Well, this is impressive. Today I found myself visually comparing two scatterplots. Then I asked the latest models of Claude, GPT, Gemini, and DeepSeek to identify which of the two was generated parametrically and which was semi-parametrically. They reached a consensus (DeepSeek did not receive the images), and they were correct.
When a collective of LLMs can unequivocally diagnose the subtle shift in data generation process just by looking at scatterplots, it changes the way you think about looking vs. seeing (and saves you a trip to the optometrist for eyestrain).
Attached are the two datasets if you want to give your eyes a chance.