Scaling laws

Minghuan Liu's field notes (Oct 2026) audit 19 public robot-scaling results: 16 show multi-scale curves, 13 reasonably control variables, 1 (Figure Helix 2.5) tests a held-out loss forecast, and 0 close the loop from forecast to real-robot success. Scaling curves yes, scaling law no. Offline loss doesn't convert cleanly into robot outcomes, there's no shared pretraining distribution or common data unit (hours aren't tokens), and the next hour's content matters more than its count.