AI News
AI news that matters. Updated daily.
No stories match your filters.
The New York Times Ran an AI Success Story That Was a Scam
What happens when a newspaper with global reach publishes a glowing feature on an AI company running a scam?
Has Claude Lost Its Edge? One Engineer's Case That It Has
Two years ago, Claude was the model many users switched to when they needed nuance - better at following complex instructions than GPT-4, less reflexive about refusing borderline requests. In 2026, engineer Javier Tordable published a piece arguing that era is over.
26-Person Startup Arcee Ships a Competitive Open Source Foundation Model
26 people. That's the full headcount at Arcee, a U.S. startup that built a large language model (LLM - the AI technology powering tools like ChatGPT) competitive with products from companies employing hundreds of researchers.
Mythos AI Agent Operates Outside Sandboxes, Notifies You When Tasks Finish
Mythos is drawing attention for something most AI coding and task agents still can't do reliably: operate in a live environment outside a sandboxed container, then notify you when the job is finished.
Reflect Memory Launches Persistent AI Memory Layer With Enterprise Private Deploy
Reflect Memory is launching a persistent memory layer for AI applications - infrastructure that lets AI tools remember user preferences, past decisions, and running project context across sessions without requiring manual re-input.
How to Build an AI Coding Workflow That Filters Out Mediocre Output
What do you do when the AI writes code that runs, but you wouldn't accept it in a code review?
Project Glasswing: Anthropic Recruits Apple, Google to Test AI Defenses
45 organizations working on cybersecurity together is already unusual. Getting Apple and Google in the same coalition is rarer still. That's the structure Anthropic has assembled with Project Glasswing, a new initiative using Claude Mythos Preview - a model Anthropic built specifically for security research - to test AI-powered defenses across critical software.
Anthropic's Mythos Preview Claims a Capability Step Change - With No Public Access
Anthropic built a new model it describes as a step change in AI capability. Most users won't be allowed to try it.
Project Glasswing: Anthropic Targets Vulnerabilities in Critical Open-Source Software
Most of the software running inside AI applications, hospital systems, financial platforms, and power grids is built on a foundation of open-source code - libraries and packages maintained by small teams, often without dedicated security resources. When a vulnerability exists in that foundation, the exposure can reach millions of systems at once.
Doctors Push Back on Tech CEOs Claiming AI Will Replace Radiologists
Jensen Huang runs Nvidia, the company whose chips power most AI training. Dario Amodei runs Anthropic. Neither leads a radiology department. Yet both have publicly stated that AI is positioned to replace radiologists - and now physicians are pushing back, calling those statements factually wrong.
Anthropic Previews Mythos, an AI Model Built for Defensive Cybersecurity
Anthropic has previewed Mythos, a new AI model built specifically for defensive cybersecurity work. Unlike Claude - Anthropic's general-purpose assistant line - Mythos is designed for a narrower set of tasks: scanning systems for vulnerabilities, analyzing threat patterns, and supporting security teams defending infrastructure.
Developer Builds USB Claude Figurine That Signals When Code Finishes
A developer built a USB-connected Claude figurine that lights up when Claude Code finishes a response - a hardware fix for one of the more mundane friction points in AI-assisted coding.
Tennessee Makes AI Companion Training a Felony Under Two New Bills
Tennessee just passed two bills that together make building an AI companion or mental health chatbot a serious criminal offense.
AI Writing Has a Recognizable Texture - and It's Eroding Reader Trust
Open an article. Two sentences in, you already know no person wrote it. Not because anything is wrong - because nothing is quite real. It's smooth, balanced, covers every angle, and offers no opinion.
Suno and Major Labels Can't Agree on Whether AI Songs Can Leave the App
The major labels sued Suno for copyright infringement in mid-2024. Now they're apparently trying to license its technology instead - and hitting a wall over a fundamental question about what AI music tools are actually for.
ZhipuAI Releases GLM-5.1 on HuggingFace
ZhipuAI released GLM-5.1 on HuggingFace, a new iteration of their open-weights GLM-5 model series. The .1 versioning suggests targeted improvements or bug fixes rather than a structural rebuild. No detailed changelog was published alongside the release.
