optimal

Definition

Best under a stated constraint — and the constraint is the part that matters. Compute-optimal scaling asks how to split a fixed training budget between parameters and tokens, and its answer says nothing about inference: a model that is compute-optimal to train may be the wrong choice to serve millions of times, which is why many deployed models are deliberately trained past that point on more tokens than the optimum prescribes.

Explainers using this term

• Scaling Skepticism — A Genealogy of the "Just Make It Bigger" CritiqueTraining Compute-Optimal Large Language Models
• Optimal Transport — The Mathematics of Moving DistributionsSinkhorn Distances: Lightspeed Computation of Optimal Transport
• Scaling Laws from Scratch — Why Making Models Bigger Makes Them Smarter (and When It Doesn't)Training Compute-Optimal Large Language Models