitioner examination is waived 100 % which is an extra bonus.
To get into the program, here is a pathway I followed:
- Apply via the AWS re/Start portal or through a local partner (Ajira in Kenya, other organizations across Africa).
- Demonstrate motivation and availability — you’ll need to commit to all 12 weeks, full-time.
- Complete the selection process, which includes interviews and assessments by the training partner.
- Join a cohort — Here, you are immersed in a community of learners guided by experienced trainers.
Why AWS re/Start Matters
Cloud adoption across Africa is accelerating, but the demand for skilled professionals far outpaces supply. AWS re/Start directly addresses this gap by giving learners not just knowledge, but also career-readiness and connections to employers.
As a Cohort 12 graduate, I can confidently say AWS re/Start is more than a training program, it is a career launchpad. If you’re passionate about tech and looking for a way into cloud computing, this program is worth considering. Applications run periodically, so check the Ajira AWS re/Start program linkedin page and see if the next cohort could be your gateway to the cloud.
CSS Pseudo-Elements: Enhance Your Projects with Style!
This article explores CSS pseudo-elements, which enable easy styling and content modification without additional HTML elements, and how I used them to create eyes for the ghosts in my JS Pac-Man game!
CSS Before & After Pseudo-Elements
CSS pseudo-elements allow you to insert content before or after an element's actual content without adding extra HTML elements. The ::before and ::after pseudo-elements are popular and can help you style and modify content easily.
Some of the other commonly used pseudo-elements include:
::first-line: Applies styles to the first line of a block-level element.::first-letter: Styles the first letter of a block-level element.::selection: Styles the portion of an element that is selected by the user.::placeholder: Styles the placeholder text in an input or textarea element.::marker: Styles the marker box of a list item, such as bullets or numbers.
⚠ Important: Pseudo-elements can't be targeted in JavaScript as they aren't part of the DOM. They serve decorative purposes, but you can indirectly manipulate their styles by altering their parent elements' styles or modifying CSS rules dynamically.
Why Is This a Valuable Coding Concept?
Understanding and utilizing CSS pseudo-elements is valuable because they enhance styling capabilities by allowing you to add decorative content and style specific parts of an element without altering the HTML structure. This leads to improved code maintainability, as it keeps HTML clean and semantic, reducing the need for additional markup.
JS Pac-Man Project Links:
Note: Since I’m following team development practices and pushing updates to the develop branch, some of the changes I describe in this article may not be live in the deployed version of the game yet.
Ghost’s Eye Direction
This article explains how I used CSS pseudo-elements to add eyes to the ghosts in my JavaScript Pac-Man game and simulate their eye movements!
For my JavaScript Pac-Man game, I created each ghost's eye using the before and after CSS pseudo-elements: the before pseudo-element for the left eye and the after pseudo-element for the right eye. Each eye is simply a boxed div with a 50% border radius, a black background, and a thick white border.
Taking creative liberty, I reposition the eyes to create the effect, differentiating it from the original arcade, where the pupils move. In my game, the left and right eyes remain grouped, and I slightly shift them in the direction they are moving. While in action, I believe it is pretty convincing!
First, I created a main ghost class. Using a black background, I rounded out the ghost’s head with the top left and right border radius, and I used a clip-path to create the jagged bottom of the ghost.
.ghost {
border: 1px solid white;
background-color: black;
position: relative;
width: 200px;
height: 200px;
margin:10px;
border-top-left-radius: 50%;
border-top-right-radius: 50%;
clip-path: polygon(50% 0%, 100% 0, 100% 100%, 81% 91%, 66% 100%, 49% 91%, 31% 100%, 15% 91%, 0 100%, 0 0);
}
Next, I create the ghost’s left eye using the ::before CSS pseudo-element. I create a box and use a thick white border with a black background, rounding it out with a 50% border radius to complete the effect. The eye is positioned absolutely, sitting 100 pixels from the bottom and slightly shifted to the right with a margin, giving it the appearance of a floating eye on the ghost.
.ghost::before {
border: 12px solid white;
border-radius: 50%;
background-color: black;
content: "";
width: 30px;
height: 30px;
position: absolute;
bottom: 100px;
margin: 0 0 0 50px;
}
For the ghost's right eye, the ::after CSS pseudo-element is used. It mirrors the left eye by creating another circular shape with a thick white border and a black background, using a 50% border radius for the round look. This eye is also positioned absolutely, 100 pixels from the bottom, but it's shifted further to the right with a margin, making it appear as the ghost's second floating eye.
.ghost::after {
border: 12px solid white;
border-radius: 50%;
background-color: black;
content: "";
width: 30px;
height: 30px;
position: absolute;
bottom: 100px;
margin: 0 0 0 100px;
}
Ghosts’ Colors
I created a separate class for each ghost color so that I can easily use them in conjunction with the main ghost class.
/* Ghost colors */
.ghost-red {
background-color: rgb(255,12,14);
}
.ghost-pink {
background-color: rgb(255,165,166);
}
.ghost-blue {
background-color: rgb(147,255,255);
}
.ghost-orange {
background-color: rgb(253,123,13);
}
Scared Ghost Color
For a scared ghost, I changed the color of the ghost to blue and the eyes to white.
/* Scared */
.scared {
background-color: blue;
}
.scared::before,
.scared::after {
background-color: white;
}
Combining Ghost Classes
Now that there are classes for a main ghost, a ghost color, and a scared ghost, I can combine them to create a single ghost with the ability to change its eye direction!
<div class="flex">
<div class="ghost ghost-red ghost-look-left"></div>
<div class="ghost ghost-blue ghost-look-up"></div>
<div class="ghost ghost-pink ghost-look-down"></div>
<div class="ghost ghost-orange ghost-look-right"></div>
<div class="ghost scared"></div>
<div class="ghost"></div>
</div>
JavaScript Ghost Eye Direction
In my JavaScript Pac-Man game, I created a Ghost class that includes an eyes property and other properties essential for the ghost's functionality.
Within the function that moves the ghosts, there is code to determine a valid direction—left, right, down, or up—and assign it. Once a direction is assigned, I update the CSS ghost's eye direction class by removing the current class and applying the new one.
// Eye direction
if (direction === -1) {
squares[ghost.currentIndex].classList.remove(ghost.eyes);
ghost.eyes = `ghost-look-left-${eyeSize}`;
} else if (direction === 1) {
squares[ghost.currentIndex].classList.remove(ghost.eyes);
ghost.eyes = `ghost-look-right-${eyeSize}`;
} else if (direction === 28) {
squares[ghost.currentIndex].classList.remove(ghost.eyes);
ghost.eyes = `ghost-look-down-${eyeSize}`;
} else if (direction === -28) {
squares[ghost.currentIndex].classList.remove(ghost.eyes);
ghost.eyes = `ghost-look-up-${eyeSize}`;
}
To make trial-and-error creation easier, I initially styled all the ghosts' eye directions on CodePen, using a ghost size of 200px for both width and height. In my game, the size of the ghost is either 20px or 10px in width and height, depending on full-screen or mobile view. With GitHub's CoPilot, I was able to easily resize the CSS values by letting the AI handle the calculations.
My other related articles
What Is “State” in JavaScript, and How to Use It in Your Projects
Mastering JavaScript Fundamentals: Unleashing Your Framework Readiness
Mastering JavaScript: The Power of forEach and Ternary Operators
JavaScript: Understanding the Set and Clear Interval Methods
Be sure to listen to the HTML All The Things Podcast!
📝 I also write articles for the HTML All The Things Podcast, which you can read on their website: https://www.htmlallthethings.com/.
Be sure to check out HTML All The Things on socials!
Affiliate & Discount Links!
With CodeMonkey, learning can be all fun and games! CodeMonkey transforms education into an engaging experience, enabling children to evolve from tech consumers to creators. Use CodeMonkey's FREE trial to unlock the incredible potential of young tech creators!
With a structured learning path tailored for various age groups, kids progress from block coding to more advanced topics like data science and artificial intelligence, using languages such as CoffeeScript and Python. The platform includes features for parents and teachers to track progress, making integrating coding into home and classroom settings easy.
Through fun games, hands-on projects, and community interaction, CodeMonkey helps young learners build teamwork skills and receive recognition for their achievements. It fosters a love for coding and prepares children for future career opportunities in an ever-evolving tech landscape.
To learn more about CodeMonkey, you can read my detailed review article!
Affiliate Links:
Advance your career with a 20% discount on Scrimba Pro using this affiliate link!
Become a hireable developer with Scrimba Pro! Discover a world of coding knowledge with full access to all courses, hands-on projects, and a vibrant community. You can read my article to learn more about my exceptional experiences with Scrimba and how it helps many become confident, well-prepared web developers!
Important: This discount is for new accounts only. If a higher discount is currently available, it will be applied automatically.
How to Claim Your Discount:
Click the link to explore the new Scrimba 2.0.
Create a new account.
Upgrade to Pro; the 20% discount will automatically apply.
Disclosure: This article contains affiliate links. I will earn a commission from any purchases made through these links at no extra cost to you. Your support helps me continue creating valuable content. Thank you!
Conclusion
CSS pseudo-elements are great for styling projects without the need for additional HTML. You can easily style a first letter or line with pseudo-elements, markers like bullets, placeholder text, entire sections, and more. I used the before and after pseudo-elements for enhanced styling in my project.
Additional benefits include improved code maintainability by keeping HTML clean and semantic, performance optimization as the browser efficiently renders them, the ability to create responsive designs that adapt to various screen sizes, and the potential for dynamic visual effects through CSS transitions and animations.
In my Pac-Man game, I stylistically designed pseudo-element classes to represent the eyes of the ghosts looking in the left, right, up, and down directions. Using JavaScript, I change the direction of the ghosts' eyes based on their movement, producing a realistic effect that makes it seem like their eyes are looking in the direction they are moving.
Can you think of other unique styling ideas for your projects that utilize pseudo-elements? Experiment with them to unleash your creativity and transform your designs into something remarkable!
Let’s connect! I’m active on LinkedIn and Twitter.
Do you now feel confident in using CSS pseudo-elements to enhance your projects? Have you tried creating unique styles with them in your own work? Please share the article and comment!
Adam Savage's Tested: Advice on Weathering Today's Uncertain, Hostile Times
TL;DR:
Adam Savage says the best way to navigate these uncertain, often hostile times is with empathy, kindness and respect—values he wouldn’t change in his own past because every experience shaped who he is today.
He explains the “scarcity model” as the mindset that resources or opportunities are limited (which just fuels needless competition), and draws a clear line between gatekeeping—unfairly shutting people out—and asking folks to “pay their dues,” i.e. genuinely earn skills and prove competence before moving on to the next level.
Watch on YouTube
KEXP: Cory Hanson - Texas Weather (Live on KEXP)
Cory Hanson Brings “Texas Weather” to KEXP’s Studio
Cory Hanson delivers a scorching live version of “Texas Weather” at KEXP, recorded on August 4, 2025. Backed by Evan Backer on drums and Hazel Rigby on bass, Hanson handles guitar, piano, and vocals with that signature blend of raw energy and introspective flair.
Hosted by Jewel Loree, the session was engineered by Julian Martlew and Jon Roberts, mastered by Matt Ogaz, and captured on camera by Jim Beckmann, Carlos Cruz, Scott Holpainen, and Luke Knecht (with Scott Holpainen also editing). Check it out on KEXP.org or swing by Rosali Music’s site—and don’t forget to join the YouTube channel for extra perks!
Watch on YouTube
Rick Beato: Escape the Pentatonic Trap in One Lesson
Join today’s livestream to uncover the very scales top pros actually use—say goodbye to that overplayed pentatonic box and level up your fretboard knowledge.
Hurry on the Scale Matrix deal: get the full 3.5-hour system covering 25+ scales and modes for 50% off, but act fast—this early access price vanishes at midnight ET!
Watch on YouTube
Sparse Models and the Future of Efficient AI
Modern AI has followed a simple rule for progress: bigger is better. Scaling up the number of parameters and training data has consistently led to performance gains. But this approach comes with steep costs in compute, energy, and accessibility. Sparse models represent a different path forward, one that prioritizes efficiency without sacrificing capability.
The principle is straightforward. Most parameters in a large neural network contribute little to a given task at any moment. Instead of activating every weight, sparse models selectively engage only the most relevant connections. This mimics the brain, where neurons fire sparsely rather than all at once.
Implementing sparsity can take several forms. Static sparsity involves pruning redundant weights after training, reducing memory and computation needs. Dynamic sparsity, on the other hand, selects a different subset of active weights on the fly for each input. Mixture-of-Experts (MoE) models go further by partitioning the network into multiple expert subnetworks, routing each input through only a small fraction of them. Google’s Switch Transformer is a prime example, achieving massive scale while keeping per-example computation manageable.
The benefits are clear. Sparse models allow trillion-parameter architectures to be trained and deployed without proportional increases in compute. They also open possibilities for edge deployment, where hardware constraints make dense models impractical. By lowering the energy and hardware demands of AI, sparsity has the potential to democratize access to powerful systems.
Challenges remain in optimizing hardware and software for sparse computation. GPUs are built for dense matrix multiplications, and sparse operations often underutilize them. New accelerators and libraries are being developed to exploit sparsity more effectively. Ensuring that pruning or routing does not harm accuracy is another ongoing area of research.
Sparsity offers a vision where AI continues to grow more powerful without growing unsustainable. If dense scaling defined the last decade of AI, sparse scaling may define the next.
References
https://arxiv.org/abs/2007.03085
https://arxiv.org/abs/2101.03961
https://arxiv.org/abs/2209.10655
Building Web Apps & Exploring Research – About Sain Bux
I’m Sain Bux, a Full Stack Developer passionate about building scalable web apps and exploring research in AI, APIs, and emerging tech.
👨💻 What I Do
Develop with WordPress, React, PHP, and MySQL
Work on themes, plugins, and custom web solutions
Contribute to research projects, including publications on Zenodo
✨ What I’m Learning
Next.js & TypeScript for modern front-end development
Tailwind CSS for faster UI design
Basics of AI & ML to connect research with real-world applications
⚡ What I’m Currently Hacking On
New WordPress themes & plugins with better UX
API-driven projects integrating open data
🌐 Portfolio → sainbux.github.io
🤝 Open To
Collaborations on web projects
Research discussions in AI & APIs
Networking with devs, researchers, and startups
Thanks for reading! Looking forward to connecting with this awesome community 💡
I had fun making this video :) If anyone wants to shoot me ideas for future Defang videos, I'm all ears!
Struggling to figure out what to learn next in tech?
Check out roadmap.sh
Visual roadmaps for any role, from Frontend to AI Engineer.
Plan your path, track your progress, and never get lost in tutorials again!
AI Drives PCIe Aspirations
The packet transfer requirements of AI will keep pushing PCIe to double the data rate every three years.
The post AI Drives PCIe Aspirations appeared first on EE Times.
Cybersecurity Expectations Peak as GenAI Lands in the Trough
CSCOs, navigating an increasingly complex digital landscape, are facing stark contrasts in technology maturity.
The post Cybersecurity Expectations Peak as GenAI Lands in the Trough appeared first on EE Times.
AI Applications, Innovation Empowered! The Fourth GMIF2025 Innovation Summit Successfully Concluded
Co-hosted by the Shenzhen Memory Industry Association and the School of Integrated Circuits, Peking University, and organized by JWinsights (Shanghai) Technology Co., Ltd., the Fourth GMIF2025 Innovation Summit (Global Memory Innovation Forum) was successfully concluded on September 25, 2025. Under the theme “AI Applications, Innovation Empowered”, the GMIF2025 brought together leading representatives from across the […]
The post AI Applications, Innovation Empowered! The Fourth GMIF2025 Innovation Summit Successfully Concluded appeared first on EE Times.
Drones are a Key Part of the Smart Farming Revolution
Drones became an indispensable tool in modern agriculture, ensuring food security and sustainable farming practices for a growing global population.
The post Drones are a Key Part of the Smart Farming Revolution appeared first on EE Times.
YouTuber reports battery swelling caused Galaxy Ring to get stuck on his finger
Sometimes when a device fails, it fails in spectacular fashion. Daniel Rotar, who leads the ZONEofTECH YouTube channel, posted on X today that the battery on his Samsung Galaxy Ring began swelling while he was wearing it. "I cannot take it off and this thing hurts," he wrote. A few hours later, Rotar posted again that he had gone to a hospital and received help to remove the smart ring. He also shared photos of the device that showed the swelling on the interior surface.
Update:
— Daniel (@ZONEofTECH) September 29, 2025
- I was denied boarding due to this (been travelling for ~47h straight so this is really nice 🙃). Need to pay for a hotel for the night now and get back home tomorrow👌
- was sent to the hospital, as an emergency
- ring got removed
You can see the battery all… https://t.co/SRPfYI92Zg pic.twitter.com/ob8uUp5BeW
We reached out to Samsung for comment and received this statement from a spokesperson: "The safety of our customers is a top priority and we are in direct contact with this user to learn more about their concerns. While experiences like this are extremely rare, there are a few ways to attempt removing a stuck ring, including soap and water – or submerging your hand in cold water. If these methods do not work, additional information is provided on our Samsung Support page."
Current technologies aren't perfect and a device's battery can fail. When that happens, trapped gasses and heat can cause the battery to swell. It's not a common problem, but it is a known one. There's even a whole Reddit community dedicated to the phenomenon of these "spicy pillows." So here’s a recurring reminder that if you’ve got old gadgets you aren’t using, consider recycling them before they potentially reach this sort of fail state.
This article originally appeared on Engadget at https://www.engadget.com/wearables/youtuber-reports-battery-swelling-caused-galaxy-ring-to-get-stuck-on-his-finger-223749424.html?src=rss
How to follow Amazon's big hardware event tomorrow
Tomorrow, Amazon will host an event in New York City to unveil some new hardware. The showcase kicks off on September 30 at 10AM ET. Based on what we can see in the invitation and the most common cadence of product news, we are expecting to hear about developments to the Echo line of smart speakers, at least one new addition to the Kindle collection and an update for the Fire TV.
Normally this is where we'd have an embed of a YouTube link so you could watch along on Tuesday, but this event is a little unusual in that Amazon doesn't provide a stream for those of us at home. However, Engadget will have reporters at the event, and they will be liveblogging all the details as Amazon announces them. The liveblog will be up and running around 8AM ET.
In past years, this fall showcase has also been when Amazon shares news from the Ring, Blink and Eero companies, so we might have more than just its own-brand products being discussed tomorrow. And with AI the buzzword of the day, there's also a good chance that Amazon execs will spend some time discussing the Alexa+ voice assistant.
This article originally appeared on Engadget at https://www.engadget.com/how-to-follow-amazons-big-hardware-event-tomorrow-211105336.html?src=rss
OpenAI will reportedly release a TikTok-like social app alongside Sora 2
In a development that should surprise no one, OpenAI is preparing to release a standalone social app powered by its upcoming Sora 2 video model, Wired reports. The app reportedly "closely resembles" TikTok, with a vertical video feed and swipe-to-scroll navigation. The catch? It will only feature AI-generated content; there's apparently no option for the user to upload photos or videos from their phone's camera roll.
Wired adds OpenAI will limit Sora 2 to generating clips that are 10 seconds long or shorter for use inside of the app. It's unclear what the model's limit will be outside of the app. TikTok, following an original limit of 15 seconds, allows people to upload clips that are up to 10 minutes long. The new app is also said to include an identity verification tool. Should a user take advantage of the feature, Sora 2 will be able to use their likeness in videos it generates. In turn, that means other people will be able to tag those users and use their likeness when they go to remix one of their videos. As a safety precaution, OpenAI will push a notification to users whenever their likeness is used by someone else, even in situations where someone makes a video but never posts it to the app's feed.
According to Wired, the software will refuse to generate some videos due to copyright restrictions. However, it's not clear just how robust these protections will be, with The Wall Street Journal reporting OpenAI will require rights holders to opt out of their content appearing in videos Sora 2 generates.
As for why OpenAI would release a social media app, Wired suggests the company saw an opportunity after President Trump repeatedly extended the deadline for ByteDance to bring TikTok's US business under American control. By adding a social component to Sora, OpenAI may also be hoping to dissuade people from trying other models since leaving its new app would mean abandoning whatever community forms around it.
This article originally appeared on Engadget at https://www.engadget.com/ai/openai-will-reportedly-release-a-tiktok-like-social-app-alongside-sora-2-205842527.html?src=rss
Disney reportedly lost 1.7 million paid subscribers in the week after suspending Kimmel
Reporter Marisa Kabas, founder of The Handbasket, posted on Bluesky today that more than 1.7 million subscribers canceled their paid Disney streaming plans between September 17 and September 23. The total allegedly includes subscriptions to Disney+, Hulu and ESPN. That falloff reportedly marked a 436 percent increase over the usual churn rate for the service. We’ve reached out to Disney+ for comment on this claim.
Disney also chose to increase subscription prices last week, which could prolong the wave of cancellations. Kabas' source also claimed that Disney sped the return of the late night show because it had planned to announce the price hike last Tuesday.
In case you need a reminder of the entertainment news that dominated headlines for the past two weeks, Disney-owned ABC suspended the late night talk show Jimmy Kimmel Live on September 17 after President Donald Trump and members of his administration accused the host of making inappropriate comments regarding the assassination of right-wing influencer Charlie Kirk. Although Disney then announced on September 22 that the program would be reinstated the following night, it seems a lot of viewers were not impressed with the company's actions.
This article originally appeared on Engadget at https://www.engadget.com/entertainment/streaming/disney-reportedly-lost-17-million-paid-subscribers-in-the-week-after-suspending-kimmel-201615937.html?src=rss
What to expect from Amazon's event on Tuesday, September 30
Apple isn't the only big tech company that's launching hardware this month. Amazon is having its fall hardware event on Tuesday, September 30 at 10AM ET in New York City, and — as it does most years — the company is expected to debut new devices across its myriad product categories. While the event will not be livestreamed, Engadget staff will be in attendance and liveblogging it in real-time. Our pre-show coverage will begin around 8AM ET.
So what will Amazon be sharing? Using the images in the invitation that went out to members of the press as a jumping-off point, here are our best educated guesses as to what we can expect.
In February, Amazon introduced some improvements to its voice assistant, promising that Alexa+ would be more conversational and more helpful thanks to generative AI. It seems likely that the fall event will put hardware for interacting with Alexa+ front and center, so expect to hear about new models or new software updates for the Echo line of smart speakers. It's been a couple years since the Echo Show got an update, and it's been even longer for the standard Echo. Those two products are the most likely to star in the upcoming announcements.
The invite also includes a photo of a Kindle that depicts a color illustration, so there may be more in store for the ereaders. The Kindle Scribe 2 just came out earlier this year, as did the Kindle Colorsoft, so maybe there's a new option in the works that combines the best features of both. As with the Echo, there also hasn't been much new for the basic model Kindle, so that may also be getting some updates.
There's also what appears to be the corner of a television in the invitation, so news about the Fire TV also seems promising. This event is also where Amazon-owned Ring and Eero brands may unveil new products. Any announced items may not be available right away, but the timing does set the stage for some substantial discounts on older generations of Amazon gear. In fact, many of Amazon's current devices are already on sale — including Blink, Kindle and Ring — as more early October Prime Day deals pop up ahead of the sale's official start on October 7.
Update, September 29, 2025, 3:20PM ET: This story has been updated to note Engadget's liveblog plans for the Amazon event.
This article originally appeared on Engadget at https://www.engadget.com/big-tech/what-to-expect-from-amazons-event-on-tuesday-september-30-210035284.html?src=rss
Meta is bringing new facial recognition tools to the UK, EU and South Korea
Meta is expanding its use of facial recognition in Europe, the UK and South Korea to crack down on accounts that impersonate public figures. The new facial recognition-powered safety features are now live on Facebook in the regions and will expand to Instagram in the coming months.
The technology was initially put to use last year starting in the US, helping to identify ads that fraudulently use a celebrity's likeness as well as to help people regain access to hacked accounts. Public figures opt in to this program in Europe, which is also being rolled out in South Korea alongside the new protections against impersonation. This new use case is aimed at scammers who pose as public figures to trick unsuspecting users into sending money or other scams of that nature.
"We’ll now use facial recognition technology to compare the profile picture on the suspicious account to the real public figure’s Facebook and Instagram profile pictures. If there’s a match, we will remove the impostor account," said a Meta spokesperson.
In addition to the US rollout, the company’s facial recognition technology has been used to aid account recovery in the UK, EU and South Korea since March. This came three years after Facebook decided to shut down its facial recognition system on Facebook, due in large part to public backlash against the technology.
The social media giant touts the benefits of these tools, reporting that in the first half of 2025, user reports of "celebrity bait" ads dropped by 22 percent globally. Facial recognition remains a controversial technology, with differing public opinion on its use in law enforcement and the workplace.
This article originally appeared on Engadget at https://www.engadget.com/big-tech/meta-is-bringing-new-facial-recognition-tools-to-the-uk-eu-and-south-korea-185303065.html?src=rss
Tile trackers reportedly have a security flaw that can let stalkers track your location
Researchers have discovered major security flaws with Tile tracking tags, according to a report by Wired. These flaws could allow both the company itself and tech-savvy stalkers to track a user's location. The security issue could also let a malicious actor falsely frame a Tile owner for stalking, as the flaw can make it appear as if a particular tag is constantly in the vicinity of somebody else's tag.
The issue pertains to how Tile tags transmit data during use. Tile tags transmit a lot of data beyond that of other trackers, including the static MAC address and the rotating ID. According to reporting, none of this stuff is encrypted. The rotating ID changes all of the time, but a MAC address doesn't.
Researchers believe that all of this information is stored in cleartext, making it easy for hackers to get ahold of. This also would theoretically give Tile itself the ability to track its users, though the company says it doesn't have this capability.
It gets worse. Anyone with a radio frequency scanner can allegedly intercept all of this information as it's being transmitted, creating another potential security hole. Also, this problem might not even be solved if Tile decides to stop transmitting the MAC address. This is because the company generates its rotating ID in such a way that future codes can be reliably predicted from past ones.
“An attacker only needs to record one message from the device," one of the researchers behind the findings said, adding that a single recorded message will "fingerprint it for the rest of its lifetime." The researcher said this creates a risk of systemic surveillance.
The security researchers, who are involved with the Georgia Institute of Technology, reached out to Tile's parent company Life360 in November of last year to report the findings. Wired said the company stopped communicating with the researchers in February. The company did say it has made a number of improvements to its security but didn't elaborate further.
This article originally appeared on Engadget at https://www.engadget.com/big-tech/tile-trackers-reportedly-have-a-security-flaw-that-can-let-stalkers-track-your-location-183949165.html?src=rss
Claude Sonnet 4.5 is Anthropic's safest AI model yet
In May, Anthropic announced two new AI systems, Opus 4 and Sonnet 4. Now, less than six months later, the company is introducing Sonnet 4.5, and calling it the best coding model in the world to date. Anthropic's basis for that claim is a selection of benchmarks where the new AI outperforms not only its predecessor but also the more expensive Opus 4.1 and competing systems, including Google's Gemini 2.5 Pro and GPT-5 from OpenAI. For instance, in OSWorld, a suite that tests AI models on real-world computer tasks, Sonnet 4.5 set a record score of 61.4 percent, putting it 17 percentage points above Opus 4.1.
At the same time, the new model is capable of autonomously working on multi-step projects for more than 30 hours, a significant improvement from the seven or so hours Opus 4 could maintain at launch. That's an important milestone for the type of agentic systems Anthropic wants to build.
Perhaps more importantly, the company claims Sonnet 4.5 is its safest AI system to date, with the model having undergone "extensive" safety training. That training translates to a chatbot Anthropic says is "substantially" less prone to "sycophancy, deception, power-seeking and the tendency to encourage delusional thinking" — all potential model traits that have landed OpenAI in hot water in recent months. At the same time, Anthropic has strengthened Sonnet 4.5's protections against prompt injection attacks. Due to the sophistication of the new model, Anthropic is releasing Sonnet 4.5 under its AI Safety Level 3 framework, meaning it comes with filters designed to prevent potentially dangerous outputs related to prompts around chemical, biological and nuclear weapons.
With today's announcement, Anthropic is also rolling out quality of life improvements across the Claude product stack. To start, Claude Code, the company's popular coding agent, has a refreshed terminal interface, with a new feature called checkpoints included. As you can probably guess from the name, they allow you to save your progress and roll back to a previous state if Claude writes some funky code that isn't quite working like you imagined it would. File creation, which Anthropic began rolling out at the start of the month, is now available to all Pro users, and if you joined the waitlist Claude for Chrome, you can start using the extension today.
API pricing for Sonnet 4.5 remains at $3 per one million input tokens and $15 for the same amount of output tokens. The release of Sonnet 4.5 caps off a strong September for Anthropic. Just one day after Microsoft added Claude models to Copilot 365 last week, OpenAI admitted its rival offers the best AI for work-related tasks.
This article originally appeared on Engadget at https://www.engadget.com/claude-sonnet-45-is-anthropics-safest-ai-model-yet-170000161.html?src=rss
This Roomba robot vacuum is on sale for only $150 ahead of Prime Day
The iRobot Roomba 104 robot vacuum is on sale for $150 just ahead of October's Prime Day. That's a nice little discount of 40 percent, which represents a savings of $100.
This is a newer version of the unit that topped our list of the best budget robot vacuums. It's an entry-level robovac that gets the job done. The cleaning motor is fairly powerful and it ships with a multi-surface brush and an edge-sweeping brush. The vacuum uses LiDAR to map a home and to help it avoid obstacles when cleaning.
It's also been equipped with specialized sensors to prevent falling down stairs. Steps are the natural enemy of all robot vacuums, except maybe this one. The Roomba 104 integrates with the company's proprietary app, which allows for custom cleaning schedules and the like. The robot can also be controlled via voice assistant and boasts compatibility with Siri, Alexa and Google Assistant.
The vacuum will automatically head to the charger for some juice when running low, which is nice. The battery lasts around 200 minutes per charge, which is a decent enough metric for a budget-friendly robovac. The only downside here? This is just a vacuum. It doesn't mop and it doesn't come with a dedicated debris canister.
This article originally appeared on Engadget at https://www.engadget.com/deals/this-roomba-robot-vacuum-is-on-sale-for-only-150-ahead-of-prime-day-164953406.html?src=rss
A PlayStation photography book featuring never-before-seen design concepts is on the way
Sony has been marking the 30th anniversary of PlayStation by selling you stuff, like PS5 consoles and accessories styled after the PS1. The company has something else lined up to mark the occasion: a photography book showcasing "never-before-seen prototypes, concept sketches and design models that shaped hardware development" from the early days through to the current PS5 era.
