@afeinman my position changed slighly away from that recently, some specialized models have shown not only ability to play chess at a decent level (easily beating average players, at about 1800 ELO), on games they never saw, but experiments indicate they build an internal representation of the board from the sequence of moves, which i think could qualify as "understanding". It seems possible that building a sufficiently large model to predict text could require infering a model of the world.
@afeinman yeah, i know about this argument, though i’m no philosophie expert, my understanding is that in that setup it’s not the user who follows the instruction that understands chinese, but the system of the user + the rules contained in the rule, can be argued to understand chinese, and for the same reason, if you can’t say a computer understands the world, the AI model, when ran in a computer, could be considered to have some understanding of the world.
Add comment