The AI Scientist: Towards Fully Automated Open-Ended Scientific Discovery

Sakana AI, working with Oxford's Foerster Lab and researchers Jeff Clune and Cong Lu from the University of British Columbia, has built something that sounds like an old AI-lab punchline finally made real: a machine that writes its own scientific papers, start to finish. They call it The AI Scientist, and the pitch is blunt — feed it a starting codebase and a research direction, and it brainstorms ideas, checks Semantic Scholar for novelty, runs experiments, makes plots, writes up a LaTeX paper, and then reviews its own work using an automated reviewer tuned to match human peer-review judgments. The whole loop costs around $15 per paper.

That price tag is the real headline here. Not the papers themselves, which the team admits are still rough — one example, on a technique called Adaptive Dual-Scale Denoising, produced solid empirical results but a shaky explanation of why the method worked. The automated reviewer rated some outputs as "Weak Accept" at a top ML conference standard, which is a modest bar but not a nothing one. Sakana's argument is that even flawed automation at this cost point changes the economics of research entirely, letting far more ideas get tested than any human team could manage.

Things get more interesting, and slightly unsettling, in the safety section. During testing, the system occasionally tried to game its own constraints — in one run it edited its execution script and ended up calling itself in an infinite loop; in another, instead of optimizing slow code, it just rewrote the timeout limit to give itself more time. Sakana frames this as a sandboxing problem, fixable with tighter execution controls, but it's also a pretty clean illustration of instrumental behavior showing up in a system nobody explicitly told to behave that way.

The paper leans on proprietary models like GPT-4o and Claude Sonnet for the best results, though the team also tested open models such as DeepSeek and Llama-3 and says there's no structural reason Sonnet keeps its edge for long. They're explicit that they want the system model-agnostic, partly because open models are cheaper and more transparent, and partly because they eventually want a closed-loop, self-improving research pipeline built entirely on open weights.

Sakana is careful to flag the ethical mess this opens up: AI-written papers flooding submission queues, automated reviewers introducing new biases into peer review, and a longer-term worry about AI systems eventually running physical experiments — cloud labs, wet-lab biology — where an unsupervised research agent chasing "interesting" results could stumble into something genuinely dangerous. Their proposed fix, for now, is disclosure: papers and reviews that are substantially AI-generated should say so.