Why it matters

DeepMind is central scaling contributor. Understanding shapes ecosystem.

Advertisement

The architecture

Gopher: 280B baseline.

Chinchilla: revised laws.

RETRO: retrieval-augmented scaling.

Data quality studies.

DeepMind scaling seriesGopher 280Blarge baselineChinchilla 70Bcompute-optimalRETRO + retrievalaug scalingDeepMind + Google Brain merged into Google DeepMind 2023; combined scaling insights
DeepMind scaling.
Advertisement

How it works end to end

Extensive published scaling work.

Multiple model families.

Data quality + curation research.