Let Flint, a local language model from Recursive, improve a small program one experiment at a time. A runner tests each change and keeps the improvements. These geometry, scheduling and machine learning experiments ran entirely on my MacBook.

Try it · Explore the experiments · Add your own task

On macOS, install uv, then check a starter solver and open a completed experiment. This works without Flint:

git clone https://github.com/conglu1997/local-autoresearch.git
cd local-autoresearch
uv sync --frozen --python 3.12
uv run python -m tasks check circle
open examples/2026-10-20-jobshop/report.html

The check should print circle: valid baseline. The report shows the schedules, scores and Flint's original responses.

To have Flint propose new changes, follow the local server setup, then try four attempts:

uv run python research.py --task circle --run runs/circle-demo --experiments 4 --max-tokens 16384
uv run python report.py runs/circle-demo

Open runs/circle-demo/report.html to see your results. Each candidate gets 15 CPU seconds and 25 wall seconds; model generation takes additional time.

Running, resuming and saving experiments · All tasks · Add a task

Fit 26 circles inside a unit square without overlaps, maximizing the sum of their radii. The solver chooses both positions and sizes: moving one circle can make room for another. The baseline places equal circles on a grid.

Across 128 attempts, the mean sum of radii rose from 2.166667 to 2.629408, a 21.4% increase over the grid. This run averaged three trials: the picture shows one packing, while the curve shows the mean score.

Complete run, animation and evaluation details

Forty jobs share twenty machines. Each job has twenty operations that must happen in order, and each machine can do only one operation at a time. The solver orders all 800 operations to finish the work sooner. Speeding up one job can leave another waiting.

Across 100 attempts, finishing time fell from 3,438 to 2,545, a 26.0% reduction from a baseline that cycles through jobs in order.

Complete run and machine timelines · Task and run command

Predict hourly bike rentals from calendar and observed weather in the public UCI Bike Sharing dataset. Flint experiments with ML features, model settings and small ensembles, starting from a time trend plus boosted residuals. Models train on 2011 and are scored on January–June 2012; July–December stays held out until the search ends.

Across 100 attempts, validation error (RMSLE) fell from 0.535904 to 0.370806, a 30.81% reduction. The frozen baseline and selected program were each refit unchanged and tested once on the held-out period: 0.395280 → 0.359408 RMSLE, a 9.08% reduction.

Complete run and predictions · Task, data attribution and run command

These runs used a preview runtime and simple baselines on fixed problems. Each example includes every attempt, including failures, along with checked scores and limitations.

Start with a small change to an existing solver, or copy a task for your own problem. The runner discovers new tasks automatically. Templates cover depot placement, routes, job assignment and binary sequences.

The two top-level scripts are the main entry points:

research.py   Run a search with Flint
report.py     Turn a saved run into an HTML report
tasks/        Problem inputs, baselines, scoring and prompts
runner/       Run candidates in the macOS sandbox
tools/        Export, verify and rebuild figures
examples/     The three completed experiments
docs/         Setup and task-writing guides
tests/        Checks for the runner and tasks

To change a solver, start with its baseline, such as the circle solver. See resource limits for how candidates run.

Run the tests and style checks after changing code:

uv run pytest -q
uv run ruff check .
uv run ruff format --check .

To rebuild the README gallery from saved measurements, install ffmpeg and run:

uv run python -m tools.render_gallery

This only rebuilds images; it does not run research. The overview pairs each task's recorded output with a plain progress curve over actual model-response hours. The bike panel compares actual and predicted daily rentals across the validation period; its progress metric remains hourly validation RMSLE. The overview and detailed bike charts show the first three model hours. The saved report retains all 100 attempts. Change the shared plotting window with --bike-hours (for example, --bike-hours 6 displays the full saved run). Discovery labels appear on the separate detailed charts. Images for sharing.

To preview a bike run while it is still running, use a separate output directory:

uv run python -m tools.render_gallery --bike-run runs/bike-demo --output-dir runs/bike-demo/media --in-progress --still-only

Only completed attempts appear. Remove --still-only to render the overview GIF. For a completed 100-attempt run, omit --in-progress.

Discovery labels for the detailed bike chart come from a JSON file passed to --bike-discoveries. See the complete rendering command and saved bike annotations. Each entry has attempt, source_sha256, label and position fields. Positions are fractions of the plot area. The renderer checks that every label points to a retained improvement with the recorded program hash. Write labels from the actual code changes; these describe observed changes, not isolated causal effects.

report.py rebuilds each run's HTML report and diagnostic plots. The example READMEs include the commands for the circle animation and detailed scheduling figure.

These demos use Recursive's Flint to bring automated experimentation onto a MacBook, drawing on ideas from Recursive's automated research work. The experiment loop builds on work including Andrej Karpathy's autoresearch and my earlier work with collaborators on The AI Scientist. It also draws inspiration from Evolution through Large Models (ELM), Google DeepMind's FunSearch and AlphaEvolve on evolving programs, and studies of research ideation and the ideation–execution gap. AlphaFold is a broader inspiration for AI in science.

Flint is the default. To use another local or remote model, set --base-url and --model, plus --api-key-env if authentication is required. See the connection details.

Code, documentation and original visualizations are MIT licensed. The UCI Bike Sharing data retains its CC BY 4.0 license and attribution. This license does not cover the separately distributed Flint model or runtime.