Topic · 15 episodes across 5 reviews

Can We Still Watch the Model Think? Oversight and Monitoring

← all reviews

A sobering cluster: the transcript often can't reveal misbehavior, chain-of-thought monitoring fails across languages, models can resist their own training, and more capability sometimes means less trustworthy outputs.

Covered in these reviews