Check out davidlitman.com for my bio
dl
Plenty of insights. Rather than probes, I played with the app’s logit lens and was inspired to quantitatively look into compositionality. One insight is that it appears certain tactics like forks are compositional features rather than a new feature in and of itself. You can check the logits of the fork for a piece move and compare it to the sums of the logits for the two individual threats the move creates, and the values are roughly equal.
I’m just curious if it’s common for chain of thought to have so many math quantifiers like \exists. Or am I reading that wrong?
I often wonder whether when a model says something nice if it is an emergent behavior or something it is fine tuned to say, for better or worse.
Please help me with intuition for the transformer residual stream?:
Hi,I have read the literature but could use a bit more intuition about why transformers are so well suited to interpretability. I get that you can decompose the residual stream by looking at how attention reroutes information, and I understand that rather than directly passing forward the MLPs’ outputs, the outputs are instead linearly combined with the stream...
But if this sequential adding to a shared vector scheme is so useful for learning, why don’t so many deep architectures like vanilla MLPs and CNNs or AlexNet have residual streams?
Thanks
Fix the board state but remove one of the forked pieces and look at the value of the logit lens for the same fork move without that piece through time.