Comment by bvcp

i predict the price of intelligence will continue to drop (normalise capability per $) eventually hardware production will meet demand. analog and inference chips with baked models have the potential to make inference distinct from training and substantially make the cost comparison asymmetric.