jan betley comments on Localizing goal misgeneralization in a maze-solving policy network

[jan betley]

Thanks! Indeed, shard theory fits here pretty well. I didn’t think about that while writing the post.