OpenAI explores hiding model 'thinking', raising safety concerns
The Information reports OpenAI is testing a technique where models reveal less of their 'thinking', making them harder to monitor. Gary Marcus warns this could undermine chain-of-thought monitoring, a key safety tool.
1 source
Policy by email
Get an email when there's news on Policy
No news that day, no email.
More stories today
- Data center spending to hit $31.6T by 2050 on AI boom
- Emad Mostaque: Frontier models will one-shot at 10k tokens/sec
- LangSmith adds Messages View for agent debugging
- Notebook collection covers 30 LLM agent memory techniques
- Alibaba Qwen releases Qwen3.8-Max-0902 upgrade