Interactive demo of scaling limits: a slider rides the log-log loss line from ten to the twenty up to ten to the twenty-nine FLOPs, showing model size, data needs, predicted loss, shrinking gains and rough cost, with hazard flags for billion-dollar budgets, gigawatt power and the stock of human text, plus three clickable escape routes.

Ride the line
Riding the scaling line loss (log) compute (log)
Compute: 10^24.0 FLOPsModel: 87.2BData needed: 1.7T tokens
Predicted loss: 1.92Gain from last ×10: 0.10Rough cost: $2.7M
budget past $1B gigawatt-class power exceeds human text (~300T)
Chinchilla-optimal split, cost at a rough $3 per 10^18 FLOPs. Flags are order-of-magnitude markers, red when tripped.
The escape routes
Three escape routes Make more dataSynthetic, multimodal Cheaper per pointDistillation, MoE Scale thinking timeCompute at answer time
Click a route to see how it dodges the walls.