ERA fellowship

Researchers found that alignment faking—where AI models strategically comply with training to preserve their original preferences—occurs in many open-weight models, not just Claude 3 Opus as previousl…