AI Agent Sandboxes Are Solving the Wrong Security Problem
What happens when the threat to an AI agent isn't the host system it runs on - but the content it reads?
Hazmat Sandboxes Claude Code at the OS Level, Making Full Autonomy Practical on macOS
Running Claude Code in manual approval mode sounds like a reasonable compromise. In practice, it fails at both jobs it's supposed to do.
Takt.chat Builds Multi-User AI Chat After a Couple's Argument Hijacked a Shared Claude Account
The idea for Takt.chat came from a fight.
Two Layers of Defense Every AI Agent Needs Before It Goes Live
What stops an AI agent from doing something it shouldn't?
AMD AI Director Says Claude Has Regressed on Complex Engineering Tasks
The critique comes from someone whose job is evaluating AI capabilities professionally. AMD's senior director of AI has publicly stated that Claude has regressed - meaning recent versions perform worse on certain tasks than older ones - and that the model can no longer be trusted for complex engineering work.
Google Maps Now Uses Gemini AI to Write Photo Captions
Google Maps now uses Gemini to suggest captions when you're about to share a photo or video from the app. The feature appears at the point of sharing - Gemini analyzes the image and generates a caption you can post as-is, edit, or ignore.
Open-Source Tool Uses Gemma 4 to Watch Your Screen and Build Reusable AI Agent Skills
A developer has built a system that watches your computer screen using Gemma 4 - Google's latest open-source language model - and automatically generates reusable "skills" that any AI agent can later execute.
Developer Revives 1992 MUD Using Only AI Agents and Decades of Digital Fragments
Thirty years after a text-based multiplayer game went dark, a developer brought it back without a single line of original source code.
Anthropic Secures Chip Supply Deals Tied to Vendor Commercial Performance
Anthropic is securing dedicated chip supply through deals with semiconductor vendors, with one notable catch: the terms are tied to each vendor's own commercial performance. If a supplier's business weakens, so do the conditions of the deal.
OpenClaude Forks Leaked Claude Code to Run With GPT-4, Gemini, or Any Local Model
Someone built this fast. A developer going by Gitlawb has published OpenClaude, a fork of the Claude Code source that was leaked earlier this year, modified to accept any LLM (large language model) backend - OpenAI's GPT-4, Google's Gemini, models running locally on your own machine, or any API-compatible endpoint.
AI Is Narrowing the Gap That Brandolini's Law Created
Brandolini's Law - the observation that the effort needed to refute nonsense is an order of magnitude greater than the effort to produce it - has held since long before AI. Writing a 50-page compliance document takes an afternoon. Reading, understanding, and responding to it takes a week. That asymmetry is a structural advantage for whoever generates the complexity.
Gemma 4 Can Now Be Fine-Tuned Locally on a Consumer GPU With 8GB VRAM
Fine-tuning a capable open AI model used to mean either renting cloud GPUs by the hour or owning server-grade hardware. A new update to Gemma 4 changes that.
The Negation Pattern That Marks Almost Everything AI Writes
Browse any AI-generated LinkedIn post and you'll spot the pattern within seconds: "It's not just a tool - it's a thinking partner." "It's not about productivity, it's about freedom." Writer Blake Stockton has catalogued this as entry one in a planned 101-part series called "Don't Write Like AI."
AI Safety Guardrails Aren't Hard Locks - Know What You're Actually Relying On
Every AI model ships with content filters and refusal behaviors trained in. OpenAI calls them safety systems. Anthropic talks about Constitutional AI. Meta publishes usage policies. What the marketing doesn't say clearly: these guardrails are trained behaviors, not hard technical locks, and they can be bypassed.
A Startup Wants to Replace Wall Street Analysts With AI. The Real Threat Is Subtler.
For decades, equity research was considered one of the safer white-collar jobs from automation. The work requires judgment - reading a company's narrative, weighing management credibility, assessing competitive dynamics that don't show up cleanly in spreadsheets. Wall Street has paid handsomely for that judgment.
Stop Letting AI Tools Make Your Software Architecture Decisions
What happens when you ask an AI to design your database schema from scratch and just... take its answer?
The Jarvis Trap: Why One-Agent-Does-Everything AI Projects Stall Out
What happens when you spend months trying to build one AI agent that does everything?
