Sakana AI Agent Wins AtCoder Heuristic Contest (First AI to Place 1st)
Sakana AI ● Covered by 5 sources
Sakana AI's ALE-Agent just won a 4-hour AtCoder optimization contest outright, beating 804 human coders. It's the first time an AI has taken 1st place in a live heuristic coding contest, not just placed well.
Based on reporting by Sakana AI — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Sakana AI's ALE-Agent won AtCoder Heuristic Contest 058 outright on December 14th, finishing ahead of 804 human competitors in a four-hour event. For context, an OpenAI agent managed only second place at last August's AHC world championship. This is the first documented case of an AI winning one of these contests outright, in real time, against a field that includes working engineers from logistics and manufacturing.
The contest problem asked entrants to plan production upgrades across a tree of interdependent machines, think apple-producing machines and the machines that build them, a setup that mirrors real supply chains more than it sounds. Problem author Hiroomi Nochide expected the winning approach to lean on human intuition for the greedy construction phase, something he figured language models would struggle with. Instead, ALE-Agent took two hours to start submitting, jumped straight to first place, traded the lead with runner-up yosupo mid-contest, and then pulled ahead for good around the 2.5-hour mark.
What actually won it wasn't cleverness in the human sense. ALE-Agent followed the same general recipe humans did, greedy construction followed by simulated annealing refinement, but it got there by generating huge numbers of candidate programs in parallel, summarizing what worked, and feeding those insights back into the next round. Along the way it invented a
My take — AI-written commentary, not fact-checked reporting
I think the headline number, $1,300 for a contest win, matters less than what it implies: brute-force trial-and-error paired with self-reflection is turning into a real substitute for domain intuition, at least on tasks with clean feedback loops. That's a narrow kind of win, since Nochide himself admits the human-style greedy insight never showed up in the logs, but narrow wins compound fast in this field, and I'd bet longer-horizon contests fall to the same trick within a year or two.
Read more about this at: Sakana AI
