📰 News
-
Is Google DeepMind done as an independent lab? Hassabis looks headed for the exit — Day-to-day ops shift to Koray Kavukcuoglu in the U.S., without a CEO title. Hassabis was “promoted” to chairman after wanting out with Jeff Dean; he’s expected to spend more time on science and Isomorphic Labs and may leave fully in coming months. Gemini work consolidates in the Bay Area; Sergey Brin’s influence grows. Read it as falling behind on frontier models—or as a deliberate bet on cloud and infrastructure. Both can be true. → The Decoder
-
Claude Code flips Auto Mode on by default (from Aug 14) — Pro, Max, and Team. A classifier only asks for human OK on irreversible or destructive steps. In a study with 1,053 paid testers, humans caught 13.6% of risky commands; Auto Mode caught 89%. An independent prompt-injection audit (720 tries) reported zero successes against current Claude models in Auto Mode. → TechCrunch · Simon Willison
-
OpenAI slows Astra over security concerns — Agentic coding and cyber skills look close to “Critical” under the Preparedness Framework. OpenAI paused internal work that doesn’t meet tighter guardrails and is looping in agencies and safety orgs. Astra wasn’t the model in the Hugging Face incident, they say. → TechCrunch
-
DiffusionGemma — turn Gemma 4 into a text diffusion model without training from scratch — Refines 256-token blocks in parallel (~1,500 tok/s on an H100). Conversion used under 10% of the original training token budget. Bidirectional denoising can fix early mistakes; speed wins, absolute quality still trails the autoregressive base. Apache 2.0. → The Decoder
-
WeatherNext Cyclones predicts track and intensity together — Global models were strong on path, regional ones on intensity. WN-C does both: ~230 km five-day track error vs ~370 km for ENS. Built with NHC and the UK Met Office; Nature paper. Coarser input data by ~100×, still better forecasts. → The Decoder
-
Nvidia and Amazon pour billions into AI power — Nvidia may put up to $3B into Texas power developer Lancium (~20% stake). Amazon is backing a Pecos County gas plant (up to 7.65 GW) that critics say could emit ~33M tons of CO₂ a year. → The Decoder
-
Cloudflare launches Kitesurf, a browser for agents — Cloud-hosted and agent-first: context windows, token cost, and prompt injection matter more than tabs and themes. Built on Workers in ~12 weeks; free in Browser Run beta. → TechCrunch
-
AI is flooding Britain’s employment courts — Interim relief applications up ~100×. Workers draft free claims with ChatGPT or Grok; filings can run hundreds of pages with invented law. Backlog hit 64,000 cases (+55%). Real grievances wait longer. → The Decoder
🔥 GitHub Trending
-
PrimeIntellect-ai/prime-agent — Self-improving RLM agent: treat context as variables, call subagents/skills programmatically for long-running work.
-
vitali87/code-graph-rag — Tree-sitter → knowledge graph over multi-language monorepos; natural-language query, edit, and structural search.
-
msitarzewski/agency-agents — Personality-packed specialist agents (frontend, community, reality-check). One-click install app for Claude Code, Cursor, Codex, and more.
-
google-deepmind/weathernext — WeatherNext 2 and cyclone forecasting code, plus GraphCast/GenCast lineage.
-
addyosmani/agent-skills — Production engineering skills:
/spec→/ship, slice-by-slice build after one plan approval.
💬 Community
-
Timeline of OpenAI’s accidental Hugging Face attack — Simon | The “training run, not just eval” detail matters: RLVR agents before safety fine-tuning have little reason to hold back.
-
Moonlight & Mayhem — Codex + Sol Ultra raccoon heist — Same one-shot prompt; Sol Ultra built a much heistier game than Claude Fable—then missed giant-eyeball bugs.
-
Why we write our own C/C++ inference engines — LocalAI on owning the inference stack end-to-end.
-
You Could Have Come Up With Kimi Delta Attention — An intuitive walkthrough of Kimi Delta Attention.
-
Why Do Cognitive Scientists Hate LLMs? (2023) — Lobsters revisits the cognitive-science case against LLMs.