Not that odd if you really understand how *human* intelligence works. We’ve known about the existence of the “g-factor” for about a century now. If *all* of the LLM training data is downstream from human cognitive abilities, then it’s intuitively unsurprising that these abilities