Structure and scale

Chalvatzaki's read on the GPT-6 Astra demos: scaling has not solved physical intelligence, but writing the results off as retrieval or cherry-picked demos does not hold either. What changed is that a general model can now use structure (action spaces, IK, simulators, RL) rather than replace it. Three different things get flattened into "the model controls the robot": Astra as the policy, Astra working through structure someone else provided, and Astra constructing the process that trains a specialist. The pen-spinning demo is the third kind, since it built the mesh, the Isaac Lab task and the reward, then trained a PPO policy and never touched the hand.

The numbers collapse exactly where contact physics lives: 19/20 on block-into-bowl, 2/20 on puzzle insertion, 7/100 on StationeryBench. The harness is part of the experiment: same model, 62.7% vs 99.9% on ARC-AGI-3 depending only on the interface around it. Her closing move is the one worth keeping: perhaps scale does not make structure obsolete, perhaps scale is becoming capable of using it.