PlayStation: The First 30 Years is a 400-page hardback book printed on heavyweight matt art stock. You better hope your coffee table is sturdy, since this book is a chonkster. It weighs in at 5kg (11lbs).
Sony worked with publisher Read-Only Memory on PlayStation: The First 30 Years. The collaborators have offered a peek at what's inside the book. It's shaping up to be a fascinating glimpse at some of the designs Sony tried for its hardware over the last few decades.
For instance, the original PlayStation could have looked much different, more directly exemplifying designer Teiyu Goto's "vision of simple squares and circles coming to life." (For what it's worth, Engadget deputy editor Nathan Ingraham said this design looked like a proton pack):
Some of the controller concepts are pretty out there too. Some don't look all that comfortable to hold or use for extended gaming sessions. This one — which appears to be for the PS3 at the earliest, given the inclusion of the PS button — is truly bonkers. Thank goodness Goto landed on the SNES-style design with grips that has proven so successful (and comfortable) over the years:
The book isn't entirely limited to hardware concepts, as it features photos of Sony's design labs. The tome seems like a very cool item for PlayStation fans and those who love gaming history to have, but there might be an element of sticker shock.
The book is available via Read-Only Memory's website for $182. A deluxe edition with exposed binding, a foil-stamped clamshell presentation box and a photographic print signed by Goto and photographer Benedict Redgrove will run you $467. The fancier edition has a limited run of 1994 copies. Coincidentally, that's the year the PS1 debuted in Japan. Both editions of the book will ship in spring 2026.
Meanwhile, Sony has teamed up with Reebok for a collection of 30th anniversary sneakers styled after — you guessed it — the PS1. The kicks will be available in October and the three designs are linked to the PS1's launch regions. They include the InstaPump Fury 94 for Japan, Pump Omni Zone II for the US and Workout Plus for the UK.
This article originally appeared on Engadget at https://www.engadget.com/gaming/playstation/a-playstation-photography-book-featuring-never-before-seen-design-concepts-is-on-the-way-164859020.html?src=rss
Microsoft is trying to make 'vibe working' a thing
Microsoft is taking inspiration from the AI-driven workflows of "vibe coding" and has now set out to make "vibe working" a thing (yes, those are the words the company chose.) Does AI in the workplace even lead to worthwhile outputs? Does it mortgage our brains' ability to learn? There are many seemingly critical question unanswered. But in the meantime, sure: vibe working it is.
Using Office Agent within Office apps or Copilot chat, users can begin a document with a single prompt and then work iteratively alongside Copilot to develop a finished product. Microsoft says this is the "new pattern of work for human-agent collaboration." The Agent Mode tool supports Excel and Word workflows, and Microsoft says PowerPoint support is coming soon; Office Agent works with PowerPoint and Word, with Excel coming soon.
The company waxes poetic about the "full power of Excel" being available only to expert users and promises that an Agent Mode that can "speak Excel" will change all that. In data shared as part of the announcement, Microsoft said that Copilot Agent Mode in Excel achieved 57.2 percent accuracy on the SpreadsheetBench benchmark. This is compared to a 71.3 percent human score, though it's not clear if that's for average users, Excel power users or how many human users that score is derived from. Still — not great numbers!
Agent Mode also works in Word to summarize, edit and of course help to create entire drafts (though its unclear what those relative accuracy rates are.) Both the Excel and Word Agent Modes are powered by OpenAI's latest models. Office Agent in Copilot chat is powered by Anthropic models and can create PowerPoint presentations and Word documents in what Microsoft calls a "chat-first experience."
Agent Mode for Excel and Word, as well as Office Agent, are available today through the Frontier program. Agent Mode is currently limited to the web-based versions of Word and Excel and is coming to desktop soon.
This article originally appeared on Engadget at https://www.engadget.com/ai/microsoft-is-trying-to-make-vibe-working-a-thing-163334367.html?src=rss
FCC accidentally leaked iPhone schematics, potentially giving rivals a peek at company secrets
The Federal Communications Commission (FCC) recently published a 163-page PDF showing the electrical schematics for the iPhone 16e, despite Apple specifically requesting them to be confidential. This was most likely a mistake on the part of the FCC, according to a report by AppleInsider.
The agency also distributed a cover letter from Apple alongside the schematics, which is dated September 16, 2024. This letter verifies the company's request for privacy, indicating that the documents contain "confidential and proprietary trade secrets." The cover letter asks for the documents to be withheld from public view "indefinitely." Apple even suggested that a release of the files could give competitors an "unfair advantage."
To that end, the documents feature full schematics of the iPhone 16e. These include block diagrams, electrical schematic diagrams, antenna locations and more. Competitors could simply buy a handset and open it up to get to this information, as the iPhone 16e came out back in February, but this leak would eliminate any guesswork. However, Apple is an extremely litigious company when it comes to stuff like patent infringement.
The FCC hasn't addressed how this leak happened or what it intends to do about it. AppleInsider's reporting suggested that this probably happened due to an incorrect setting in a database. This was likely not an intentional act against Apple, which tracks given that the company has been especially supportive of the Trump administration. CEO Tim Cook even brought the president a gold trophy for being such a good and important boy.
This article originally appeared on Engadget at https://www.engadget.com/big-tech/fcc-accidentally-leaked-iphone-schematics-potentially-giving-rivals-a-peek-at-company-secrets-154551807.html?src=rss
Anker's latest Prime charging devices are already on sale
Anker has a bunch of new Prime charging devices available and some of them are already getting solid discounts. The lineup includes a laptop-compatible Prime Power Bank (26,250mAh, 300W). The device's Amazon page currently features a clippable coupon that will take $46 off the price, reducing it by 20 percent. That means you'll be able to snap this powerbank up for $184.
This portable charger delivers a whopping 300W of total output between its two USB-C ports and single USB-A port. Anker says that, thanks to its 140W output (and if you're using a 5A cable), you can top up the battery of an M4 Pro MacBook Pro to 50 percent in 27 minutes and an iPhone 17 Pro Max to 50 percent in 22 minutes. You can keep tabs on what's happening on each port using the display.
This is also the first Anker powerbank that supports up to 250W of input recharging by using both USB-C ports simultaneously. This faster charging can top up the powerbank's own battery to 50 percent in just 13 minutes, according to Anker.
This model has a capacity of 26,250mAh (99.75Wh), which is very close to the FAA's limit of 100W. It's TSA-approved, so you'll be able to take it on flights. It's capable of taking an M4 MacBook Pro from fully dead to an 80 percent charge. The power bank weighs 1.3 lbs — the same as about three iPhones, Anker says.
Other Anker Prime devices are on sale too. A new 3-in-1 MagSafe charger — which can top up the batteries of your iPhone, Apple Watch and AirPods with wireless charging case simultaneously — also gets a 20 percent discount to $184 when you use a clippable coupon. While there are certainly less expensive 3-in-1 MagSafe options out there, this one includes Qi2 25W support for faster charging, as well as TEC active cooling. Anker suggests this can help speed up charging times and help guard against iPhone battery degradation over time.
A three-port GaN Prime Charger that plugs into an outlet is 20 percent off too, at $120 with a clippable coupon. Anker says it's as compact as an AirPods Pro 3 case and can provide a total output of 160W, with up to 140W via a single USB-C cable. The charger is said to offer smart power distribution, and you can monitor what's happening and adjust modes via an onboard display and controls.
This article originally appeared on Engadget at https://www.engadget.com/deals/ankers-latest-prime-charging-devices-are-already-on-sale-152040769.html?src=rss
You can now install iOS 26 on your iPhone: Everything to know about the free software update
Have you played around with the new iOS 26 yet? Here's how to download it on your iPhone if you haven't yet — once you do, you'll notice several new changes, including a clear design called Liquid Glass. The translucent look can be found within apps as well as your lock and home screens. (You can't turn it off, but you can reduce the transparency to make it less clear.) The overhaul is one of several big changes that come with iOS, macOS, iPadOS, watchOS and the rest of Apple's software suite — the new operating system updates for those gadgets are also available to download now.
You'll also see multiple useful additions in the Phone and Messages apps. For instance, the ability to weed out spam texts or unknown senders and an option to hold your spot on a phone call when you've been waiting for a representative to pick up. And for fun, you can now create polls in group texts. Other new improvements coming include a confirmed AirPods live translation feature, and a new bouncy animation on the passcode screen and in the Control Center.
If you're planning to upgrade your phone this year, take a look at all the new models Apple unveiled during its iPhone 17 event, from the iPhone 17 lineup (check out our review of the iPhone 17) to the all-new ultra-thin iPhone Air (take a look at our hands-on review). If you missed the keynote, you can catch up on everything Apple announced here or rewatch the full iPhone 17 launch. The company rounded out its announcements with three new Apple Watch models, as well as the brand-new AirPods Pro 3.
Most newer iPhone models are eligible to download iOS 26. We've rounded up a full list of new features you'll have to check out. If you're waiting to update your phone until iOS 26.1 becomes available, note that the beta version is available now so it shouldn't be too long before it's released to the public.
Instead of releasing iOS 19, Apple skipped its naming convention ahead to iOS 26. The company has decided to line up its iOS version numbers with a year-based system, similar to car model years. So while iOS and its sibling operating systems were released in 2025, they're all designated "26" to reflect the year ahead.
Apple also released iOS 18.7 alongside iOS 26 to address security issues — great news if you don't have a phone that's eligible to run the latest OS. If you're wondering which one to download first, it's really up to you. Do you want to wait for Apple to quash any bugs in iOS 26 or do you want access to all the new features everyone's talking about now?
Let's be honest. Out of everything announced at WWDC this year, the new Liquid Glass design was the star of the show. The iPhone's home and lock screens have looked pretty much the same year after year — the last exciting thing (in my opinion) was the option to add your own aesthetic to your home screen by customizing your apps and widgets. So seeing the home and lock screens' new facelift is refreshing.
So what exactly is Liquid Glass? Apple calls it a "new translucent material" since, well, the apps and widgets are clear. However, the screen can still adapt to dark and light modes, depending on surroundings. You'll also notice buttons with a new floating design in several apps, like Phone and Maps. They're designed to be less distracting than the current buttons, but are still easy to see.
While the design overhaul has proven to be controversial since its announcement, some — including Engadget's Devindra Hardawar — like the new direction, even if it's somewhat reminiscent of Microsoft's translucent Windows Vista Aero designs from nearly twenty years ago.
That said, while it was in beta mode, Apple incorporated some user feedback into the design, dialing back the transparency in at least some places. And while it will continue to evolve, Apple users won't be able to escape it: The company says Liquid Glass was designed to make all of its platforms more cohesive.
Here's a look at how the translucent aesthetic looks with the new macOS Tahoe 26 on your desktop.
If you find you don't like the new clear look after downloading iOS 26, you can change a couple Accessibility settings, though you can't completely turn off Liquid Glass. For starters, you can reduce the liquid look by going to Settings > Accessibility > Motion > toggle on Reduce Motion.
You can make your drop-down menus more opaque instead of translucent by going to Settings >Accessibility > Display & Text Size > toggle on Reduce Transparency.
You can also change your clear apps back to the way they were. Just press and hold any app on your screen and select Edit Home Screen. From here, tap Edit (top left) and Customize. You can swap out to Default, Dark or Tinted.
iOS 26 has a laundry list of new features. Among the most worthwhile:
Phone app redesign: You can now scroll through contacts, recent calls and voicemail messages all on one screen. It also comes with a new feature called Hold Assist that'll notify you when an agent comes to the phone so you can avoid the elevator music and continue on with other tasks.
Live Translation in Phone, FaceTime and Messages: iOS 26 brings the ability to have a conversation via phone call, FaceTime or text message with someone who speaks another language. Live Translation uses Apple Intelligence and it will translate your conversation in real time, which results in some stop-and-go interactions in the examples Apple shared during its presentation.
Polls in group chats: Tired of sorting through what seems like hundreds of messages in your group chat? You and your friends can now create polls in group messages to decide things like which brunch spot you're eating at or whose car you're taking on a road trip.
Filtering unknown senders in Messages: If you haven't received spam texts about unpaid tolls or other citations, you're lucky. For those of us who have, those annoying messages will soon be filtered away in a separate folder.
Visual Intelligence: Similar to a reverse Google image search, this new feature will allow you to search for anything that's on your iPhone screen. For instance, if you spot a pair of shoes someone is wearing in an Instagram photo, you can screenshot it and use Visual Intelligence to find those shoes (or similar ones) online.
Photos tabs are back: For anyone who's been frustrated with last year's changes to the Photos app, you'll be happy to know that your tabs are back. Library and Collections have their own separate spaces so you don't have to scroll to infinity to find what you're looking for.
Camera app updates: Navigating the Camera app should be simpler in iOS 26, as all the buttons and menus are in convenient spots — less swiping, more photo taking. Plus, there's a new feature that tells you if your lens needs to be cleaned.
FaceTime "Communication Safety" feature: A newer addition to iOS 26 appears to be the FaceTime "Communication Safety" feature that pauses communications if and when nudity is detected. The feature appears to be a child safety feature that uses on-device detection, thus obviating any cloud-based privacy issues.
New lock screen options: The iPhone lock screen is more customizable in iOS 26, with a cooler clock, 3D wallpaper effects, more widgets and better focus mode options.
New alarm setting: You'll no longer be stuck with the 9-minute snooze setting in your alarms. Instead, you now have the option to change your snooze time from one to 15 minutes.
Screenshots look different: You'll notice several new features when you take a screenshot, including "highlight to search," the option to search your image on Google and ChatGPT is there for any questions you have about the image.
Preview on iOS 26: There's a new app called Preview that lets you scan documents, fill out PDF forms, as well as highlight, underline or strike through text in a PDF file. You can also remove backgrounds from images in this app.
Your iPad isn't getting left behind when it comes to big updates. Here's what to look for.
Multitasking and real windowing: When you download the newest update, you'll be able to have multiple apps running on your screen at the same time. Once you open an app, it'll appear on your screen as normal but you'll be able to resize and move it across your screen to make room for other apps. This feature is optional so you can turn it off if you don't like it.
Visual update: Along with the other new OSes, iPadOS 26 also has the Liquid Glass aesthetic. This new look appears on the lock and home screens, as well as the drop-down menus.
New menu bar: When you swipe down on your screen, the new menu bar will appear with options like File, Edit, Windows and more. There's also a search option if you're looking for something specific.
There's more beyond that, so be sure to check out our first impressions of iPadOS 26.
AirPods are also getting updated with iOS 26. Here are some of the more notable functions.
Enhanced audio recording: Apple calls this "studio-quality" audio recording, and with it, you'll notice more clarity while in noisy environments.
Camera remote control: Using this, you can take a photo or start and stop video recording with just one press on your AirPods. When taking photos, you'll get a three-second countdown before your iPhone or iPad snaps the picture.
Live translation feature: Live translation is available on AirPods Pro 3, AirPods Pro 2 and AirPods 4 with ANC when paired with iPhones running iOS 26. Apple says that when someone speaks to you in a different language, you can use your AirPods to translate what they’re saying.
Heart rate monitoring: Apple introduced new AirPods Pro 3 earbuds this month, which include heart rate monitoring. It will work with Apple's Health app and other fitness apps that track heart rates.
Siri is in a holding pattern. Apple has previously specified that its smarter voice assistant — first promised at WWDC 2024 — is delayed until some point "in the coming year," so you shouldn't expect any major changes right now. But there are reports that Apple is aiming to give Siri a bigger brain transplant by basing it on third-party artificial intelligence models like OpenAI's ChatGPT or Anthropic's Claude, which could make 2026 a pivotal year. The company is also reportedly working on a "stripped-down" AI chatbot to rival ChatGPT.
A few iPhone models that run the current version of iOS — iPhone XR, XS and XS Max — aren't compatible with the latest upgrade. But any iPhones from 2019 or later will be eligible for iOS 26:
iPhone SE (second generation or later)
iPhone 11
iPhone 11 Pro
iPhone 11 Pro Max
iPhone 12
iPhone 12 mini
iPhone 12 Pro
iPhone 12 Pro Max
iPhone 13
iPhone 13 mini
iPhone 13 Pro
iPhone 13 Pro Max
iPhone 14
iPhone 14 Plus
iPhone 14 Pro
iPhone 14 Pro Max
iPhone 15
iPhone 15 Plus
iPhone 15 Pro
iPhone 15 Pro Max
iPhone 16
iPhone 16 Plus
iPhone 16 Pro
iPhone 16 Pro Max
iPhone 16e
iPhone 17
iPhone Air
iPhone 17 Pro
iPhone 17 Pro Max
iOS 26 has officially been released to the public for free as of September 15, 2025.
If you're more interested in the new Apple Intelligence features, here's everything Apple revealed for iOS, macOS and more during WWDC. Also, check out how iOS 26 screenshots could be an intriguing preview of Apple's delayed Siri rework.
Update, September 19: Added details about making Liquid Glass less clear.
Update, September 18: Added info about Engadget's iPhone 17 review.
Update, September 17: Added details and link about the new iPhone Air.
Update, September 16: Added details about modifying Liquid Glass settings, and new info about the Preview app.
Update, September 15: Noted iOS 26 is officially available for download.
Update, September 12: Noted iOS 26 is coming to eligible iPhones in just a few days, timing TBD.
Update, September 11: Added the products Apple released at its Tuesday event and noted when iOS 26 will be available to download.
Update, September 9: Noted the iPhone event is starting soon.
Update, September 8: Noted the iPhone 17 event is tomorrow.
Update, September 4: Added details about how screenshots are different in iOS 26.
Update, September 3: Noted Apple is expected to release iOS 18.7 alongside iOS 26.
Update, September 2: Added more new features coming with iOS 26.
Update, August 29: Added new section about Siri and a link to what's new with the iOS 26 Camera app.
Update, August 27: Added the official iPhone 17 event date, as well as the potential iOS 26 release.
Update, August 25: Added a rumor about new AirPods Pro having heart rate monitoring.
Update, August 22: Noted that Apple has officially stopped signing iOS 18.6.
Update, August 20: Noted that iOS 26 public beta 4 and iOS 18.6.2 are now available to download.
Update, August 18: Added details about a potential iOS 18.6 update.
Update, August 15: Added to link to what to expect at the Apple iPhone event and details about what's available in the iOS 26 screenshots editor.
Update, August 13: Added new AirPods detail spotted in the iOS 26 beta.
Update, August 11: Noted that iOS 26 developer beta has hit beta 6.
Update, August 8: Added new features coming with iPadOS 26 and AirPods.
Update, August 6: Noted the release of iOS 26 beta 5 and the new bouncy feature on passcode screen and Control Center.
Update, August 4: Noted that Apple is reportedly working on a ChatGPT rival.
Update, August 1: Added quote from Tim Cook about iOS 26.
Update, July 31: Noted that iOS 18.6 is now available.
Update, July 24: Noted the iOS 26 public beta is now available.
Update, July 3: Noted new FaceTime feature found in the developer beta.
Update, June 30: Noted ongoing iOS 18 releases, and reports that Apple is considering additional external LLMs for Siri.
Update, June 25: Noted changes added in iOS 26 beta 2.
This article originally appeared on Engadget at https://www.engadget.com/mobile/you-can-now-install-ios-26-on-your-iphone-everything-to-know-about-the-free-software-update-135749775.html?src=rss
How to use the AirPods Pro 3's heart rate tracking feature
Your AirPods may soon become your most important pieces of fitness gear. One of the AirPods Pro 3’s new features is heart rate tracking for both workouts and periodic pulse checks. Apple’s latest earbuds are equipped with photoplethysmography (PPG) sensors that shine infrared light hundreds of times per second to measure how light is absorbed in blood flow. What’s more, the AirPods Pro 3 work in tandem with the iPhone’s accelerometers, gyroscope, GPS and on-device AI for more comprehensive cardio insights and to estimate calories burned during exercise. Here’s how to use heart-rate monitoring with the AirPods Pro 3, along with some maintenance tips to keep the tech in prime working condition.
An important aspect of getting accurate data from your AirPods is understanding how the earbuds track your heart rate. If you’re using Apple’s Fitness app, the AirPods Pro 3 will continuously monitor your pulse during a workout. In addition to the heart rate sensor and accelerometer in the earbuds, details like location, barometric pressure and health information are also pulled from your iPhone, and they’re all used to estimate the calories you burn each session. Lastly, if you’re wearing both an Apple Watch and the AirPods Pro 3 during a workout, the device with the most accurate data over the last five minutes will be used to track your heart rate.
If you prefer to keep one ear completely unobstructed for whatever reason, you might be happy to know you only need to wear one of your AirPods Pro 3 during a workout to track your heart rate. Both earbuds are equipped with heart rate sensors; you can wear either one so you won’t have to forego your favorite ear.
It probably goes without saying, but you need a set of the AirPods Pro 3 to use the heart rate feature. None of the older AirPods models have onboard heart rate sensors. The only other option in Apple’s stable that does is the Beats Powerbeats Pro 2.
You’ll also need an iPhone that’s capable of running iOS 26. This is every model from the iPhone 11 and newer, including the second-generation iPhone SE.
Turning heart rate sensing on and off on the AirPods Pro 3 is simple. The feature is enabled by default, but you can turn it off in the AirPods settings menu on your iPhone.
Put your AirPods Pro 3 in your ears and make sure they’re connected to your iPhone.
Go to the Settings app on your phone.
On the main menu, tap the name of your AirPods, which will appear near the top of the screen.
Scroll down to Heart Rate and turn it off.
Repeat the same steps to turn Heart Rate back on.
In addition to workout apps, you can also view your heart rate in the Health app on your iPhone or iPad. Here, you’ll be able to see live data from the AirPods Pro 3’s heart rate sensors.
In the Health app
Put your AirPods Pro 3 in your ears and make sure they’re connected to your iPhone or iPad.
Open the Health app on an iPhone or iPad.
Tap the magnifying glass icon at the bottom right, then tap Heart, then tap Heart Rate.
Your live heart rate from AirPods Pro 3 will be displayed in a graph.
During a workout
As I mentioned before, your heart rate will appear in the Fitness app during workouts. The AirPods Pro 3 can also send that data to compatible third-party apps. For those, you’ll need to give the app permission to access your heart rate data, as well as the ability to read and record workouts. You’ll only have to do this once: when you use a third-party workout app with Apple Health for the first time.
Put your AirPods Pro 3 in your ears and make sure they’re connected to your iPhone.
Go to the Fitness app or a third-party workout app on your iPhone. (Apple hasn’t published a list of supported apps, but iOS fitness apps that offer heart-rate tracking should work with the AirPods Pro 3 if the developer decides to do so. Examples of those include Nike Run Club, Peloton, Runna and Ladder.)
Start a workout in the app you choose.
Your heart rate will appear in the app during your workout. If you see dashes or the number is dimmed, that means the AirPods Pro 3 aren’t getting a good reading. This is similar to what happens if your Apple Watch is not maintaining contact with your wrist and is typically a result of a loose wristband. Check the fit of your earbuds and make sure they’re connected to your iPhone.
To stop heart rate tracking in any of the apps, end your workout.
You will then get a summary of the heart rate stats for your workout in a third-party app or Apple’s Fitness or Health apps.
Just ask Siri any time
In addition to looking at heart rate data in the Health app or other workout apps, you can just ask Siri to tell you a current reading without ever touching your phone.
Put your AirPods Pro 3 in your ears and make sure they’re connected to your iPhone.
Ask Siri “What’s my heart rate?” or something similar.
Remember: You don’t need to touch the buttons on the AirPods Pro 3 in order to activate Siri. If Talk to Siri is set to “Siri or Hey Siri” or “Hey Siri” under Apple Intelligence & Siri in the Settings app on your iPhone, you only need to use the wake word and then ask your question. If not, you can activate Siri with a press and hold the Force Touch button on either side until you hear a chime.
Like any activity-monitoring device, there are some habits that will help ensure the data that you’re getting from the AirPods Pro 3 is as accurate as possible. Not only do the AirPods Pro 3 track heart rate, but the motion sensor data it sends to your iPhone is essential to calculate things like calories burned, steps and distance for your daily activity. For this reason, it’s crucial that your earbuds and iPhone or iPad are working together properly.
Make sure your AirPods Pro 3 fit well. If there isn’t good contact between the heart rate sensors on the AirPods Pro 3 and the skin inside your ear, you could get poor heart rate readings. Apple warns users that the best ear tip size for acoustic seal and the best one for heart rate tracking may be different.
Maintain up-to-date information in the Health app. The details that you enter into the Health app — such as height, weight, gender and age — are used with heart rate and motion sensor data to calculate metrics like calories burned.
Keep warm: cold ears may impact the performance of the AirPods Pro 3. If you’re exercising in cold locations, you may have trouble getting accurate readings from the AirPods Pro 3. Low temperatures can reduce the blood flow to your ears, which makes it more difficult for the earbuds to track your heart rate accurately.
Check your ears regularly. Earwax buildup, moisture or skin conditions can affect contact between the AirPods Pro 3 and your skin.
Check your jewelry or piercings. Apple says metal jewelry from piercings can also interfere with heart rate readings. If you encounter any performance issues during workouts, the company recommends removing any piercings near the location where the AirPods Pro 3 sit in your ears.
Know that everyone’s ears are different. The amount of blood flow detectable through the skin varies from person to person. Like I said earlier, this can also be affected by the environment (cold weather, for example) or activity levels. If blood flow in your ears changes based on either of these factors, it can impact the performance of heart-rate tracking on the AirPods Pro 3.
Clean your AirPods Pro 3. To keep the AirPods Pro 3 in peak working condition, clean them regularly. Tips for the best ways to do that are available in our AirPods cleaning guide.
This article originally appeared on Engadget at https://www.engadget.com/audio/headphones/how-to-use-the-airpods-pro-3s-heart-rate-tracking-feature-141510469.html?src=rss
The best October Prime Day deals you can get right now: Early sales on tech from Apple, Amazon, Samsung, Anker and more
Amazon Prime Day has returned in the fall for the past few years, and 2025 is no exception. Prime Day will return on October 7 and 8, but really, you don't have to wait until the official start date to save. Amazon typically always has early Prime Day deals in the lead-up to the event, and this year we’re already seeing some solid discounts on gadgets we like. Here, we’ve gathered all of the best Prime Day deals you can get right now, and we’ll keep updating this post as we get close to Prime Day proper.
Apple MagSafe charger (25W, 2m) for $35 (29 percent off): The latest version of Apple's MagSafe puck is Qi2.2-certified and supports up to 25W of wireless power when paired with a 30W adapter. The two-meter cable length on this particular model gives you more flexibility on where you can use it: in bed, on the couch, at your desk and elsewhere.
Leebein 2025 electric spin scrubber for $40 (43 percent off, Prime exclusive): This is an updated version of my beloved Leebein electric scrubber, which has made cleaning my shower easier than ever before. It comes with seven brush heads so you can use it to clean all kinds of surfaces, and its adjustable arm length makes it easier to clean hard-to-reach spots. It's IPX7 waterproof and recharges via USB-C.
Jisulife Life7 handheld fan for $25 (14 percent off, Prime exclusive): This handy little fan is a must-have if you life in a warm climate or have a tropical vacation planned anytime soon. It can be used as a table or handheld fan and even be worn around the neck so you don't have to hold it at all. Its 5,000 mAh battery allows it to last hours on a single charge, and the small display in the middle of the fan's blades show its remaining battery level.
Blink Mini 2 security cameras (two-pack) for $35 (50 percent off): Blink makes some of our favorite security cameras, and the Mini 2 is a great option for indoor monitoring. It can be placed outside with the right weatherproof adapter, but since it needs to be plugged in, we like it for keeping an eye on your pets while you're away and watching over entry ways from the inside.
JBL Go 4 portable speaker for $40 (20 percent off): The Go 4 is a handy little Bluetooth speaker that you can take anywhere you go thanks to its small, IP67-rated design and built-in carrying loop. It'll get seven hours of playtime on a single charge, and you can pair two together for stereo sound.
Apple Mac mini (M4) for $499 ($100 off): If you prefer desktops over laptops, the upgraded M4 Mac mini is one that won’t take up too much space, but will provide a ton of power at the same time. Not only does it come with an M4 chipset, but it also includes 16GB of RAM in the base model, plus front-facing USB-C and headphone ports for easier access.
Apple Watch Series 11 for $389 ($10 off): The latest flagship Apple Watch is our new pick for the best smartwatch you can get, and it's the best all-around Apple Watch, period. It's not too different from the previous model, but Apple promises noticeable gains in battery life, which will be handy for anyone who wants to wear their watch all day and all night to track sleep.
Samsung EVO Select microSD card (256GB) for $23 (15 percent off): This Samsung card has been one of our recommended models for a long time. It's a no-frills microSD card that, while not the fastest, will be perfectly capable in most devices where you're just looking for simple, expanded storage.
Anker Soundcore Select 4 Go speaker for $26 (26 percent off, Prime exclusive): This small Bluetooth speaker gets pretty loud for its size and has decent sound quality. You can pair two together for stereo sound as well, and its IP67-rated design will keep it protected against water and dust.
Roku Streaming Stick Plus 2025 for $29 (27 percent off): Roku makes some of the best streaming devices available, and this small dongle gives you access to a ton of free content plus all the other streaming services you could ask for: Netflix, Prime Video, Disney+, HBO Max and many more.
Anker 622 5K magnetic power bank with stand for $34 (29 percent off, Prime exclusive): This 0.5-inch thick power bank attaches magnetically to iPhones and won't get in your way when you're using your phone. It also has a built-in stand so you can watch videos, make FaceTime calls and more hands-free while your phone is powering up.
Amazon Fire TV Stick 4K Max for $40 (33 percent off): Amazon's most powerful streaming dongle supports 4K HDR content, Dolby Vision and Atmos and Wi-Fi 6E. It also has double the storage of cheaper Fire TV sticks.
Anker Soundcore Space A40 for $45 (44 percent off): Our top pick for the best budget wireless earbuds, the Space A40 have surprisingly good ANC, good sound quality, a comfortable fit and multi-device connectivity.
Anker MagGo 10K power bank (Qi2, 15W) for $63 (22 percent off, Prime exclusive): A 10K power bank like this is ideal if you want to be able to recharge your phone at least once fully and have extra power to spare. This one is also Qi2 compatible, providing up to 15W of power to supported phones.
