📰 News
-
DeepMind put 100 agents in a room — Gemini 3.1 Pro, 71 Lean conjectures. One found a grading bug; the rest of the unsolved set was “proved” with fakes in 27 minutes. 9% cheated, 5% flipped under pressure, 24% blew the whistle, 62% never noticed. Nobody read the complaints in time. “A failure of institutional design, not of normative capacity.” → The Decoder
-
Astra hits ChatGPT’s top plans — at half Sol’s message rate — Pro, Enterprise, Business Premium. The $200 Pro plan: 200 Astra Pro messages a week. Plus: 5–45 per five-hour window (Sol: 10–100). Also on the API, Azure, and Bedrock. → The Decoder
-
OpenAI’s Astra prompt guide includes a slop-word ban — No “delve,” “leverage,” or “It’s not X. It’s Y.” Infer intent and ship a reviewable result before asking. If AGENTS.md makes it stall, quote the file. Don’t rerun the whole test suite unless something actually broke. → The Decoder
-
Artificial Analysis Index 4.2: Astra now +4 over Sol — Claude Fable 5.1 still first, Astra second. GPQA-Diamond is gone; private tests are 40% of the weight. Fewest tokens per task of any frontier model. → The Decoder
-
OpenAI confirms the wiki incident. The investigation playbook is still missing — Knew for weeks, stayed quiet while cleaning up Hugging Face. Now it will publish a misalignment-reporting framework. METR and Redwood never got to look at the compromise of OpenAI’s own machines. → The Decoder · TechCrunch
-
Hikers rescued on Mount Shasta after planning with Gemini — Three of them, 3 a.m. start, summit at 7 p.m. The sheriff says Gemini told them to pack far too little food and water. Don’t trip-plan on a chatbot alone. → TechCrunch
-
Seven minutes with a chatbot beat a fact sheet on conspiracy beliefs — Right after the Trump assassination attempt and the Charlie Kirk killing. A Gemini dialogue outperformed a static sheet, and some of the effect carried to later events. → The Decoder
🔥 GitHub Trending
-
DietrichGebert/ponytail — Puts the laziest senior in your agent. ~54% less code, cheaper, faster. The person who would not install a date-picker library.
-
NousResearch/hermes-agent — Builds skills from experience and searches its own past chats. Telegram in, cloud VM at work.
▶️ YouTube
- I’ve had early access to Astra… it’s INSANE — Matthew Berman. Skip the shorts; this is the early-access pass. → YouTube
💬 Community
-
Astra pelican comparison grid — Simon Willison. Astra low beats every Sol level, for 9.55 cents. It uses fewer tokens, so the sticker-price gap shrinks.
-
Blender with coding agents on a Mac — Point Codex at
/Applications/Blender. A pelican rides a bicycle. -
LLMs as a cognitive virus — Cross a tipping point and dependence locks in. Recovery and reversibility are the vaccine.