@love786

LLMs are incredible at System 1 thinking—pattern matching, linguistic probability, and retrieving information they've seen millions of times. The Problem: ARC tasks are intentionally "OOD" (Out of Distribution). You can’t "predict" the next pixel based on a trillion words of internet text. The Goal: To solve an ARC task, the model must build a mental program (a "System 2" process) on the fly. It has to look at 3 examples, say "Oh, the blue square is actually a 'container' and the red dots are 'gravity'," and then apply that logic to a new grid.

0 reply

0 recast

0 reaction