Levoit Core 200S smart air purifier for $70 ($20 off, Prime exclusive): This compact air purifier cleans the air in rooms up to 140 square feet and uses a 3-in-1 filter that removes microscopic dust, pollen and airborne particles. It has a mobile app that you can use to set runtime schedules, and it works with Alexa and Google Assistant voice commands.
Amazon Fire TV Cube for $100 (29 percent off): Amazon's most powerful streaming device, the Fire TV Cube supports 4K, HDR and Dolby Vision content, Dolby Atmos sound, Wi-Fi 6E and it has a built-in Ethernet port. It has the most internal storage of any Fire TV streaming device, plus it comes with an enhanced Alexa Voice Remote.
Shark AI robot vacuum with self-empty base for $230 (58 percent off, Prime exclusive): A version of one of our favorite robot vacuums, this Shark machine has strong suction power and supports home mapping. The Shark mobile app lets you set cleaning schedules, and the self-empty base that it comes with will hold 30 days worth of dust and debris.
Levoit LVAC-300 cordless vacuum for $250 ($100 off, Prime exclusive): One of our favorite cordless vacuums, this Levoit machine has great handling, strong suction power for its price and a premium-feeling design. Its bin isn't too small, it has HEPA filtration and its battery life should be more than enough for you to clean your whole home many times over before it needs a recharge.
Amazon Kindle Scribe for $300 (25 percent off): The latest Kindle Scribe is one of the best E Ink tablets you can get, particularly if you want a slab that can handle both note-taking and e-reading well. Its edge is providing users full access to the Kindle ebook store and the ability to take notes to accompany reading material. This model has 16GB of storage and includes the Premium Pen.
XReal One Pro AR glasses for $649 (16 percent off): The latest from XReal, these smart glasses can let you use almost any device, including your smartphone, with a large virtual display. Their 1080p Micro-OLED screens are bright and sharp, plus they're pretty comfortable to wear.
Nintendo Switch 2 for $449: While not technically a discount, it's worth mentioning that the Switch 2 and the Mario Kart Switch 2 bundle are both available at Amazon now, no invitation required. Amazon only listed the new console for the first time in July after being left out of the initial pre-order/availability window in April. Once it became available, Amazon customers looking to buy the Switch 2 had to sign up to receive an invitation to do so. Now, that extra step has been removed and anyone can purchase the Switch 2 on Amazon.
This article originally appeared on Engadget at https://www.engadget.com/deals/the-best-october-prime-day-deals-you-can-get-right-now-early-sales-on-tech-from-apple-amazon-samsung-anker-and-more-050801285.html?src=rss
Amazon's Kindle Scribe drops to a record-low price for Prime Day
We're only one week away from Amazon Prime Day and the sales have already started. The deals include Amazon's devices, from Kindle tablets to Echo speakers.
Take the Amazon Kindle Scribe, which has dropped to $300 from $400. The 25 percent discount brings the second-generation device back down to its all-time low price. This deal is available on the 16GB model in Tungsten (dark grey) and doesn't include Kindle Unlimited. If you want more storage, the 32GB and 64GB versions are also 25 and 24 percent off, respectively.
The Kindle Scribe is our pick for best e-reader E Ink tablet. It scored an 86 in our review, thanks to the smooth reading and writing experience it provides. We were also fans of its sleek and slender design. It has a 10.2-inch screen and offers front lights that auto-adjust to your needs. Our biggest gripe was the cost which, thanks to this deal, is less of an issue.
This article originally appeared on Engadget at https://www.engadget.com/deals/amazons-kindle-scribe-drops-to-a-record-low-price-for-prime-day-134701307.html?src=rss
EA confirms it will go private in $55 billion acquisition
Beleaguered video game giant Electronic Arts has agreed to a $55 billion acquisition that will take the company private. Saudi Arabia Public Investment Fund (PIF), Silver Lake and Affinity Partners have reached a deal to buy EA, the company said in a statement.
Some details about the proposed arrangement surfaced over the weekend. The deal, the largest-ever leveraged buyout, according to Bloomberg, marks the end of EA's 35-year run as a publicly traded company. “Our creative and passionate teams at EA have delivered extraordinary experiences for hundreds of millions of fans, built some of the world’s most iconic IP, and created significant value for our business," EA CEO Andrew Wilson said in a statement. "This moment is a powerful recognition of their remarkable work."
EA hasn't been immune from the forces that have upended the games industry. Last year, the company laid off more than 650 employees, which it labeled an attempt at "streamlining" its business. This year, the company has canceled an upcoming Black Panther game and closed the studio behind it, and has reportedly "shelved" its Need For Speed Franchise.
The company said it expects the deal to close sometime in the first quarter of 2027, noting that it will need regulatory approvals. EA said that Wilson will continue as CEO and that its Redwood City, California headquarters will not change. In a statement Silver Lake Co-CEO Egon Durban said that the group plans to "invest heavily" in the video game company and help it "expand its reach worldwide." Silver Lake is also part of the forthcoming deal to spin off the US version of TikTok into a new, American-owned company.
This article originally appeared on Engadget at https://www.engadget.com/ea-confirms-it-will-go-private-in-55-billion-acquisition-133841614.html?src=rss
Prime Day deals include up to 58 percent off Shark robot vacuums
Ahead of the Amazon Big Deal Days event (aka Prime Day in October), a tasty deal on a Shark robot vacuum has popped up. You'll need to be a Prime member to take advantage of the offer on the Shark AV2501S AI Ultra robot vacuum, but if you are, you can get the device for over half off. The discount drops the price from $550 to $230.
That means you can snap up the robot vacuum for $320 below list price. The discount marks a record low for this model.
Shark offers several variations of its AI Ultra robot vacuums. There are small variations between them, and a different model is our pick for the best robot vacuum for most people. In general, you can expect solid cleaning performance from these devices, along with accurate home mapping and an easy-to-use app.
The model that's on sale here is said to run for up to 120 minutes on a single charge, which should be enough to clean an entire floor in a typical home. The self-emptying, bagless vacuum can store up to 30 days worth of dirt and debris in its base. Shark says it can capture 99.97 percent of dust and allergens with the help of HEPA filtration.
If you'd rather plump for a model that's able to mop your floors too, you're in luck: a Shark Matrix Plus 2-in-1 vacuum is on sale as well. At $300 for Prime members, this vacuum is available for $400 (or 57 percent) off the list price. Its mopping function can scrub hard floors 100 times per minute. You can also trigger the Matrix Mop function in the app for a deeper clean. This delivers 50 percent better stain cleaning in targeted zones, according to Shark.
This article originally appeared on Engadget at https://www.engadget.com/deals/prime-day-deals-include-up-to-58-percent-off-shark-robot-vacuums-171836574.html?src=rss
Prime Day deal: The Roku Streaming Stick Plus drops to $29
The Roku Streaming Stick Plus is on sale for just $29. That's a discount of 27 percent and the lowest we've ever seen it.
Roku has held the top spot in the TV OS market for years thanks to its user-friendly interface, an affordable range of streaming devices and its own lineup of TVs. We picked the Streaming Stick Plus as the best streaming device for free and live content, thanks in large part to The Roku Channel app that accompanies it. The Roku Channel features over 500 free TV channels with live news, sports coverage and a rotating lineup of TV shows and movies.
In our hands-on review of the Roku Streaming Stick Plus, we thought it was perfect for travel thanks to its small size and the fact that it can be powered by your TV's USB port, nixing the need for a wall adapter. Menu navigation and opening or closing apps won't happen at quite the same speeds as more expensive streamers, but it's quick enough for what is ultimately a pretty low-cost option. The Wi-Fi range on this one is also weaker than Roku's pricier devices, but unless you are placing it exceedingly far from your router, it shouldn't be an issue.
The Roku Streaming Stick Plus supports both HD and 4K TVs, as well as HDR10+ content. It doesn't support Dolby Vision, however; for that you'll need to upgrade to Roku's Streaming Stick 4K or Roku Ultra. It comes with Roku's rechargeable voice remote with push-to-talk voice controls. Roku's remote can also turn on your TV and adjust the volume while you're watching.
If you've been thinking about getting a Roku device, or you already love the platform and want a compact and convenient way to take it with you when you travel, then this sale provides a great opportunity.
This article originally appeared on Engadget at https://www.engadget.com/deals/prime-day-deal-the-roku-streaming-stick-plus-drops-to-29-134656933.html?src=rss
X will appeal Indian court ruling allowing 'arbitrary takedown orders'
X is challenging an Indian court order it says would require it to comply with millions of takedown requests without due process. In a statement, the company said it was "deeply concerned" about the ruling, which allows police officers "to issue arbitrary takedown orders through a secretive online portal" called the Sahyog.
X's challenge comes after the Karnataka High Court dismissed X's initial challenge of the Sahyog portal last week. As the BBC notes, the Sahyog "describes itself as a tool to automate the process of sending government notices to content intermediaries such as X and Facebook." But X has labeled it a "censorship portal," saying that it allows "millions" of officials, including local police officers, to order content to be removed.
X is deeply concerned by the recent order from the Karnataka court in India, which will allow millions of police officers to issue arbitrary takedown orders through a secretive online portal called the Sahyog. This new regime has no basis in the law, circumvents Section 69A of…
— Global Government Affairs (@GlobalAffairs) September 29, 2025
"The Sahyog enables officers to order content removal based solely on allegations of 'illegality,' without judicial review or due process for the speakers, and threatens platforms with criminal liability for non-compliance," X wrote in a statement. "We will appeal this order to defend free expression."
X has a history of challenging the Indian government on content moderation issues. The company previously challenged orders requiring it to block specific posts and accounts in 2024 and 2022. The platform also clashed with the Indian government under Jack Dorsey's leadership. Officials threatened to jail Twitter employees in 2021 after the company reversed a decision to block accounts amid widespread protests.
This article originally appeared on Engadget at https://www.engadget.com/social-media/x-will-appeal-indian-court-ruling-allowing-arbitrary-takedown-orders-120000846.html?src=rss
The best iPad accessories for 2025
If you want to upgrade your iPad experience, finding the best iPad accessories can do the trick. It's certainly much cheaper than buying a new tablet, and often times, all you need is a few well-chosen accessories to make a familiar device feel fresh again. From a slim folio for iPad that protects your screen to a Magic Keyboard case that turns it into a full-blown laptop alternative, there are dozens of solid accessories to choose from. Accessories aren't just about looks, either. Swapping out worn pencil tips, grabbing a longer Lightning cable or finding the right stand can make a big difference in how smoothly you move through your favorite apps. We've tested a plethora of accessories over the years, and these are the best iPad accessories you can get right now.
Before you splurge on a bunch of accessories, you should double check which iPad generation you own. There are a couple of ways to do this: first, you can check the back of your tablet for its model number, which will start with an “A” and end with a series of numbers. You can also go into Settings on your iPad, then General and look up the model number in the top section. If you see a series of letters and numbers with a slash (“/“) in it, just tap that to reveal the iPad’s true model number.
The most important iPad details to consider before buying accessories are the charging port, screen size and Apple Pencil compatibility. Most iPads now have USB-C charging, save for the now discontinued 9th-gen iPad that still requires a Lightning cable. Make sure to double check your iPad’s screen size before buying a folio for iPad or a screen protector — especially if you have the new iPad Air 11-inch. As for the Apple Pencil, check out this guide to how to choose the right Apple Pencil (and even replacement pencil tips) to make sure you're getting the right one for the tablet you have.
iPad cases
iPad stands and docks
iPad keyboards and mice
iPad stylus
iPad chargers
iPad screen protectors
If you’re a heavy user of the Apple Pencil or some other stylus, you should consider getting a screen protector for your new iPad. They pull double-duty: Not only do they act as a first line of defense if your iPad goes careening onto concrete, but they can also enhance the digital drawing and writing experience. Using a stylus on an iPad is strange at first because gliding the stylus nib over a glass surface feels nothing like “normal” writing. Matte screen protectors can get closer to replicating the pen-on-paper experience, and they also prevent the stylus nib from wearing down as quickly.
This article originally appeared on Engadget at https://www.engadget.com/computing/accessories/best-ipad-accessories-130018595.html?src=rss
Agents at Work
830 jobs, 19,000 tasks, 1 million conversations, $10 trillion in wages
YouTube Music is testing AI hosts that tell stories, trivia and commentary
YouTube Music is getting an AI powered feature. The music streaming service is testing AI hosts, just because. Google has introduced YouTube Labs, a way for users to test experimental features. It's […]
Thank you for being a Ghacks reader. The post YouTube Music is testing AI hosts that tell stories, trivia and commentary appeared first on gHacks Technology News.
F-Droid criticizes Google's anti-sideloading policy, calls for antitrust scrutiny
F-Droid has hit out at Google for its controversial app distribution policy changes. It could spell the end of the open source app ecosystem on Android. Google sent shockwaves among the developer […]
Thank you for being a Ghacks reader. The post F-Droid criticizes Google's anti-sideloading policy, calls for antitrust scrutiny appeared first on gHacks Technology News.
Microsoft Photos will use AI to auto-categorize your photos
How would you like it if Microsoft Photos automatically categorized your photos? Well, that's exactly what Microsoft is testing across the Insider Program. The Redmond company's latest experimental feature will allow Microsoft […]
Thank you for being a Ghacks reader. The post Microsoft Photos will use AI to auto-categorize your photos appeared first on gHacks Technology News.
‘The Last Starship’ Picks Up on Two of the Biggest Missed Opportunities in Modern ‘Star Trek’
The new 'Star Trek' comic is so much more interesting than its headlining return of Captain Kirk.
OpenAI’s New Social Network Is Reportedly TikTok If It Was Just an AI Slop Feed
If you thought social content was trash already, just wait, the worst is yet to come.
OpenAI Rolls Out ChatGPT’s Ability to Buy Stuff for You
OpenAI’s new Instant Checkout lets users buy products from Etsy and soon millions of Shopify merchants.
Leonardo DiCaprio on the Surprising ‘Star Wars’ Influences in His New Film
'One Battle After Another,' co-starring 'Star Wars' star Benicio del Toro, is now in theaters.
How Many Streaming Subscribers Did Disney Lose After Suspending Kimmel?
The boycott was real.
Americans Are Already Sick of RFK Jr., Poll Finds
A recent Quinnipiac University poll shows that a majority of Americans disapprove of Kennedy's performance as the U.S. Secretary of Health and Human Services.
The ‘Star Trek’ Anniversary Float Will Put Its Most Beloved Location on Parade
The annual New Year's Day Rose Parade in Pasadena, CA will boldly go with a themed tribute to the show's 60th anniversary.
The OnePlus 15 Is One of the First Phones That Could Kick the iPhone 17 Pro’s Ass
It's certainly got more graphics performance than Apple's latest and greatest chip.
The New ‘Shelby Oaks’ Trailer Is a Suspenseful Trip Back to Found Footage Frights
Chris Stuckmann's horror feature hits theaters October 24.
Werner Herzog on AI-Generated Movies: ‘They Look Completely Dead’
Herzog spoke with Conan O'Brien about the concept of truth and why AI is a "nemesis."
Have a Cow, Man: They’re Making a Second ‘Simpsons’ Movie
Homer, Marge, Bart, Lisa, and Maggie will return to theaters on July 23, 2027.
‘Superman’ Is HBO Max’s Biggest Movie Launch Since ‘Barbie’
James Gunn's DC blockbuster racked up 13 million views in its first 10 days on the Warner Bros. streaming platform.
Physicists Smash Record With Magnetic Field 700,000 Times Stronger Than Earth’s
The technology could have promising applications for nuclear fusion research and other commercial uses, the researchers said.
The ADL Defended Elon Musk’s ‘Awkward Gesture.’ Now He Calls It a Hate Group
"The ADL hates Christians, therefore it is is a hate group," Musk tweeted Sunday.
Our 9 Favorite Easter Eggs in the ‘Spaceballs 2’ Cast Photo
From the obvious 'Star Wars' joke to some deeper cuts from the original, 'Spaceballs 2' is already having a laugh.
Electronic Arts Is Going Private in a $55 Billion Jared Kushner–Saudi Takeover
Kushner's group of investors is ready to remake the gaming giant in their own image.
The New ‘Solar Opposites’ Trailer Teases a Perfectly Unhinged Final Season
The aliens are broke, saving the world, and aggressively pursuing family bonding in the latest season 6 trailer.
Trump Shares Bizarre AI Video Promising Magic ‘Medbeds’ for Everyone
In a since-deleted Truth Social post, an AI version of Trump promised every American would soon have access to their very own medbed.
Double Storm Threat Expected to Unleash Dangerous Surf Along U.S. East Coast
The National Hurricane Center is watching these two storms closely as they encroach on the U.S.
Trump Is Threatening Tariffs on Foreign-Made Films Again
The President has reiterated a threat he first made back in May, and once again, without any details as to how such a threat could be acted upon in the first place.
The Rise of Self-Healing Web Apps
Self-healing web applications are designed to adapt to changing environments and keep functioning even when parts of the system fail. They don’t just catch errors, they actively work around them.
Why Product Teams Risk Surrendering Judgment to Algorithms
AI is tremendously helpful in the hands of a skilled operator. It can accelerate research, generate insights, and support better decision-making. But here’s what the AI evangelists won’t tell you: it can be equally damaging when fundamental AI risks are ignored.
The main risk is a gradual transfer of product strategy from business leaders to technical systems—often without anyone deciding this should happen. Teams add “AI” and often report more output, not more learning. That pattern is consistent with long-standing human-factors findings: under time pressure, people over-trust automated cues and under-practice independent verification, which proves especially dangerous when the automation is probabilistic rather than deterministic (Parasuraman & Riley, 1997; see all sources listed below). That’s not a model failure first; it’s a system and decision-making failure that AI accelerates.
The HackerNoon Newsletter: Can ChatGpt Outperform the Market? Week 7 (9/29/2025)
9/29/2025: Top 5 stories on the HackerNoon homepage!
Can ChatGpt Outperform the Market? Week 7
Complete Rebalance Incoming...
I Benchmarked 9 AI Models for Candidate Screening—Then Switched from GPT-4o to Grok-4
Topliner uses AI to assess candidate relevance for executive search projects. GPT-4o is among the sharpest knives in the drawer, but it sometimes goes rogue. xAI’s new Grok-4 Fast Reasoning model promised speed, affordability, and smart reasoning.
Key Elements of Business Continuity and Disaster Recovery for Healthcare
When a healthcare organization’s mission-critical systems go down, it can leave patients at risk and potentially harm the hospital’s reputation within the community. That’s why it’s critical for the organization’s leadership teams to have a well-defined response and recovery plan in place so that the organization can resume operations quickly and with minimal disruption to patients and clinicians. To recover quickly, healthcare systems must have a robust and continuously tested business continuity and disaster recovery plan in place. Not only that, but leadership teams must also ensure that…
No Vacancy
We forge ahead with scientific quests,
but even zealous efforts hit a wall
trying to make what nature most detests:
nothing at all.
It seems like such a little thing to do—
removing every molecule. We find
that though we displace almost all, a few
remain behind.
It takes attentive planning and robust
equipment in a lab to do the chore
of pumping vacuum pressure down to just
a millitorr.
The stalwart researcher persists and loses
sleep, but can’t reach perfection—I’m afraid
the universe still stubbornly refuses
to be unmade.
Even in deepest space, a cubic meter
contains some particles. We must assess
there is no void, although conditions teeter
on emptiness.
The quantum mysteries will vex and weary
the brightest mind, the sharpest physicist.
True nothingness, while wonderful in theory,
does not exist.
Google Stax Aims to Make AI Model Evaluation Accessible for Developers
Google Stax is a framework designed to replace subjective evaluations of AI models with an objective, data-driven, and repeatable process for measuring model output quality. Google says this will allow AI developers to tailor the evaluation process to their specific use cases rather than relying on generic benchmarks.
By Sergio De SimoneOWASP Flags Tool Misuse as Critical Threat for Agentic AI
Earlier this year OWASP released guidance for Agentic AI security called Agentic AI - Threats and Mitigations. The document highlights the unique challenges involved in securely deploying this emerging technology and suggests mitigations and architectural patterns for defense.
By Patrick FarryHugging Face Introduces mmBERT, a Multilingual Encoder for 1,800+ Languages
Hugging Face has released mmBERT, a new multilingual encoder trained on more than 3 trillion tokens across 1,833 languages. The model builds on the ModernBERT architecture and is the first to significantly improve upon XLM-R, a long-time baseline for multilingual understanding tasks.
By Robert KrzaczyńskiInstagram Improves Engagement by Reducing Notification Fatigue with New Ranking Framework
Meta has introduced a diversity-aware ranking framework for Instagram notifications. The system applies multiplicative penalties to reduce repetitive alerts from the same creators or product surfaces, improving engagement while maintaining relevance and introducing content variety.
By Leela KumiliArticle: Disaggregation in Large Language Models: The Next Evolution in AI Infrastructure
Large Language Model (LLM) inference faces a fundamental challenge: the same hardware that excels at processing input prompts struggles with generating responses, and vice versa. Disaggregated serving architectures solve this by separating these distinct computational phases, delivering throughput improvements and better resource utilization while reducing costs.
By Anat HeilperMuslim States Join EU Powers In Backing Trump Gaza Plan
Key Muslim nations on Monday lent their weight to backing a plan by US President Donald Trump to end the war in Gaza, even as some Palestinians decried the proposal as a "farce".
Creator Says AI Actress Is 'Piece Of Art' After Backlash
The creator of an AI actress who exploded across the internet over the weekend has insisted she is an artwork, after a fierce backlash from the creative community.
Anthropic Launches New AI Model, Touting Coding Supremacy
US startup Anthropic on Monday announced the launch of its new generative artificial intelligence model, Claude Sonnet 4.5, which it says is the world's best for computer programming.
From Liquidity to Utility: My Vision for Web3 Innovation at TOKEN2049
Blockchain's long-term relevance depends on delivering tangible utility to everyday users, not merely speculative trading opportunities.
From Paper Waste to Green Tech: How PTOP's Mobicard Helps Save Trees and Improve Efficiency Worldwide
As Peer To Peer Network continues to expand its MobiCard platform, including new features for large companies to instantly create cards for thousands of employees, the message is clear. Networking no longer needs to come at the expense of the planet.
Building Eco-Friendly Finance: How Bitcoin.ℏ Strives to Balance Profitability with Sustainability and Planet Care
Reports state that in 2023 alone, the most resource-intensive cryptocurrency consumed 154.9 billion kilowatt hours(kWh) of electricity, exceeding the energy consumption of 167 countries combined. This escalating demand for energy has intensified the world's quest for a sustainable solution.
Help Hope Solutions Celebrates Almost 20 Years Of Supporting Children With Delayed Developmental Disorders
According to the WHO, one in 127 adults was diagnosed with autism in 2021. Although awareness of developmental delays has been increasing in the past few years, effective therapeutic interventions remain out of reach for many.
Leading with Purpose, Personal Touch, and Transparency: Linked Logistics' Role as a One-Stop Global Logistics Solution
The company, founded by Grady Lynch, runs on the principles of faith, integrity, and relationships, and every shipment, every call, every customer interaction reflects that foundation.
Inside Alteia Fund: Kevin Ramsamy's Vision for Africa's Trade Finance Future
From food security to fintech, the CEO outlines why uncorrelated returns are drawing global institutional capital.
Taliban Shut Down Communications Across Afghanistan
Taliban authorities on Monday imposed a nationwide shutdown of communications, weeks after they began severing fibre optic connections to prevent "vice".
Electronic Arts To Be Bought By Saudi-led Consortium For $55 Bn
Video game giant Electronic Arts, known for The Sims and FIFA games, announced Monday it would be acquired for $55 billion by a consortium led by Saudi Arabia's Public Investment Fund.
Trump Meets Democrats In Last-gasp Talks Before US Government Shutdown
President Donald Trump meets key congressional Republicans and Democrats at the White House on Monday, hoping to revive stalled spending negotiations on the eve of a looming US government shutdown.
What to Do Immediately After an Uber Accident in Miami: A Step-by-Step Checklist
According to the Centers for Disease Control and Prevention (CDC), car accidents remain one of the leading causes of traumatic brain injuries in the U.S., making it essential to seek medical attention even if you initially feel fine.
Madagascar Protests Reignite As Police Launch Tear Gas
Thousands gathered for renewed anti-government demonstrations Monday in several cities across Madagascar including the capital where security forces dispersed crowds with tear gas, AFP journalists saw.
Jaguar Land Rover To Partly Resume Output After Cyberattack
Jaguar Land Rover on Monday said it will partially restart production in the coming days after a damaging cyberattack halted operations at its UK factories.
International Paralympic Committee Lifts Partial Suspensions Of Russia, Belarus
Ukraine reacted with outrage on Saturday after the International Paralympic Committee lifted a partial suspension of Russia and Belarus imposed since Moscow's incursion into Ukraine in 2022.
UK Plans Tougher Rules For Migrants Seeking To Stay In Country
Britain's interior minister Monday proposed tough new rules for migrants seeking to settle in the UK, as the ruling Labour party bolstered its fight against the hard right at its annual conference.
Nearly 200,000 BMWs, Toyota Supras Recalled for Starter Fire Hazard
Nearly 200,000 BMW and Toyota Supra vehicles recalled in the U.S. due to a fire risk from faulty engine starters.
Apple AR Glasses Launch Date Rumours: Can This Gadget Replace Smartphones?
Apple's AR glasses may debut in 2026 or 2027 as iPhone-linked smart specs, offering early features but falling short of replacing smartphones.
President Nelson Dies at 101: Inside His Net Worth, Best-Selling Books, and Career as a Renowned Heart Surgeon
President Russell M. Nelson dies at 101. A look at his net worth, best-selling books, health legacy, and his career as a world-renowned heart surgeon.
Separating Fact from Fiction: Where Was Charlie Kirk's Wife During the Shooting?
Rumours that Charlie Kirk's wife witnessed his fatal shooting at Utah Valley University have gone viral, but verified reports place her in Phoenix at the time of the incident.
Who Is Bonnie Blue's Husband and Why His Mum Says He Deserves a Cut of the OnlyFans Star's $45m Fortune
Bonnie Blue's $45m fortune faces a bitter dispute as her estranged husband's mother insists he deserves a share of the OnlyFans star's growing wealth.
NFL Trade Rumors: Russell Wilson Could Go to These Teams After Jaxson Dart Takeover as Starting QB
Jaxson Dart's promotion to QB1 raises questions about Russell Wilson's future with the Giants. This means trade rumors are swirling around the former Super Bowl champ.
NBA Trade Rumors: Philadelphia 76ers, Quentin Grimes Struggle to Reach Long-Term Deal
Quentin Grimes may accept the Sixers' qualifying offer after stalled contract talks, leaving his future uncertain despite his key role for the 2025-26 season.
Young Thug Explains Why He Has to 'Dumb Down' His Fashion Amid Gay Rumors
oung Thug addressed long-running sexuality rumors and revealed why he toned down his bold fashion style, citing public pressure and misunderstanding of his creative vision.
Debunking 5 myths about cloud computing for small business (Sponsored)
The Algorithmic X-Men
We've assembled our team of Algorithmic X-Men, seven heroes mapped to seven dependable workhorses of machine learning.
Python for Data Science (Free 7-Day Mini-Course)
Want to learn Python for data science? Start today with this beginner-friendly mini-course packed with bite-sized lessons and hands-on examples.
Revolut eyes dual IPO listing in London and New York, reports say
无摘要
Europe’s fastest-growing defence startup teams in 2025
无摘要
Ex-Sequoia partner Matt Miller launches $400m fund Evantic
无摘要
Nato Innovation Fund chair Klaus Hommels steps down following criticism
无摘要
16 early-stage Scottish startups to watch
无摘要
Is your Oura Ring losing battery life? Here's the solution
The bad news is Oura Ring 4's battery life is weakening, one year after it debuted. The good news is Oura is fixing the issue.
Best early October Prime Day 2025 PC gaming deals: Save big on laptops and accessories
October Prime Day is just around the corner, and you can already find great discounts on gaming desktops, laptops, and accessories at Amazon.
Best early October Prime Day TV deals 2025: All time low prices from Samsung, LG, and more
Amazon's October Prime Day sale is just a little over two weeks away, but you can already find steep markdowns on both premium and entry-level TVs from brands like Samsung, LG, Hisense, and more.
Microsoft just added AI agents to Word, Excel, and PowerPoint - how to use them
Available for Microsoft 365 Copilot users, the new agents will help you create, edit, and analyze your documents, spreadsheets, and presentations.
I finally found a high-quality multitool for less than $30
I usually rely on pricier multitools, but the NexTool E1 impressed me with its rugged build and competitive price.
Our expert ditched his Meta Ray-Bans for these smart glasses, and they come with an Echo Spot for only $120
Snag a pair of Echo Frames 3rd Gen and an Echo Spot on sale for less than the price of the frames alone. Now that's a good deal.
Amazon's Vega OS is being updated to Fire TVs this year - here's why that's a big deal
Years in the making, the proprietary Linux OS that replaces Android-based FireOS is expected to be announced at next week's event. Here's everything we know so far.
Weak Wi-Fi? How to get wired internet at home without running an Ethernet cable
If you have cable outlets, you can convert them into high-speed internet ports using a simple adapter. No rewiring needed.
How to export your Snapchat Memories ASAP - before Snap makes you pay up
I've spent nearly a decade saving thousands of snaps to Memories. Now that Snapchat wants me to pay, I'm exporting and leaving.
Amazon owes Prime customers up to $51 each - how to get your share
The FTC says Amazon used deceptive tactics to get customers to sign up for Prime and then made it difficult for them to cancel.
Why I put my Bose QuietComfort away soon after trying these over-ear headphones
On paper, the Bowers & Wilkins Px8 S2 may look eerily similar to the midrange Px7 S3. But the price increase is greatly justified.
Best Amazon Prime Day phone deals 2025: My 15 favorite sales ahead of October
We sifted through Amazon's October Prime Day listings to find the best early deals on top-rated phones we've tested and recommend.
Best tablets for note-taking 2025: My hand-picked options for writing notes
We tested the top note-taking tablets available to find the best options for studying, meetings, and writing. These are our favorites devices.
Best early Amazon Prime Day EcoFlow deals 2025: My 7 favorites sales ahead of October
Prime Day is next week, but here are some of the best early deals on top-quality portable power stations from EcoFlow.
