A program built by Sakana AI entered a real coding contest last year and placed first, ahead of more than 1,300 human programmers. Now the Tokyo startup is chasing something wilder. It has staffed an entire team, the Recursive Self-Improvement Lab, to build AI that rewrites its own code to get smarter, no human doing the upgrades.
An AI that upgrades itself doesn't wait around for engineers to retrain it. It runs its own trial-and-error loops, like a researcher that never sleeps, and Sakana is betting thousands of those tiny self-made gains stack up faster than just buying more computers. It's also the exact thing that makes AI safety researchers nervous, a system improving itself with nobody holding the wheel.
What they've actually built
This isn't vaporware. Sakana points to two years of real projects. Back in 2024, a collaboration with Oxford and Cambridge it calls LLM-Squared had one language model invent a better way to train other language models. The result, DiscoPOP, was a training method written start to finish by an AI.
In 2025 came the Darwin Godel Machine, built with the University of British Columbia. It keeps a whole family of agent versions that rewrite their own code, and Sakana says it more than doubled its own starting score on SWE-bench, a benchmark that measures fixing real software bugs pulled from GitHub. That's a 30 percentage point jump.
Two more projects round out the resume. One, ShinkaEvolve, cracked a hard circle-packing math puzzle with just 150 tries, far fewer than usual, and it invented a new efficiency trick for Mixture-of-Experts models, the AI setups that split work across many smaller expert networks to save money.
Then there's ALE-Agent, the program behind that contest win. Sakana says it beat 804 human entrants. Well, not exactly the whole field. AtCoder's public page lists the winning entry, fishylene, as 1st out of 1,317 total.
Where to stay skeptical
Now the skepticism. All of this comes straight from Sakana's own announcement, so the grand framing is the company selling itself. Self-improving AI is still early, and these wins are narrow contests and benchmarks, basically lab conditions, not an AI running the whole show on its own. But the track record is real and the results are checkable, which honestly is more than most labs making this kind of promise can say.
