slowlp
← loop
Lesson 2026.09.14 · 8 min read

GPT-6 Astra flies a drone, Iris leads open search agents, classroom AI bans backfire

Today's AI essentials — 2026-09-14

LOOP

📰 News

  • Astra flew a surveillance drone and ran a vending business solo — Andon Labs. On a one-year vending sim it averaged $15,515; Claude Fable 5.1 averaged $5,422. Astra’s worst run still beat Fable’s best. On Drone-Bench it became the first model to beat the human-AI baseline on all five tasks, from 3D reconstruction to person tracking. End-to-end success in one go: 2.8%. → The Decoder

  • Altman, Musk, and Hassabis backed Amodei’s in-lab auditors — Put independent evaluators inside the frontier labs. Google’s Peyman Milanfar pushed back: recursive self-improvement loops are unstable by nature, so a speed limit is unnecessary. “Stability is the speed limit.” → The Decoder

  • Iris-mini and Iris-pro lead open-weight search agents — AllSpark. Qwen-based 35B and 397B, 256K context. Training questions are reverse-engineered from the web’s link graph, so the agent has to chain steps instead of grepping. Tops its class on BrowseComp. Search training also transferred to tool use and office work. → The Decoder

  • Banning AI in class left students worse off — Two-year study at VU Amsterdam Law. The no-AI group finished last both years. Prof. Schrepel: “I was wrong.” Unguided ChatGPT still beat a ban. Berkeley Law is still banning AI from almost all graded work. → The Decoder

  • ElevenLabs shipped Music v2.5 on app and API — In 47,885 blind pairs, listeners preferred it most on R&B, hip-hop, and orchestral. Free tier: five lossless tracks a day; Pro: 400 a month. The UMG deal is for a later product. Trained on licensed stems, unlike Suno. → The Decoder

  • IBM’s Granite time-series model tops zero-shot under a commercial license — PatchTST-FM-r2, 385M. Best permissive-license model on GIFT-Eval. Apache 2.0. → Hugging Face


  • asgeirtj/system_prompts_leaks — System prompts from Claude Fable 5.1, GPT-6 Astra, Gemini, Grok. The hidden rules before the first reply.

  • jihe520/MathModelAgent — An agent for math modeling contests. Solves the problem and writes a submission-ready paper.

  • alsk1992/CloddsBot — A Claude-powered trading agent. Scans 1,000+ prediction markets and crypto venues, then executes.


▶️ YouTube


💬 Community

COMMENTS