📰 News
-
OpenAI’s Astra is the first model it rates ‘critical’ for cyber — It finds zero-days and chains them. Perfect on ExploitBench; on 20 recent V8 bugs it blows past GPT-5.6 Sol. Browser sandbox escape and a path to root showed up under Daybreak Blue, not the consumer setup. Refusal rate 91.5% vs Sol’s 59%. The scare is recurrent depth: the same layers loop, so some “thinking” lives in numbers, not readable CoT. Pachocki calls monitoring “fragile.” Altman says the summer went to safety and Astra finished training a while ago. → The Decoder
-
DOJ tells the NYT court that training is fair use — Copying for training is not the same as what the model emits, and outputs often lack substantial similarity. Joan Didion copying Hemingway to learn sentences is the analogy. The Copyright Office report, they say, isn’t binding. A 20-page Trump-admin brief — not a ruling. → The Decoder
-
Gemini 3.8 Flash: Google’s third budget model in six weeks — DeepSWE 73.7%, a hair under Opus 5 (74.0%). $0.75/$3.75, doubling in January 2027. $0.58 per task sits on the Pareto front, but it burns more tokens, so jobs cost ~40% more than 3.7. Frontier 3.5 Pro and 4 still missing. Cyber stays inside Fairwind. → The Decoder
-
Pentagon adds ChatGPT Mil and Grok to GenAI.mil — Gemini was the only tenant. 1.7 million users. Claude is still out after refusing unrestricted DoD use; a late-August court said the “supply-chain risk” label was unlawful. → The Decoder
-
World Labs Atlas builds 3D worlds from a handful of photos — Fei-Fei Li’s lab. Camera paths go in as geometry, not prose; up to a minute of 1440p. Point clouds, Gaussian splats, real-to-sim for robots. Human raters picked Atlas 75–94% of the time over MiniMax and Gemini Omni. Early access for partners. → The Decoder
-
Thirty more Tumbler Ridge suits hit OpenAI — Teachers, a principal, students who were in the building. First aiding-and-abetting claims. WSJ: staff wanted to call Canadian police; leadership deactivated the account instead. Plaintiffs name Chris Lehane as the person who stood them down; OpenAI says that’s false. Altman is a defendant. → TechCrunch
🔥 GitHub Trending
-
Gitlawb/openclaude — One coding-agent CLI for cloud and local backends. OpenAI-compatible APIs, Gemini, Ollama, MCP, slash commands. VS Code extension bundled.
-
Imbad0202/academic-research-skills — Claude Code skills for papers. Lit search, citations, logic checks — it will not write the paper. Copilot, not pilot.
-
THU-MAIC/OpenMAIC — Tsinghua’s multi-agent classroom. One prompt, a whole course. v1.0 adds a chat workbench that plans and revises the curriculum.
-
firecrawl/pdf-inspector — Rust library that tells scanned from text PDFs in 10–50ms. Skip OCR on the ones that don’t need it; under 200ms locally.
-
browser-use/video-use — Edit video with a coding agent. Cuts filler, grades color, burns subs, spins Manim overlays. Drop footage in a folder, chat, get final.mp4.
▶️ YouTube
-
Anthropic went CRAZY (Mythos/Fable 5.1) — Matthew Berman. Hands-on with yesterday’s Fable and Mythos 5.1 drop. → YouTube
-
Data centers might get solved — Matthew Berman short. A pitch that the buildout bottleneck can break. Lands on the same day Trump called data-center protests a gift to China. → YouTube
💬 Community
-
Claude’s system prompt really doesn’t want song lyrics — Fable 5.1 bans lyrics, poems, book passages. Pre-1929 is fine. Landed days after Sony and Warner sued over lyrics databases.
-
Paint.NET: Claude rewrote Direct2D in 180k lines — Rick Brewster. Direct2D on WINE was a dead end, so Claude clean-roomed it. “Vibe coded. I cannot review 180,000 lines.”
-
llm-gemini 0.34 — Gemini 3.8 Flash lands in llm. A “make me a cool thing in html” toy in 13 seconds, 1.8 cents. Pelicans included.