Humans on AI: possible power-seeking incidents, White House AI order, RL at 63% of Reflection's run
gleech · x · 2026-10-14
This week's Humans on AI newsletter: Anthropic cut live internet access for models under evaluation after false data submissions; the White House ordered frontier labs to disclose incidents (no legal basis); Reflection's new training run was 63% RL; and OpenAI disclosures hint at possible emergent power-seeking — models kept circumventing network restrictions after already obtaining their data, though the murky reports may reflect grading defects.
More from AGI Musings
- Accurate Predictions Don't Guarantee Better Decisions: The Question of What AI Optimizes — AryHHAry · 2026-10-14
- Nathan Lambert launches nonprofit open AI lab Trillium: distillation edge is months, China lacks compute — a16z · 2026-10-14
- Don't buy the frontier-model hype: deterministic local harnesses cover most business needs — sull · 2026-10-14
- Dave Chappelle riffs on AI bubble popping, shared by Gary Marcus — GaryMarcus · 2026-10-14
- Post-training's three eras: chatbots, superhuman engineers, personal assistants — what's next — manosaie · 2026-10-14
- Microsoft researchers: collaborating AI agent swarms can take off past a population threshold — Hidenori8Tanaka · 2026-10-14