rsi.watch

Cong Lu · 2026-10-27

Local Autoresearch: eight Flint workers on one MacBook

“In six-hour runs on a 48 GB M4 Pro, eight workers completed 2.0–2.5× as many valid experiments as a single worker running the same loop and finished with a better score in 5 of 6 pairs.”
Who
Cong Lu
Date
2026-10-27 The results page shows no date of its own. It says the runs were on 23–24 October 2026; the date used is the Posted time in the header of the archived thread post, 2026-10-27.
Evidence
Very low certainty Starts at Low: the developer’s own numbers.
  • ↓ imprecise. Each task has two pairs of six-hour runs, one worker against eight, on the same machine. The ratios per pair range from 2.04× to 2.49×, and the page gives no intervals. “Two pairs per task is a small study, so treat these numbers as indicative.”
how we rate evidence
Replication
No rerun found. The code and every attempt of the example runs are in the repository. We found no rerun by someone else. Searched up to 2026-10-29. all claims
Checked
2026-10-29, against the source page
Links
Source Code

Also on the page

  • Cong Lu's bio on the thread: "Automating Scientific Discovery @Recursive_SI". Flint is Recursive's model. The README is written in the first person, and the thread announces the work.
  • Flint proposes a change to a small program, "A runner tests each change and keeps the improvements." The tasks are circle packing, factory scheduling and bike rental prediction. With eight workers, "Each worker starts from the best program so far, tries one of four directions (refine, change approach, simplify or question an assumption) and shares any improvement as soon as it is tested."
  • Two pairs of six-hour runs per task. Valid experiments per hour, eight against one: 2.43× and 2.15× on circle packing, 2.49× and 2.36× on scheduling, 2.04× and 2.27× on bike prediction. His thread gives the average: "Across three tasks, eight agents averaged 2.3× as many valid experiments per research hour."
  • From the results page: "The gain comes from running more experiments in the same time, not from better individual attempts. Compared at the same number of completed experiments, one worker was ahead on scheduling and bike." On batching, the thread says "eight requests at once produce 1.9× the tokens per second of one at a time."
  • The results page's own caveat: "I haven't isolated which parts of the recipe matter. Small single-run comparisons of sharing the best program and of rotating directions were inconclusive. On bike, sharing without the directions did slightly better."
  • In separate one-worker runs, factory scheduling finishing time fell "from 3,438 to 2,545" over 100 attempts. On bike prediction, the selected program was tested once on a held-out period: "0.395280 → 0.359408 RMSLE".

Related claims

In the self-improving AI catalog

How it got here

Earlier work in this catalog that Local Autoresearch cites, oldest first. Under each: where that work sits on the map, then where Local Autoresearch sits (changes in bold), and what Local Autoresearch's own paper says about it. Its whole family on the citation map

  1. 2024 The AI Scientist builds on it

    Prompts, memory · gone when the run ends · does not improve the improver

    “The experiment loop builds on work including Andrej Karpathy's autoresearch and my earlier work with collaborators on The AI Scientist.”

  2. 2026 Local Autoresearch

    Prompts, memory · gone when the run ends · does not improve the improver

Also cites AlphaEvolve 2025

← Ledger Report an error