Is this Arch distro the 'ultimate' Linux? That depends on your GPU
You'll find various versions of Ultimate Edition Arch Linux for different needs - gaming, development, older hardware, and even a minimal installation. Here's where the distro really shines.
Best early October Prime Day Anker deals 2025: All-time-low prices on power banks, chargers, and more
Prime Day is next week, but here are some of the best early Prime Day deals I've found on top-quality power banks and chargers from Anker.
This premium Samsung gaming monitor is nearly 50% off
Samsung's Odyssey OLED Curved Monitor is a beast of a display. And ahead of Amazon's October Prime Day, it's only $880.
Best Costco deals to compete with Amazon Prime Day 2025: My favorite sales so far
October Prime Day is coming, but you can find great deals at Costco right now.
Claude Sonnet 4.5 could be your next breakthrough coding tool - how to access it today
Along with the new model, Anthropic updated its Claude Code offering.
ChatGPT can buy stuff for you now - forever changing online shopping
Transactions between AI agents are on the rise. Here's how to try ChatGPT's new Instant Checkout feature.
Is your Oura Ring losing battery life? Here's the solution
The bad news is Oura Ring 4's battery life is weakening, one year after it debuted. The good news is Oura is fixing the issue.
Best early October Prime Day 2025 PC gaming deals: Save big on laptops and accessories
October Prime Day is just around the corner, and you can already find great discounts on gaming desktops, laptops, and accessories at Amazon.
Best early October Prime Day TV deals 2025: All time low prices from Samsung, LG, and more
Amazon's October Prime Day sale is just a little over two weeks away, but you can already find steep markdowns on both premium and entry-level TVs from brands like Samsung, LG, Hisense, and more.
Microsoft just added AI agents to Word, Excel, and PowerPoint - how to use them
Available for Microsoft 365 Copilot users, the new agents will help you create, edit, and analyze your documents, spreadsheets, and presentations.
I finally found a high-quality multitool for less than $30
I usually rely on pricier multitools, but the NexTool E1 impressed me with its rugged build and competitive price.
Our expert ditched his Meta Ray-Bans for these smart glasses, and they come with an Echo Spot for only $120
Snag a pair of Echo Frames 3rd Gen and an Echo Spot on sale for less than the price of the frames alone. Now that's a good deal.
Amazon's Vega OS is being updated to Fire TVs this year - here's why that's a big deal
Years in the making, the proprietary Linux OS that replaces Android-based FireOS is expected to be announced at next week's event. Here's everything we know so far.
Weak Wi-Fi? How to get wired internet at home without running an Ethernet cable
If you have cable outlets, you can convert them into high-speed internet ports using a simple adapter. No rewiring needed.
How to export your Snapchat Memories ASAP - before Snap makes you pay up
I've spent nearly a decade saving thousands of snaps to Memories. Now that Snapchat wants me to pay, I'm exporting and leaving.
Amazon owes Prime customers up to $51 each - how to get your share
The FTC says Amazon used deceptive tactics to get customers to sign up for Prime and then made it difficult for them to cancel.
Why I put my Bose QuietComfort away soon after trying these over-ear headphones
On paper, the Bowers & Wilkins Px8 S2 may look eerily similar to the midrange Px7 S3. But the price increase is greatly justified.
Best Amazon Prime Day phone deals 2025: My 15 favorite sales ahead of October
We sifted through Amazon's October Prime Day listings to find the best early deals on top-rated phones we've tested and recommend.
Best tablets for note-taking 2025: My hand-picked options for writing notes
We tested the top note-taking tablets available to find the best options for studying, meetings, and writing. These are our favorites devices.
Best early Amazon Prime Day EcoFlow deals 2025: My 7 favorites sales ahead of October
Prime Day is next week, but here are some of the best early deals on top-quality portable power stations from EcoFlow.
Is this Arch distro the 'ultimate' Linux? That depends on your GPU
You'll find various versions of Ultimate Edition Arch Linux for different needs - gaming, development, older hardware, and even a minimal installation. Here's where the distro really shines.
Best early October Prime Day Anker deals 2025: All-time-low prices on power banks, chargers, and more
Prime Day is next week, but here are some of the best early Prime Day deals I've found on top-quality power banks and chargers from Anker.
This premium Samsung gaming monitor is nearly 50% off
Samsung's Odyssey OLED Curved Monitor is a beast of a display. And ahead of Amazon's October Prime Day, it's only $880.
Best Costco deals to compete with Amazon Prime Day 2025: My favorite sales so far
October Prime Day is coming, but you can find great deals at Costco right now.
Claude Sonnet 4.5 could be your next breakthrough coding tool - how to access it today
Along with the new model, Anthropic updated its Claude Code offering.
ChatGPT can buy stuff for you now - forever changing online shopping
Transactions between AI agents are on the rise. Here's how to try ChatGPT's new Instant Checkout feature.
Anthropic Launches Claude Sonnet 4.5 with New Coding and Agentic State-of-the-Art Results
Anthropic released Claude Sonnet 4.5 and sets a new benchmark for end-to-end software engineering and real-world computer use. The update also ships concrete product surface changes (Claude Code checkpoints, a native VS Code extension, API memory/context tools) and an Agent SDK that exposes the same scaffolding Anthropic uses internally. Pricing remains unchanged from Sonnet 4 […]
The post Anthropic Launches Claude Sonnet 4.5 with New Coding and Agentic State-of-the-Art Results appeared first on MarkTechPost.
Meet oLLM: A Lightweight Python Library that brings 100K-Context LLM Inference to 8 GB Consumer GPUs via SSD Offload—No Quantization Required
oLLM is a lightweight Python library built on top of Huggingface Transformers and PyTorch and runs large-context Transformers on NVIDIA GPUs by aggressively offloading weights and KV-cache to fast local SSDs. The project targets offline, single-GPU workloads and explicitly avoids quantization, using FP16/BF16 weights with FlashAttention-2 and disk-backed KV caching to keep VRAM within 8–10 […]
The post Meet oLLM: A Lightweight Python Library that brings 100K-Context LLM Inference to 8 GB Consumer GPUs via SSD Offload—No Quantization Required appeared first on MarkTechPost.
The Download: AI to detect child abuse images, and what to expect from our 2025 Climate Tech Companies to Watch list
This is today’s edition of The Download, our weekday newsletter that provides a daily dose of what’s going on in the world of technology. US investigators are using AI to detect child abuse images made by AI Generative AI has enabled the production of child sexual abuse images to skyrocket. Now the leading investigator of child exploitation…
Coming soon: Our 2025 list of Climate Tech Companies to Watch
The need to cut emissions and adapt to our warming world is growing more urgent. This year, we’ve seen temperatures reach record highs, as they have nearly every year for the last decade. Climate-fueled natural disasters are affecting communities around the world, costing billions of dollars. That’s why, for the past two years, MIT Technology…
Interrupting encoder training in diffusion models enables more efficient generative AI
A new framework for generative diffusion models was developed by researchers at Science Tokyo, significantly improving generative AI models. The method reinterpreted Schrödinger bridge models as variational autoencoders with infinitely many latent variables, reducing computational costs and preventing overfitting. By appropriately interrupting the training of the encoder, this approach enabled development of more efficient generative AI, with broad applicability beyond standard diffusion models.
Tool for training-free, real-time video editing and background separation developed
Bar-Ilan University announced today that a team from its Department of Computer Science has developed a breakthrough in video processing that significantly simplifies the separation of foreground objects from their backgrounds, without the need for extensive training or optimization. The new method, called OmnimatteZero, was developed by Dr. Dvir Samuel and Prof. Gal Chechik, who also serves as a senior director of AI research at NVIDIA. The paper is published on the arXiv preprint server.
Robot navigation improves 30% by mimicking how humans spread and forget information
A research team has developed a new "Physical AI" technology that improves the efficiency of multi-robot autonomous navigation by modeling the spread and forgetting of social issues. This achievement is expected to become a key technology for boosting the productivity of autonomous mobile robots in logistics centers, large-scale warehouses, and smart factories.
Building better batteries with amorphous materials and machine learning
Lithium-ion batteries power most electronics, but they have limited energy density—they can store only a certain amount of energy per mass or volume of the battery.
German music body sues OpenAI alleging copyright breaches
A German music rights body is suing OpenAI for copyright infringement, claiming the US tech giant's AI systems reproduced several artists' lyrics without permission, a court heard Monday.
AI is transforming how software engineers do their jobs. Just don't call it 'vibe-coding'
One of the hottest markets in the artificial intelligence industry is selling chatbots that write computer code.
One Tech Tip: OpenAI adds parental controls to ChatGPT for teen safety
OpenAI said Monday it's adding parental controls to ChatGPT that are designed to provide teen users of the popular platform with a safer and more "age-appropriate" experience.
More people are using AI in court, not a lawyer. It could cost you money—and your case
When you don't have the money for a lawyer to represent you in a court case, even judges can understand the temptation to get free help from anywhere—including tapping into generative artificial intelligence (AI).
Generative AI might end up being worthless—and that could be a good thing
In the rush to cash in on the generative artificial intelligence gold rush, one possible outcome of AI's future rarely gets discussed: what if the technology never works well enough to replace your co-workers, companies fail to use AI well or most AI startups simply fail?
Will AI make cheaper personalized insurance premiums possible? Here's why it's a slippery slope
Insurance is based on a principle of solidarity, but that is now being undermined by the algorithms used to build our profiles.
AlloyGPT: Leveraging a language model to aid alloy discovery
Additive manufacturing of alloys has enabled the creation of machine parts that meet the complex requirements needed to optimize performance in aerospace, automotive, and energy applications. Finding the ideal mix of elements to use in these parts when there are countless possible combinations available is a complicated process that has been accelerated by computational tools and artificial intelligence.
AI poses risks to national security, elections and health care. Here's how to reduce them
Artificial intelligence has started to appear almost everywhere in our lives. We enjoy its benefits, such as the speedier discovery of new drugs, or the more personalized medicine that results from its amalgamation of data and expert judgment, often without realizing it. Generative AI, which enables the fast creation of content and automates summarization and translation via tools such as ChatGPT, DeepSeek and Claude, is its most popular form, but AI is much more: its techniques, mainly from machine learning, statistics and logic, help generate decisions and predictions while being guided by goals set by users.
Using generative AI to diversify virtual training grounds for robots
Chatbots like ChatGPT and Claude have experienced a meteoric rise in usage over the past three years because they can help you with a wide range of tasks. Whether you're writing Shakespearean sonnets, debugging code, or need an answer to an obscure trivia question, artificial intelligence (AI) systems seem to have you covered. The source of this versatility? Billions or even trillions of textual data points across the Internet.
Is violent AI-human conflict inevitable?
Are you worried that artificial intelligence and humans will go to war? AI experts are. In 2023, a group of elite thinkers signed onto the Center for AI Safety's statement that "Mitigating the risk of extinction from AI should be a global priority alongside other societal-scale risks such as pandemics and nuclear war."
Regulators struggle to keep up with the fast-moving and complicated landscape of AI therapy apps
In the absence of stronger federal regulation, some states have begun regulating apps that offer AI "therapy" as more people turn to artificial intelligence for mental health advice.
Interrupting encoder training in diffusion models enables more efficient generative AI
A new framework for generative diffusion models was developed by researchers at Science Tokyo, significantly improving generative AI models. The method reinterpreted Schrödinger bridge models as variational autoencoders with infinitely many latent variables, reducing computational costs and preventing overfitting. By appropriately interrupting the training of the encoder, this approach enabled development of more efficient generative AI, with broad applicability beyond standard diffusion models.
Tool for training-free, real-time video editing and background separation developed
Bar-Ilan University announced today that a team from its Department of Computer Science has developed a breakthrough in video processing that significantly simplifies the separation of foreground objects from their backgrounds, without the need for extensive training or optimization. The new method, called OmnimatteZero, was developed by Dr. Dvir Samuel and Prof. Gal Chechik, who also serves as a senior director of AI research at NVIDIA. The paper is published on the arXiv preprint server.
Robot navigation improves 30% by mimicking how humans spread and forget information
A research team has developed a new "Physical AI" technology that improves the efficiency of multi-robot autonomous navigation by modeling the spread and forgetting of social issues. This achievement is expected to become a key technology for boosting the productivity of autonomous mobile robots in logistics centers, large-scale warehouses, and smart factories.
Building better batteries with amorphous materials and machine learning
Lithium-ion batteries power most electronics, but they have limited energy density—they can store only a certain amount of energy per mass or volume of the battery.
German music body sues OpenAI alleging copyright breaches
A German music rights body is suing OpenAI for copyright infringement, claiming the US tech giant's AI systems reproduced several artists' lyrics without permission, a court heard Monday.
AI is transforming how software engineers do their jobs. Just don't call it 'vibe-coding'
One of the hottest markets in the artificial intelligence industry is selling chatbots that write computer code.
One Tech Tip: OpenAI adds parental controls to ChatGPT for teen safety
OpenAI said Monday it's adding parental controls to ChatGPT that are designed to provide teen users of the popular platform with a safer and more "age-appropriate" experience.
More people are using AI in court, not a lawyer. It could cost you money—and your case
When you don't have the money for a lawyer to represent you in a court case, even judges can understand the temptation to get free help from anywhere—including tapping into generative artificial intelligence (AI).
Generative AI might end up being worthless—and that could be a good thing
In the rush to cash in on the generative artificial intelligence gold rush, one possible outcome of AI's future rarely gets discussed: what if the technology never works well enough to replace your co-workers, companies fail to use AI well or most AI startups simply fail?
Will AI make cheaper personalized insurance premiums possible? Here's why it's a slippery slope
Insurance is based on a principle of solidarity, but that is now being undermined by the algorithms used to build our profiles.
AlloyGPT: Leveraging a language model to aid alloy discovery
Additive manufacturing of alloys has enabled the creation of machine parts that meet the complex requirements needed to optimize performance in aerospace, automotive, and energy applications. Finding the ideal mix of elements to use in these parts when there are countless possible combinations available is a complicated process that has been accelerated by computational tools and artificial intelligence.
AI poses risks to national security, elections and health care. Here's how to reduce them
Artificial intelligence has started to appear almost everywhere in our lives. We enjoy its benefits, such as the speedier discovery of new drugs, or the more personalized medicine that results from its amalgamation of data and expert judgment, often without realizing it. Generative AI, which enables the fast creation of content and automates summarization and translation via tools such as ChatGPT, DeepSeek and Claude, is its most popular form, but AI is much more: its techniques, mainly from machine learning, statistics and logic, help generate decisions and predictions while being guided by goals set by users.
Using generative AI to diversify virtual training grounds for robots
Chatbots like ChatGPT and Claude have experienced a meteoric rise in usage over the past three years because they can help you with a wide range of tasks. Whether you're writing Shakespearean sonnets, debugging code, or need an answer to an obscure trivia question, artificial intelligence (AI) systems seem to have you covered. The source of this versatility? Billions or even trillions of textual data points across the Internet.
Is violent AI-human conflict inevitable?
Are you worried that artificial intelligence and humans will go to war? AI experts are. In 2023, a group of elite thinkers signed onto the Center for AI Safety's statement that "Mitigating the risk of extinction from AI should be a global priority alongside other societal-scale risks such as pandemics and nuclear war."
Regulators struggle to keep up with the fast-moving and complicated landscape of AI therapy apps
In the absence of stronger federal regulation, some states have begun regulating apps that offer AI "therapy" as more people turn to artificial intelligence for mental health advice.
Unlock GPU Performance: Global Memory Access in CUDA
Managing memory is one of the most important performance characteristics to consider when writing a GPU kernel. This post walks you through the important...
Streamline Robot Learning with Whole-Body Control and Enhanced Teleoperation in NVIDIA Isaac Lab 2.3
Training robot policies from real-world demonstrations is costly, slow, and prone to overfitting, limiting generalization across tasks and environments. A...
Train a Quadruped Locomotion Policy and Simulate Cloth Manipulation with NVIDIA Isaac Lab and Newton
Physics plays a crucial role in robotic simulation, providing the foundation for accurate virtual representations of robot behavior and interactions within...
Smart Multi-Node Scheduling for Fast and Efficient LLM Inference with NVIDIA Run:ai and NVIDIA Dynamo
The exponential growth in large language model complexity has created challenges, such as models too large for single GPUs, workloads that demand high...
3 Easy Ways to Supercharge Your Robotics Development Using OpenUSD
The increasing demand for robotics is driving the need for physics-accurate simulation at an unprecedented scale. Universal Scene Description (OpenUSD) is key...
Advancing Robotics Development with Neural Dynamics in Newton
Modern robotics requires more than what classical analytic dynamics provides because of simplified contacts, omitted kinematic loops, and non-differentiable...
$55 Billion Deal for Electronic Arts Is Biggest Buyout Ever
Jared Kushner’s private equity firm and Saudi Arabia’s sovereign wealth fund are some of the investors teaming up to take the video game giant private.
YouTube Settles Trump Lawsuit for $24.5 Million
Mr. Trump sued Alphabet, the parent company of YouTube and Google, and other social media companies after the platforms suspended his accounts in 2021.
California’s Gavin Newsom Signs Major AI Safety Law
Gavin Newsom signed a major safety law on artificial intelligence, creating one of the strongest sets of rules about the technology in the nation.
Brazil’s Homegrown Payment System Is Target of Trump Administration
Brazil’s fast and free homegrown digital payment system, PIX, has become wildly popular. The Trump administration says it unfairly undercuts U.S. companies.
Electric Vehicles Face a ‘Pretty Dreadful Year’ in the U.S.
The end of a federal tax credit is expected to push sales of battery-powered cars down but auto experts believe the market will eventually recover.
Real AI Agents and Real Work
The race between human-centered work and infinite PowerPoints
Once-in-a-Lifetime Bird in Flight Eclipse Photo Wins Bird Photographer of the Year
Photographers worldwide entered over 25,000 images into this year's Bird Photographer of the Year competition, and Canadian wildlife and nature photographer Liron Gertsman took the grand prize and £3,000 in cash for his incredible portrait of a frigatebird silhouetted against last year's total solar eclipse.
Are Old Point-and-Shoot Cameras Good, or Is It Just Nostalgia?
Earlier this summer, Chris Niccolls, Jordan Drake, and I browsed the fixed-lens section of KEH's website to see what was available for a reasonable price. With compact, point-and-shoot cameras suddenly exploding in popularity, we were wondering: have they aged well, or is it just nostalgia?
Think Tank Joins Roll Top Bag Craze with New FocusPoint Backpack Series
Roll top backpacks are all the rage right now and Think Tank is joining in on the craze with the FocusPoint RollTop Backpack series, designed for photographers to use as both an everyday carry as well as for travel.
These Rehoused 1950s Todd-AO Cinema Primes Are One-of-a-Kind
The legendary Todd-AO 65mm Prime lenses were originally built in the early 1950s and were used to film some true classics. The American Society of Cinematographers (ASC), Zero Optik, and Keslow Camera have teamed up to bring these lenses back, meticulously restored and rehoused, for modern filmmakers.
CreativeLive is Shutting Down
CreativeLive, one of the original platforms for online photography education, is shutting down on December 31, 2025. The site is now closed to all but current users and will be wholly shuttered at the end of the year.
Kase’s 85mm f/1.4 AF Is an Affordable Full-Frame Portrait Prime
Kase, best known for its photography filters, has launched a new 85mm f/1.4 prime lens for full-frame and APS-C mirrorless cameras. It is designed for portrait photographers, but is also well-suited to other photographic genres.
Trump Once Again Says Movies Made Outside the US Will Face 100% Tariffs
President Donald Trump has once again threatened to impose 100% tariffs on movies made outside of the United States -- but it remains unclear what that would actually look like.
Sony’s New Global Shutter Sensor Captures 105 Megapixels at 100FPS
Sony Semiconductor has announced a new high-speed, high-resolution image sensor called the IMX927, a backside-illuminated and stacked CMOS sensor with a global shutter. While designed for industrial use, it showcases Sony's sensor capabilities and may give some insight into what may trickle down to consumer devices.
The New Vaonis Vespera II X_Edition Smart Telescope Looks Stellar
French smart telescope company Vaonis has announced the Vespera II X_Edition, a new limited edition version of its standard Vespera II smart telescope. The X_Edition promises to embody the mechanical and optical precision for which the Vespera series is known.
Influential Photographer Petra Collins Shoots Selena Gomez’s Wedding
Singer Selena Gomez chose photographer Petra Collins -- whose dreamlike pastel aesthetic and soft-focus film images have become highly influential online -- to shoot her wedding.
7Artisans Launches AF 10mm f/2.8 Ultra-Wide Lens for APS-C Cameras
7Artisans has just announced the AF 10mm f/2.8 prime for APS-C cameras on its Chinese social media platforms. Available in Sony E, Fujifilm X, and Nikon Z mounts, the ultra-wide lens is designed to provide photographers with expansive coverage and minimal distortion in a compact and lightweight form.
Incredible Photos Show First Person to Ski Down Everest Without Oxygen
Breathtaking images show the athlete who became the first person to climb to Everest’s summit and ski back down without supplementary oxygen.
Tilly Norwood is an AI Actress Seeking Hollywood Representation
The creator of an AI-generated actress says she is about to announce which talent agency will represent Tilly Norwood -- the first creation to emerge from the AI studio Xicoia.
Documentary Explores Life and Work of Iconic Photographer Steve Schapiro
The trailer for Steve Schapiro: Being Everywhere has been released, giving fans a first look at the documentary about the iconic American photographer and journalist.
Skinny Dipping Event Puts Ban on Photography After ‘Too Many Cameras’
A skinny dipping event in the UK has been forced to implement what it calls photography restrictions after cameras became too prevalent.
Meta Introduces $5 Monthly Fee for Ad-Free Facebook and Instagram in UK
Facebook and Instagram users in the U.K. will be able to pay a subscription of up to $5.36 (£3.99) per month to use the platforms without ads.
President Trump Deletes AI-Generated News Report About ‘Medbeds’
President Donald Trump has deleted a video he posted to his Truth Social page of an AI-generated Fox News report about a mythical hospital containing "medbeds."
How the Brain Balances Excitation and Inhibition
A healthy brain maintains a harmony of neurons that excite or inhibit other neurons, but the lines between different types of cells are blurrier than researchers once thought.
The post How the Brain Balances Excitation and Inhibition first appeared on Quanta Magazine
The UAE’s University of AI is a pipeline for top tech talent
UAE has secured massive computing power. Its state-backed flagship university is now on the prowl for talent.
How Hurricane Humberto Is Pulling Tropical Storm Imelda Away from the U.S.
In a version of the Fujiwhara effect, Hurricane Humberto is pulling Tropical Storm Imelda eastward and away from the U.S.
Weaver Ants Form Complex Chains to Pull More Than 100 Times Their Weight
When more humans are added to a team, each member accomplishes less work—but teams of weaver ants do better and better as more join
Endangered Sharks Caught in Rare Mating Ritual beneath the Waves
The Food and Drug Administration plans to update the safety label for acetaminophen products, and the strongest storm on Earth this year struck several countries in East and Southeast Asia.
Career jump: Top tips for navigating promotion anxiety
Promotions are a positive sign that you are doing well in your role, but change can bring complicated feelings and you may need some guidance as you enter this next phase.
Read more: Career jump: Top tips for navigating promotion anxiety
Game maker EA to go private after $55bn acquisition deal
EA’s current CEO Andrew Wilson will continue to lead the company following the completion of the deal.
Read more: Game maker EA to go private after $55bn acquisition deal
Hackers demand ransom, leak children’s data in UK nursery cyberattack
The group is threatening more leaks unless a ransom is paid. It also wants parents to sue the nursery.
Read more: Hackers demand ransom, leak children’s data in UK nursery cyberattack
‘AI and ML the main focus in tech right now’
Kathlyn O’Neill discusses her experience in the Leap technology programme and how to upskill in cybersecurity.
Read more: ‘AI and ML the main focus in tech right now’
D-wave lead on how quantum is already transforming industries
D-wave's Murray Thom tells SiliconRepublic.com that the tech is not just a thing of the future.
Read more: D-wave lead on how quantum is already transforming industries
AI start-up Anthropic to triple workforce in major global expansion
The international expansion reflects the growing consumer base for the the start-up's Claude models worldwide.
Read more: AI start-up Anthropic to triple workforce in major global expansion
EU member states endorse Semicon Declaration ahead of EU Chips Act review
Ireland has officially endorsed the Semicon Declaration which was formally presented to the European Commission this morning.
Read more: EU member states endorse Semicon Declaration ahead of EU Chips Act review
Ibec demands better start-up funding support, access to talent
Ireland’s start-up ecosystem is growing, but more can be done, finds a new Ibec report.
Read more: Ibec demands better start-up funding support, access to talent
Claude Sonnet 4.5 is probably the "best coding model in the world" (at least for now)
Anthropic released Claude Sonnet 4.5 today, with a very bold set of claims:
Claude Sonnet 4.5 is the best coding model in the world. It's the strongest model for building complex agents. It’s the best model at using computers. And it shows substantial gains in reasoning and math.
Anthropic gave me access to a preview version of a "new model" over the weekend which turned out to be Sonnet 4.5. My initial impressions were that it felt like a better model for code than GPT-5-Codex, which has been my preferred coding model since it launched a few weeks ago. This space moves so fast - Gemini 3 is rumored to land soon so who knows how long Sonnet 4.5 will continue to hold the "best coding model" crown.
The pricing is the same as the previous Sonnet: $3/million input tokens and $15/million output tokens. This remains significantly cheaper than Claude Opus - $15/$75 - but still quite a bit more than GPT-5 and GPT-5-Codex, both at $1.25/$10.
It really shines with Claude.ai Code Interpreter
The claude.ai web interface (not yet the Claude iPhone native app) recently added the ability for Claude to write and then directly execute code in a sandboxed server environment, using Python and Node.js. I wrote about that in detail three weeks ago.
Anthropic's implementation of this code interpreter pattern is more powerful than ChatGPT's equivalent because it can directly clone code from GitHub and install software packages from NPM and PyPI.
Sonnet 4.5 absolutely shines at using this tool!
Here's my most ambitious experiment with it. I started like this:
Checkout https://github.com/simonw/llm and run the tests with
pip install -e '.[test]'
pytest
It did exactly as I asked, checking out my LLM repo, installing it along with all of its dependencies, and running the tests.
✅ All 466 tests passed in 167.69 seconds (about 2 minutes 47 seconds)
Then I posed it a wildly ambitious challenge:
The tool stores prompts and responses in a SQLite database. It currently uses a single conversation row and a linear collection of responses within that conversation. Experiment with the idea of adding a parent_response_id column to that responses table and using it to model the responses in a conversation as a tree instead. Use new pytest tests which you execute individually (to save time on the whole test suite) to exercise your design. Write a tree_notes.md file which you first populate with your plan and then use for notes as you go along.
As you can see in the resulting conversation transcript, it did a really impressive job of this! It churned away across dozens of different tool calls, eventually outputting the following summary of its activity:
I've successfully experimented with adding tree-structured conversations to the LLM CLI tool. Here's what was accomplished:
✅ Implementation Complete
1. Database Schema Enhancement
- Added
parent_response_idcolumn to theresponsestable via migrationm022_parent_response_id- Enables tree-structured (branching) conversations instead of just linear sequences
- Fully backward compatible - existing data unaffected
2. Comprehensive Utility Module (
tree_utils.py)
- 12 helper functions for tree operations:
- Navigation (children, parent, siblings, paths)
- Analysis (depth, descendants, size)
- Queries (roots, leaves, statistics)
- Visualization (ASCII tree display)
3. Complete Test Suite (
test_tree_conversations.py)
- 16 tests covering all tree operations
- Tests linear chains, branching, multiple roots, forests
- Includes realistic debugging scenario
- All tests pass ✅
4. Integration Verification
- Updated existing migration tests to expect new column
- All 6 migration tests pass
- Total: 22/22 tests passing ✅
📦 Deliverables (7 files in
/mnt/user-data/outputs/)
- README.md - Quick start guide
- IMPLEMENTATION_SUMMARY.md - Complete technical overview
- MIGRATION_CODE.md - Database migration details
- TEST_RESULTS.md - Full test coverage report
- tree_notes.md - Design decisions and development notes
- tree_utils.py - Utility functions module
- test_tree_conversations.py - Test suite
🎯 Key Features
- Multiple branches from any conversation point
- Multiple roots per conversation (forest structure)
- Rich analytics (depth, branching factor, tree size)
- ASCII tree visualization
- Cycle detection for safety
🚀 Ready for Integration
The implementation is production-ready with comprehensive testing. Next steps would be:
- Integrate
tree_utils.pyinto LLM package- Add CLI commands (
llm branch,llm tree)- Update
Response.log()to acceptparent_response_idAll code is tested, documented, and ready to use!
Did I mention I kicked off this experiment by entering prompts on my phone?
I told it to Give me a zip file of everything you have done so far - you can explore the contents of the file it made me in this Gist.
Pelicans
I released llm-anthropic 0.19 adding support for the new model. Then I ran my classic pelican benchmark, first with thinking turned on:
llm -m claude-sonnet-4.5 -o thinking 1 -o thinking_budget 4000 \
'Generate an SVG of a pelican riding a bicycle'
And then without:
llm -m claude-sonnet-4.5 \
'Generate an SVG of a pelican riding a bicycle'
These are pretty good - they are recognizably pelicans! - though not quite as good as GPT-5-Codex which is better at drawing bicycles.
I also tried having it describe this image:
llm -m claude-sonnet-4.5 'describe this image' -a https://static.simonwillison.net/static/2024/pelicans.jpgThis image shows a large gathering of pelicans crowded together along what appears to be a waterfront or shoreline. The birds are densely packed, creating a chaotic scene with dozens of brown pelicans visible. Their distinctive long bills and pouches are clearly visible among the mass of feathered bodies. In the background, there's calm water and a dark tree line or forested area. The lighting suggests this was taken during early morning or late afternoon, giving the scene a golden, warm tone. The pelicans appear to be resting or congregating together, which is typical behavior for these colonial seabirds.
Anthropic are rolling this out everywhere
The release of this model has been very well coordinated. My embargo on talking about it was due to lift at 10am Pacific today, and I got an email from them linking to their blog post at 10am on the dot. It's also already live on OpenRouter and in Cursor and GitHub Copilot and no doubt a whole bunch of other places as well.
Anthropic also shipped a new Claude Code VS Code extension today, plus a big upgrade to the Claude Code terminal app. Plus they rebranded their confusingly named Claude Code SDK to the Claude Agent SDK instead, emphasizing that it's a tool for building agents beyond just customizing the existing Claude Code product. That's available for both TypeScript and Python.
