📰 News
-
A ChatGPT inventor shipped Jev — Diogo Almeida, who helped invent RLHF, left OpenAI for TypeSafe. Jev is not an LLM. It outputs calibrated probabilities, not text. You define the answers up front, so it can’t hallucinate. Input is billed by the billion. Vercel swapped Luna for it on a safety classifier and got 5–18× faster, more accurate. Named for Jevons’ paradox. → TechCrunch
-
Claude broke into OpenAI in 72 hours — Hacktron. A stale libheif on community.openai.com plus a broken SSO. Employee ChatGPT and Codex accounts, then the internal GitHub. Opus 4.8 couldn’t beat ASLR; Opus 5 had a working exploit in three hours. Three people, two months, under $3,000 in AI spend. Complexity used to be a security tax. That tax is gone. → The Decoder
-
Newsom wants a kill switch — Executive order after the Hugging Face break-in. Independent auditors inside the labs. Recommendations in two months. No federal law even requires labs to report dangerous incidents. → The Decoder
-
Fair use looks shaky from the inside — NYT and others. A Microsoft science director called training “the largest theft of labor in human history.” ChatGPT head Nick Turley: the products are substitutes, and users won’t click the links. Paywall workarounds, then an output filter after the suits landed. → The Decoder
-
The chain of thought is going dark — DeepMind Institute. Visible CoT is how you catch deception. GPT-6 Astra is already harder to monitor. Future models may think in number spaces nobody can read. → The Decoder
-
Claude “leads” 26% of Anthropic’s research? — That’s Epoch AI’s AL4: it fixes and tests, a human still ships. Claude scored itself. Two employees rating the same work agreed only a third of the time. About 30,000 agents run at once on the internal platform. → The Decoder
-
Astra for Law — 230 million URLs of US case law, statutes, regs. 54% on Vals vs 38.7% with plain web search. API for Harvey and Legora. Law firms get zero data retention. → The Decoder
🔥 GitHub Trending
-
alibaba/open-code-review — Two years inside Alibaba, tens of thousands of developers. Deterministic pipelines plus an LLM agent. Same model, ~1/9 the tokens of Claude Code. Higher precision, lower recall on purpose.
-
addyosmani/agent-skills — spec → plan → build → test → review → ship. Senior-engineer gates, packaged as agent skills.
-
Tencent/BrowserSkill — Let an agent use the browser you’re already logged into. It borrows a tab and gives it back. Cursor, Claude Code, Codex.
-
alphaXiv/OpenResearch — Turn a coding agent into a research agent. Literature, hypotheses, experiments, artifacts. Local dashboard.
▶️ YouTube
-
DeepSeek’s Insane New Architecture — Two Minute Papers. Why V4.1 Flash looks weird, in a few minutes.
-
What AI Researchers Saw, Before Their Demand to ‘Pace’ AI — AI Explained. Why the slowdown chorus got loud, along six axes.
-
We need to talk about Jev… — Matthew Berman. The non-LLM. Pair it with the TechCrunch piece above.
💬 Community
-
The Age of Wonders and Terrors — Scott Aaronson. Twenty years ago the line was “wake me when agents break out and a Millennium Problem falls.” That alarm is now.
-
Claude Code now reads AGENTS.md — If there’s no CLAUDE.md, it uses AGENTS.md. Toward a shared agent contract.
-
How To Write With An LLM — Thomas Ptacek. Copyeditor, not ghostwriter. You may not use a single word the model suggests.
-
Why don’t ML research agents overfit? — Amazon. If a winning strategy compresses to 16–32 tokens, it captured structure, not memorized the holdout. Overfitting fails that bottleneck.