Recursive self-improvement · February 2026 · Sakana AI
The AI Scientist: automating open-ended discovery end to end
Sakana AI's pipeline ideates, codes, runs experiments, and writes the paper — then reviews itself. Notes on what an autonomous research loop says about supervision.
Our whole thesis is you supervise, AI executes. The AI Scientist is the most complete attempt we've seen to push the "AI executes" side all the way to its limit — an agent that runs the entire research loop with no human in the middle — so it's the sharpest test of where supervision has to live.
What the paper does
Sakana AI's AI Scientist automates open-ended scientific discovery end to end. Given a research direction, it:
- generates novel research ideas,
- writes the code to test them,
- runs the experiments and visualizes results,
- writes a full paper describing the findings, and
- runs a simulated peer review to score its own work.
The loop can repeat — using prior results to seed new ideas — so it behaves like a miniature, self-driving research community. They applied it across diffusion modeling, transformer language modeling, and learning dynamics, at a reported cost of under ~$15 per paper. A later version, v2, produced a workshop paper that passed real peer review.
What I take from it
The automated reviewer is the quietly important part. A generator without an evaluator drifts. The system validates an automated reviewer that approaches human agreement, and that reviewer is what lets the loop run open-ended without a person scoring every output. The quality of your evaluator sets the ceiling on how far you can let the generator run unattended.
Cost-per-idea changes the economics of being wrong. At ~$15 a paper, you can afford a high failure rate. That inverts the usual incentive: the expensive resource stops being compute and becomes the human attention needed to tell good output from plausible-looking output.
The lesson for product work: don't ask "can the agent do the task?" Ask "how cheaply can it be wrong, and how good is the thing that judges it?"
The supervision angle
The AI Scientist is exciting and a little uncomfortable for the same reason: it shows that the executing side can be almost fully automated, which means all the leverage moves to supervision — choosing directions, setting the evaluator, and deciding what to trust. That's the design problem we care about. The work we ship keeps the human as the evaluator-of-last-resort by construction; this paper is the version that removes that human, and it's instructive precisely because of what then has to be true about the reviewer.
Code and report are open-source. Read it alongside the critiques — the SIGIR Forum evaluation is a useful, skeptical counterweight to the headline claims.