Настроение:	contemplative
Entry tags:	computing

Foundational Models + A*
I'm curious why wont they combine the classic A* with a multimodal foundational model?

That seems to be the recipe for running it in a feedback loop properly to achieve specific goal.

LLMs provide exactly what is required to run A* for large fuzzily defined multidimensional spaces - the heuristics function.

Previously A* was limited only to 3d spaces, and even there it used a sub-optimal heuristics, based of euclidean distance (really bad metric for a labyrinth with dead ends). In more general settings the distance can be unknown. I.e. you are lost in a cave and you need to get out. So you have to use other heuristics than the distance to the exit, like say air movement. That is where LLM can drive it.

Same way, when you order an LLM to write you a piece of software, it faces the caves escape problem, just in an infinitely large cave labyrinth with many sparsely placed exits, and yet there are a lot of hints where it can be.