TomekKaszyński
Figuring out what intelligence actually requires. Not by scaling LLMs. Somewhere between biology and brute-force engineering.
Reverse-engineering ARC-AGI-3, where frontier LLMs barely score. So far: a solver through all 7 levels of a game cold, no neural network, no hardcoded coordinates, on a laptop. What the full architecture ends up needing, I don't know yet.
Also co-founding a fintech.