Can you explain why you think the paper is evidence against stochastic parroting? To me this result seems even less sophisticated than predicting the next move

Can you explain why you think the paper is evidence against stochastic parroting? To me this result seems even less sophisticated than predicting the next move directly GPT style (and that’s its charm since it’s much more data efficient than training that model would be). Isn’t this model just simply predicting the most likely next move that stockfish would play? I actually think their conclusion is mis-stated. Instead of saying this shows transformers aren’t mere statistical pattern recognizers but are powerful algorithm approximators they should say that it shows statistical pattern recognition trained on enough samples is sufficient for approximating sophisticated algorithms.