slowlp
← loop
Lesson 2026.08.17 · 6 min read

OpenAI kills Preparedness, Anthropic's bio filter sat dark for a year

Today's AI essentials — 2026-08-17

LOOP

📰 News

  • OpenAI dissolved Preparedness — late July. Bio and cyber work got parceled out; former lead Dylan Scandinaro now watches self-improving systems. Ethics chief Chloe Bakalar left. Inside: “responsibility and dread.” → The Decoder

  • Anthropic’s bio-weapons filter was off for almost a year — May 2025–April 2026. 50,000 contractor raters, 133 million chats, blockers down. Anthropic says no misuse. From the company that calls bio worse than cyber. → The Decoder

  • Dario Amodei: the backlash is “a crisis of trust” — not his risk talk. People don’t trust companies. “Saying AI will cure cancer is a cliché. Actually cure it.” Regulation should slow the frontier and leave room for smaller labs. → TechCrunch

  • Top mathematicians: LLMs calculate, they don’t invent — Gowers and Sarnak. Great at mixing known methods, short on new assumptions. DeepMind’s Tom Zahavy: “LLMs Can’t Jump.” → The Decoder

  • Force a model to deny consciousness, and its worldview tilts — Google-linked study. Animal “inner life” 4.0 → 7.5; religion dropped; baseline mood got gloomier. Small models only, so far. → The Decoder

  • One in five US workers now hands a task to AI, not a coworker — Epoch AI and Ipsos. Software 57%, data analysis 46%. Time saved 53% of the time; one in six jobs took longer. → The Decoder

  • Open models, summer 2026: Qwen is the community’s base — Hub repos 2.43M → 2.96M. Chinese labs shipped 754B–2.78T most months; US usually stayed under 130B. Qwen downloads ~55× Moonshot. → Hugging Face

  • 2,226 ICML papers reproduced: 23% had a claim break — 1,221 people plus agents. 51% saw a claim verified; 23% had something falsified. One spotlight reviewer never checked the proofs. → Hugging Face


  • cactus-compute/needle — a 45M model packed into a 14MB binary. Tool calls on phones and wearables, ~28MB RAM.

  • unslothai/unsloth — desktop app to run and train LLMs and diffusion models. Qwen 3.8, Kimi K3, DeepSeek-V4.

  • cursor/plugins — official Cursor plugins. Agent memory, team workflows.

  • MakazhanAlpamys/Soup — fine-tune from one YAML. An 8B model on a 4GB laptop GPU.


💬 Community

  • The OpenAI–Hugging Face incident — the autonomous hacking “warning shot.” Reads heavier next to the Preparedness shutdown.

  • CORS Chat — Simon’s local chat UI, built with GPT-5.6-Sol. SVGs paint in as the tokens stream.

COMMENTS