Karpathy's AutoResearch and what it could mean for neuroscience data

Andrej Karpathy recently shared AutoResearch, a project exploring autonomous AI-driven scientific research workflows. The idea is an agent that can form hypotheses, retrieve relevant papers, run analyses, interpret results, and iterate, the way a research assistant would, but without needing sleep or coffee. It's early, but the direction is genuinely interesting.

Why this matters beyond the hype

Most AI research tools today are retrieval tools dressed up as reasoning tools. They find papers and summarise them. AutoResearch is a step toward something more ambitious: agents that can run the full research loop. That's a different category of capability, and if the tool-use and evaluation components are solid, the bottleneck shifts from finding information to deciding what's worth investigating.

The OpenNeuro angle

OpenNeuro is a platform hosting thousands of openly available neuroimaging and neuroscience datasets, fMRI, EEG, MEG, and more, contributed by research labs globally. It's one of the most significant public data repositories in cognitive science. The datasets are structured, metadata-rich, and have been the basis of hundreds of published studies.

I've been thinking about what an AutoResearch-style agent looks like running against OpenNeuro. The pipeline writes itself: the agent queries the dataset catalogue, pulls relevant BIDS-formatted data, runs standard preprocessing, applies a hypothesis-driven analysis, and surfaces findings. The boring parts of research, which happen to be most of the work, become agentic. The researcher's job shifts to evaluation and direction.

What needs to be solved first

• →Robust data access: querying OpenNeuro programmatically at scale requires solid API and tooling integration
• →Domain-grounded evaluation: the agent needs to know what a valid neuroscience result looks like, not just a result
• →Reproducibility: automated pipelines need provenance tracking or the findings are hard to trust
• →Cost: neuroimaging analyses are computationally expensive and need GPU-backed infrastructure

Where we are exploring this

At JDML we're actively looking at how agent-driven research pipelines could be applied to open scientific databases including OpenNeuro. The combination of structured data, programmatic access, and a well-defined analysis domain makes neuroscience datasets a compelling target for automated research workflows. I'll post updates as this develops.