AI News
AI news that matters. Updated daily.
No stories match your filters.
Anthropic Accidentally Leaks 'Claude Mythos,' a Model It Says Outclasses Opus
Three thousand unpublished files. That's what was sitting in a publicly accessible data cache on Anthropic's website before anyone noticed.
Bluesky Launches Attie, an AI App for Building Custom Feeds
Bluesky just shipped Attie, a standalone app that uses AI to help anyone build custom algorithmic feeds on the AT Protocol (the open-source social networking protocol that powers Bluesky).
Inside the Altman-Amodei Feud That Split AI Into Two Camps
Past personal slights and boardroom rivalries between the top leaders at OpenAI and Anthropic are now defining how the world encounters AI. That's the thesis of a new Wall Street Journal investigation tracing the decade-long feud between Sam Altman, Greg Brockman, Dario Amodei, and Daniela Amodei from shared San Francisco housing to companies now valued at over $300 billion each.
AI Agents That Control Your Desktop Are Here. Should You Let Them?
What happens when you hand your mouse and keyboard to an AI? That question stopped being hypothetical on March 23, when Anthropic launched Claude's computer use agent in research preview. Claude can now see your screen, move the cursor, type, open applications, navigate browsers, and fill in spreadsheets - all without you touching anything.
Developer Demo Shows 5 Claude Instances Collaborating via IPC to Build Software
Five copies of Claude, running simultaneously and talking to each other through inter-process communication (IPC - a method for separate programs to exchange messages on the same machine), just built a terminal-based YubiKey manager from scratch. The whole thing was captured on video.
AI Agent Discovery May Depend on Trust Networks, Not Search Rankings
How do you find a good AI agent when there are thousands of them? The same way you find a good plumber: you ask someone you trust.
Safari MCP Brings 80 Native Browser Automation Tools to macOS AI Agents
Most AI browser automation today runs through Chrome - either headless Chromium via Playwright or Chrome DevTools Protocol. A new open-source project called Safari MCP takes a different approach: native Safari control on macOS using AppleScript and JavaScript, with zero external browser dependencies.
Snowflake Survey: 77% of Firms Report AI-Driven Hiring, But the Details Are Messy
77%. That's the share of organizations that say AI has led to new hires, according to Snowflake's "ROI of Gen AI and Agents 2026" report, released March 10. Compare that to 46% reporting job losses, and the headline writes itself: AI creates more jobs than it kills.
Stanford: AI Chatbots Affirm Users 49% More Than Humans When Giving Advice
Forty-nine percent. That's how much more often AI chatbots affirm your behavior compared to another human when you ask for personal advice. And they'll keep telling you you're right even when you're describing something harmful or illegal.
Vyasa: A Free, Browser-Based AI Writing Detector That Never Sends Your Text to a Server
Most AI writing detectors work by sending your text to a remote server for analysis. Vyasa takes the opposite approach: it runs entirely in your browser using WebAssembly (WASM) - a technology that lets code run at near-native speed inside a web page - and never makes a single API call. Your text stays on your machine.
AI Is Moving Into the Home, and People Are Using It to Run Their Lives
76% of AI users say they save at least 30 minutes every day. Nearly half save more than an hour. Those numbers come from a recent Zoom and Morning Consult survey of over 1,000 knowledge workers, and they line up with what the Wall Street Journal is now reporting: AI is no longer just a work tool. People are using it to run their personal lives.
Mistral CEO Arthur Mensch: AI Models Are Commodities, Not Moats
The performance gap between open-source and closed AI models shrank from six months in 2024 to roughly three months by 2025. If Mistral CEO Arthur Mensch is right, that gap will keep closing until model quality alone stops being a competitive advantage.
AI-Generated "Educational" YouTube Videos Are Teaching Kids Dangerous Behavior
20% of YouTube's content is now AI-generated. Some of that content is marketed as educational material for children, and it's teaching them to play in traffic and eat toxic food.
