📰 News
-
DeepMind leadership reset — Hassabis to Alphabet Chief Scientist; Jeff Dean exits — Demis Hassabis steps back from day-to-day ops to focus on AGI strategy; Koray Kavukcuoglu runs DeepMind. After 27 years, Jeff Dean leaves with Sanjay Ghemawat, Oriol Vinyals, and Quoc Le to found Discovery Loop, a PBC aimed at automating ML experiments and scientific research. → The Decoder
-
Google Assistant dies September 2026; Gemini takes Android and Wear OS — Phones, tablets, watches, headphones, Android Auto. Cars with “Google built-in” keep Assistant for now. Open question: can a probabilistic LLM match the old deterministic assistant for lights-and-timers reliability? → The Decoder
-
Meta ships Muse Code beta — a terminal agent for large codebases — Powered by Muse Spark. Plans, writes, validates; big jobs fan out to sub-agents in isolated worktrees. Positioning against Codex and Claude Code, with cost as a selling point. → TechCrunch
-
Anthropic confirms it is hiring an AI chip design team — Co-design hardware with models for speed and efficiency. Reports of Samsung talks; same “custom silicon” path as OpenAI’s Jalapeño, Google TPUs, Meta MTIA. → TechCrunch
-
Mistral’s Shieldstral (3B) matches much larger safety models — Runtime plain-language yes/no rules instead of fixed taxonomies. Text F1 84.9%, tying GPT-OSS-Safeguard-20B (~7× larger). Apache 2.0 open weights. → The Decoder
-
SpaceX compute ambitions: math that can exceed two million Nvidia Rubin GPUs — Musk: >2GW by end of 2027 (now ~1.4GW), all on Nvidia. Full Vera Rubin expansion implies well past a million chips. AI-segment revenue jump driven more by Colossus leases (likely Anthropic) than Grok alone. → The Decoder
-
Black Forest Labs makes FLUX 3 Video generally available — HD/FHD clips up to 20s with native audio and lip-sync in 14+ languages. BFL’s Elo puts it ahead of Seedance 2.0 and Minimax H3. → The Decoder
-
Hark previews Handoff, a browser-use agent — Navigates sites without APIs via structure + vision. Claims next-action prediction, not next-token; waitlist for late-summer launch. → TechCrunch
🔥 GitHub Trending
-
cloudflare/computer — Virtual filesystem inside a Durable Object; give your agent a computer (preview).
-
huangruiteng/loopx — State kernel for long-running agent teams: goals, gates, evidence, quota, handoffs across Codex/Claude Code.
-
addyosmani/agent-skills — Production engineering skills packaged for coding agents (
/spec→/ship). -
obra/superpowers — Spec → plan → subagent TDD methodology that auto-triggers in common agent harnesses.
-
esengine/DeepSeek-Reasonix — DeepSeek-native terminal coding agent tuned for prefix-cache stability; single Go binary.
▶️ YouTube
-
The Billion Dollar AI Race Just Broke — Two Minute Papers | How open models like Qwen 3.8 Max are shaking the frontier race.
-
Master Codex with these 15 Tips — Matthew Berman | Browser/computer use, scheduling, skills, thread delegation.
-
Open-source is WINNING — Matthew Berman | Qwen 3.8 benches and the open-source catch-up story.
💬 Community
-
One-shotting a Raccoon Heist game with Claude Fable 5 — Simon | Dump a 2022 GPT-3 concept tweet into Fable 5 + Claude Code for web; get a playable game.
-
LLM 0.32 + reasoning traces — Reasoning on stderr, server-side tools, OpenAI Responses, GPT-5.6 Luna as default.
-
Why we write our own C/C++ inference engines — Lobsters | LocalAI on owning the inference stack end-to-end.
-
You Could Have Come Up With Kimi Delta Attention — A walkthrough that makes Kimi Delta Attention feel inventable from first principles.