Scale laws told us that loss falls predictably with compute, parameters, and tokens. They did not tell us that all tokens are equal, because in the regime that produced them, they nearly were. That regime is ending.
The experiment
We held compute fixed and varied only the composition of the training mixture, sweeping from raw web text toward a corpus engineered for teaching signal — worked examples, contrastive pairs, and explanations that name the reasoning step rather than merely performing it.
The curve bent. At matched compute, the teaching-dense mixture reached a validation loss the raw mixture needed roughly three times the tokens to touch. The scaling law was not violated; it was re-parameterised. The effective token count of a lesson is larger than its literal length.
Why this is not a free lunch
Teaching signal is expensive to manufacture and easy to counterfeit. A corpus that looks instructional but repeats the same shallow pattern collapses back onto the raw curve. The gain lives in variety of difficulty, not in surface polish — a point we return to throughout this line of work.