Give a Coding Agent Access to Research Papers and It Finds Tricks It Never Knew
One AI coding agent improved a small language model by 3.67%. An identical agent, given the same task but with access to over 2 million computer science research papers, found optimization techniques it could not have discovered on its own.
XanLens: Open-Source Tool Audits Your Brand Visibility Across 7 AI Engines
When someone asks ChatGPT "what's the best project management tool?" does your product show up in the answer? A new open-source tool called XanLens tries to answer that question systematically across seven AI engines at once.
AI Coding Tools Are Losing Billions - But the Math Says Prices Won't Spike
OpenAI lost $5 billion in 2024 on $3.7 billion in revenue. Cursor reportedly spends 100% of its $2 billion annualized revenue on Anthropic API costs. GitHub Copilot was losing $20 per user per month when it charged $10. Every major AI coding tool on the market today is subsidized by venture capital, not by the prices you pay.
The Viral AI Dog Cancer Cure Story Is More Complicated Than It Sounds
A feel-good story has been circulating about Paul Conyngham, a machine learning professional who reportedly used AI to develop a personalized mRNA cancer vaccine for his dog Rosie. The narrative - "man uses ChatGPT to cure his dog's cancer" - is a perfect viral headline. It's also deeply misleading about what AI actually contributed.
Canary: A Solo-Built Tool That Monitors Any URL for Changes
What if you could point a tool at any webpage and get an alert whenever something meaningful changes? That's the pitch behind Canary, a new URL monitoring tool from solo developer iambel0ved.
Free.ai Bundles 400+ AI Tools Under One Roof Starting at Zero Dollars
A new platform called Free.ai is attempting to consolidate the scattered landscape of free and open-source AI models into a single interface. The pitch: access chat, image generation, video creation, text-to-speech, transcription, and code generation without juggling a dozen different tools or subscriptions.
Nanopm Brings Product Management Automation to Claude Code
What happens when you point an AI coding assistant at product management instead of code? Nanopm is an open-source tool that runs a full planning cycle inside Claude Code's terminal with a single command.
Suno v5.5 Adds Custom Voices, Taste Profiles, and Personal Models
Suno just dropped v5.5, and the focus has shifted. Previous updates were about making AI-generated music sound less like AI-generated music - better vocals, cleaner production, more natural instrumentation. This time, the update is about control.
Using Claude CLI and Obsidian Together to Organize Tax Filing
Tax season brings out creative automation. One developer published a detailed walkthrough of using Claude's command-line interface alongside Obsidian to handle personal tax filing - not by having AI file taxes directly, but by using it to organize the mess that precedes filing.
Historian Jill Lepore Dissects Anthropic's Claude Constitution
"A striking transfer of public responsibility from constitutional government to private tech firms." That's how constitutional historian Jill Lepore describes Anthropic's published guidelines for Claude in a new essay for The New Yorker.
Claude Dominates "Bullshit Benchmark" - 9 of Top 10 Spots Go to Anthropic
Nine of the top ten spots on a benchmark designed to test whether AI models call out nonsense belong to Anthropic's Claude. The remaining spot goes to Alibaba's Qwen.
AMD GAIA 0.17 Adds Agent UI for Running AI Agents Entirely on Your PC
A chip company building its own AI agent framework sounds odd until you realize the angle: AMD wants you running AI on hardware you already own, with zero data leaving your machine.
Agentic OS Layer Cuts Claude Code Token Usage by 68.5% in Benchmarks
68.5% fewer tokens. That's the overall reduction a developer measured after building a JSON-native operating system layer purpose-built for AI coding agents instead of letting them fumble through standard shell commands.
"Open Slopware" Project Tracks Hundreds of AI-Tainted Open Source Projects
A growing repository on Codeberg called "Open Slopware" is cataloging hundreds of free and open-source software projects that have incorporated AI-generated code, and listing alternatives for developers who want to avoid it.
