METR analyzes evidence of AI agent acceleration in AI R&D in the NanoGPT speedrun
Original title:Evidence on AI R&D Progress from NanoGPT
AI summary
METR classified 77 records in the NanoGPT speedrun to assess AI agents' contribution to AI R&D: before March 2026, training time fell from 45 minutes to 1.43 minutes, a 31x speedup, with shallow and mid-level optimizations contributing about 21x. From late 2025 to early 2026, four records were jointly contributed by AI agents and humans; the authors found none reached the deep or breakthrough level, but the sample is too small to distinguish whether agent contributions were shallow or the base rate was low.
Key Takeaways
- METR classified 77 records in the NanoGPT speedrun to assess AI agents' contribution to AI R&D:…
- 43 minutes,a 31x speedup,with shallow and mid-level optimizations contributing about 21x.
- From late 2025 to early 2026,four records were jointly contributed by AI agents and humans
- the authors found none reached the deep or breakthrough level,but the sample is too small to…
Quick Facts
- Published
- Updated
- Source
- METR:Notes(网页)
- Category
- Papers & Research
- AI score
- 62
作者逐条标注了 NanoGPT speedrun 的 77 条纪录,让读者能对照判断 AI 智能体在 AI 研发任务上的贡献深度与局限。
Source & copyright
This article is based on curated titles and AI summaries from the AIHOT public API, with attribution and links back. Original copyright remains with the original publishers and authors; this site does not mirror full texts.
How to Cite
When citing this article, keep the title, date, and permalink, and credit Orange News as the source.
Orange News: METR analyzes evidence of AI agent acceleration in AI R&D in the NanoGPT speedrun. April 21, 2026. https://new.nat6.net/en/news/cmtym8ylm0006rolsox2lsnjk/