Overview
Scaling laws are the empirical observation that a model's performance improves predictably as you increase three things: the amount of training data, the amount of compute, and the number of model parameters. Crucially, the improvement follows a smooth curve, so labs can forecast how much better a model will get before they build it.
This finding reframed AI research as, in part, an engineering and capital problem: if bigger reliably means better, then whoever can marshal the most data and compute has an edge. It's a reinforcing loop — capability funds revenue, which funds more compute, which funds more capability.
How it works
Add data
More high-quality training data improves the model, up to a point set by the other factors.
Add compute
More training compute lets the model learn more from that data.
Add parameters
Larger models have more capacity to capture patterns — balanced against data and compute.
Forecast
Because the curve is smooth, labs predict a bigger model's capability in advance and invest accordingly.
A concrete example
A lab can run small experiments, fit the scaling curve, and forecast that a model 10x larger will hit a particular capability — then raise the capital to build it. That predictability is why frontier AI became a race to assemble the most compute.
Limits & risks
- Scaling has limits — data and useful compute aren't infinite, and returns eventually bend.
- Bigger models cost more to train and run, raising energy and capital barriers.
- Capability gains don't automatically bring safety or reliability.
Frequently asked questions
What are scaling laws in AI?
The empirical pattern that model performance improves smoothly and predictably as you increase training data, compute, and parameters.
Will scaling continue forever?
No — data and economically useful compute are finite, and returns eventually diminish, which is why labs also pursue better algorithms and training methods.