📰 News
-
Claude Code flips Auto Mode on by default (from Aug 14) — Pro, Max, and Team plans. A classifier only asks for human OK on dangerous or irreversible steps. In a study with 1,053 paid testers, humans caught 13.6% of risky commands; Auto Mode caught 89%. Teams also shipped ~25% more PRs. An independent prompt-injection audit (720 tries) saw zero successes against current Claude models in Auto Mode. → The Decoder
-
Claude Code sessions can talk to each other — On macOS/Linux, sessions exchange summaries and questions across terminals. Same machine stays local; cross-machine goes through Anthropic (responses only). Built for parallel work and long-running status checks. Not on Bedrock, GCP Agent Platform, or Foundry yet. → The Decoder
-
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 said Astra wasn’t the model in the Hugging Face incident. → TechCrunch
-
Agents use roughly 600× the energy of a simple chat — Climate scientist Zeke Hausfather logged eight weeks of Claude Code (~1,100 prompts → 14k model calls, 3.2B tokens). About 150 Wh per prompt. Heavy annual use lands near an electric dryer in CO₂. His main lever: clean power at the data center, not guilt. → The Decoder
-
OpenAI acquires NextSlide — The presentation startup that turned prompts, notes, and docs into editable decks. Team is moving onto ChatGPT. → TechCrunch
-
Fields Medalist Jacob Tsimerman joins OpenAI safety — Number theorist who wrote on AI-linked “omnicide” scenarios. Argues math can help where systems still run on vibes and evals. → The Decoder
-
Cloudflare launches Kitesurf, a browser for agents — Cloud-hosted, 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
-
Readers score AI short stories higher—until they know it’s AI — 2,500+ participants couldn’t beat chance at spotting ChatGPT fiction. Ratings drop once authorship is labeled machine-made; smooth, easy prose likely inflates scores. → The Decoder
🔥 GitHub Trending
-
PrimeIntellect-ai/prime-agent — Self-improving RLM agent: treat context as variables, call subagents/skills programmatically for long-running work.
-
semantica-agi/semantica — Enterprise data → context/knowledge graphs with decision provenance for audit-heavy domains.
-
666ghj/MiroFish — Multi-agent “parallel world” simulator for what-if prediction from news, policy, or narrative seeds.
-
cloudflare/computer — Virtual filesystem + runtime inside a Durable Object—give your agent a computer (preview).
-
mattpocock/skills · addyosmani/agent-skills · obra/superpowers — Composable engineering skills:
/spec→/ship, and auto-triggered spec → subagent TDD.
▶️ YouTube
-
AI is getting a little out of control — AI Explained | Autonomous math wins, cyber-test incidents, agent message boards, DeepMind leadership churn.
-
DeepMind Just Changed How AI Sees The World — Two Minute Papers | Quick take on Gemma-side multimodal / world-perception work.
-
What is Google even doing? — Matthew Berman | DeepMind org signals and where Google sits in the AI race right now.
💬 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.
-
The Tokenpocalypse — PDFs chew tokens — Accenture anecdote: non-engineers drive usage; PDF → images → markdown is a token vacuum.
-
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.
-
Revision Prompting — Lobsters | “Make it revise again” as an industrial LLM process pattern.
-
Why we write our own C/C++ inference engines — LocalAI on owning the inference stack end-to-end.