Projection sampling preserves capabilities but diminishes gains from alignment fine-tuning

Projection sampling maintains the core capabilities of AI models but reduces the benefits gained from alignment fine-tuning, according to recent analysis shared on LessWrong.