Scaling laws are the closest thing AI research has to an engineering curve. Plot a model's error against the compute spent training it, use logarithmic axes on both, and the result is not a scatter of dots — it is close to a straight line. Error falls as a power of the resources invested: more compute, more data, more parameters, and the model gets predictably better. The word 'law' is doing slightly too much work here. These are empirical regularities observed across many training runs, not physics. But they have held across several orders of magnitude, which is more than most things in software can claim.
Two things get lost in the retelling. The first is that a line which looks gentle on a logarithmic axis is brutal on a linear one: each further increment of quality costs several times what the last one did, which is how frontier training budgets went from millions to hundreds of millions and are still climbing. The second is that a scaling law predicts the loss — how well the model predicts the next token — and not the thing anyone actually cares about. Capabilities arrive unevenly, sometimes appearing to jump, and a perfectly smooth curve in the loss can sit underneath a very bumpy curve in what the model can do.
What makes scaling laws genuinely useful is that they let a lab spend money on purpose. Before committing to a run that cannot be taken back, you fit the curve on small models and extrapolate: this budget, divided this way between data and parameters, should land about here. The most consequential finding was that the field had been building models too large and feeding them too little — that for a fixed budget, a smaller model trained on far more data comes out better. Nearly everything built since has been shaped by that arithmetic. The laws set the limits too: if quality scales with data and the good data is running out, the curve eventually meets a problem that spending cannot solve.
The human translation is a forecast you can act on but should not fall in love with. Scaling laws are the reason the last few years felt inevitable to the people inside them: the curve kept holding, so the next result was never quite a surprise. They are also the reason so many confident predictions you read are about compute rather than about consequences. A line that has held for a decade is genuinely good evidence. It is not a promise, and it has never once told anyone what the model would actually turn out to be able to do.