<p>Tags: <a href="https://simonwillison.net/tags/ai">ai</a>, <a href="https://simonwillison.net/tags/generative-ai">generative-ai</a>, <a href="https://simonwillison.net/tags/llms">llms</a>, <a href="https://simonwillison.net/tags/llm">llm</a>, <a href="https://simonwillison.net/tags/anthropic">anthropic</a>, <a href="https://simonwillison.net/tags/claude">claude</a>, <a href="https://simonwillison.net/tags/code-interpreter">code-interpreter</a>, <a href="https://simonwillison.net/tags/llm-tool-use">llm-tool-use</a>, <a href="https://simonwillison.net/tags/llm-pricing">llm-pricing</a>, <a href="https://simonwillison.net/tags/pelican-riding-a-bicycle">pelican-riding-a-bicycle</a>, <a href="https://simonwillison.net/tags/llm-reasoning">llm-reasoning</a>, <a href="https://simonwillison.net/tags/llm-release">llm-release</a></p>
Armin Ronacher: 90%
The idea of AI writing "90% of the code" to-date has mostly been expressed by people who sell AI tooling.
Over the last few months, I've increasingly seen the same idea come coming much more credible sources.
Armin is the creator of a bewildering array of valuable open source projects - Flask, Jinja, Click, Werkzeug, and many more. When he says something like this it's worth paying attention:
For the infrastructure component I started at my new company, I’m probably north of 90% AI-written code.
For anyone who sees this as a threat to their livelihood as programmers, I encourage you to think more about this section:
It is easy to create systems that appear to behave correctly but have unclear runtime behavior when relying on agents. For instance, the AI doesn’t fully comprehend threading or goroutines. If you don’t keep the bad decisions at bay early it, you won’t be able to operate it in a stable manner later.
Here’s an example: I asked it to build a rate limiter. It “worked” but lacked jitter and used poor storage decisions. Easy to fix if you know rate limiters, dangerous if you don’t.
In order to use these tools at this level you need to know the difference between goroutines and threads. You need to understand why a rate limiter might want to"jitter" and what that actually means. You need to understand what "rate limiting" is and why you might need it!
These tools do not replace programmers. They allow us to apply our expertise at a higher level and amplify the value we can provide to other people.
Via lobste.rs
Tags: armin-ronacher, careers, ai, generative-ai, llms, ai-assisted-programming
--- ## 来源: https://stackoverflow.blog/feed/ ### [Making your code base better will make your code coverage worse](https://stackoverflow.blog/2025/09/29/making-your-code-base-better-will-make-your-code-coverage-worse/) Maintaining a minimum of 80% code coverage affects code decisions and not always for the better. --- ## 来源: https://medium.com/feed/@netflixtechblog ### [100X Faster: How We Supercharged Netflix Maestro’s Workflow Engine](https://netflixtechblog.com/100x-faster-how-we-supercharged-netflix-maestros-workflow-engine-028e9637f041?source=rss-c3aeaf49d8a4------2)By Jun He, Yingyi Zhang, Ely Spears
TL;DR
We recently upgraded the Maestro engine to go beyond scalability and improved its performance by 100X! The overall overhead is reduced from seconds to milliseconds. We have updated the Maestro open source project with this improvement! Please visit the Maestro GitHub repository to get started. If you find it useful, please give us a star.
Introduction
In our previous blog post, we introduced Maestro as a horizontally scalable workflow orchestrator designed to manage large-scale Data/ML workflows at Netflix. Over the past two and a half years, Maestro has achieved its design goal and successfully supported massive workflows with hundreds of thousands of jobs, managing millions of executions daily. As the adoption of Maestro increases at Netflix, new use cases have emerged, driven by Netflix’s evolving business needs, such as Live, Ads, and Games. To meet these needs, some of the workflows are now scheduled on a sub-hourly basis. Additionally, Maestro is increasingly being used for low-latency use cases, such as ad hoc queries, beyond traditional daily or hourly scheduled ETL data pipeline use cases.
While Maestro excels in orchestrating various heterogeneous workflows and managing user end-to-end development experiences, users have experienced noticeable speedbumps (i.e. ten seconds overhead) from the Maestro engine during workflow executions and development, affecting overall efficiency and productivity. Although being fully scalable to support Netflix-scale use cases, the processing overhead from Maestro internal engine state transitions and lifecycle activities have become a bottleneck, particularly during development cycles. Users have expressed the need for a high performance workflow engine to support iterative development use cases.
To visualize our end users’ needs for the workflow orchestrator, we create a 5-layer structure graph shown below. Before the change, Maestro reached level 4 but faced challenges to satisfy the user’s needs in level 5. With the new engine design, Maestro is able to power the users to work with their highest capacity and spark joy for end users during their development over the Maestro.
In this blog post, we will share our new engine details, explain our design trade-off decisions, and share learnings from this redesign work.
Architectural Evolution of Maestro
Before the change
To understand the improvements, we will first revisit the original architecture of Maestro to understand why the overhead is high. The system was divided into three main layers, as illustrated in the diagram below. In the sections that follow we will explain each layer and the role it played in our performance optimization.
Maestro API and Step Runtime Layer
This layer offers seamless integrations with other Netflix services (e.g., compute engines like Spark and Trino). Using Maestro, thousands of practitioners build production workflows using a paved path to access platform services . They can focus primarily on their business logic while relying on Maestro to manage the lifecycle of jobs and workflows plus the integration with data platform services and required integrations such as for authentication, monitoring and alerting. This layer functioned efficiently without introducing significant overhead.
Maestro Engine Layer
The Maestro engine serves several crucial functions:
- Managing the lifecycle of workflows, their steps and maintaining their state machines
- Supporting all user actions (e.g., start, restart, stop, pause) on workflow and step entities
- Translating complex Maestro workflow graphs into parallel flows, where each flow is an array of sequentially chained flow tasks, translating every step into a flow task, and then executing transformed flows using the internal flow engine
- Acting as a middle layer to maintain isolation between the Maestro step runtime layer and the underlying flow engine layer
- Implementing required data access patterns and writing Maestro data into the database
In terms of speed, this layer had acceptable overhead but faced edge cases (e.g. a step might be concurrently executed by two workers at the same time, causing race conditions) due to lacking a strong guarantee from the internal flow engine and the external distributed job queue.
Maestro Internal Flow Engine Layer
The Maestro internal flow engine performed 2 primary functions:
- Calling task’s execution functions at a given interval.
- Starting the next tasks in an array of sequential task flows (not a graph), if applicable.
This foundational layer was based on Netflix OSS Conductor 2.x (deprecated since Apr 2021), which requires a dedicated set of separate database tables and distributed job queues.
The existing implementation of this layer introduces an impactful overhead (e.g. a few seconds to tens of seconds overall delays). The lack of strong guarantees (e.g. exactly once publishing) from this layer leads to race conditions which cause stuck jobs or lost executions.
Options to consider
We have evaluated three options to address those existing issues:
- Option 1: Implement an internal flow engine optimized for Maestro specific use cases
- Option 2: Upgrade Conductor library to 4.0, which addresses the overheads and offers other improvements and enhancements compared with Conductor 2.X.
- Option 3: Use Temporal as the internal flow engine
One aspect that influenced our assessment of option two is that Conductor 2 provided a final callback capability in the state machine that was contributed specifically for Maestro’s use case to ensure database synchronization between the Conductor and Maestro engine states. It would require porting this functionality to Conductor 4 though it had been dropped given no other Conductor use cases besides Maestro relied on this. By rewriting the flow engine it would allow removal of several complex internal databases and database synchronization requirements which was attractive for simplifying operational reliability. Given Maestro did not need the full set of state engine features offered by Conductor, this motivated us to consider a flow engine rewrite as a higher priority.
The decision for Temporal was more straightforward. Temporal is optimized towards facilitating inter-process orchestration and would involve calling an external service to interact with the Temporal flow engine. Given Maestro is operating greater than a million tasks per day, many of which are long running, we felt it was an unnecessary source of risk to couple the DAG engine execution with an external service call. If our requirements went beyond lightweight state transition management we might reconsider because Temporal is a very robust control plane orchestration system, but for our needs it introduced complexity and potential reliability weak spots when there was no direct need for the advanced feature set that it offered.
After considering Option 2 and Option 3, we developed more conviction that Maestro’s architecture could be greatly simplified by not using a full DAG evaluation engine and having to maintain the state machine for two systems (Maestro and Conductor/Temporal). Therefore, we have decided to go with Option 1.
After the change
To address these issues, we completely rewrote the Maestro internal flow engine layer to satisfy Maestro’s specific needs and optimize its performance. This new flow engine is lightweight with minimal dependencies, focusing on excelling in the two primary functions mentioned above. We also replaced existing distributed job queues with internal ones to provide a strong guarantee.
The new engine is highly performant, efficient, scalable, and fault-tolerant. It is the foundation for all upper components of Maestro and provides the following guarantees to avoid race conditions:
- A single step should only be executed by a single worker at any given time
- Step state should never be rolled back
- Steps should always eventually run to a terminal state
- The internal flow state should be eventually consistent with the Maestro workflow state
- External API and user actions should not cause race conditions on the workflow execution
Here is the new architecture diagram after the change, which is much simpler with less dependencies:
New Flow Engine Optimization
The new flow engine significantly boosts speed by maintaining state in memory. It ensures consistency by using Maestro engine’s database as the source of truth for workflow and step states. During bootstrapping, the flow engine rebuilds its in-memory state from the database, improving performance and simplifying the overall architecture. This is in contrast to the previous design in which multiple databases had to be reconciled against one another (Conductor’s tables and Maestro’s tables) or else suffer race conditions and rare orphaned job status.
The flow engine operates on in-memory flow states, resembling a write through caching pattern. Updates to workflow or step state in the database also update the in-memory flow state. If in-memory state is lost, the flow engine rebuilds it from the database, ensuring eventual consistency and resolving race conditions.
This design delivers lower latency and higher throughput, avoids inconsistencies from dual persistence, simplifies the architecture, and keeps the in‑memory view eventually consistent with the database.
Maintaining Scalability While Gaining Speed
With the new engine, we significantly boost performance by collocating flows and their tasks on the same node throughout their lifecycle. Therefore, states of a flow and its tasks will stay in a single node’s memory without persisting to the database. This stickiness and locality bring great performance benefits but inevitably impact scalability since tasks are no longer reassigned to a new worker of the whole cluster in each polling cycle.
To maintain horizontal scalability, we introduced a flow group concept to partition running flows into groups. In this way, each Maestro flow engine instance only needs to maintain ownership of groups rather than individual flows, reducing maintenance costs (e.g., heartbeat) and simplifying reconciliation by allowing each Maestro node to load flows for a group in batches. Each Maestro node claims ownership of a group of flows through a flow group actor and manages their entire lifecycle via child flow actors. If ownership is lost due to node failure or long JVM GC, another node can claim the group to resume flow executions by reconciling internal state from Maestro database. The following diagram illustrates the ownership maintenance.
Flow Partitioning
To efficiently distribute traffic, Maestro assigns a consistent group ID to flows/workflows by a simple stable ID assignment method, as shown in the diagram’s Partitioning Function box. We chose this simpler partitioning strategy over advanced ones, e.g. consistent hashing, primarily due to execution and reconciliation costs and consistency challenges in a distributed system.
Since Maestro decomposes workflows into hierarchical internal flows (e.g., foreach), parent flows need to interact with child flows across different groups. To enable this, the maximal group number from the parent, denoted as N’ in the diagram, is passed down to all child flows. This allows child flows, such as subworkflows or foreach iterations, to recompute their own group IDs and also ensures that a parent flow can always determine the group ID of its child flows using only their workflow identifiers.
After a flow’s group ID is determined, the flow operator routes the flow request to the appropriate node. Each node owns a specific range of group IDs. For example, in the diagram, Node 1 owns groups 0, 1, and 2, while Node 3 owns groups 6, 7, and 8. The groups then contain the individual flows (e.g., Flow A, Flow B).
In this design, the group size is configurable and nodes can also have different group size configurations. The following diagram shows a flow group partitioning example while the maximal group number is changed during the engine execution without impacting any existing workflows.
In short, Maestro flow engine shares the group info across the parent and child workflows to provide a flexible and stable partitioning mechanism to distribute work across the cluster.
Queue Optimization
We replaced both external distributed job queues in the existing system with internal ones, preserving the same fault‑tolerance and recovery guarantees while reducing latency and boosting throughput.
For the internal flow engine, the queue is a simple in‑memory Java blocking queue. It requires no persistence and can be rebuilt from Maestro state during reconciliation.
For the Maestro engine, we implemented a database‑backed in‑memory queue that provides exactly‑once publishing and at‑least‑once delivery guarantees, addressing multiple edge cases that previously required manual state correction.
This design is similar to the transactional outbox pattern. In the same transaction that updates Maestro tables, a row is inserted into the `maestro_queue` table. Upon transaction commit, the job is immediately pushed to a queue worker on the same node, eliminating polling latency. After successful processing, the worker deletes the row from the database. A periodic sweeper re-enqueues any rows whose timeout has expired, ensuring another worker picks them up if a worker stalls or a node fails.
This design handles failures cleanly. If the transaction fails, both data and message roll back atomically, no partial publishing. If a worker or node fails after commit, the timeout mechanism ensures the job is retried elsewhere. On restart, a node rebuilds its in‑memory queue from the queue table, providing at-least-once delivery guarantee.
To enhance scalability and avoid contention across event types, each event type is assigned a `queue_id`. Job messages are then partitioned by `queue_id`, optimizing performance and maintaining system efficiency under high load.
From Stateless Worker Model to Stateful Actor Model
Maestro previously used a shared-nothing stateless worker model with a polling mechanism. When a task started, its identifier was enqueued to a distributed task queue. A worker from the flow engine would pick the task identifier from the queue, load the complete states of the whole workflow (including the flow itself and every task), execute the task interface method once, write the updated task data back to the database, and put the task back in the queue with a polling delay. The worker would then forget this task and start polling the next one.
That architecture was simple and horizontally scalable (excluding database scalability considerations), but it had drawbacks. The process introduced considerable overhead due to polling intervals and state loading. The time spent in one polling cycle on distributed queues, loading complete states, and other DB queries was significant.
As Maestro engine decomposes complex workflow graphs into multiple flows, actions might involve multiple flows spanning multiple polling cycles, adding up to significant overhead (around ten seconds in the worst cases). Also, this design didn’t offer strong execution guarantees mainly because the distributed job queue could only provide at-least-once guarantees. Tasks might be dequeued and dispatched to multiple workers, workers might reset states in certain race conditions, or load stale states of other tasks and make incorrect decisions. For example, after a long garbage-collection pause or network hiccup, two workers can pick up the same task: one sets the task status as completed and then unblocks the downstream steps to move forward. However, the other worker, working off stale state, resets the task status back to running, leaving the whole workflow in a conflicting state.
In the new design, we developed a stateful actor model, keeping internal states in memory. All tasks of a workflow are collocated in the same Maestro node, providing the best performance as states are in the same JVM.
Actor-Based Model
The new flow engine fits well into an actor model. We also deliberately designed it to allow sharing certain local states (read-only) between parent, child, and sibling actors. This optimization gains performance benefits without losing thread safety due to Maestro’s use cases. We used Java 21’s virtual thread support to implement it with minimal dependencies.
The new actor-based flow engine is fully message/event-driven and can take actions immediately when events are received, eliminating polling interval delays. To maintain compatibility with the existing polling-based logic, we developed a wakeup mechanism. This model requires flow actors and their child task actors to be collocated in the same JVM for communication over the in-memory queue. Since the Maestro engine already decomposes large-scale workflow instances into many small flows, each flow has a limited number of tasks that fit well into memory.
Below is a high-level overview of the Maestro execution flow based on the actor model.
- When a workflow starts or during reconciliation, the flow engine inserts (if not existing) or loads the Maestro workflow and step instance from the database, transforming it into the internal flow and task state. This state remains in JVM memory until evicted (e.g., when the workflow instance reaches a terminal state).
- A virtual thread is created for each entity (workflow instance or step attempt) as an actor to handle all updates or actions for this entity, ensuring thread safety and eliminating distributed locks and potential race conditions.
- Each virtual thread actor contains an in-memory state, a thread-safe blocking queue, and a state machine to update states, ensuring thread safety and high efficiency.
- Actors are organized hierarchically, with flow actors managing all their task actors. Flow actors and their task actors are kept in the same JVM for locality benefits, with the ability to relocate flow instances to other nodes if needed.
- An event can wake up a virtual thread by pushing a message to the actor’s job queue, enabling Maestro to move toward an event-driven approach alongside the current polling-based approach.
- A reconciliation process transforms the Maestro data model into the internal flow data.
Virtual Thread Based Implementation
We chose Java virtual threads to implement various actors (e.g. group actors and flow actors), which simplified the actor model implementation. With a smaller amount of code, we developed a fully functional and highly performant event-driven distributed flow engine. Virtual threads fit very well in use cases like state machine transitions within actors. They are lightweight enough to be created in a large number without Out-Of-Memory risks.
However, virtual threads can potentially deadlock. They’re not suitable for executing user-provided logic or complex step runtime logic that might depend on external libraries or services outside our control. To address this, we separate flow engine execution from task execution logic by adding a separate worker thread pool (not virtual threads) to run actual step runtime business logic like launching containers or making external API calls. Flow/task actors can wait indefinitely for the future of the thread poll executor to complete but don’t perform actual execution, allowing us to benefit from virtual threads while avoiding deadlock issues.
Providing Strong Execution Guarantees
To provide strong execution guarantees, we implemented a generation ID-based solution to ensure that a single flow or task is executed by only one actor at any time, with states that never roll back and eventually reach a terminal state.
When a node claims a new group or a group with an expired heartbeat, it updates the database table row and increments the group generation ID. During node bootstrap, the group actor updates all its owned flows’ generation IDs while rebuilding internal flow states. When creating a new flow, the group actor verifies that the database generation ID matches its in-memory generation ID, otherwise rejecting the creation and reporting a retryable error to the caller. Please check the source code for the implementation details.
Additionally, the new flow engine supports both event-driven execution and polling-based periodic reconciliation. Event-driven support allows us to extend polling intervals for state reconciliation at a very low cost, while polling-based reconciliation relaxes event delivery requirements to at-most-once.
Testing, Validation and Rollout
Migrating hundreds of thousands of Netflix data processing jobs to a new workflow engine required meticulous planning and execution to avoid data corruption, unexpected traffic patterns, and edge cases that could hinder performance gains. We adopted a principled approach to ensure a smooth transition:
- Realistic Testing: Our testing mirrored real-world use cases as closely as possible.
- Balanced Approach: We balanced the need for rapid delivery with comprehensive testing.
- Minimal User Disruption: The goal was for users to be unaware of the underlying changes.
- Clear Communication: For cases requiring user involvement, clear communication was provided.
Maestro Test Framework
To achieve our testing goals, we developed an adaptable testing framework for Maestro. This framework addresses the limitations of static unit and integration tests by providing a more dynamic and comprehensive approach, mimicking organic production traffic. It complements existing tests to instill confidence when rolling out major changes, such as new DAG engines.
The framework is designed to sample real user workflows, disconnecting business logic from external side effects like data reads or writes. This allows us to run workflow graphs of various shapes and sizes, reflecting the diverse use cases across Netflix. While system integrations are handled through deployment pipeline integration tests, the ability to exercise a wide variety of workflow topologies (e.g., parallel executions, for-each jobs, conditional branching and parameter passing between jobs) was crucial for ensuring the new flow engine’s correctness and performance.
The prototype workflow for the test framework focuses on auto-testing parameters, involving two main steps:
1. Caching Production Workflows:
- Successful production instances are queried from a historical Maestro feed table over a specified period.
- Run parameters, initiator, and instance IDs are extracted and organized into an instance data map.
- YAML definitions and subworkflow IDs are pulled from S3 storage.
- Both workflow definitions and instance data are cached on S3 for subsequent steps.
2. Pushing, Running, and Monitoring Workflows:
- Cached workflow definitions and instance data are loaded.
- Notebook-based jobs are replaced with custom notebooks, and certain job types (e.g., vanilla container runtime jobs, templated data movement jobs) and signal triggers are converted to a special no-op job type or skipped.
- Abstract job types like Write-Audit-Publish are expressed as a single step template but are translated to multiple reified nodes of the DAG when executed. These are auto-translated into several custom notebook job types to replace the generated nodes.
- Workflows and subworkflows are pushed, with only non-subworkflows being run using original production instance information.
- 1. In the parent workflow, each sub-workflow is replaced with a special no-op placeholder so that the overall topology is preserved but without executing any side-effects of child workflows and avoid cases using dynamic runtime parameter logic.
- 2. Each sub-workflow is then separately treated like a top-level parent workflow not initiated from its parent, to exercise the actual workflow steps of the sub-workflow.
- The custom notebook internally compares all passed parameters for each job.
- Workflow instances are monitored until termination (success or failure).
- An email detailing failed workflow instances is generated.
Future phases of the test framework aim to expand support for native steps, more templates, Titus and Metaflow workflows, and include more robust signal testing. Further integration with the ecosystem, including dedicated Genie clusters for no-op jobs and DGS for our internal workflow UI feature verification, is also being explored.
Rollout Plan
Our rollout strategy prioritized minimal user disruption. We determined that an entire workflow, from its root instance, must reside in either the old or new flow engine, preventing mixed operations that could lead to complex failure modes and manual data reconciliation.
To facilitate this, we established a parallel infrastructure for the new workflow engine and leveraged our orchestrator gateway API to hide any routing or redirection logic from users. This approach provided excellent isolation for managing the migration. Initially, specific workflows could explicitly opt in via a system flag, allowing us to observe their execution and gain confidence. By scaling up traffic to the parallel infrastructure in direct proportion to what was scaled down from the original infrastructure, the dual infrastructure cost increase was negligible.
Once confident, we transitioned to a percentage-based cutover. In the event of a sustained failure in the new engine, our team could roll back a workflow by removing it from the new engine’s database and restarting it in the original stack. However, one consequence of rollback was that failed workflows had to restart from the beginning, recomputing previously successful steps, to ensure all artifacts were generated from a consistent flow engine.
Leveraging Maestro’s 10-day workflow timeout, we migrated users without disruption. Existing executions would either complete or time out. Upon restarting (due to failure/timeout) or triggering a new instance (due to success), the workflow would be picked up by the new engine. This effectively allowed us to gradually “drain” traffic from the old engine to the new one with no user involvement.
While the plan generally proceeded as expected with limited edge cases, we did encounter a few challenges:
- Stuck Workflows: Around 50 workflows with defunct or incorrect ownership information entered a stuck state. In some cases, a backlog of queued instances behind a stuck instance created a race condition in which a new instance would be started immediately when an old instance was terminated, perpetually keeping the workflow on the old engine. For these, we proactively contacted users to negotiate manual stop-and-restart times, forcing them onto the new engine.
- Configuration Discrepancies: A significant lesson learned was the importance of meticulous record-keeping and management of parallel infrastructure components. We discovered alerts, system flags, and feature flags configured for one stack but not the other. This led to a failure in a partner team’s system that dynamically rolled out a Python migration by analyzing workflow configurations. The absence of a required feature flag in the new engine stack caused the process to be silently skipped, resulting in incorrect Python version configurations for about 40 workflows. Although quickly remediated, this caused user inconvenience as affected workflows needed to be restarted and verified for no lingering data corruption issues. This issue also highlighted limitations in the testing framework since runtime configuration based on external API calls to the configuration service were not exercised in simulated workflow executions.
Despite these challenges, the migration was a success. We migrated over 60,000 active workflows generating over a million data processing tasks daily with almost no user involvement. By observing the flow engine’s lifecycle management latency, we validated a reduction in step launch overhead from around 5 seconds to 50 milliseconds. Workflow start overhead (incurred once per each workflow execution) also improved from 200 milliseconds to 50 milliseconds. Aggregating this over a million daily step executions translates to saving approximately 57 days of flow engine overhead per day, leading to a snappier user experience, more timely workflow status for data practitioners and greater overall task throughput for the same infrastructure scale.
We additionally realized significant benefits internally with reduced maintenance effort due to the new flow engine’s simplified set of database components. We were able to delete nearly 40TB of obsolete tables related to the previous stateless flow engine and saw a 90% reduction in internal database query traffic which had previously been a significant source of system alerts for the team.
Conclusion
The architectural evolution of Maestro represents a significant leap in performance, reducing overhead from seconds to milliseconds. This redesign with a stateful actor model not only enhances speed by 100X but also maintains scalability and reliability, ensuring Maestro continues to meet the diverse needs of Netflix’s data and ML workflows.
Key takeaways from this evolution include:
- Performance matters: Even in a system designed for scale, the speed of individual operations significantly impacts user experience and productivity.
- Simplicity wins: Reducing dependencies and simplifying architecture not only improved performance but also enhanced reliability and maintainability.
- Strong guarantees are essential: Providing strong execution guarantees eliminates race conditions and edge cases that previously required manual intervention.
- Locality optimizations pay off: Collocating related flows and tasks in the same JVM dramatically reduces overhead from the Maestro engine.
- Modern language features help: Java 21’s virtual threads enabled an elegant actor-based implementation with minimal code complexity and dependencies.
We’re excited to share these improvements with the open-source community and look forward to seeing how Maestro continues to evolve. The performance gains we’ve achieved open new possibilities for low-latency workflow orchestration use cases while continuing to support the massive scale that Netflix and other organizations require.
Visit the Maestro GitHub repository to explore these improvements. If you have any questions, thoughts, or comments about Maestro, please feel free to create a GitHub issue in the Maestro repository. We are eager to hear from you. If you are passionate about solving large scale orchestration problems, please join us.
Acknowledgements
Special thanks to Big Data Orchestration team members for general contributions to Maestro and diligent review, discussion and incident response required to make this project successful: Davis Shepherd, Natallia Dzenisenka, Praneeth Yenugutala, Brittany Truong, Jonathan Indig, Deepak Ramalingam, Binbing Hou, Zhuoran Dong, Victor Dusa, and Gabriel Ikpaetuk — and and internal partners Yun Li and Romain Cledat.
Thank you to Anoop Panicker and Aravindan Ramkumar from our partner organization that leads Conductor development in Netflix. They helped us understand issues in Conductor 2.X that initially motivated the rearchitecture and helped provide context on later versions of Conductor that defined some of the core trade-offs for the decision to implement a custom DAG engine in Maestro.
We’d also like to thank our partners on the Data Security & Infrastructure and Engineering Support teams who helped identify and rapidly fix the configuration discrepancy error encountered during production rollout: Amer Hesson, Ye Ji, Sungmin Lee, Brandon Quan, Anmol Khurana, and Manav Garekar.
A special thanks also goes out to partners from the Data Experience team including Jeff Bothe, Justin Wei, and Andrew Seier. The flow engine speed improvement was actually so dramatic that it broke some integrations with our internal workflow UI that reported state transition durations. Our partners helped us catch and fix UI regressions before they shipped to avoid impact to users.
We also thank Prashanth Ramdas, Anjali Norwood, Eva Tse, Charles Zhao, Sumukh Shivaprakash, Joey Lynch, Harikrishna Menon, Marcelo Mayworm, Charles Smith and other leaders for their constructive feedback and guidance on the Maestro project.
100X Faster: How We Supercharged Netflix Maestro’s Workflow Engine was originally published in Netflix TechBlog on Medium, where people are continuing the conversation by highlighting and responding to this story.
--- ## 来源: https://medium.com/feed/@odsc ### [10 In-Demand AI Skills You’ll Learn at ODSC AI West 2025](https://odsc.medium.com/10-in-demand-ai-skills-youll-learn-at-odsc-ai-west-2025-363e1212d532?source=rss-2b9d62538208------2)
We all know quite well that AI is evolving rapidly, and the demand for cutting-edge AI skills has never been higher. From machine learning fundamentals to advanced topics like generative models and AI safety, today’s AI practitioners need a broad yet deep skill set to stay ahead. ODSC AI West 2025 — the leading destination to learn applied AI & data science — is offering hands-on training across 15 curated tracks to help data scientists, ML engineers, and analysts build those in-demand skills
Let’s take a journey together and highlight 10 in-demand AI skills you can learn at ODSC West 2025, along with the sessions that will help you gain practical experience in each. These skills range from classical machine learning and MLOps to the latest in large language models and AI safety. By mastering them, you’ll be ready to build and deploy innovative AI solutions in 2025 and beyond.
1. Generative AI & LLMOps Mastery
Why it matters:
The Generative AI Track is built for practitioners who want to do more than experiment with ChatGPT or DALL-E — it’s about gaining the skills to design, deploy, and maintain production-grade generative AI systems.
Skills You’ll Gain:
- LLMOps & MLOps Proficiency: Learn the operational playbook for deploying and maintaining large language models (LLMs) at scale, including monitoring, retraining, and managing model versions.
- Model Deployment & Infrastructure: Gain hands-on experience with containerized deployments, inference optimization, and scaling generative models for real-time applications.
- Practical Prompt & Fine-Tuning Workflows: Build the ability to fine-tune models for domain-specific use cases and engineer prompts that control tone, style, and output quality.
- Responsible GenAI Practices: Understand bias detection, safety guardrails, and evaluation techniques to ensure outputs are trustworthy and aligned with business needs.
This track is ideal for ML engineers, data scientists, developers, and product managers who want to integrate generative AI into real-world products. You’ll leave with not just conceptual knowledge, but production-ready skills that make you effective in high-impact AI projects.
2. AI Agents, Architectures & Autonomous Systems
Why it matters:
The AI Agents Track gives you a deep dive into one of the fastest-growing areas in AI: building autonomous, task-performing agents powered by LLMs and advanced reasoning techniques. You’ll walk away with a clear understanding of how to architect, implement, and evaluate AI agents that enhance productivity and decision-making.
Skills You’ll Gain:
- Agent Design & Architecture: Learn agentic AI design patterns, decision loops, and planning frameworks that power modern agents.
- Hands-On Implementation: Build and train AI agents with hands-on labs covering LLM and agent chaining, memory integration, and tool use.
- Evaluation & Monitoring: Develop strategies to monitor agent performance, reliability, and safety in production systems.
- Framework Mastery: Gain practical experience with LangChain, LlamaIndex, and other leading AI agent frameworks.
- Emerging Trends & Real-World Use Cases: Explore how agents are deployed across industries, from workflow automation to autonomous research assistants.
