Glossary · Term

Muon-AdamW

← all terms

Definition

Plain language

A combination of two optimizers used in modern small-GPT training scripts.

As stated in the literature

A training-optimization pairing combining the Muon optimizer with AdamW, used in the Karpathy nano-GPT baseline that anchors the Autoresearch challenge.

Why it matters: Pairing optimizers exploits the strengths of each on the parameter groups they handle best, often improving small-model training efficiency.

For example, a training script might use Muon for the matrix weights and AdamW for embeddings and biases.

Heard on the show

“… take a small GPT training script — multi-head attention, rotary embeddings, flash attention, the Muon-AdamW optimizer combo — and you have five minutes of wall-clock training on a single GPU to minimize …”
Episode 053 — An AI Agent Swapped In Focal Loss And Beat A Human-Tuned Training Script

Mentioned in 1 episode

  1. 053
    An AI Agent Swapped In Focal Loss And Beat A Human-Tuned Training Script

Related terms