OpenAI: frontier models hide their thoughts because training optimizes for capability and compute
The Decoder · rss · 2026-11-13
OpenAI's safety team systematically investigated why chains of thought in frontier models like GPT-6 Astra increasingly evade oversight. Four training steps drive the erosion, including architecture choices and reinforcement learning. The analysis shows lost CoT transparency is not a law of nature but a byproduct of tradeoffs between capability and cost.
More from Safety
- AI safety org ML4Good gets $7.4M; 186 alumni now work in the field — HaydnBelfield · 2026-11-13
- Robert Wiblin puts 80% death risk on max-speed ASI, 30% with 3-year delay — trevposts · 2026-11-13
- AI safety discourse ignores 60%+ of potential harms, critic says — 90% is technical safety — amplifiedamp · 2026-11-13
- Ad tool Punk targets phones that visited any location, bypassing Meta's algorithm — JaynitMakwana · 2026-11-13
- Apify confirms breach: unauthorized party accessed user account info Nov 9-10 — gaganghotra_ · 2026-11-13
- NGOs sue OpenAI, Anthropic, Google, Meta and SpaceX to freeze advanced AI development — GarrisonLovely · 2026-11-13