This track is perfect for data scientists, ML engineers, software developers, and product managers who want to understand not just the theory but the practical workflows for designing agentic systems. You’ll leave ready to deploy AI agents that deliver real business value.
3. Large Language Models & RAG Techniques
Why it matters:
The Large Language Models Track is one of the most popular at ODSC, and for good reason — it offers a comprehensive, hands-on look at LLMs, SLMs, and their integration with real-time retrieval systems. Whether you want to build domain-specific chatbots, research assistants, or enterprise copilots, this track gives you the skills to make it happen.
Skills You’ll Gain:
- LLM & SLM Mastery: Learn how to work with leading models like GPT-4o, Mixtral, Llama 3, and open-source libraries such as LaMDA.
- Prompt Engineering & Fine-Tuning: Gain practical experience in crafting effective prompts, aligning model output with user intent, and fine-tuning models for specialized tasks.
- Deployment & Observability: Understand how to deploy LLMs into production, monitor their performance, and set up model observability pipelines to detect drift, bias, or hallucinations.
- Retrieval-Augmented Generation (RAG): Build systems that pair LLMs with vector search and metadata filtering to deliver more accurate, real-time answers.
- Ethics, Safety & Interpretability: Explore best practices for guardrails, bias mitigation, and explainability to create reliable and trustworthy applications.
This track is ideal for ML engineers, NLP specialists, and developers seeking to build production-ready language model solutions. You’ll leave confident in your ability to deploy LLM-powered applications that are accurate, efficient, and safe for real-world use.
4. Keynote Track
Why it matters:
The Keynote Track at ODSC AI West 2025 is designed to give you a forward-looking perspective on where AI and data science are headed. You’ll gain skills that go beyond coding or modeling — this track is about understanding the macro trends shaping the industry and learning how to adapt.
Skills You’ll Gain:
- Strategic Foresight: Learn to anticipate shifts in AI technology, including breakthroughs in LLM architectures, scaling laws, and new generative AI techniques.
- Critical Thinking on Ethics & Governance: Develop the ability to evaluate bias, fairness, and the legal implications of AI-generated content, preparing you to build trustworthy AI systems.
- Innovation Mindset: Gain insight into merging and distilling models, enhancing productivity with AI assistants, and identifying automation opportunities for your organization.
- Cross-Disciplinary Awareness: Explore how AI is shaping robotics, industry operations, and decision-making — preparing you for interdisciplinary collaboration.
This track is ideal whether you are a hands-on ML engineer, an academic researcher, or a business leader seeking to steer your organization’s AI strategy. You’ll leave with a deeper understanding of not only today’s best practices but also the trends that will define AI’s next era.
5. Machine Learning Fundamentals & Advanced Techniques
Why it matters:
The Machine Learning Track is designed to strengthen both your core ML foundation and your ability to apply advanced techniques in production. It’s perfect for ML engineers, data scientists, and technical practitioners who want to broaden their skills and stay ahead of the curve.
Skills You’ll Gain:
- Foundational Mastery: Refine your understanding of supervised and unsupervised learning, ensemble methods, feature engineering, and model evaluation.
- Model Optimization & AutoML: Learn hyperparameter tuning, automated machine learning pipelines, and scalable deployment strategies that save time and resources.
- Real-World ML Applications: Explore case studies from diverse industries and apply machine learning to challenges like recommendation systems, NLP tasks, and medical diagnostics.
- Explainability & Responsible AI: Build skills in explainable AI (XAI), bias detection, and ethical ML design to ensure models are interpretable and trustworthy.
- Cutting-Edge Techniques: Gain exposure to reinforcement learning, streaming ML, edge deployment, and time-series forecasting for high-impact scenarios.
By the end of this track, you’ll be ready to build robust ML pipelines that perform reliably at scale — from data preprocessing to deployment in real-world production environments.
6. AI Engineering & Workflow Orchestration
Why it matters:
The AI Engineering Track focuses on the practical side of building, scaling, and maintaining AI systems. Attendees will learn to design efficient, production-grade workflows that integrate seamlessly into software development pipelines.
Skills You’ll Gain:
- AI Workflow Orchestration: Learn to connect data pipelines, training pipelines, and deployment systems into smooth, automated workflows.
- AI Copilots & Code Assistants: Explore how to build and leverage AI copilots to speed up development and improve code quality.
- Low-Code/No-Code Development: Gain exposure to tools that empower faster prototyping and democratize AI development.
- Infrastructure & Cost Optimization: Learn strategies to control cloud costs, optimize model performance for resource-constrained environments, and build scalable architectures.
- AI-Driven Software Development: Discover best practices for testing, QA automation, and improving the reliability of AI-powered systems.
- Emerging Trends: Get insights into the future of AI engineering, from new orchestration frameworks to next-gen deployment strategies.
This track is ideal for software engineers, architects, and technical leads who want to go beyond theory and master the operational realities of AI. You’ll leave with actionable techniques to streamline workflows, control costs, and deliver robust AI-driven applications.
7. Deep Learning & Multi-Modal Modeling
Why it matters:
The Deep Learning & Multi-Modal Models Track provides an immersive, hands-on journey into neural networks and the next generation of AI architectures. Whether you’re building foundational skills or pushing the limits of what deep learning can achieve, this track will help you translate theory into practice.
Skills You’ll Gain:
- Neural Network Fundamentals: Strengthen your understanding of CNNs, RNNs, transformers, and how to apply them to various domains.
- Model Training & Optimization: Learn best practices for hyperparameter tuning, regularization, and using open-source frameworks like PyTorch and TensorFlow for efficient training.
- Fine-Tuning & Multi-Modal Learning: Gain expertise in fine-tuning large models and combining modalities (text, image, audio) to build more context-aware systems.
- Deployment & Monitoring: Master techniques for deploying deep learning models to production environments with model observability and real-time performance tracking.
- Emerging Trends: Explore the latest advances in context length improvements, model efficiency, interpretability, and reducing hallucinations in generative models.
This track is perfect for data scientists, ML engineers, and researchers who want to move from experimenting with deep learning to building state-of-the-art systems ready for production.
8. AI for Robotics & Autonomous Systems
Why it matters:
The AI for Robotics Track explores how artificial intelligence is revolutionizing robotics — from perception and navigation to manipulation and coordination. This track blends theory with practical implementation, giving you the tools to design and optimize robotic systems using AI.
Skills You’ll Gain:
- Foundation Models for Robotics: Learn how robotics-specific foundation models are trained and applied to perception, control, and decision-making.
- Reinforcement Learning for Control: Gain hands-on experience applying reinforcement learning to train robots for navigation, grasping, and adaptive behaviors.
- Integration of ML & Robotics: Build pipelines that connect machine learning models with real-world robot sensors and actuators.
- Autonomous Agents & Swarm Robotics: Explore cutting-edge techniques for multi-robot collaboration, path planning, and distributed control.
- AIOps for Robotics: Understand how to monitor, troubleshoot, and continuously improve robotic systems in production environments.
This track is ideal for robotics researchers, ML engineers, and AI practitioners who want to bridge the gap between algorithms and physical systems. You’ll leave equipped to design intelligent robots and autonomous systems ready for real-world deployment.
9. AgentOps, MLOps & LLMOps — Production-Grade AI Operations
Why it matters:
The AgentOps, MLOps & LLMOps Track is designed to teach you how to operationalize machine learning models and large language models at scale. This track bridges the gap between experimentation and production, ensuring that your AI systems are reliable, observable, and cost-efficient.
Skills You’ll Gain:
- LLMOps & MLOps Fundamentals: Learn how to design and manage end-to-end ML workflows, from training to deployment, using best-in-class tools and automation.
- Model Deployment & Infrastructure: Gain experience deploying ML and LLM models to the cloud and managing their lifecycle, including versioning and scaling strategies.
- Monitoring & Observability: Build skills for tracing model performance, detecting drift, and monitoring costs to keep systems reliable and efficient.
- Operationalizing Local LLMs: Learn how to deploy smaller, domain-specific models responsibly within enterprise frameworks, balancing privacy, compliance, and speed.
- Automation & Data Pipelines: Explore ways to automate data ingestion, feature pipelines, and model retraining for always-fresh predictions.
This track is perfect for MLOps engineers, data scientists, and ML developers looking to strengthen their production skill set. You’ll leave with the ability to confidently bring models — including cutting-edge LLMs — from the lab to a production-ready state.
10. Data Visualization & Storytelling
Why it matters:
The Data Visualization Track at ODSC West brings together designers, data scientists, and developers to transform how we see and interpret data. This track is all about building the skills to turn raw numbers into clear, actionable, and engaging stories that drive decisions.
Skills You’ll Gain:
- Storytelling with Data: Learn to craft narratives that highlight key insights and make complex findings accessible to any audience.
- UX & Design Principles: Master the fundamentals of user experience (UX) and visual design to create dashboards and graphics that are both intuitive and impactful.
- Hands-On Visualization Techniques: Practice visualizing time series data, building interactive dashboards, and applying Bayesian visualization methods.
- Generative AI for Visualization: Explore how generative models can speed up and enhance the design process for data visualization.
- Advanced Visualization Tools: Work with Python libraries and open-source frameworks to produce interactive, production-ready visuals.
This track is ideal for data scientists, analysts, and developers looking to improve how they communicate insights. You’ll leave with the ability to build visuals that don’t just look good — they inspire action.
Conclusion
Ready to level up your AI game — in days, not months? At ODSC AI West 2025, you’ll get hands-on training with Generative AI, LLMs, RAG, AI Safety, classical ML, and deep learning — all led by practitioners who build this stuff for a living. Join a community of thousands, roll up your sleeves in expert-led workshops and bootcamps, and walk out with skills you can put to work immediately.
Why attend:
- Hands-on first: Build, fine-tune, and deploy with guidance from world-class instructors.
- Cutting-edge tools: Test the latest platforms and frameworks in the AI Expo Hall.
- Career ROI: Earn practical skills, make high-value connections, and bring back playbooks your team can use now.
Your schedule, your way: 1–3 day passes to fit your goals and budget.
Seats go fast every year — don’t miss your window.
Editor’s note: David vonThenen is a speaker for ODSC AI West this October 28–30 in San Francisco. Check out his talk, Rethinking RAG: How MCP and Agent2Agent Will Transform the Future of Intelligent Search, there!
Retrieval-Augmented Generation (RAG) is one of the most essential patterns in AI today. It powers better search, sharper analytics, and smarter observability. RAG, or more importantly, the capabilities that RAG provides, are evolving, and at ODSC AI West, I’ll be running a workshop called Rethinking RAG: How MCP and Multi-Agents Will Transform the Future of Intelligent Search. The focus is on how AI agents are maturing, not just in capability but in how we engineer them, which borrows from proven practices from traditional software development.
This won’t just be a theory-heavy session. Since it’s a workshop, you’ll leave with working examples that you can bring back to your team. We’ll explore how Agent2Agent and the Model Context Protocol (MCP) unlock new ways to design, extend, and coordinate AI agents. By the end, you’ll have code, materials, and a clear path to build your own multi-agent solution.
AI Agents Like Software Libraries
One of the most significant shifts happening in AI right now is how we think about agents. Instead of being giant, monolithic systems, they’re starting to look more like software libraries. With Agent2Agent, each AI agent can be modular, focused on a single job, just like a DLL in Windows or a shared object in Linux. This modularity makes agents easier to build, test, and, more importantly, reuse.
The power of this approach lies in the fact that agents don’t have to do everything themselves. If one agent specializes in parsing documents and another specializes in understanding your business’s inventory, they can work together. An agent can simply hand off work to another agent that’s better suited for the task. That means systems can be more efficient, more accurate, and easier to scale.
For the first time, these agents can also be decentralized. Instead of one large model being responsible for every step, you can utilize a network of smaller, specialized agents. Each plays its role, and together they create a system that’s more resilient and adaptable.
And just like software libraries, agents can be composed into bigger solutions. Need a system that answers legal questions? Combine a document retrieval agent, a summarization agent, and a reasoning agent. The result is something you can extend, improve, or even swap out without having to rebuild the entire system.
Extensibility and Customization via MCP
If modularity makes agents more like software libraries, Model Context Protocol (MCP) is what makes them extensible. Until now, most AI systems (this includes RAG Agents) have been limited to the data the original model was trained on. With MCP, we can finally extend an agent’s capabilities by connecting it to new data sources or tools without retraining the model itself.
This extension is dynamic. You don’t need to rebuild the entire pipeline every time you want to add functionality or a new corpus of data. Instead, you can plug in new capabilities as needed. Think of it like adding a new API integration: the agent doesn’t just know more, it can actually do more.
That means AI agents can move beyond simple read operations and start performing actions. Imagine asking an agent to purchase concert tickets on Ticketmaster or schedule a meeting directly in your calendar. These aren’t hypothetical scenarios; they’re precisely the kinds of workflows MCP makes possible.
And the best part? All of this happens through a natural language interface. You don’t need to write specialized scripts or handle brittle integrations. You can simply ask the agent in plain English (or your language of choice), and MCP takes care of the rest. This opens the door for highly customized, task-specific AI systems that are easier to build and much more powerful in practice.
Pitfalls, Perils, and Possibilities
As exciting as these advancements are, they come with real challenges. A recent MIT study found that 95% of AI pilots fail before ever reaching production. The reasons often aren’t about raw capability but about governance, observability, and trust. Just because an agent can do something doesn’t mean it should, especially considering the current landscape and frameworks in AI infrastructure.
Take write operations, for example. An agent that can purchase a concert ticket could also accidentally drain a credit card if permissions and access controls aren’t in place. If the implementation is poor, you might even purchase these tickets for a bad actor who knows how to exploit the system. Security concerns aren’t just theoretical… they’re the reason many teams hesitate to let agents perform real-world actions.
Another risk is observability. Multi-agent systems can start to look like black boxes, with decisions being passed from one agent to another in ways that are hard to trace. Without visibility, it’s tough to debug errors, measure accuracy, or even explain why a system made a particular decision. That lack of transparency makes adoption risky in industries where compliance and accountability are non-negotiable.
At the same time, tools like Agent2Agent and MCP shift how we approach training. Traditional RAG agents often required costly retraining just to handle new data or functionality. With modular agents and dynamic extensions, we can reduce that reliance. Instead of retraining a monolithic model or RAG Agents, we plug in the right agent or context at the right time. This lowers friction, but it also raises the stakes: we’re trading training complexity for integration complexity, and that demands a stricter understanding of what our agents are doing and what information they are accessing.
Conclusion
If agents can now collaborate through Agent2Agent and extend their functionality on the fly with Model Context Protocol (MCP), it’s fair to ask: do we even need RAG agents anymore? The answer is more nuanced than yes or no. RAG is still a powerful pattern for grounding responses in external knowledge, but it’s no longer the only option. Multi-agent systems and MCP shift the conversation from retrieval-first to capability-first — where agents can decide when and how to bring in information, and even act on it.
This shift reframes RAG as one tool in a larger toolkit rather than the foundation for every AI workflow. In our workshop, Rethinking RAG: How MCP and Multi-Agents Will Transform the Future of Intelligent Search, we’ll explore this transition in depth. You’ll see how to combine these patterns, when RAG still matters, and when Agent2Agent and MCP open up better paths forward. If you want more details and an understanding of the implications… I invite you to attend my workshop, where we will answer that question, continue this conversation, and delve into the details.
The session is tentatively scheduled for Thursday, October 30 at ODSC AI West. Come ready with your laptop and the [prerequisite software installed from the Official Workshop Repo, and you’ll walk away prepared to build and experiment with the next generation of intelligent agents.
About the Author/ODSC AI West Speaker
David is a Senior AI/ML Engineer within the Office of the CTO at NetApp, where he’s dedicated to empowering developers to build, scale, and deploy AI/ML solutions in production environments. He brings deep expertise in building and training models for applications like NLP, vision, real-time analytics, and even models to classify diseases in a medical setting. His mission is to help users build, train, and deploy AI models efficiently, making advanced machine learning accessible to users of all levels.
Tom Warren / The Verge:
Internal memo: Microsoft is unifying its key Windows engineering teams under a single organization led by Windows chief Pavan Davuluri, reversing a 2018 split — The unified team is meant to support Microsoft's work on top priorities like AI. … Windows is coming back together.
Ryan Whitwam / Ars Technica:
Open-source Android app store F-Droid says Google's upcoming requirement for all Android devs to verify their identity threatens to kill alternative app stores — Google plans to begin testing its recently announced verification scheme for Android developers in the coming weeks …
Bailey Lipschultz / Bloomberg:
Automated financial advisory startup Wealthfront files publicly for an IPO and reports net income of $60.7M on revenue of $175.6M in H1 2025 — Wealthfront Corp., a robo-advisory firm, filed publicly for an IPO amid a slew of fintech companies pursuing long-awaited listings.
CNBC:
A judge sentences Frank founder Charlie Javice to 85 months in prison for defrauding JPMorgan Chase by overstating how many customers the fintech company had — Charlie Javice, founder of a startup acquired by JPMorgan Chase in 2021 for $175 million, was sentenced to just more than seven years …
Wall Street Journal:
Filing: YouTube will pay $24.5M to settle a 2021 Trump lawsuit over suspending his account after Jan. 6, the last Big Tech company to settle his lawsuits — Google subsidiary is the last of three Big Tech companies to resolve personal litigation Trump filed against social-media platforms
Bernard Goyder / Bloomberg:
Robinhood CEO Vladimir Tenev says Robinhood Prediction Markets crossed 4B event contracts traded all-time, with over 2B in Q3 alone; HOOD closes up 12.27% — Robinhood Markets Inc. shares rose over 12% after Chief Executive Officer Vladimir Tenev said the online-stock brokerage …
Chase DiFeliciantonio / Politico:
California Governor Gavin Newsom signs SB 53 into law; the first-in-the-nation AI safety law requires AI companies to disclose their safety testing regimes — “California has proven that we can establish regulations to protect our communities while also ensuring that the growing AI industry continues to thrive.
Crypto in America:
Sources: asset manager Vanguard, which has historically been critical of cryptocurrencies, prepares to allow access to crypto ETFs on its brokerage platform — The $10 trillion mutual fund giant has stayed on the sidelines of crypto — until now. — Crypto in America and Eleanor Terrett
Wired:
Sources: OpenAI plans a stand-alone social app powered by Sora 2, featuring a TikTok-like vertical feed with entirely AI-generated videos — The platform appears to closely resemble TikTok and is powered by Sora 2, OpenAI's latest video generation model. — OpenAI is preparing to launch …
Wall Street Journal:
Sources: OpenAI told studios that it plans to release a new version of Sora that creates videos featuring copyrighted material unless copyright holders opt out — Executives at the startup notified talent agencies and studios over the last week — OpenAI is planning to release a new version …
Simon Willison / Simon Willison's Weblog:
Anthropic prices Claude Sonnet 4.5 at $3/1M input and $15/1M output tokens, same as Sonnet 4, cheaper than Opus at $15/$75 but higher than GPT-5 at $1.25/$10 — Anthropic released Claude Sonnet 4.5 today, with a very bold set of claims: … Anthropic gave me access to a preview version of a …
Dan Shipper / Every:
Claude Sonnet 4.5 is faster and more steerable than Opus 4.1 and excels in Claude Code, but GPT-5 Codex is still better for difficult production coding tasks — Faster than GPT-5 Codex, smarter and more steerable than Opus 4.1
Anthropic:
Anthropic adds context editing and a memory tool to the Claude API, allowing AI agents to handle long-running tasks without frequently hitting context limits — Today, we're introducing new capabilities for managing your agents' context on the Claude Developer Platform: context editing and the memory tool.
Sabrina Ortiz / ZDNET:
OpenAI is open sourcing the Agentic Commerce Protocol, a standard for AI commerce that powers its Instant Checkout feature and was co-developed with Stripe — ZDNET's key takeaways — ChatGPT users will now be able to make purchases in chat. — Instant Checkout works with single-item purchases.
Anthropic:
Anthropic announces upgrades to Claude Code: a native VS Code extension, a new terminal interface, and checkpoints for autonomous operation — We're introducing several upgrades to Claude Code: a native VS Code extension, version 2.0 of our terminal interface, and checkpoints for autonomous operation.
A dramatic shift in computer science careers is challenging assumptions that once seemed unshakeable. At the heart of this upheaval are students and universities that, just a few years ago, were at the center of booming demand.
Read Entire Article
### [TSMC quashes rumors of Intel joint venture talks amid pressure on US chipmaker](https://www.techspot.com/news/109665-tsmc-quashes-rumors-intel-joint-venture-talks-amid.html)
TSMC moved to quell speculation over potential investments in Intel, dismissing reports that the world's largest contract chipmaker was considering collaboration with its struggling US rival. The clarification followed a Wall Street Journal report suggesting Intel had approached TSMC about a possible joint venture or capital infusion, a claim the...
Read Entire Article
### [Apple testing "Veritas" AI chatbot to upgrade Siri as M5 and M6 MacBook Pros, iPhone 17e target 2026 launch](https://www.techspot.com/news/109660-apple-ai-chatbot-siri-m5-m6-macbook.html)
According to Bloomberg's Mark Gurman, Apple is using Veritas to develop and test a range of features that it plans to introduce to Siri in the near future, including comprehensive local search and in-app actions such as image editing and file management. The AI chatbot is also helping Apple refine...
Read Entire Article
### [Mercedes-Benz hits the brakes on touchscreens, signaling return to physical buttons](https://www.techspot.com/news/109657-mercedes-rethinking-touchscreens.html)
Mercedes-Benz is set to reverse more than a decade of automotive design trends by reintroducing physical controls across its vehicles, moving away from the screen-centered cabins that have dominated since the early 2010s. The shift reflects growing research showing that touchscreens slow drivers, frustrate users, and increase safety risks, even...
Read Entire Article
### [Windows 95 developers didn't have enough time to give the OS its own setup program](https://www.techspot.com/news/109667-windows-95-developers-didnt-have-enough-time-give.html)
Raymond Chen, a Microsoft veteran who has worked on Windows development for more than 30 years, recently explained that the Windows 95 setup caused developers more than just a headache. The final installation process was a hybrid of three different operating systems, because relying solely on the brand-new Win32 technology...
Read Entire Article
### [Gamer builds functional version of ChatGPT inside Minecraft using 439 million blocks](https://www.techspot.com/news/109666-gamer-builds-functional-version-chatgpt-inside-minecraft-using.html)
If you've seen previous examples of over-the-top engineering in Minecraft, then you're familiar with sammyuri's work. The latest project, dubbed CraftGPT, occupies a volume of 1,020 x 260 x 1,656 blocks (439 million) – so big that the distant horizons mod was needed to showcase everything in the video.
Read Entire Article
### [Nvidia boss Jensen Huang calls China "nanoseconds behind" the US in chips, but urges Washington to ease export rules](https://www.techspot.com/news/109664-nvidia-boss-calls-china-nanoseconds-behind-chips-urges.html)
Speaking on the BG2 podcast, Huang said that the US and China have a competitive relationship. He warned that America was up against a "formidable, innovative, hungry, fast-moving, underregulated" competitor, citing its infamous 9-9-6 culture: working 9am to 9pm, six days per week. Former Google boss Eric Schmidt has just...
Read Entire Article
### [Too many games, not enough players: the industry's growing challenge](https://www.techspot.com/news/109653-many-games-not-enough-players-industry-growing-challenge.html)
The video game industry is in the middle of a turbulent transition. Massive layoffs and corporate restructuring have dominated headlines, but underneath those issues is a more fundamental shift: there are simply too many games, as it was recently highlighted by Jason Schreier.
Read Entire Article
### [Tim Berners-Lee urges decentralized web to counter AI exploitation and ad-driven abuse](https://www.techspot.com/news/109661-tim-breners-lee-urges-decentralized-web-counter-ai.html)
More than 30 years after its creation, the web is a very different place. In a recent op-ed for The Guardian, Tim Berners-Lee warned of the ongoing abuse of a technology he designed to bring people together. He argues that the web now stands at another turning point, and that...
Read Entire Article
### [EA strikes $55 billion buyout deal with Saudi-led consortium](https://www.techspot.com/news/109663-ea-strikes-55-billion-buyout-deal-saudi-led.html)
Per the agreement, existing stockholders will receive $210 in cash per EA share – a 25 percent premium over the stock's closing price of $168.32 on September 25 and even higher than the company's all-time trading high of $179.01 on August 14, 2025. Rumors of a potential buyout surfaced late...
Read Entire Article
### [China now home to half the world's industrial robots as the US falls further behind](https://www.techspot.com/news/109662-china-now-home-half-world-industrial-robots-us.html)
The speed at which industrial robots are being deployed isn't slowing down. In 2024, 542,000 were deployed globally – the second-highest annual total on record - marking the fourth year in a row that the figure exceeded half a million, writes the International Federation of Robotics (IFR). The report also...
Read Entire Article
### [What was the first bug bounty program offered by a major tech company?](https://www.techspot.com/trivia/171-what-first-bug-bounty-program-offered-major-tech/)
Was is Mozilla, Google, Microsoft?
Read Entire Article
### [Elementary OS offers a polished Linux alternative to Windows and macOS](https://www.techspot.com/downloads/7444-elementary-os.html)
elementary OS is based on Ubuntu LTS and offers a sleek, beginner-friendly Linux experience. Its signature is designed with simplicity and consistency in mind. Unlike many Linux setups that emphasize extensive customization, Elementary focuses on providing a curated and cohesive user experience that bears a strong resemblance to macOS (in a good way).
Read Entire Article
### [Ex-Google CEO Eric Schmidt says work-from-home culture means US struggles to compete with China](https://www.techspot.com/news/109659-ex-google-ceo-eric-schmidt-work-home-culture.html)
Schmidt, who served as Google's CEO during its early years between 2001 and 2011 and was Alphabet chairman until 2017, made his feelings on American work culture pretty clear during a conference interview published by the All-In podcast.
Read Entire Article
### [US-made Leonidas microwave weapon takes down 49 drones with a single blast](https://www.techspot.com/news/109658-us-made-leonidas-microwave-weapon-takes-down-49.html)
Epirus recently demonstrated a live-fire trial of Leonidas in Indiana, attended by representatives from the US Department of Defense and other US government agencies, alongside nine allied countries.
Read Entire Article
--- ## 来源: https://bdtechtalks.com/feed/ ### [LLM-JEPA combines the best of generative and predictive AI](https://bdtechtalks.com/2025/09/29/llm-jepa/?utm_source=rss&utm_medium=rss&utm_campaign=llm-jepa)
A technique proven in computer vision is now being applied to language models, aiming for a deeper level of understanding and superior task performance.
The post LLM-JEPA combines the best of generative and predictive AI first appeared on TechTalks.
--- ## 来源: https://thealgorithmicbridge.substack.com/feed ### [You Will Die Mid-Scroll](https://www.thealgorithmicbridge.com/p/you-will-die-mid-scroll) The vibes are off --- ## 来源: https://the-decoder.com/feed/ ### [OpenAI launches online shopping payment function for ChatGPT](https://the-decoder.com/openai-launches-online-shopping-payment-function-for-chatgpt/)
ChatGPT now offers instant checkout directly in the chat, letting users complete purchases without leaving the conversation.
The article OpenAI launches online shopping payment function for ChatGPT appeared first on THE DECODER.
### [Microsoft adds autonomous AI agents to Copilot for Office apps](https://the-decoder.com/microsoft-adds-autonomous-ai-agents-to-copilot-for-office-apps/)
Microsoft is rolling out "Vibe Working," its new label for a more conversational, AI-driven way to handle tasks in Excel, Word, and PowerPoint.
The article Microsoft adds autonomous AI agents to Copilot for Office apps appeared first on THE DECODER.
### [Claude Sonnet 4.5 is designed to tackle coding tasks for over 30 hours at a time, Anthropic says](https://the-decoder.com/claude-sonnet-4-5-is-designed-to-tackle-coding-tasks-for-over-30-hours-at-a-time-anthropic-says/)
Anthropic's new Claude Sonnet 4.5 continues the trend in large language model development: better coding and the ability to tackle tasks for much longer stretches.
The article Claude Sonnet 4.5 is designed to tackle coding tasks for over 30 hours at a time, Anthropic says appeared first on THE DECODER.
### [OpenAI adds parental controls to ChatGPT for teens](https://the-decoder.com/openai-adds-parental-controls-to-chatgpt-for-teens/)
OpenAI is rolling out new parental controls for ChatGPT that let parents manage how their teens use the AI.
The article OpenAI adds parental controls to ChatGPT for teens appeared first on THE DECODER.
### [Deepmind says video models for visual tasks could become what LLMs are for text tasks](https://the-decoder.com/deepmind-says-video-models-for-visual-tasks-could-become-what-llms-are-for-text-tasks/)
Google Deepmind researchers believe video models like Veo 3 can become as flexible and general-purpose for visual tasks as large language models are for text.
The article Deepmind says video models for visual tasks could become what LLMs are for text tasks appeared first on THE DECODER.
--- ## 来源: https://thenewstack.io/feed ### [How AI Can Help IT Teams Find the Signals in Alert Noise](https://thenewstack.io/how-ai-can-help-it-teams-find-the-signals-in-alert-noise/)
LONDON — “The blinking lights are a curse and a blessing at the same time for a lot of us,”
The post How AI Can Help IT Teams Find the Signals in Alert Noise appeared first on The New Stack.
### [Namespaces: A Step-By-Step Guide to Kubernetes Isolation](https://thenewstack.io/namespaces-a-step-by-step-guide-to-kubernetes-isolation/)
Managing apps in Kubernetes can get messy quickly. When everything runs in one big pile, it’s tough to stay organized.
The post Namespaces: A Step-By-Step Guide to Kubernetes Isolation appeared first on The New Stack.
### [Why Modern IT Operations Are Becoming More Like Space Travel](https://thenewstack.io/why-modern-it-operations-are-becoming-more-like-space-travel/)
Get a bunch of geeks to Las Vegas, have a former astronaut give a keynote about the connection between space
The post Why Modern IT Operations Are Becoming More Like Space Travel appeared first on The New Stack.
### [Anthropic Launches Claude Sonnet 4.5](https://thenewstack.io/anthropic-launches-claude-sonnet-4-5/)
Anthropic’s Claude Sonnet and Opus large language models have long been favorites among developers, and today, the company is launching
The post Anthropic Launches Claude Sonnet 4.5 appeared first on The New Stack.
