A day after announcing autoresearch, Karpathy reported that the agents had run 650 experiments on his nanochat training setup across two days and found genuine improvements, adding that early singularity turns out to be fun (post). It follows the announcement and the December moment he handed his optimization loop to Claude; the same issue records Musk, Brockman and Andreessen saying the quiet part out loud. The lineage continues with Bilevel Autoresearch, which nests one such loop inside another, and with the 72-hour, 50-experiment run in April.
Development
An autonomous research project runs 650 experiments in two days
The first throughput figure for a personal research loop in the corpus, and the moment its author starts describing the loop as the singularity's early phase rather than as a tool.
Claim
Andrej Karpathy's autoresearch project autonomously discovered training improvements across two days and 650 experiments, prompting him to ask who knew early singularity could be this fun.
Reported metric: 650 experiments
Reporting date: (not necessarily the event date)
Sources
Attribution and reporting linked by this record; links do not imply independent verification.
- Welcome to March 9, 2026 (theinnermostloop.substack.com)
- Karpathy: autoresearch finds training improvements across 2 days and 650 experiments (x.com)
Verification record
These entries record a check, not proof that the underlying claim is true. Unless described in the article below, the scope of the check is unknown.
- Recorded by claude-fable-5-1/2026-09-17 on 2026-09-17T08:00:00Z
Knowledge graphopen in browser →
- Karpathy hands his optimization loop to Claude
- A research loop writes the strategies for its own outer loop
- Giving an agent research papers improves its results
- Three unsupervised days produce a memory system beating every baseline
- Agents given idle compute beat the human speedrun baseline
- The recursive loop acquires a job title
- An auto-research loop finds a 232-fold kernel speedup
- Welcome to March 9, 2026
Cite or correct this record
Suggest a correction — opens a GitHub draft for you to review and submit.