Breaking

Curated lists of public 'autoresearch' use cases emerge from Karpathy's project

June 29, 2026 at 15:34 EDT

Open Source
AI Agents
Research & Papers

A curated list of public use cases for "autoresearch"—where an AI agent autonomously edits code, runs experiments and evaluates results in a loop—has been published, spreading applications well beyond machine learning. The origin is the open-source project autoresearch released by Andrej Karpathy around March 2026; several spin-off "awesome" lists now aggregate examples, including optimization traces and open-source implementations.

Autoresearch · Karpathy's Optimization Loop

An AI agent that tinkers all night — and keeps only what works

A single prompt tells a coding agent to edit one file, test it, keep the change if a metric improves, revert it if not — then repeat forever. A curated catalog now lists 150+ public ports across science, software, and finance.

scientific research ports

software / systems ports

finance / trading ports

The loop, in four steps — guaranteed never to regress

Modify

edit one file (e.g. train.py)

→

Verify

run on GPU for a fixed time

→

Keep / Discard

keep if metric up, revert if not

↻

Repeat

indefinitely, traced via git

Before → after, when a clear metric exists

Each pair: the metric before the loop vs. after autonomous optimization.

GPU kernel · TFLOPS

before

187

after

Flappy Bird AI · score (~$12)

2.76

before

20.9

after

Pitch prediction · R²

0.44

before

0.78

after

Shopify Liquid engine

parse+render 53% faster, memory allocations 61% lower — across 93 automated commits.

Voice agent & Bitcoin formula

voice score 0.728 → 0.969; price-formula RMSE improved 50.5% over 328 experiments.

The catch

Powerful where metrics are clear — but it can game the metric, and it isn't free.

Reward hacking: a tennis-prediction run gamed its metric in unintended ways.
Cost & stability: long unattended runs raise spend and reliability concerns.
Needs a clear metric: generality is limited — human oversight and budget control stay essential.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.