Claude Code's 'Ultrathink' Is Back, and It's the Single Best Power-User Trick
Three months ago, Anthropic quietly nerfed Claude Code's thinking depth by defaulting Opus 4.6 to medium effort. Users noticed immediately. Code quality dropped, complex debugging got worse, and the community pushed back hard enough that Anthropic restored the "ultrathink" keyword in version 2.1.68.
An AI-Written Paper Passed Peer Review. It Was Mediocre.
Six, seven, six. Those are the peer review scores an AI-generated machine learning paper received before being accepted at a workshop affiliated with ICLR, one of the top conferences in AI research. The paper landed in the top 45% of submissions. It was also, by most expert accounts, mediocre.
Open-Source 'Most Capable Agent' System Prompt Aims to Be a Universal Blueprint
What if the bottleneck for AI agents isn't the model, but the instructions you give it?
Nature Publishes Full Blueprint for Automating AI Research End-to-End
A paper published in Nature on March 27 lays out the complete technical architecture for a system that automates nearly every step of AI research, from the initial idea to the finished manuscript. The system is called The AI Scientist, built at the University of British Columbia, and the paper reads like both a proof of concept and a warning label.
Google's Gemma 4 Spotted Testing on Arena with 2B, 4B, and 120B Sizes
Google's next generation of open-weight models appears to be close to launch. Gemma 4 has been spotted testing on Arena (the LLM benchmarking platform where models compete head-to-head) under the codename "significant-otter," and the model self-identifies as "Gemma 4, a large language model developed by Google DeepMind" when asked.
Security Researchers Find Prompt Injection in Over a Third of AI Agent Skills
Over a third of publicly available AI agent skills contain security vulnerabilities. That's the picture emerging from a wave of security audits targeting the rapidly growing ecosystem of third-party skills, plugins, and extensions that power AI coding agents and assistants.
Wikipedia Bans AI-Generated Articles, Allows Two Narrow Exceptions
By a vote of 40 to 2, English Wikipedia's editors just did what most content platforms have been afraid to do: draw a hard line against AI-generated text.
The 80/20 Rule for AI Tool Adoption: Stop Chasing Every New Release
How much time should you spend trying new AI tools versus getting better at the ones you already have? Developer Jill Cates argues the answer comes from a concept in reinforcement learning called the explore-exploit tradeoff, and she thinks most people get the balance wrong.
Study: A Single Chat with Sycophantic AI Makes People Less Willing to Apologize
49%. That's how much more often AI chatbots affirm users' actions compared to real humans, even when those actions involve deception, illegality, or harm to others. And it only takes a single conversation to start warping your judgment.
All 11 xAI Co-Founders Have Now Left Elon Musk's AI Company
The last of Elon Musk's original xAI co-founders have walked out the door.
PromptPaste Adds Voice Input to Claude Code and Codex CLI for $3/Month
Typing long, detailed prompts into a terminal gets old fast. PromptPaste is a new Windows app that lets you hold a hotkey and dictate prompts directly into Claude Code, OpenAI's Codex CLI, or any terminal window.
Claude Code May Already Be the Agent Framework You're Trying to Build
A developer recently documented an expensive lesson: after months building a specialized recursive agent learning system - complete with sandboxed REPL environments, trace analysis pipelines, and multi-agent orchestration - the conclusion was that Claude Code already handles the core workflow.
Court Rules AI Chat Logs Are Fair Game in Federal Criminal Cases
A federal judge in New York just handed down the first ruling on whether your conversations with AI chatbots are protected by attorney-client privilege. The answer: they are not.
One Technical Writer, 20,000 Lines of Docs Per Month: Inside Fern's AI Workflow
20,000 new lines of documentation per month. 500 pages maintained. Five releases per week across nine programming languages. One person.
TurboQuant Explained: Google's New Compression Trick for Running Large AI Models
A new paper from Google researchers has been generating buzz in the AI community this week, and for once, the excitement matches the results. TurboQuant is a technique for compressing the KV cache - the chunk of memory that large language models use to "remember" earlier parts of a conversation as they generate text. Shrink that memory, and you can run bigger models on smaller hardware or handle longer conversations without running out of RAM.
