I think part of the trick is that LLMs actually do a lot of impressive "intelligent" thinking /but it happens during training/.

So if you leave the training time annd expensive out, it looks like it's a lot cheaper to produce intelligence than it really is.