Scaling law

Scaling laws made large language models predictable. Train a few small models, plot how error drops as data grows, and you can forecast how a much bigger run will perform before spending the money on it.

Figure says the same pattern holds for teaching robots from human video. It trained four models across an eightfold range of data and reports that it forecast the largest run’s error to four decimal places in advance.

That only measures prediction accuracy inside the model, not chores finished in a real bedroom. The link between the two is still the open question.