Emmanuel Ameisen：模型如何表征概念、提前规划与接受监督

Anthropic interpretability researcher Emmanuel Ameisen joins Lenny Rachitsky to discuss internal model mechanisms. The publisher highlights conceptual representations, planning across multiple tokens, and capabilities that develop through training, examining what the next-token-prediction framing can leave out.

The outline also connects interpretability with alignment and safety, the Golden Gate Claude experiment, and paths into the field. Findings and analogies reflect the guest and publisher. The full episode remains unreviewed; available YouTube automatic captions await listening checks, and the paid newsletter transcript has not been acquired.