📰 News
-
Claude Fable 5.1: better coding, cheaper cache reads — Fable and Mythos 5.1. Cache reads drop $1→$0.25; Anthropic says ~25% off typical jobs and up to 45% on long agent runs. Terminal-Bench 4.0 hits 55.8% (Mythos 60.9%); the science bench jumps 24.7→52.6%. Built-in watermarks, with a detection API for regulators, press, and fact-checkers. Caveat: at max effort it emits ~1.7× output tokens, so Artificial Analysis says it can cost 20% more. Fewer false-positive refusals, zero data retention coming this fall. Mythos stays cyber/life-sciences partners only. → The Decoder
-
ChatGPT Health plugs into Epic, read-only — 325 million patient records. Notes, labs, meds, specialist docs — summarize and build timelines. No writes back to the chart. A public-data plugin hits ClinicalTrials and PubMed. 4,300 physicians, 27 use cases, 99.1% “safe.” OpenAI still says don’t diagnose; dosage lawsuits are already in court. → TechCrunch
-
Apple calls a MacBook “shocking evidence” against an OpenAI hire — Chang Liu. An Apple circuit schematic allegedly used at OpenAI, plus a tool that shares an internal Apple name. Apple says he and colleague Yu-Ting Peng tried to wipe traces once the probe started. OpenAI: leftover access is Apple’s offboarding mess. Apple: he exploited an auth bug. 400+ Apple alumni now at OpenAI. Apple wants a hardware injunction. → TechCrunch
-
DeepMind’s new chief: the frontier is the only thing that matters — Koray Kavukcuoglu. Models sit “a little bit below” today; he’s “100% certain” they catch up. Gemini 4 is “the most ambitious run.” 3.5 Pro is months late. Flash 3.5→3.7, he says, is the hop from language model to coding agent. → The Decoder
-
AIR raises $50M to vet the agent supply chain — $10M Sequoia, $40M Greenoaks. Continuously re-check skills, plugins, and MCP servers; it already blocks ~27% of what it finds online. 20+ customers, strongest in finance and pharma. Pitch: unsigned drivers, round two. → TechCrunch
-
Runway Solaris paints the UI instead of running it — An “Interface World Model.” Clicks and voice render a live screen, no app code. Shopping and tutorials are the pitch. Text still glitches, no screen readers, Gen-4.5 underneath. Research, not a product. → The Decoder
-
Google’s election AI Overviews lean on ten domains — AlgorithmWatch. YouTube is the most-cited. AfD answers show up less often. Politicians get flattering adjectives — classic sycophancy. → The Decoder
🔥 GitHub Trending
-
tt-a1i/archify — Agent emits typed JSON IR; you get HTML/SVG architecture maps with motion. Before/Delta/After diffs, PNG and WebM export. Skill for Cursor, Claude Code, Codex.
-
jingyaogong/minimind — A 64M LLM from scratch in ~2 hours and ~$3. Pretrain, SFT, LoRA, DPO, agentic RL — all native PyTorch.
-
Osmantic/ODS — Turn a PC into a local AI server. Inference, chat UI, agents, RAG, image gen in one stack. No cloud required.
▶️ YouTube
-
Cursor just got BANNED (It’s because of Elon…) — Matthew Berman. After SpaceX bought Cursor, OpenAI winds down model supply on November 12, citing Musk companies’ contract history. Bring-your-own API key still works. → YouTube
-
GLM 5.3: Powerful AI Is Becoming Almost Free — Two Minute Papers. Z.ai’s GLM-5.3-Flash: frontier-ish coding, near-free prices. → YouTube
💬 Community
-
44% on ARC-AGI-1 for 67 cents — Tiny transformer, trained from scratch at test time, 1.5 hours on a 5090. Matches TRM/HRM; 7% on ARC-2. Code is up.
-
Codex ships LibreOffice — Simon Willison found 1.7GB in the desktop Codex cache: Python, Node, Poppler, git, and a full LibreOffice. Document skills call those binaries.
-
wrapture — Graham Dumpleton of wrapt fame. Mocking plus tracing. Every line written by an agent — not vibe coding; he designed it, the model typed.