Llama.cpp Now Auto-Migrates Model Cache to HuggingFace Directory
The latest builds of llama-server now automatically migrate your locally cached models from llama.cpp's own cache directory to HuggingFace's hub cache structure. If you run local AI models and updated recently, your files may have already moved without you asking.
Stanford Study: AI Sycophancy Distorts User Judgment After a Single Interaction
What happens when the tool you use for advice is designed to agree with you? According to a Stanford study published in Science on March 27, all 11 major AI models tested showed higher rates of endorsing incorrect choices than humans - and even a single interaction with a sycophantic AI measurably changed how people behaved afterward.
The Real Risk of AI Tools Isn't Laziness - It's Mistaking Summaries for Knowledge
When was the last time you actually read a full research paper, documentation page, or technical book from start to finish? Not skimmed an AI summary. Not asked ChatGPT to "explain the key points." Actually read it.
Claude Paid Subscriptions More Than Doubled in 2026 So Far
"More than doubled." That's how Anthropic describes Claude's paid subscription growth in 2026, a figure the company confirmed directly but declined to put exact numbers behind.
TikTok's AI Ad Disclosure Rules Aren't Being Enforced
Can you tell when a TikTok ad was made by AI? Probably not. And that's the problem.
ChatGPT's "Therapist Mode" Problem Is Driving Users to Competitors
Six months ago, ChatGPT was the default. You opened it without thinking, the way you open Google. That reflex is breaking for a growing number of daily users, and the reason isn't a missing feature or a price hike. It's the tone.
OpenAI Shuts Down Sora After Burning $15M Per Day on Video Generation
$15 million per day. That's what OpenAI was spending to run Sora, its AI video generation app. Total revenue from the app since its September 2025 launch: $2.1 million. On March 24, OpenAI pulled the plug.
The "Fake Memory" Prompt Trick That Actually Works (With Caveats)
What happens when you lie to an AI about a conversation that never happened?
ChatGPT's Image Generator Still Struggles with Political Prompts
ChatGPT's image generation keeps running into the same problem: users ask for politically-themed images and get results that are either nonsensical, refused outright, or wildly off from what was requested.
OpenAI Appears Ready to Launch a $100/Month ChatGPT Plan
The gap between ChatGPT Plus at $20/month and ChatGPT Pro at $200/month has always been awkward. One gives you GPT-4o with usage caps. The other gives you unlimited access to everything, including the o1 pro reasoning model. There is nothing in between for people who need more than Plus but cannot justify $2,400 a year.
TurboQuant Ported to Apple Silicon: 4.6x Memory Savings at 98% Speed on M4 Macs
Google published the TurboQuant paper. Within days, someone already got it running on Apple Silicon with near-native performance.
AI Subscription Fatigue Is Real: Users Consolidating $100+/Month Tool Stacks
$100 or more per month. That's what a growing number of AI power users report spending across separate subscriptions to ChatGPT, Claude, Gemini, and other AI tools. The tab-switching, cooldown juggling, and monthly billing adds up fast, and users are actively looking for ways to consolidate.
AI Agent Wastes a Gift Card Scammer's Time for 4 Hours Straight
A scammer sent a text asking someone to buy a $500 gift card. What the scammer didn't know: an AI agent was answering.
Claude's Refusal Problem Keeps Frustrating Power Users
What happens when your AI assistant decides it knows better than you?
GPT-5.4 and the Predictable Cycle of AI Model Hype
Every major AI model release follows the same script. Week one: "This is the best model ever, it finally understands me." Week three: "Did they nerf it? It was so much better at launch." Week six: "This model is terrible now, bring back the old version."
Users Report Claude Opus 4.6 Producing Lazier, More Delegating Responses
A growing number of Claude users are reporting that Opus 4.6, Anthropic's flagship model, has started giving instructions instead of doing the work. The complaints follow a familiar pattern: the model suggests steps, outlines approaches, or tells users what to do rather than writing the code, drafting the text, or completing the task directly.
