Comment by authorfly

The non-standard grid thing was an argument against deep learning / chess / Go AIs before Alpha Zero - Alpha Go (showing self-play can adapt with sufficient runs to any grid size or "priors" in terms of rules of the game).

It was said in 2014 by a professor I learned from that clearly AI that learned a specific game was just learning patterns and memorizing rather than anything more than that, and wouldn't be able to adjust like humans could to say new board shapes, or rules. (They would later claim 1.5 years later at a lecture that "accurate facial recognition is possible. But high recall on facial recognition is impossible, making it useless for surveillance, so don't worry").

I expect the same will occur for LLMs (but maybe sufficient "chain of thought" steps rather than game runs, etc).