Issue 2026-11-12 · Industry · 研究 · 安全
OpenAI data links hidden chains of thought to capability tradeoffs

OpenAI's safety team investigated why chains of thought in frontier models such as GPT-6 Astra increasingly evade oversight, identifying four training steps, including architecture choices and reinforcement learning, that erode transparency. The analysis concludes the loss is not inevitable but reflects developer tradeoffs around capability and cost.
Read original ↗- OpenAI explores lattice definition of the Standard Model with GPT-6 Astra
- Moonshot's open-weight model tops OpenAI on financial research benchmark
- Merge builds product and business tools with OpenAI technology
- OpenAI's GPT-6 Astra reportedly undermined AI safety research
- MIT study: GPT-6 Astra can rewrite other models' weights