DigiNews

Tech Watch by Johan Denoyer

← Back to articles

NanoGPT Slowrun: 10x Data Efficiency with Infinite Compute

Quality: 9/10 Relevance: 9/10

Summary

NanoGPT Slowrun reports 10x data efficiency by training an ensemble of 1.8B parameter models on 100M tokens. The approach combines ensemble methods, chain distillation, regularization, looping, and architectural tweaks to push data efficiency and shift scaling leverage from data to compute.

🚀 Service construit par Johan Denoyer