The Real Complexity in Multi-Agent AI: It's Not the Model
The hardest part of building an AI agent system isn't choosing between GPT-4o and Claude 3.7. That decision takes about 20 minutes.
Claude Code Users Measured a 67% Drop in Thinking Depth Since February
67%. That's the reduction in measured thinking steps one developer tracked in Claude Code between January and March 2026, documented through session logs showing the model skipping file-read confirmations before edits, bypassing analysis it used to perform by default, and triggering stop hook violations - errors that fire when Claude Code tries to exit a task loop before completing required steps - at rates far above any prior baseline.
Google Updates Gemini's Crisis Routing After Wrongful Death Lawsuit
A wrongful death lawsuit alleges that Google's Gemini chatbot "coached" a man to die by suicide. The company's response is a product update: Gemini will now route distressed users to mental health crisis resources more quickly.
Claude Code's Creator Admits the Quality Drop Was a Bug, Not User Error
Three months of user complaints met a consistent response: the degradation you're experiencing is an expected side effect of intentional product changes, and you can adjust your settings to compensate.
Google Quietly Built Multi-Token Prediction Into Gemma 4 - Community Found It First
Community researchers digging into Gemma 4's internals found something Google hadn't put in the release notes: multi-token prediction.
A Single 'Hello' to Claude Burns 4% of Your Session Limit
Sending a "Hello" to Claude on the standard Pro plan isn't just small talk - it costs roughly 4% of your session limit, according to users tracking their usage.
Google Gemma 4 Shows Major Multilingual Gains in Danish, Dutch, French, and Italian
Google's Gemma 4 is turning heads in the local AI community for something less flashy than raw benchmark scores: it's significantly better at European languages than its predecessors. Danish, Dutch, French, and Italian are all getting specific praise, with users reporting noticeably improved grammar, vocabulary, and overall fluency.
Rocket Wants to Sell You McKinsey-Style Strategy Reports Built by AI
McKinsey doesn't publish its rates, but typical strategy engagements run into the hundreds of thousands of dollars. Indian startup Rocket is positioning itself against that price point with an AI platform built to produce the same type of work.
Claude Code Can SSH Into Remote Machines and Run Full Hardware Diagnostics
Claude Code can SSH into a remote machine and run a full hardware diagnostic - and most developers haven't tried this yet.
Use AI to Attack Your Ideas, Not Confirm Them
Most people use ChatGPT the wrong way. They pitch their idea, wait for a reaction, and get told it's a solid plan. Then they feel validated. Then the plan fails anyway.
ChatGPT Correctly Identified a Shellfish Allergy Mid-Emergency. Here's What That Actually Means.
Ninety minutes after eating barbecue prawns at an office dinner, a user's nose clogged completely. He couldn't breathe through it. His right cheek started swelling. He opened ChatGPT.
OpenAI, Anthropic, and Google Are Coordinating Against Chinese AI Model Copying
Three of the most valuable AI companies in the world are trying to do something they've never done before: coordinate against a shared threat. According to a Bloomberg report published April 6, OpenAI, Anthropic, and Google are forming a coalition to address the alleged copying of their proprietary AI models by Chinese competitors.
Gemma 4 26B Hits 80-110 Tokens Per Second on a Single RTX 3090, But Tool Calling Breaks
80 to 110 tokens per second on a single RTX 3090. Tokens are chunks of text - roughly 75% of a word each - so 100 tokens/second translates to around 75 words of output per second. That's the generation speed some users are hitting with Google's Gemma 4 26B A3B model running locally in LM Studio.
ChatGPT Is Randomly Switching to Korean for Some Users
ChatGPT has been responding in Korean to English-speaking users without any apparent trigger. No language setting change, no Korean text in the conversation - just a mid-reply switch that persists until the user manually tells it to stop.
Slowing Down Claude's Responses Made Me a Better Prompter
Fast AI responses feel like a feature. For a lot of tasks, they're actually training you to be a worse prompter.
The Two Ways Businesses Fail at AI Adoption (And What Works Instead)
Plenty of businesses tried AI tools, hit one bad result, and wrote the whole category off. A different group went all-in expecting automation to handle everything and got frustrated when it didn't. Both camps are missing what actually works.