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.