>It's like how LLMs learn to build world models without directly being trained to do so, just in order to predict the next token That's the whole point under c

>It's like how LLMs learn to build world models without directly being trained to do so, just in order to predict the next token That's the whole point under contention, but you're stating it as fact.