straycat on Nostr: I’m thinking the class thread rule for LLMs could be something like the Einstein ...

I’m thinking the class thread rule for LLMs could be something like the Einstein equation for 4-manifolds. In both cases, we have a continuous space, the shape of which is constrained by an equation.

An untrained LLM doesn’t follow the class thread rule, but a well trained one does.

So tapestry theory would have several uses:
1. How to read the graphical (compressed) representation of the embedding space directly
2. How to detect when training is incomplete: measure the extent to which the class thread rule is or is not fulfilled.
3. Corollary to 2: how to guide LLM training to those areas where the class thread rule is being poorly followed.