Comment by thomasahle

> Latent / embedding-space reasoning seems a step in the right direction

Might be good for reasoning, but it's terrible for interpretation / AI-safety.