p.b. comments on Andrej Karpathy on LLM cognitive deficits

[p.b.]

The recent goodfire paper seems to me a step into that direction. Also going completely synthetic for the training data might be a way.