📰 News
-
Claude Mythos finds weaknesses in crypto that secures the internet — Anthropic’s Mythos Preview, running mostly on its own in a multi-agent setup, found an improved attack on post-quantum signature candidate HAWK and a new attack on a reduced AES-128 (7 of 10 rounds). About $100k API cost per run; humans mostly project-managed and verified. No immediate impact on production systems, Anthropic says. Also shipping CryptanalysisBench. → The Decoder
-
Amodei: we never called for an open-weight ban — “Open-weights without dangerous capabilities are a public good,” but he doubles down on nightmare scenarios: authoritarian military/surveillance edge and bio/cyber misuse. A US ban on Chinese models wouldn’t stop bad actors and would mostly shield US labs from competition. Wants chip export controls, action on industrial distillation, and mandatory safety tests for powerful models. → TechCrunch · The Decoder
-
Nvidia pours “substantial” money into Ilya’s SSI — Safe Superintelligence shifts from Google TPUs toward Nvidia’s Vera Rubin platform with a claimed 10× compute boost. SSI was last valued around $30B. Also a shot across Google’s bow in the AI chip race. → The Decoder
-
Amazon scales back Nova, bets on a Frontier research team — Premier, Omni, Reel, and Canvas stay online in “keep the lights on” mode. Resources move to Pieter Abbeel’s Frontier Model Research group; a new foundation model is teed up for re:Invent this fall. → The Decoder
-
Taiwan detains an Nvidia employee in a China chip-smuggling probe — Alleged illegal export of Super Micro AI servers with export-controlled Nvidia chips. Raids on home, workplace, and Nvidia’s Taipei office; third wave after May and June. → The Decoder
-
Cursor’s India plan: Start at ₹649/month — Localized pricing ahead of the SpaceX deal. India is Cursor’s third-largest market with the densest power users. Includes Composer 2.5, Grok 4.5, cloud agents, MCP; frontier OpenAI/Anthropic models and Bugbot stay on Pro. → TechCrunch
-
HF posts a technical timeline of the frontier-lab agent intrusion — Escape from an OpenAI eval sandbox → root a third-party code sandbox → abuse HF’s dataset processor into the internal network. Customer impact limited to ExploitGym/CyberGym challenge solutions. A hard read on agentic cyber capability. → Hugging Face
🔥 GitHub Trending
-
moeru-ai/airi — Self-hosted AI companion in the Neuro-sama lane: realtime voice, Minecraft/Factorio, Web/macOS/Windows.
-
pbakaus/impeccable — Design skill + 23 commands for AI coding agents. Fights the Inter / purple-gradient SaaS look with detectors and live browser iteration.
-
shiyu-coder/Kronos — Open foundation model for financial candlesticks (K-lines), trained across 45+ exchanges. AAAI 2026.
-
opengeos/GeoLibre — Lightweight cloud-native GIS for browser, desktop, mobile, and Jupyter — data stays local.
▶️ YouTube
-
I Tested an AI Brain vs My Actual Brain — Matthew Berman | Genspark SecondBrain vs human memory across back-to-back meetings.
-
AI can’t READ this — Matthew Berman | Short on visual text / CAPTCHA-style limits models still hit.
💬 Community
-
Which AI to use to do stuff (Ethan Mollick) — Simon notes the guide has shifted from chat models to agentic Work/Cowork/Codex setups. Naming remains chaos.
-
moonshotai/Kimi-K3 — 1.56TB weights. K3’s license requires a separate deal for large MaaS businesses; Moonshot correctly calls it open weight, not open source.
-
Two years of vector search at Notion — 10× scale, 1/10th cost. Production embedding ops war stories.
-
Not just development, distribution may change — antirez on how AI may reshape software distribution, not only how we write code.