Interactive scaling-laws calculator using the Chinchilla equation: sliders for model size and training tokens update the predicted loss, its floor, model-size penalty and data penalty, plus a diagram of fitting the law on small runs and extrapolating to a large one.

The law, live
loss = 1.69 + 406.4 / N0.34 + 410.7 / D0.28
Model size N: 2.2B parameters
Training data D: 4.0T tokens
Predicted loss: 2.081,778 tokens per parameterovertrained on purpose: small and cheap to serve
floor 1.69 model too small 0.27 too little data 0.12
Why labs trust it
Fit small, predict big loss (log) compute (log) cheap small runs predicted before spending
The exponents are fitted on runs that cost pennies, then trusted for the run that costs millions.