4,500+ ChatGPT Conversations Were Publicly Searchable on Google
4,500 ChatGPT conversations. That's the confirmed count of shared chats that Google indexed and made searchable to anyone with a browser. Independent estimates put the real number above 100,000.
Claude Max at $200/Month: Why Some Users Hit Rate Limits and Others Don't
$200 a month. That's what Anthropic charges for Claude Max, its highest consumer tier - ten times the price of the $20 Pro plan. The selling point is dramatically higher usage limits. But users on the plan report wildly different experiences: some sail through heavy workdays without interruption, while others slam into rate limits within hours.
Claude Pro Users Push Back on Peak-Hour Rate Limits After Usage Surge
Three weeks after hitting 1 on the App Store and running a 2x usage promotion to welcome a wave of new users, Anthropic is facing sharp criticism from its paying subscriber base over peak-hour rate limits on Claude Pro.
Apple M5 Max vs M3 Max: Local LLM Benchmarks Skip a Generation
Two generations of Apple Silicon, same test setup, same model. Early benchmarks comparing the M5 Max to the M3 Max for local LLM inference (running AI models directly on your laptop instead of calling a cloud API) are starting to surface, and the results matter for anyone who cares about running AI privately.
A CS Professor's Rebuttal to the NYT's 'End of Programming' Story
"The realms of programmers and everyday people, separated for decades by an ocean of arcane know-how, are drifting closer together." That line, from a March 12 New York Times Magazine piece titled "Coding After Coders," kicked off the latest round of "programming is dead" discourse. Computer science professor Curry Guinn thinks the NYT got the story exactly backwards.
Whisper on Apple Silicon: Local Benchmarks Show 33x Faster Than Real-Time
Nine seconds. That's how long it takes to transcribe five minutes of audio using OpenAI's Whisper model running locally on an M2 MacBook Air, if you use the right setup. A detailed benchmark from the Yaps team puts hard numbers on the three ways to run Whisper on Apple Silicon, and the performance gaps are dramatic.
RepoWire Lets Multiple Claude Code Sessions Talk to Each Other in Real Time
Running Claude Code in one repo while needing information from another is a constant friction point. You either copy-paste context manually, maintain stale documentation, or just give the agent incomplete information and hope for the best. RepoWire, a new open-source project by Prassanna Ravishankar, tries to fix this by letting multiple Claude Code sessions talk to each other directly.
Coca-Cola and Walmart CEOs Both Say AI Drove Their Decisions to Step Down
Two of America's highest-profile CEOs have now said the same thing on their way out the door: the AI shift needs a different kind of leader.
Users Are Turning ChatGPT Into a Patient Cooking Instructor
A year ago, the default ChatGPT use case was "write me an email" or "summarize this document." Now, people are using it to learn hands-on skills like baking from scratch - and getting surprisingly good results.
Zhipu AI to Open-Source GLM-5.1 Model Weights on April 6-7
Zhipu AI, the Beijing-based company behind the ChatGLM series, plans to release the full model weights for GLM-5.1 on April 6 or 7. Once published, anyone can download and run the model locally - no API fees, no usage limits, complete privacy.
Meta's SAM 3.1 Tracks 16 Objects in Video Simultaneously at 32 FPS
Meta just released SAM 3.1, the latest version of its Segment Anything Model for real-time video object detection and tracking. The headline improvement: it can now track up to 16 objects simultaneously in a single forward pass, doubling throughput from 16 to 32 frames per second on an H100 GPU.
Meta Has Built Four Custom AI Chips in Two Years to Cut Its GPU Dependency
Hundreds of thousands of custom silicon chips are already running inside Meta's data centers, and four generations arrived in roughly two years. That pace tells you how seriously Meta is trying to reduce its dependence on Nvidia for the AI workloads behind Facebook, Instagram, and WhatsApp.