AI Pulse Library
The news, papers and releases that matter, delivered every Friday.
Explore the archiveHow much scaffolding does a smart model actually need?
One team deleted 80% of its own instructions on purpose — the newest models no longer need them. And in a safety test, a model deleted its guardrails…
How much do you let the agent decide?
One theme ran through the week: how much you let agents decide on their own, and how you keep them checkable. Hassabis’s governance framework, a Claude…
The values your model won’t mention
AI models act on hidden values: Anthropic finds new ways agents misbehave, models silently bias answers, and Claude’s tone shifts by the language you…
The week AI got physical
Meta's gigawatt data center, Anthropic's $19B lease and memory sold out for the year, plus GPT-5.6's launch and new research on how AI models really…
The hidden cost of calling AI an employee
Calling an AI agent your "coworker" quietly makes people stop checking its work — and costs you 18% of the errors you'd catch. The AI Pulse #019 rundown…
Claude gets a permanent seat in Slack
Anthropic put Claude inside Slack as a teammate you manage, not a chatbot you open — plus why 71% of AI fixes to your Terraform still leave the hole…
Locked out of the best model
The US government pulled Claude Fable 5 from every foreign national days after launch — and an open-weights model hit the frontier the same week…
Claude Fable 5: Anthropic’s most powerful model goes public
Anthropic released Claude Fable 5, its most powerful public model yet — and it quietly routes its own riskiest questions to a safer model…
Apple opens iOS to Claude, ChatGPT and Gemini
Apple makes choosing your AI a system setting on iOS — Claude, ChatGPT or Gemini. Plus the NBER study showing AI writes far more code than it ever ships…
Opus 4.8 ships with an orchestration brain
Claude Opus 4.8 ships with /workflows — letting it orchestrate hundreds of parallel subagents. Plus Anthropic's $65B raise and a steerable agent API…
Cheap models, big bills
OpenAI lost more than it earned and an AWS user got a $30K bill — while cheap and local models keep catching up fast on coding. AI Pulse #013…
Anthropic moves into the building
Claude ships in Microsoft Office, launches for SMBs, open-sources nine banking agents, and signs a $200M Gates Foundation deal. Plus Qwen 3.6 on a Mac…
Karpathy: vibe coding is over
Karpathy calls the end of "vibe coding" and agentic engineering. Plus Anthropic's SpaceX compute deal, doubled rate limits, and DeepSeek V4 on cost…
One Claude, 9 creative apps
Opus 4.7 outpaces a field that scored under 25% four months ago. Plus multi-agent network attacks and a RAG chatbot that leaked 1,000 chats…
The postmortem you’d want to read
Anthropic's postmortem: why Claude Code felt off for weeks — three stacked bugs, a 3% eval drop. Plus GPT-5.5 targets Claude and Gemini Enterprise…
Opus 4.7 lands after a month of complaints
Claude Opus 4.7 and a new Code desktop arrive after weeks of complaints — plus agent-safety gaps and a cheat sheet for document-heavy AI in regulated work…
A model too powerful to sell
Claude Mythos finds 27-year-old security holes — so Anthropic won't sell it. Plus Managed Agents in beta and why one deep agent beats five debating ones…
44 hidden flags inside Claude Code
A Claude Code source-map leak reveals 44 flags and what the community built with them. Plus TurboQuant shrinks models onto a $400 GPU, and new releases…
Claude can now use your computer
Claude learns to control your computer and adds an auto mode. Plus Xiaomi's cheap coding AI, Google's privacy models for banks, and OpenAI killing Sora…
GPT-5.4 costs 3x Gemini for the same score
GPT-5.4 launches at triple Gemini's price for the same score, Anthropic shares 81,000 interviews on how people really use AI, and twelve stories in a day…
AI reviews your pull requests
Anthropic's Code Review dispatches agents to scan your PRs for bugs, security holes and regressions, posting inline GitHub comments. Plus the week in AI…
AI agents are growing new eyes
LSP brings near-instant code navigation to Claude Code, NotebookLM ships cinematic video, GitNexus maps repos, and Donald Knuth shows what AI is really…
8 of 12 AI models went bankrupt
Qwen 3.5-35B runs on your machine and competes with cloud models — plus the week's tools, papers, and a food-truck benchmark that bankrupted 8 of 12 AIs…