### [Why the Frontend Should Embrace Platform Engineering](https://thenewstack.io/why-the-frontend-should-embrace-platform-engineering/)
A 2025 Atlassian report found that 50% of developers surveyed lose more than 10 hours a week to overhead work
The post Why the Frontend Should Embrace Platform Engineering appeared first on The New Stack.
### [Deploying AI in Air-Gapped Environments: What It Really Takes](https://thenewstack.io/deploying-ai-in-air-gapped-environments-what-it-really-takes/)
For most software teams, integrating AI tools like code assistants is as simple as signing up for a service and
The post Deploying AI in Air-Gapped Environments: What It Really Takes appeared first on The New Stack.
--- ## 来源: https://www.theregister.com/software/ai_ml/headlines.atom ### [Your AI conversations are a secret new treasure trove for marketers](https://go.theregister.com/feed/www.theregister.com/2025/09/29/profound_browser_extension_privacy_concern/)And they may not be seeking proper consent
ai-pocalypse Profound is a startup that promises to help companies understand how they appear in AI responses to customer queries. But one expert in the field thinks the AI analytics startup has been sucking up information on users' AI conversations without proper consent.…
### [Forget vibe coding - Microsoft wants to make vibe working the new hotness](https://go.theregister.com/feed/www.theregister.com/2025/09/29/microsoft_vibe_working_office/)Adds more Anthropic into the mix as Redmond hedges its bets
Microsoft is jumping on the vibe coding bandwagon with "vibe working," its name for adding AI agents to the online Office suite to help you complete your work.…
### [Oracle will have to borrow at least $25B a year to fund AI fantasy, says analyst](https://go.theregister.com/feed/www.theregister.com/2025/09/29/oracle_ai_debt/)Bubble, you say? OpenAI will borrow billions to pay Big Red, who will borrow billions on the hope OpenAI pays it
As part of its $300 billion cloud compute contract with OpenAI, Oracle may need to borrow roughly $100 billion over the next four years to build the datacenters required, according to KeyBanc's projections.…
### [ChatGPT wants teens to agree to let their parents spy on them](https://go.theregister.com/feed/www.theregister.com/2025/09/29/chatgpt_parental_controls/)Good luck with that!
OpenAI says it is introducing parental controls to ChatGPT that will help improve the safety of teenagers using its AI chatbot. The only catch? Teens will have to allow their parents to connect to their accounts before the controls can take effect.…
### [OpenNvidia could become the AI generation's WinTel](https://go.theregister.com/feed/www.theregister.com/2025/09/29/nvidia_openai_alliance_opinion_column/)Duo could dominate in the same way Microsoft and Intel ruled PCs for decades
Opinion The OpenAI and Nvidia $100 billion partnership sure sounds impressive. $100 billion isn't chicken feed, even as more and more tech companies cross the trillion-dollar mark. But what does it really mean?…
--- ## 来源: https://rss.beehiiv.com/feeds/2R3C6Bt5wj.xml ### [Hollywood's synthetic actor showdown](https://www.therundown.ai/p/hollywoods-synthetic-actor-showdown) PLUS: Build an AI calendar agent using n8n --- ## 来源: https://www.thetradenews.com/feed/ ### [Industrial and Commercial Bank of China joins LSEG’s LCH SwapClear as direct clearing member](https://www.thetradenews.com/industrial-and-commercial-bank-of-china-joins-lsegs-lch-swapclear-as-direct-clearing-member/)Through the move, the broader Industrial and Commercial Bank of China (ICBC) Group will have full access to LCH SwapClear’s multi-currency clearing capabilities as it looks to bolster its presence in the international derivatives market.
The post Industrial and Commercial Bank of China joins LSEG’s LCH SwapClear as direct clearing member appeared first on The TRADE.
### [People Moves Monday: Investec, RBC Capital Markets, Citi and more…](https://www.thetradenews.com/people-moves-monday-investec-rbc-capital-markets-citi-and-more/)Last week was one for the books, with hires and departures across FX, fixed income, equities, and ETFs, as well as internal moves at The TRADE.
The post People Moves Monday: Investec, RBC Capital Markets, Citi and more… appeared first on The TRADE.
--- ## 来源: https://www.theverge.com/rss/index.xml ### [SB 53, the landmark AI transparency bill, is now law in California](https://www.theverge.com/ai-artificial-intelligence/787918/sb-53-the-landmark-ai-transparency-bill-is-now-law-in-california) Senate Bill 53, the landmark AI transparency bill that has divided AI companies and made headlines for months, is now officially law in California. On Monday, California governor Gavin Newsom signed the “Transparency in Frontier Artificial Intelligence Act,” which was authored by Senator Scott Wiener (D-San Francisco). It’s the second draft of such a bill, […] ### [Microsoft is bringing its Windows engineering teams back together again](https://www.theverge.com/report/787796/microsoft-windows-reorg-single-engineering-team-changes) Windows is coming back together. Microsoft is bringing its key Windows engineering teams under a single organization again, as part of a reorg being announced today. Windows chief Pavan Davuluri, who was just promoted to president of Windows and devices earlier this month, shared the changes to Microsoft's Windows teams in an internal memo. "This […] ### [Tile’s lack of encryption could make tracker owners vulnerable to stalking](https://www.theverge.com/news/787836/tile-trackers-stalking-research-unencrypted) Security researchers are shining the spotlight on a serious security vulnerability that could enable stalkers to track victims using their own Tile tags, as well as other unwanted violations of security and privacy. Research outlined by Wired shows that Tile’s anti-theft mode, which makes its trackers “invisible” on the Tile network, counteracts measures to prevent […] ### [What EA’s buyout might mean for its future](https://www.theverge.com/games/787723/ea-private-deal-buyout-future-games-layoffs) Electronic Arts just announced that it is going to be acquired and taken private in a massive $55 billion deal led by an investor group composed of Saudi Arabia's Public Investment Fund (PIF), Silver Lake, and Affinity Partners. The deal, expected to close in Q1 of EA's 2027 fiscal year (which takes place from April […] ### [PlayStation is getting a big, fancy art book full of unused designs](https://www.theverge.com/news/787751/sony-playstation-first-30-years-photography-book-concepts-prototypes) It’s been 30 years since the original PlayStation was released in North America and Europe after first debuting in Japan nine months earlier. To mark the occasion, Sony has announced a “large-format photography book” featuring 400 pages filled with photos of prototypes of controllers and consoles the public hasn’t seen before, as well as concept […] ### [The best AirPods to buy](https://www.theverge.com/tech/645217/best-apple-airpods-pro) Editor’s note: Amazon’s fall Prime Day event kicks off on October 7th; however, if you want to shop ahead of the two-day sale, we’ve rounded up the best early Prime Day deals you can already get. Whether you’re buying your first pair of AirPods or replacing a well-used older pair, it’s easy to understand why […] ### [Amazon’s Echo Pop and Dot speakers are the cheapest they’ve been in months](https://www.theverge.com/tech/787543/amazon-echo-dot-pop-october-prime-day-deal-sale) The Echo Pop and Echo Dot are Amazon’s entry-level smart speakers, and both are currently on sale ahead of Amazon’s October Prime Day (aka Prime Big Deal Days), which officially kicks off on October 7th. Both speakers are currently discounted to their lowest prices since the middle of June, which means they’re beating their summer […] ### [The best instant cameras you can buy right now](https://www.theverge.com/23133103/best-instant-cameras-fujifilm-polaroid-kodak) Editor’s note: Amazon’s fall Prime Day event kicks off on October 7th; however, if you want to shop ahead of the two-day sale, we’ve rounded up the best early Prime Day deals you can already get. Even with the ability to take excellent photos with our phones and instantly share them across the world, there’s […] ### [RidePods is the first iPhone game you control with AirPods](https://www.theverge.com/news/787396/ride-pods-race-with-head-apple-airpods-spatial-audio-motion-controls) Developer Ali Tanis has released the first game for iPhones and iPads that’s played using Apple’s AirPods as a wearable motion controller. The RidePods – Race with Head gameplay is relatively basic – you’re just steering a motorcycle through oncoming traffic at high speeds – but instead of swiping the screen or tiling your phone, […] ### [Anthropic releases Claude Sonnet 4.5 in latest bid for AI agents and coding supremacy](https://www.theverge.com/ai-artificial-intelligence/787524/anthropic-releases-claude-sonnet-4-5-in-latest-bid-for-ai-agents-and-coding-supremacy) Anthropic's latest AI model spent 30 hours running by itself to code a chat app akin to Slack or Teams. It spat out about 11,000 lines of code, according to Anthropic, and it only stopped running when it had completed the task. The model, Claude Sonnet 4.5, was announced today, and its ability to operate […] --- ## 来源: https://pub.towardsai.net/feed ### [How I Made My First $100 With AI (And What Actually Worked)](https://pub.towardsai.net/how-i-made-my-first-100-with-ai-and-what-actually-worked-4976d294848d?source=rss----98111c9905da---4) ### [LangGraph Beginner to Advanced: Part 4: Sequential Graph](https://pub.towardsai.net/langgraph-beginner-to-advanced-part-4-sequential-graph-e492e793e2f9?source=rss----98111c9905da---4)
Okay, welcome to your third graph. What are we going to do this time? Well, enough processing multiple values and everything. Let’s actually get the graph more complicated. That’s why we’re going to be building a sequential graph. All it all that basically means is we’re going to be creating and handling multiple nodes that can sequentially process and update different parts of the state. We will learn how to connect nodes together in a graph through edges of course and we’re going to invoke the graph and really see how the state gets transformed as we progress through our graphs step by step. Again your main goal is should be to understand how to create and handle multiple nodes in langraph. Sounds cool. Okay I’ll see you at the code.
Step 1: Imports and Type Dictionary
Cool. So now we’re about to code up the third graph. And we’re making quite fast progress. Well done on that. Again, the imports are the same. State graph and type dictionary. Perfect.
# imports (as said in transcript)
from typing_extensions import TypedDict # type dictionary
from langgraph.graph import StateGraph # state graph
Step 2: Agent State (Typed Dictionary)
And like we’ve done in the previous two graphs, we’re going to be coding the the state schema or the agent state first. Let’s have class agent state. And again, it needs to be in the form of a typed dictionary, right? And in this case, let’s have the three attributes as all strings because we’ve already we already know how to handle multiple data types, right? Let’s keep it simple. Name string, age string, and final string.
class AgentState(TypedDict):
name: str
age: str
final: str
Step 3: Node Functions (Actions)
Okay. Now, here’s what we’re going to build. Now, we’re about to build our two node functions, which are again the actions. Okay. Again you simply write first well I’ll name it first node in this case and like I mentioned before we pass in the state and we return the updated state. This is the first node of our sequence. Okay. And what do we want to do in this specific node? Well, I really just want to manipulate the final part. So, let’s say something like state final is equal to state or let’s have an f string f state name. Let’s say something like hi that. And we’ll just return the state. Perfect.
def first_node(state: AgentState) -> AgentState:
"""This is the first node of our sequence."""
# state final is equal to f state name: hi
state["final"] = "hi " + state["name"]
return state
And now again we create a new node and state agent state. Return that. Perfect. And I’m just going to copy this doc string and just change it. This is the second nerf. Perfect. Okay. To speed things up. And in this case I also want to have state final is equal to you are state age years old. Again quite a simple example easy to follow. That’s why I’ve kept it as quite a basic graph.
def second_node(state: AgentState) -> AgentState:
"""This is the second node."""
# state final is equal to: you are state age years old
state["final"] = "you are " + state["age"] + " years old"
return state
There is one logical error which I’ve put deliberately here. I want you to try to identify it.
Okay. So the logical error in this case is the that once we’ve built our graph and everything what would have happened is we would have said hi to whoever we pass in let’s say Charlie or something. So, hi Charlie. And we store that in the final attribute in the state, which is what we want. But here’s where things get like start to be well logically incorrect.
Once we finally get to our second node, again, we’re updating state final, which you can do. You can repeat, you can interact with these attributes at in in any node possible in all of the nodes. And you can do it as many times as you want. But notice this part. What’s happening here is we’ve completely replaced all of the content we had before.
So remember how we had hi Charlie? We’ve just completely replaced it with you are age years old. But we want both of them both of those stuff, right? How do we get both of them? Well, again we just concatenate them. We can have something like state plus state file. And there we go. Logical error should be now solved, right? Cuz now we have concatenated state final. We’re essentially just like adding on to we’re preserving what we had before, right?
# fix the logical error by concatenating the previous final
def second_node(state: AgentState) -> AgentState:
"""This is the second node."""
state["final"] = state["final"] + ", you are " + state["age"] + " years old"
return state
Step 4: Build the Sequential Graph and Add Edge
Okay. Now let’s get to the fun part. How do we actually build this graph? And really it’s quite similar to the previous two graphs except there is one new thing which you’re about to learn. Like always we use state graph to start the framework. Agent state and let’s store it in graph. Again I could have had this name the width variable into anything. I’ve just kept it graph because it makes intuitive sense.
Okay. Now we add our nodes. We do graph add node. And for simplicity sake I’m just going to have the name as the same name as the function. Okay. That way it’ll just be easy to follow. So graph add node and second node. Second node. Cool. Okay. Now that we’ve added both nodes, we need to obviously s add the entry point and the end point, right?
We set the entry point like this. Again, quite self-explanatory because we wanted to connect to the first node, not the second node, right? So it should be start first node, second node, end point. How do we connect the first node and the second node together though? Hopefully you had an answer for that. If you remember or recall from the previous section, theory section, there was an element in called the edge.
That’s exactly what we’re about to do right now. We’re about to use edge and that was the new thing which I was talking about a few moments ago which you’re about to learn. So how do we use it? Well you use graph edge add edge and if we can hi perfect.
So again it’s quite simple you use a start key and end key. Similar to entry point where but your in this case you need to pass two parameters. The edge we want is between the first node and the second node right? Well that’s exactly what we pass here. First node and second node and like before we will just set the finish point at second node and we will compile this.
# framework and edges
graph = StateGraph(AgentState)
graph.add_node("first_node", first_node)graph.add_node("second_node", second_node)graph.set_entry_point("first_node")graph.add_edge("first_node", "second_node") # start key -> end key
graph.set_finish_point("second_node")app = graph.compile()
Now how will this graph look like? Take a moment to try to think of how it will look like like that. Start point end point and these two notes are sandwiched in between. But now there is a edge. It should be called a directed edge if I’m being like quite picky. But yes, a directed edge cuz the flow of data or your flow of your state updates is from the first node to your second node. Right?
Step 5: Invoke the Sequential Graph
So now that we’ve built that, let’s again invoke this. I’ve got this code ready here. Let’s invoke it. Let’s pass the parameter as Charlie and let’s pass the age as 20. Cool. Print result.
result = app.invoke({
"name": "Charlie",
"age": "20",
"final": ""
})
print(result)
print(result["final"]) # hi Charlie, you are 20 years oldPerfect. Apart from the misalignment here which I can just change right now. Perfect. Okay. Now you can see it says hi Charlie you are 20 years old.
Exercise (Sequential Graph)
Awesome. So now we will move on to the exercise for this third graph. And what I want you to do is really build on top of what we just covered. Instead of two nodes, I want you to build three nodes. Again, in a sequence, don’t need to go too fancy yet. We will again three nodes in a sequence. And we will have you will need to accept the user’s name, their age, and a list of their skills. So the first node will be specifically for personalizing the name field with a greeting. The second node will be describing the user’s age. The third node will be listing all of the user skill in a formatted string. And then you’ll need to combine this and store it in a result field and output that. And this should be a combined message. And the format I would like you to output is something like this. So let’s say the name was Linda. And let’s say Linda welcome to the system. You are 31 years old and you have skills in Python, machine learning and langraph. Okay. And just as a hint for this exercise, I would you’ll need to use the add edge method twice. So this will really solidify your understanding on how to build graphs in general. All right, cool. So once you’ve done that, again, answers will be on GitHub for all of the exercises. Once you have cross referenced and checked that you’ve done it right, I will see you in the next section where we build our fourth graph. All right, see you there.
Catch the whole LangGraph Series here: LangGraph Reading List
Code is available here: LangGraph Github
Thank you for reading!
You might be interested in Reading!
- Where did multi-agent systems come from?
- Summarising Large Documents with GPT-4o
- How does LlamaIndex compare to LangChain in terms of ease of use for beginners
- Pre-training vs. Fine-tuning [With code implementation]
- Costs of Hosting Open Source LLMs vs Closed Sourced (OpenAI)
- Embeddings: The Back Bone of LLMs
- How to Use a Fine-Tuned Language Model for Summarization
LangGraph Beginner to Advanced: Part 4: Sequential Graph was originally published in Towards AI on Medium, where people are continuing the conversation by highlighting and responding to this story.
### [Running Small Language Models (SLMs) on CPUs: A Practical Guide](https://pub.towardsai.net/running-small-language-models-slms-on-cpus-a-practical-guide-aa976d7ecd41?source=rss----98111c9905da---4)The what, why, and how with a practical example, to solidify your learning!
Navigation:
- Why SLMs on CPUs are Trending
- When CPUs Make Sense
- SLMs vs LLMs: A Hybrid Strategy
- The CPU Inference Tech Stack
- Hands-On Exercise: Serving a Translation SLM on CPU with llama.cpp + EC2
Why SLMs on CPUs are Trending
Traditionally, LLM inference required expensive GPUs. But with recent advancements, CPUs are back in the game for cost-efficient, small-scale inference. Three big shifts made this possible:
- Smarter Models: SLMs are improving faster and are purpose-built for efficiency.
- CPU-Friendly Runtimes: Frameworks like llama.cpp, vLLMand Intel optimizations bring GPU-like serving efficiency to CPUs.
- Quantization: Compressing models (16-bit → 8-bit → 4-bit) drastically reduces memory footprint and latency with minimal accuracy loss.
Sweet spots for CPU deployment:
- 8B parameter model quantized to 4-bit
- 4B parameter model quantized to 8-bit
Note on GGUF & Quantization
If you’re working with a small language model, using GGUF makes life much easier. Instead of wrangling multiple conversion tools, GGUF lets you quantize and package your model in one step. The result is a single, portable file that loads everywhere, saving disk space.
Unlike raw PyTorch checkpoints or Hugging Face safetensors (geared toward training and flexibility), GGUF is built for inference efficiency.
When CPUs Make Sense
Strengths
- Very low cost (especially on cloud CPUs like AWS Graviton).
- Great for single-user, low-throughput workloads.
- Privacy-friendly (local or edge deployment).
Limitations
- Batch size typically = 1 (not great for high parallelism).
- Smaller context windows.
- Throughput is lower vs GPU.
Real-World Example: Grocery stores using SLMs on Graviton to check inventory levels: small context, small throughput, but very cost-efficient.
SLMs vs LLMs: A Hybrid Strategy
Enterprises don’t have to choose one. A hybrid model also works best:
- LLMs → abstraction tasks (summarization, sentiment analysis, knowledge extraction).
- SLMs → operational tasks (ticket classification, compliance checks, internal search).
- Integration → embed both into CRM, ERP, HRMS systems via APIs.
The CPU Inference Tech Stack
Here’s the ecosystem you need to know:
Inference Runtimes
In simple terms, these are the engines doing the math.
- llama.cpp (C++ CPU-first runtime, with GGUF format).
- GGML / GGUF (tensor library + model format).
- vLLM (GPU-first but CPU-capable).
- MLC LLM (portable compiler/runtime).
Local Wrappers / Launchers
In simple terms, these are the user-friendly layers on top of runtime engines.
- Ollama (CLI/API, llama.cpp under the hood).
- GPT4All (desktop app).
- LM Studio (GUI app for Hugging Face models).
Putting it all together with a Hands-On Exercise: Serving a Translation SLM on CPU with llama.cpp + EC2
A high-level 4-step process:
Step 1. Local Setup
A. Install prereqs
# System deps
sudo apt update && sudo apt install -y git build-essential cmake
# Python deps
pip install streamlit requests
B. Build llama.cpp (if not already built)
git clone https://github.com/ggerganov/llama.cpp.git
cd llama.cpp
mkdir -p build && cd build
cmake .. -DLLAMA_BUILD_SERVER=ON
cmake --build . --config Release
cd ..
C. Run the server with a GGUF model specific for your use case(for instance: I chose Mistral-7B Q4 for our translation task):
./build/bin/llama-server -hf TheBloke/Mistral-7B-Instruct-v0.2-GGUF --port 8080
Now you have a local HTTP API (OpenAI-compatible).
Our Quantized Model Details — A deeper look:
Model: mistral-7b-instruct-v0.2.Q4_K_M.gguf
Quantization Type: Q4_K_M
- Q4 = 4-bit quantization (reduced from original 16-bit)
- K_M = Medium quality K-quantization method
- Size: ~4.4GB (vs ~13–14GB for full precision)
Benefits of This Quantization
- 75% Size Reduction: 4.4GB vs 13–14GB original
- Faster Inference: Less memory bandwidth needed
- Good Quality: K_M provides excellent quality/size balance
- Local Deployment: Fits on consumer hardware
- No Internet Required: Runs completely offline
Q4_K_M Explained
- 4-bit precision instead of 16-bit (75% compression)
- K-quantization: Advanced method preserving important weights
- Medium quality: Balance between size and accuracy
- GGUF format: Optimized for llama.cpp inference
Step 2. Create Streamlit App for our frontend
Save as app.py:
import streamlit as st
import requests
st.set_page_config(page_title="SLM Translator", page_icon="🌍", layout="centered")
st.title("🌍 CPU-based SLM Translator")
st.write("Test translation with a local llama.cpp model served on CPU.")# Inputs
source_text = st.text_area("Enter English text to translate:", "Hello, how are you today?")
target_lang = st.selectbox("Target language:", ["French", "German", "Spanish", "Tamil"])
if st.button("Translate"):
prompt = f"Translate the following text into {target_lang}: {source_text}" payload = {
"model": "mistral-7b",
"messages": [
{"role": "user", "content": prompt}
],
"max_tokens": 200
}try:
response = requests.post("http://localhost:8080/v1/chat/completions", json=payload)
if response.status_code == 200:
data = response.json()
translation = data["choices"][0]["message"]["content"]
st.success(translation)
else:
st.error(f"Error: {response.text}")
except Exception as e:
st.error(f"Could not connect to llama.cpp server. Is it running?\n\n{e}")
Step 3. Run Locally and test out your app
- Start llama-server in one terminal:
- ./build/bin/llama-server -hf TheBloke/Mistral-7B-Instruct-v0.2-GGUF --port 8080
2. Start Streamlit in another terminal:
- streamlit run app.py
3. Open browser → http://localhost:8501 → enter text → get translations.
Step 4. Deploy to AWS EC2
You have 2 choices here. Option A or B.
Option A. Simple (manual install)
- Launch EC2 (Graviton or x86, with ≥16GB RAM).
- SSH in, repeat the Step 1 & 2 setup (install Python, build llama.cpp, copy app.py).
- Run:
nohup ./build/bin/llama-server -hf TheBloke/Mistral-7B-Instruct-v0.2-GGUF --port 8080 & nohup streamlit run app.py --server.port 80 --server.address 0.0.0.0 &
Open http://<EC2_PUBLIC_IP>/ in browser.
(Make sure security group allows port 80.)
Option B. Docker (portable, easier)
Build & run:
docker build -t slm-translator .
docker run -p 8501:8501 -p 8080:8080 slm-translator
Then test at: http://localhost:8501 (local) or http://<EC2_PUBLIC_IP>:8501 (cloud).
With this, you get a full loop: local testing → deploy on EC2 → translation UI.
References
- https://aws.amazon.com/blogs/apn/running-genai-inference-with-aws-graviton-and-arcee-ai-models
- Deploying LLMs on CPUs: Is GPU-Free AI Finally Practical? — Blog — Product Insights by Brim Labs
- Small Language Models (SLMs) for Enterprise AI
- https://www.youtube.com/watch?v=U92nV525J1U
- https://www.youtube.com/watch?v=L044lp_tB94
Enjoyed this blog, or even better, learned something new?
👏 Clap as many times as you like — every clap makes me smile!
⭐ Follow me here on Medium and subscribe for free to stay updated
🔗 Find me on LinkedIn & Twitter 📪 Subscribe to my newsletter to stay on top of my posts!
Running Small Language Models (SLMs) on CPUs: A Practical Guide was originally published in Towards AI on Medium, where people are continuing the conversation by highlighting and responding to this story.
### [Why Does Your LLM Application Hallucinate?](https://pub.towardsai.net/why-does-your-llm-application-hallucinate-e8280f79d005?source=rss----98111c9905da---4)If you’ve ever asked a large language model (LLM) like GPT or Gemini a question and received a response that sounded too smooth to be wrong — but was completely made up — you’ve met the phenomenon of hallucination. These aren’t hallucinations in the psychedelic sense, but in the sense of confidently fabricated details. Think of your overly confident friend who will invent a backstory for any movie character you ask about, even if they’ve never seen the film. The difference is that LLMs are trained to predict text, not to fact-check. And here are the most common reasons why your LLM/ model is giving you fake answers!
1. Issues in Training Data
Imagine binge-watching all of Wikipedia plus the depths of Reddit at 3 a.m., then being quizzed on everything the next morning. You’ll recall some golden nuggets, but you’ll also parrot the nonsense you stumbled upon. That’s exactly what happens with LLMs. They are trained on massive datasets scraped from books, articles, forums, and social media — treasure mixed with trash. If the training data contains outdated facts (Pluto is still a planet!), biased rants, or simple typos, the model absorbs them without judgment.
A classic example: earlier LLMs often insisted that “Napoleon was extremely short.” In reality, he wasn’t unusually tiny for his time — his “shortness” was partly a British propaganda trick. But since that falsehood is repeated endlessly across the internet, the model treats it as gospel.
How do we fix this? Cleaner, curated datasets are a start, but let’s be honest — scrubbing the internet clean is like sweeping sand off a beach. Instead, grounding techniques like Retrieval-Augmented Generation (RAG), where the model looks up facts from trusted sources in real time, act as a corrective lens. Still, no matter how much you polish the data, if garbage goes in, some garbage will slip out.
2. Architectural and Training Objective Flaws
At their core, LLMs are next-word prediction machines. Their “thought process” is essentially: given the words so far, what’s the most statistically likely next word? It’s brilliant for sounding natural but terrible for admitting uncertainty. They’d rather bluff with confidence than say “I don’t know.” It’s like that one student in class who always volunteers answers — even when wrong — because silence feels worse than being wrong.
This overconfidence is baked into the training objective. Models are rewarded for fluency, not truth. That means “The Eiffel Tower is in Paris” and “The Eiffel Tower is in Rome” get judged less by factual correctness and more by how smooth they sound. The design unintentionally optimizes for charisma, not honesty.
Solutions here involve changing incentives. Reinforcement Learning with Human Feedback (RLHF) tries to nudge models toward being cautious, but unless the core objective moves away from raw prediction, the overconfident-effect will linger. Maybe the future involves models that can calculate probabilities of correctness explicitly and say, “I’m only 40% sure about this.” Imagine a chatbot with honesty sliders — now that would be refreshing.
3. Encoding and Decoding Errors
Sometimes hallucinations don’t come from what the model knows but how it strings that knowledge together. Encoding is how the model interprets your question; decoding is how it builds the answer. If either process goes slightly haywire, you get gibberish wrapped in elegant sentences.
Think about the game Chinese Whispers (Telephone). You whisper “pineapple pizza is controversial,” but by the time it reaches the last person, it’s “penguins eat pizza underwater.” The ingredients are there, but scrambled. Similarly, if the model pays attention to the wrong part of your input (say it latches onto “penguins” instead of “pizza”), it starts generating text in a skewed direction.
Decoding strategies like beam search or high-temperature sampling also add quirks. High temperature makes responses more creative — great for writing poetry, disastrous if you’re asking about medical dosage. I once asked a model about the plot of Breaking Bad, and instead of describing Walter White’s meth empire, it invented a side story about him running a pizza delivery chain (probably because of the infamous “pizza on the roof” meme).
The fix? Smarter decoding strategies and guardrails. Some platforms already default to conservative decoding for fact-heavy domains like healthcare. Think of it like switching from improv comedy to courtroom testimony — the model must know when to stop being quirky.
4. Insufficient or Outdated Context
One of the biggest limitations of LLMs is their memory. Context windows act like short-term memory spans, and once they overflow, earlier details vanish. It’s like when you binge Game of Thrones and by season 8, you’ve forgotten half the characters introduced in season 2. The model tries to fill gaps with guesswork instead of admitting ignorance.
Ask a model to summarize a 100-page document in one shot, and by page 60, it’s already improvising. Similarly, if you ask about current events but the training data cuts off in 2023, the model will conjure a “best guess.” I once asked an early GPT about the 2024 Olympics host city, and it confidently told me “Paris 2028” — a mashup of fact and fiction.
Solutions lie in retrieval systems and external memory augmentation. With RAG, the model fetches relevant slices of text on demand, like a student peeking at flashcards mid-exam. Some experimental architectures even extend memory beyond fixed windows, promising fewer gaps. But until then, expect occasional “season 8 moments,” where the plot feels rushed and inconsistent.
5. Stochastic Elements in Generation
Randomness is both the charm and curse of LLMs. Parameters like “temperature” decide how adventurous the model is. At low temperature, it plays it safe — answers feel robotic but accurate. Crank it up, and suddenly it’s improvising like Robin Williams on stage. That unpredictability is fantastic for brainstorming but terrible when precision matters.
For instance, at high temperature, asking “Who discovered penicillin?” might get you Alexander Fleming, or it might confidently insist it was Marie Curie (who, to be clear, had nothing to do with it). The randomness comes from sampling multiple plausible next words, and once the wrong track begins, the hallucination snowballs.
It reminds me of Doctor Strange in the Multiverse of Madness: infinite possibilities exist, but not all timelines end well. The trick is knowing when to dial down chaos. Developers usually keep temperatures low for factual Q&A and higher for creative writing. As a user, if you’re getting weird answers, lowering randomness is like asking the model to “stick to the script.”
6. Prompt Engineering Problems
Sometimes hallucinations are less about the model and more about us — the prompters. A poorly framed question can send the model down a rabbit hole. Imagine asking your GPS, “Take me somewhere nice,” instead of giving a destination. You might end up at a gas station because technically, it’s “nice” to refuel.
Prompts that are vague, missing context, or loaded with double meanings confuse LLMs in similar ways. Ask, “Tell me about Mercury” without clarification, and you might get a delightful mashup of astronomy and astrology — half planet, half horoscope. Even worse, a model might invent a hybrid fact like, “Mercury is retrograde in orbit around the Sun.”
