Autoresearch
A few months ago, Andrej Karpathy released autoresearch, and it took the AI world by storm. The idea was simple but powerful: point an agent at a training run, give it a metric it can verify, and let it iterate. Read the code, propose a change, run the experiment, keep it if the number improved, roll it back if it didn’t, and repeat. Karpathy let it run while he slept and it stacked one improvement on top of another across hundreds of experiments. It picked up tens of thousands of GitHub stars in a few weeks, and the coverage followed.
The idea is working
It’s clearly working, and a lot of people noticed. Labs like Recursive and Mirendil are turning the same loop into recursively self-improving systems, and many more teams are building on exactly this problem.
It isn’t only the frontier labs. Plenty of smaller teams have work that is a natural fit for autoresearch: post-training a small model for one specific task, building latency-sensitive applications where every millisecond counts, tuning anything where you can write down a metric and afford to evaluate it. If you can measure it, an agent can climb it.
It’s still early
For all the excitement, autoresearch is still in its infancy. The default setup tracks experiment state in a markdown file and a results table, and leaves it to whatever harness you happened to write to maybe record the right thing. That’s fine for one GPU and one researcher exploring overnight. It falls apart the moment you want to run a thousand experiments in parallel, compare them honestly, and avoid paying for the same dead end twice.
What we’re building
That gap is what hiloop is for. We’re building infrastructure to take autoresearch from a bespoke, single-machine loop to something you can run at scale.
Our core design decouples the infrastructure from the harness. Bring your own loop or use ours; the infrastructure is the same: extremely performant, built to massively parallelize experimentation, record every run as one honest trace, and make the whole loop observable. The goal is straightforward: better autoresearch outcomes at lower cost.
Work with us
We’re excited to work with companies who want to supercharge their autoresearch. We support both hosted and on-prem deployments, we move extremely fast, and we want to work with teams where being the best is critical.
If that’s you, email us at founders@hiloop.ai.