> LLMs were never designed for this. […] AI researchers have known this from the start and won't be surprised by this because it was never intended to be able t

> LLMs were never designed for this. […] AI researchers have known this from the start and won't be surprised by this because it was never intended to be able to do this. Sure, but it doesn’t help when well-respected researchers, like Ilya Sutskever, go around talking about how OpenAI’s LLMs have intelligence. There have been plenty of commenters on HN who, without a hint of irony, talk about how “well maybe self-attention is the mechanism of consciousness”. And all the scaling papers suggest (suggest in the sense that they seem to want to draw that inference, not that I am endorsing it) that LLMs have no limit to scaling with more parameters and training tokens. Still other (serious, and well-cited) papers run benchmarks that include math and logic tests… why would anyone do that if they genuinely believed that LLMs are just stochastic next-token predictors? So this is a little more than “you’re holding it wrong”. The entire AI/ML industry is telling you to hold it that way, then acting shocked when they discovered gambling in the establishment.