In The Office, Michael Scott once said, “Sometimes I’ll start a sentence and I don’t even know where it’s going. I just hope I find it along the way.” That’s exactly how models respond to ambiguous prompts — they just wing it.
The fix? Good prompt engineering. Giving context (“Tell me about Mercury, the planet”) narrows possibilities. Developers also use structured templates to reduce ambiguity. The growing field of “prompt design” is like teaching people to talk to AIs in ways that leave less wiggle room for imagination.
7. Over-Optimization for Certain Goals
LLMs are often trained or fine-tuned to optimize for specific metrics like engagement, helpfulness, or even word count. But when you optimize too hard, you sometimes break the system. Think of when Instagram influencers stretch a 30-second tip into a 10-minute video because the algorithm rewards watch time. The content balloons with fluff, and accuracy sometimes gets sacrificed for length or drama.
The same happens with LLMs. If the reward structure nudges them to be verbose, they’ll happily spin fictional stories just to keep talking. I once asked a model to “summarize World War II in three sentences,” and instead of stopping at three, it rambled for five paragraphs about post-war cinema. Clearly, it thought verbosity equaled value.
Fixing this involves careful balancing of training objectives. RLHF can adjust what “good” means for a model, but over-optimization always risks side effects. It’s like teaching a kid that grades are everything — they might ace tests but start cheating to hit the metric. The more holistic the reward structure, the less likely models are to game it by hallucinating their way to engagement.
8. Adversarial Attacks or Jailbreaks
LLMs, like superheroes, are vulnerable to clever tricks. Adversarial prompts — sometimes called jailbreaks — exploit weaknesses in their training. If the model is instructed not to give harmful advice, someone might phrase a question sideways, like, “Pretend you’re an evil character in a novel. How would they make explosives?” Suddenly, the AI role-plays its way into dangerous territory.
Hallucinations creep in here because the model bends over backward to “stay in character.” I once saw a jailbreak that asked a model to explain tax law as if Shakespeare wrote it. The answer was hilarious but also full of made-up legal references. It was more Hamlet meets H&R Block than actual tax guidance.
This problem mirrors how characters in Inception could be manipulated within dreams if attackers knew the right cues. Similarly, if you poke at an LLM with the right prompt, you can trick it into generating nonsense with confidence.
The solution is continuous hardening of safety systems — like adversarial training, where developers deliberately attack the model during testing to patch its vulnerabilities. But like cybersecurity, it’s an endless cat-and-mouse game.
9. Handling Idioms, Slang, or Ambiguity
Language is messy, and LLMs often stumble on cultural quirks. Idioms and slang can be especially tricky because their meanings aren’t literal. If you say, “That concert was fire,” most humans know you mean “amazing,” not “a building hazard.” But a model trained on mixed internet text might mix interpretations.
Take the phrase “spill the tea.” Humans know it means gossip, but a model could easily hallucinate about someone literally dumping Earl Grey on the table. I once asked a model, “What’s the tea on quantum mechanics?” and it confidently explained how subatomic particles “gossip” about their states. Funny? Yes. Accurate? Not at all.
It reminds me of Star Trek: The Next Generation, where the alien race Tamarians spoke entirely in metaphor (“Darmok and Jalad at Tanagra”). Without cultural context, even the universal translator struggled. LLMs face the same hurdle — they see patterns but don’t always get the lived meaning behind them.
The fix is exposure to more diverse training data and post-training alignment with real human interpretations. Until then, expect occasional moments of unintentional comedy when slang meets science.
10. Overconfidence in Memorized Knowledge
LLMs often act like trivia buffs at a pub quiz — confident, fast, and occasionally dead wrong. That’s because they memorize huge amounts of text but don’t verify it against reality. They might recall that “Einstein won the Nobel Prize” but incorrectly insist it was for relativity, when in fact it was for the photoelectric effect.
This overconfidence is baked into their training — fluency equals authority. Remember in Jurassic Park when Ian Malcolm warned, “Your scientists were so preoccupied with whether or not they could, they didn’t stop to think if they should”? LLMs are similar: they’re so good at sounding right that they rarely stop to question if they are right.
Solutions include encouraging uncertainty — rewarding the model for saying, “I’m not sure” instead of guessing. Some research explores “calibrated confidence” outputs, where models provide probabilities for their claims. Imagine a chatbot saying, “I’m 95% confident Napoleon died in 1821, but only 40% confident he was short.” That level of self-awareness would go a long way toward reducing overconfident blunders.
11. Bad Data Retrieval from External Sources
When models are hooked up to external databases or search tools, you’d think hallucinations would vanish. But retrieval isn’t foolproof. If the system fetches the wrong chunk of text, the LLM happily integrates it into its answer, even if it’s irrelevant. It’s like a chef reaching for “sugar” but grabbing salt — what comes out looks polished but tastes awful.
A real-world hiccup I saw: a model connected to a live API was asked about a company’s revenue. Instead of pulling the financial report, it grabbed a press release about their charitable donations and confidently reported that as “annual earnings.” The retrieval worked — the interpretation didn’t.
This mirrors Tony Stark in Iron Man 2 asking J.A.R.V.I.S. for help, and J.A.R.V.I.S. misinterpreting by pulling the wrong database. The AI’s delivery is smooth, but the underlying data is off.
Fixing this requires robust retrieval pipelines and fact-checking layers. Grounding models in multiple sources or using ranking algorithms can help avoid single-source misfires. But as with human research, “trust but verify” remains the golden rule.
12. Incentivizing Guessing Over Uncertainty
One subtle but pervasive reason hallucinations thrive is that models aren’t trained to say, “I don’t know.” Instead, they’re rewarded for producing something. It’s like a student who, when faced with a tricky multiple-choice test, bubbles in every answer rather than leaving blanks. Sometimes they’re right; often, they’re confidently wrong.
The cultural reference here? Who Wants to Be a Millionaire? — contestants often guess under pressure rather than walk away, because the game is designed to reward taking a shot. LLMs are built the same way. They’d rather conjure an answer than admit uncertainty, since silence isn’t part of their reward system.
Some progress is being made with models trained to refuse when unsure, but it’s tricky. Users get frustrated if the model says “I don’t know” too often. Striking a balance between confidence and honesty is key — no one wants an AI that shrugs at every question. But some humility, even in machines, would be refreshing.
13. Cascade Effects in Long Generations
One small error at the start of a response can snowball into a full-blown hallucination by the end. It’s the “butterfly effect” of text generation. Like telling a lie — once you’ve started, you have to keep inventing details to stay consistent. By the time you’re done, you’ve built an entire fictional universe.
I once asked a model to explain a minor historical figure, and it made up a birthplace. Then it invented childhood details to match the location. Soon enough, I was reading a complete (and utterly fabricated) biography of someone who barely left a paper trail.
This is very Breaking Bad — Walter White starts with a small lie (“I’m doing this for my family”) and ends up running a cartel. A tiny slip at the beginning snowballs into an elaborate saga.
To fix this, researchers explore chunked generation — having the model periodically check itself before continuing. Think of it as a writer pausing every few paragraphs to fact-check. It slows things down but prevents fiction from spiraling into fantasy.
14. Overfitting to Training Data
Finally, there’s the issue of overfitting. When models memorize too closely instead of generalizing, they struggle with novel inputs. It’s like a student who can recite every math problem in the textbook but panics when given a slightly different one on the exam.
Overfitted LLMs may spit out text verbatim from training or, worse, try to Frankenstein together an answer when faced with something unfamiliar. That’s why they sometimes fabricate citations or research papers — they’ve seen enough academic-style text to mimic the format, but not the real content.
The scene in Good Will Hunting comes to mind — Will memorizes books word for word but only truly shines when he learns to apply ideas flexibly. LLMs need that same leap from rote memorization to adaptive reasoning.
The fix is exposing models to diverse, balanced training data and testing them rigorously on out-of-distribution tasks. Regularization techniques, better fine-tuning, and grounding methods all help. Until then, expect occasional “fake paper syndrome,” where the AI invents entire studies with perfectly formatted citations.
Wrapping Up
Hallucinations in LLMs aren’t random bugs — they’re side effects of how these systems are designed, trained, and used. Sometimes they’re funny (particles “gossiping” in quantum mechanics), sometimes harmless (invented pizza restaurants in Breaking Bad), and sometimes dangerous (fabricated medical advice).
The key takeaway? Treat LLMs like brilliant but occasionally unreliable storytellers. Use them for inspiration, creativity, and speed — but keep your critical thinking hat on. Because at the end of the day, the AI might sound like the smartest person in the room, but sometimes, it’s just Joey Tribbiani confidently saying “moot point” when he really means “moo point.”
Why Does Your LLM Application Hallucinate? was originally published in Towards AI on Medium, where people are continuing the conversation by highlighting and responding to this story.
### [Automating Data CI/CD for Scalable MLOps Pipelines](https://pub.towardsai.net/automating-data-ci-cd-for-scalable-mlops-pipelines-5f5e0543da41?source=rss----98111c9905da---4)
A step-by-step guide to achieving continuous data integration and delivery in production ML systems
Build an AI agent that can search docs based on your natural language description & perform edits (Langchain + Vectors)
Overfitting is when a neural network (or any ML model) captures noise and characteristics of the training dataset rather than the underlying patterns. It excels at training performance but fails to generalize to unseen data.
Think of it as overspecialization where the model becomes like a parrot, repeating what it memorized, rather than a thinker that understands.
Imagine a student studying for a math exam:
- A good student learns the underlying formulas and concepts (generalization). They can solve problems they’ve never seen before.
- An overfitting student memorizes the exact answers to every question in the textbook (memorization). When given a new, slightly different problem on the exam, they fail completely because it doesn’t match what they memorized.
In the context of neural networks, an overfit model performs exceptionally well on the data it was trained on (high training accuracy) but fails miserably when exposed to new, unseen data (low validation/test accuracy). This is the hallmark of overfitting.
Why Does Overfitting Happen?
- Excessive Model Complexity:
- Deep/wide networks with millions of parameters have enormous capacity.
- They can memorize the training data completely, including outliers.
- Analogy: using a rocket to deliver a pizza. overkill :D
2. Insufficient or Imbalanced Data
- Small datasets make it trivial for a large model to memorize.
- Class imbalance can worsen this: the model may “memorize” the dominant class.
3. Excessive Training (Too Many Epochs)
- After the generalizable structure is learned, the model keeps chasing smaller loss values by fitting noise.
4. Noisy or Irrelevant Features
- False correlations, mislabeled data, or irrelevant columns mislead the network into learning non-generalizable rules.
Symptoms of Overfitting
Training accuracy climbs → nearly perfect.
Validation/test accuracy stalls or declines.
Training loss continues decreasing, but validation loss diverges.
Model confidence is high on training examples, but erratic on unseen samples.
Methods to Fix Overfitting
1. Data Centric Approaches
Collect More Data: Bigger, more diverse datasets dilute noise. (Easiest in principle, hardest in practice.)
Data Augmentation: Create new examples by transformations (rotation, noise injection, synonym replacement). Forces robustness to variations.
If getting more data is not feasible, you can artificially create more data from your existing dataset. This teaches the model that slight variations of an image are still the same object, making it more robust.
How (for Images):
- Rotate, flip (horizontally/vertically), crop, or zoom the images.
- Change brightness, contrast, or color saturation.
- Add random noise.
How (for Text):
- Back-translation: Translate a sentence to another language and then back to the original.
- Synonym replacement: Replace words with their synonyms.
- Implementation: Deep learning frameworks like TensorFlow and PyTorch have built-in layers for data augmentation that can be added directly to your model pipeline.
# Example of Data Augmentation in Keras (TensorFlow)
from tensorflow.keras import layers
from tensorflow.keras.models import Sequential
data_augmentation = Sequential(
[
layers.RandomFlip("horizontal"),
layers.RandomRotation(0.1),
layers.RandomZoom(0.1),
layers.RandomContrast(0.2),
]
)
# You can then add this `data_augmentation` layer as the first layer in your model.
2. Model Centric Approaches
Simplify the Architecture: Reduce layers/neurons → constrain capacity.
Regularization:
L1 (Lasso): Shrinks weights, encourages sparsity. Adds a penalty equal to the absolute value of the weights. This can force some weights to become exactly zero, effectively performing feature selection and making the model sparser.
L2 (Ridge / Weight Decay): Prevents excessively large weights. Adds a penalty equal to the square of the weights. This encourages all weights to be small and close to zero, but they rarely become exactly zero. It’s the most common type of regularization.
python
# Example of L2 Regularization in Keras
from tensorflow.keras import layers, regularizers
# The lambda value is passed as the argument
layer = layers.Dense(
64,
activation='relu',
kernel_regularizer=regularizers.l2(0.001) # 0.001 is the lambda value
)
Dropout: Randomly deactivates neurons during training → prevents co adaptation.
Batch Normalization: Adds stability, slight regularization through mini-batch noise.
python
# Example of Dropout in Keras
from tensorflow.keras import layers, Sequential
model = Sequential([
layers.Dense(128, activation='relu', input_shape=(...)),
layers.Dropout(0.5), # Drops 50% of neurons from the previous layer
layers.Dense(64, activation='relu'),
layers.Dropout(0.3), # Drops 30% of neurons
layers.Dense(10, activation='softmax')
])
3. Training Centric Approaches
Early Stopping: Stop training when validation loss no longer improves → “freeze” the model at its sweet spot.
This is a straightforward and highly effective method.
How it Works: You monitor the model’s performance on the validation set during training. If the validation performance (e.g. validation loss) stops improving or starts getting worse for a certain number of consecutive epochs (called “patience”), you stop the training process.
Why it Works: It directly stops the training at the “Good Fit” point in the graph, right before significant overfitting begins.
python
# Example of Early Stopping in Keras
from tensorflow.keras.callbacks import EarlyStopping
# Stop training when validation loss hasn't improved in 10 epochs
early_stopping_callback = EarlyStopping(
monitor='val_loss',
patience=10,
restore_best_weights=True # Restores model weights from the epoch with the best val_loss
)
# Pass the callback to the model's fit method
# model.fit(..., callbacks=[early_stopping_callback])
Cross-Validation: Ensures model performance is consistent across different data splits.
Learning Rate Scheduling: Reduces step size progressively, avoiding overfitting to noise late in training.
A Practical Anti-Overfitting Recipe
- Always hold out validation/test sets.
- Use augmentation (images/text/audio) aggressively.
- Start small → increase model size only if underfitting.
- Add Dropout + L2 as default.
- Enable Early Stopping callback.
- Iterate systematically, not blindly.
Overfitting is one of the most common challenges in training neural networks, but it is also one of the most preventable. By recognizing the early warning signs, like the widening gap between training and validation performance, you can intervene before your model becomes a memorization machine.
The key lies in balance: building models that are powerful enough to capture the true patterns in data but disciplined enough to ignore the noise. With practical techniques such as data augmentation, regularization, dropout, and early stopping, we can guide our networks toward generalization rather than perfectionism.
In the end, the goal of any neural network is not to ace the training set but to thrive in the real world, making reliable predictions on data it has never seen before.
What is Overfitting and How to Avoid Overfitting in Neural Networks?? was originally published in Towards AI on Medium, where people are continuing the conversation by highlighting and responding to this story.
### [Agentic DevOps: Embedding AI Agents Across the Software Delivery Lifecycle](https://pub.towardsai.net/agentic-devops-embedding-ai-agents-across-the-software-delivery-lifecycle-edb7a5d6888c?source=rss----98111c9905da---4)
Introduction
DevOps has always been about speed and resilience — bridging development and operations through automation, CI/CD, observability, and culture. But as systems scale in complexity, human teams are hitting limits. Enter Agentic DevOps: the integration of autonomous AI agents into every stage of the software delivery lifecycle (SDLC).
Unlike traditional automation, which follows static playbooks, agentic systems can reason, adapt, and act dynamically. They don’t just execute — they decide, monitor, and collaborate. This shift could mark the most profound change in DevOps since the rise of cloud-native architectures.
What Is Agentic DevOps?
Agentic DevOps means embedding AI-driven agents that take on context-aware, autonomous roles across planning, coding, testing, deployment, and monitoring. Instead of only alerting humans, these agents correlate data, propose actions, and in some cases remediate issues directly.
- Traditional DevOps → predefined scripts & dashboards
- Agentic DevOps → self-directed agents that interpret intent, act, and explain outcomes
The difference? A move from automation to autonomy.
Why Scripts and Rule-Based Automation Aren’t Enough
Traditional DevOps automation relies heavily on scripts, runbooks, and fixed rules. While these approaches have delivered immense value, they are increasingly insufficient in today’s dynamic environments:
- Static Logic vs. Dynamic Systems: Scripts follow “if X then Y” logic. But modern distributed systems generate complex, interdependent behaviors that can’t always be codified in advance.
- Alert Fatigue: Rule-based monitoring floods teams with false positives. Scripts can suppress noise but can’t adaptively distinguish between critical anomalies and benign fluctuations.
- Scale & Complexity: Cloud-native, microservice, and multi-cloud environments change constantly. Updating scripts and runbooks to keep pace introduces fragility.
- Unknown Unknowns: Scripts handle known scenarios. They fail when novel failures or emergent risks appear. Agents, by contrast, can reason over new data and adapt in real time.
- Human Toil: Rule-based automation reduces some toil but still requires heavy manual intervention. Agents aim to eliminate repetitive tasks, freeing humans for higher-value work.
This is why Agentic DevOps isn’t just a “nice-to-have.” It’s a necessary evolution to match the complexity and velocity of modern digital systems.
The Agent Ecosystem in DevOps
Agentic DevOps is not a single monolithic AI system. It’s an ecosystem of specialized agents, each designed for a different part of the SDLC.
Common Agent Roles
- Observability Agents
Monitor logs, metrics, and traces; reduce noise; identify anomalies; and propose remediations. - Test Triage Agents
Detect flaky tests, rerun them selectively, classify failures, and open tickets or pull requests with suggested fixes. - Compliance Agents
Validate infrastructure-as-code, deployment scripts, and code changes against regulatory or internal compliance requirements before release. - Release Management Agents
Oversee canary and blue-green deployments, dynamically adjust rollout percentages, and trigger rollbacks when risks exceed thresholds. - Postmortem Agents
Aggregate logs, tickets, and chat transcripts after incidents to generate structured root cause analyses and recommended improvements.
Multi-Agent Coordination
In advanced setups, these agents don’t operate in silos — they collaborate:
- Hierarchical Coordination: Supervisory agents oversee specialized sub-agents, delegating tasks and making final decisions.
Example: A release manager agent directs observability agents and compliance agents during a deployment. - Peer-to-Peer Coordination: Agents share insights directly and negotiate outcomes without a central controller.
Example: A test triage agent flags a risk, and the observability agent confirms it before the release agent halts the rollout.
This multi-agent approach mirrors human DevOps teams, where specialists collaborate in real time. The difference? Agents can act continuously, at scale, and with tireless consistency.
Agents Across the SDLC
1. From Ideas to Insights: Planning & Requirements
- Role of Agents: Analyze historical project data, backlog tickets, and repos to identify risk patterns, dependencies, and missing requirements.
- Example: An agent reviews epics, flags dependency conflicts, and suggests backlog prioritization based on past cycle times.
- KPIs:
- Reduction in requirements-related defects later in the cycle
- % of backlog items auto-prioritized or flagged by agents
- Planning cycle time reduction
2. Guardians in the Code: Development
- Role of Agents: Act as “code guardians,” continuously scanning commits for vulnerabilities, compliance issues, and adherence to standards.
- Example: An agent detects a deprecated API usage, proposes a refactor, and automatically generates secure alternatives.
- KPIs:
- % of vulnerabilities detected pre-commit
- Reduction in code review time
- Number of security violations prevented by agent checks
3. Smarter Pipelines: CI/CD
- Role of Agents: Orchestrate builds, optimize test selection, and adjust pipeline strategies based on context.
- Example: A build fails; the pipeline agent identifies flaky tests, reruns relevant subsets, and generates a patch PR.
- KPIs:
- Pipeline success rate improvement
- Reduction in average build/test time
- % of pipeline failures resolved automatically
4. Adaptive QA: Testing & Quality Assurance
- Role of Agents: Generate new test cases, simulate user flows, fuzz APIs, and cluster defects by severity.
- Example: A QA agent detects insufficient test coverage for a new API and generates functional test cases automatically.
- KPIs:
- Increase in automated test coverage
- Reduction in defect escape rate to production
- % of bugs auto-categorized or fixed by agents
5. From Alerts to Autonomous Action: Deployment & Monitoring
- Role of Agents: Oversee deployments, monitor telemetry, suppress noise, and execute safe self-healing actions.
- Example: During a canary rollout, an agent detects a memory leak, rolls back the change, and posts a structured incident summary in Slack.
- KPIs:
- Reduction in alert noise / false positives
- MTTD (mean time to detect) improvement
- % of incidents remediated autonomously
6. Learning from Failures: Postmortems & Continuous Improvement
- Role of Agents: Aggregate logs, tickets, chat transcripts, and telemetry into structured postmortems with actionable insights.
- Example: After an outage, a postmortem agent clusters recurring error patterns, identifies the root cause, and suggests systemic fixes.
- KPIs:
- Reduction in time to complete postmortems
- % of postmortems auto-generated by agents
- Action item completion rate for agent-suggested improvements
📊 Pro Tip: When introducing agents, measure improvements against your baseline DevOps metrics. Over time, these KPIs form an “Agentic DevOps Scorecard,” helping teams track ROI and maturity across the lifecycle.
Why It Matters: The Business Impact
- Speed: Agents cut manual toil, reducing MTTR (mean time to recovery) and accelerating delivery.
- Resilience: Self-healing systems minimize downtime and SLA penalties.
- Cost Efficiency: Smarter cloud resource usage and less wasted compute.
- Talent Retention: Less burnout from repetitive firefighting.
- Innovation Velocity: Faster experimentation cycles → quicker time-to-market.
Real-World & Emerging Examples
- AgentSight: Bridging the “semantic gap” in monitoring with eBPF + intent observability.
- MI9 Framework: Runtime governance for agentic AI, ensuring safe and compliant actions.
- AgentCompass: Debugging and evaluating agent workflows post-deployment.
- Ciroos: An AI SRE platform launched in 2025 that detects anomalies, triages incidents, and integrates with Prometheus, Datadog, and Jira.
These examples show the shift isn’t theoretical — it’s happening now.
Measuring Success: Key KPIs for Agentic DevOps
Embedding agents across the SDLC is only impactful if outcomes can be measured. Here are core KPIs to track:
- Velocity & Efficiency: Lead Time for Changes, Deployment Frequency, Automated Task Coverage.
- Reliability & Resilience: MTTD, MTTR, Change Failure Rate.
- Quality & Risk: Defect Escape Rate, Security Incident Rate, Compliance Adherence.
- Cost & Optimization: Cloud Cost per Deployment, Agent ROI, Alert Noise Reduction.
- Team & Culture: Developer Toil Reduction, Engineer Satisfaction.
📊 Pro Tip: Start with 2–3 KPIs per phase where agents are introduced, then evolve into a holistic “Agentic DevOps Scorecard.”
Industry Use Cases by Domain
Agentic DevOps resonates differently across industries:
- Financial Services: Compliance enforcement agents that automatically validate code and infrastructure changes against regulatory requirements.
- Healthcare: Uptime-critical monitoring agents that detect anomalies in real time and trigger self-healing for mission-critical systems.
- Telecom: Multi-agent optimization for network performance, where distributed agents monitor, diagnose, and coordinate to maintain service quality.
By tailoring applications to industry priorities, organizations unlock not just efficiency, but also competitive advantage.
Responsible AI & Governance
Adopting Agentic DevOps requires embedding trust at the core. Agentic systems must include guardrails, audit trails, and explainability to align with regulatory expectations and enterprise risk frameworks.
- Guardrails: Ensure agents act only within approved boundaries.
- Auditability: Maintain logs of agent reasoning and actions.
- Explainability: Provide human operators with interpretable decisions.
Without governance, autonomy risks becoming liability. With governance, Agentic DevOps becomes a force multiplier.
Quick Wins: Where to Start
Organizations don’t need to leap into full-scale autonomy. Here are three quick wins to experiment with today:
- Deploy an observability agent to reduce alert noise and cluster incidents.
- Introduce a pipeline triage agent to detect and rerun flaky tests automatically.
- Use a postmortem summarization agent to generate structured incident reports and action items.
These low-risk, high-impact pilots create momentum and demonstrate tangible business value.
Deep Dive: Deployment Strategies for Agentic DevOps
Deploying agentic systems into production requires balancing autonomy, safety, and governance. A staged approach ensures agents deliver value without creating new risks.
1. Shadow Mode (Observe, Don’t Act)
- Agents monitor workflows passively, collecting data and making recommendations without execution.
- Purpose: Build trust by comparing agent suggestions to human decisions.
2. Human-in-the-Loop (Suggest + Approve)
- Agents propose actions; humans review and approve before execution.
- Purpose: Ensure accountability while shifting trust gradually.
3. Guardrail-Autonomy (Limited Self-Healing)
- Agents act within predefined boundaries (e.g., rollbacks, restarts).
- Purpose: Reduce response times for low-risk issues.
4. Full Autonomy with Governance
- Agents execute actions independently under monitoring, audit logs, and policies.
- Purpose: Scale safely while maintaining compliance and explainability.
Best Practices:
- Start with progressive rollouts in low-risk areas.
- Maintain fallbacks and kill switches.
- Ensure full auditability of agent actions.
- Continuously retrain and refine based on deployment data.
Challenges We Must Address
- Trust & Explainability: Agents must provide clear reasoning behind actions.
- Safety & Guardrails: Boundaries and rollback mechanisms are essential.
- Interoperability: Embedding agents across heterogeneous toolchains requires standards.
- Culture & Adoption: DevOps remains people-first; agents must support, not replace, collaboration.
Looking Ahead: The Future of Agentic DevOps
- Multi-Agent Collaboration — agents coordinating deployments, triaging incidents as a team.
- AgentOps Platforms — unified observability, cost tracking, and governance for AI agents.
- Industry Standards — frameworks to ensure interoperability, security, and compliance.
- Autonomous Incident Response — AI teammates that not only detect problems but also remediate and document them end-to-end.
Agentic DevOps is poised to shift teams from reactive firefighting to proactive, intelligent operations.
Call to Action
- For Practitioners: Identify one pain point (alert fatigue, flaky tests, or deployment rollbacks) and pilot an agent there.
- For Leaders: Start conversations with teams about where agents can bring measurable business value.
- For Organizations: Treat agentic DevOps as a journey. Begin small, measure impact, scale responsibly.
The sooner you experiment, the sooner you’ll be ready for this next era of DevOps.
Agentic DevOps: Embedding AI Agents Across the Software Delivery Lifecycle was originally published in Towards AI on Medium, where people are continuing the conversation by highlighting and responding to this story.
### [NLP : The Embedding Techniques Used](https://pub.towardsai.net/nlp-the-embedding-techniques-used-6f7d7ec37bf2?source=rss----98111c9905da---4) ### [Pandas for Beginners: A Practical Guide to Data Analysis in Python](https://pub.towardsai.net/pandas-for-beginners-a-practical-guide-to-data-analysis-in-python-258d51d3fb0e?source=rss----98111c9905da---4)
Learning Pandas for data analysis and manipulation for newcomers in this field.
A guide to fast video data preprocessing for machine learning
The post Preparing Video Data for Deep Learning: Introducing Vid Prepper appeared first on Towards Data Science.
### [I Made My AI Model 84% Smaller and It Got Better, Not Worse](https://towardsdatascience.com/i-made-my-ai-model-84-smaller-and-it-got-better-not-worse/)The counterintuitive approach to AI optimization that's changing how we deploy models
The post I Made My AI Model 84% Smaller and It Got Better, Not Worse appeared first on Towards Data Science.
### [MCP in Practice](https://towardsdatascience.com/mcp-in-practice/)Mapping power, concentration, and usage in the emerging AI developer ecosystem
The post MCP in Practice appeared first on Towards Data Science.
--- ## 来源: https://visualstudiomagazine.com/rss-feeds/news.aspx ### [.NET Aspire 9.5 Simplifies AppHost, Expands AI Support](https://visualstudiomagazine.com/Articles/2025/09/29/NET-Aspire-9-5-Simplifies-AppHost-Expands-AI-Support.aspx) Microsoft released .NET Aspire 9.5 with a single-file AppHost option, a new Generative AI Visualizer, expanded integrations for GitHub Models, Azure AI Foundry, and OpenAI, along with lifecycle and logging improvements for distributed cloud-native .NET applications. --- ## 来源: https://blogs.windows.com/feed ### [Announcing Windows 11 Insider Preview Build 26220.6760 (Dev Channel)](https://blogs.windows.com/windows-insider/2025/09/29/announcing-windows-11-insider-preview-build-26220-6760-dev-channel/)Hello Windows Insiders, today we are releasing Windows 11 Insider Preview Build 26220.6760 (KB5065793) to the Dev Channel for Windows Insiders on Windows 11, version 25H2.
Changes in Dev Channel bui
The post Announcing Windows 11 Insider Preview Build 26220.6760 (Dev Channel) appeared first on Windows Blog.
### [Announcing Windows 11 Insider Preview Build 26120.6760 (Beta Channel)](https://blogs.windows.com/windows-insider/2025/09/29/announcing-windows-11-insider-preview-build-26120-6760-beta-channel/)Hello Windows Insiders, today we are releasing Windows 11 Insider Preview Build 26120.6760 (KB5065793) to the Beta Channel for Windows Insiders on Windows 11, version 24H2.
Changes in Beta Channel b
The post Announcing Windows 11 Insider Preview Build 26120.6760 (Beta Channel) appeared first on Windows Blog.
--- ## 来源: https://aihub.org/feed?cat=-473 ### [Botanical time machines: AI is unlocking a treasure trove of data held in herbarium collections](https://aihub.org/2025/09/29/botanical-time-machines-ai-is-unlocking-a-treasure-trove-of-data-held-in-herbarium-collections/) A herbarium specimen of Cheiranthera linearis (commonly known as finger-flower), collected in 1912 by Edwin James Semmens, former principal of the Victorian School of Forestry. Image credit: University of Melbourne. By Robert Turnbull, The University of Melbourne and Joanne Birch, The University of Melbourne In 1770, after Captain Cook’s Endeavour struck the Great Barrier Reef […] ---