Intel / article

Evidence on AI R&D Progress from NanoGPT

Source: metr.org
Date: Published Apr 21, 2026

Context

A public optimization challenge provides a history against which to compare an 's contribution. METR's NanoGPT speedrun analysis examines attempts to train a small to a fixed quality target on specified hardware in less time. That makes improvements inspectable through code changes and a shared objective, instead of relying only on accounts of an agent being helpful.

The article identifies four records credited to agents alongside human collaborators as of April 2026. Those records establish useful improvements, but leave human involvement, failed attempts and spending largely unknown. They also concern a small training task. The record is a starting point for studying research contributions; assessing research automation requires knowing the effort behind the successful result.

Terms in this piece · Glossary

AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
LLM — A large language model — the neural network behind tools like Claude and ChatGPT, trained on huge amounts of text to predict what comes next.
inference — Running a trained model to get answers — the phase where AI is actually used, as opposed to trained.
pretraining — The first, biggest phase of building a model: training it on enormous amounts of text so it learns language, facts, and reasoning in general.

Read the source metr.org