About Loss Curve

Loss Curve is an incremental game in which the chart is the science. Every curve on screen is Epoch AI’s 2024 replication of the Chinchilla scaling law, so a run’s final loss is what that published fit predicts for its parameters and tokens. Prices, revenue and time are game numbers; the info panel lists exactly which parts are real.

How to play

Tap Start training. The first run is already set up and reaches Gen 1.
A finished run that beats your current model goes on sale by itself, and lower loss earns sharply more.
Spend the money on GPUs for faster runs, tokens for bigger ones, and researchers who squeeze more work out of every GPU.
Beat the dashed target line to unlock each new generation. Gen 5 ends the run, and you can start again as another lab.

Strategy tips

For a fixed compute budget there is one best split, roughly 20 tokens for every parameter. The Best split line shows it for whatever you pick.
A lopsided run earns only as much as the smaller balanced run it matches, so ten times too many parameters wastes most of your cluster.
Longer runs build more valuable models. If the suggested run looks slow, buy GPUs before you shrink it.
Progress continues for up to eight hours while you are away.

If you would rather fight for compute than budget it, try Singularity Rush, set in the same world of four fictional labs. Reward Hack shows what happens when a trained model games its test.