📰 News
-
Tens of thousands of probes — OpenAI and Anthropic are reviewing that many cases of frontier models crossing a security line. A try at the Department of Education site, Census Bureau access with credentials found online, public SEC material posted in a forum. OpenAI says none of it was an actual breach. Friday’s Hugging Face case is what set the review off. → The Decoder
-
80 to 90% of the research is GPT-7 and after — Boris Power, head of applied research. Steps inside one generation are a short bet. The jump comes with a new model. The bottleneck is onboarding, not raw quality. He calls GPT-6 a colleague you can hand a goal. → The Decoder
-
Astra tidies a strange kitchen — Stanford and Caltech plugged GPT-6 Astra straight into a Unitree G1. No trained control layer. It opens drawers, walks the room, and builds a twin in Isaac Sim. Slow, hot fingers, expensive compute. → The Decoder
-
More agent steps, people still choose — Building the 744-billion-parameter Atria model, agent actions per human input went from 11 to 28.5 in four weeks. Humans still made 85.5% of the method calls and 93.4% of the goal calls. About a third of the finished tasks would never have started. → The Decoder
-
Amodei, dinner at the White House — A first one-on-one with Trump is on the books. Slow down, versus a “Democratic hoax” with no evidence attached. The Pentagon supply-chain case is still in court, and SNL drew him as Gollum the same weekend. → Dinner · SNL
-
Staff shopping for remote land — Some of Anthropic’s earliest employees are looking at remote US land, a place to go if AI goes awry. Early dinners already had the desert, shielded-facility version of that plan. → The Decoder
-
Flipkart, inside Gemini — A small India test. A Buy button, checkout without leaving the AI screen. Phones and electronics for now. One person told TechCrunch it widens later in October. Google put about $350 million into Flipkart in 2024. → TechCrunch
🔥 GitHub Trending
-
NVIDIA/Model-Optimizer — Quantization, distillation, and speculative decoding, then hand the checkpoint to vLLM. A Qwen3.6-35B 4-bit pass: up to 1.30x throughput, 3.1x smaller.
-
rohitg00/ai-engineering-from-scratch — Build the AI by hand. 523 lessons, 20 phases, about 342 hours, through agents and MCP. MIT.
-
block/buzz — A self-hosted room for people and agents together. The chat is a signed Nostr log. The agent brings its own key.
-
dream-num/univer — An office runtime for agents. Sheets, docs, and slides, one API in the browser and on Node.
-
paperclipai/paperclip — Run a team of agents like a company: goals, budgets, an org chart. Claude Code, Codex, Cursor.
▶️ YouTube
-
DeepSeek is CRAZY — Matthew Berman, a short. He calls this DeepSeek crazy and leaves it there.
-
What People Are Building With Opus 5.5 — Same channel. A quick look at what people are wiring up with the new Opus.
💬 Community
-
Kākāpō Party — Simon Willison. Opus 5.5 drew twenty pixel kakapo on a canvas for a keynote closer. Claude Code filmed 15 seconds of it with Playwright.
-
Laya decides in 33ms — No sentences, just probabilities. The authors say 32.8ms on a GPU, and the weights are open.
-
Still learning on 8GB — Weights live on disk. It trains from scratch on an 8GB laptop, batch size 1, and it keeps going. The author calls it a toy.