“I’m completely exhausted… barely alive.” Polish programmer becomes first in the world to beat OpenAI’s ChatGPT at algorithmic programming competition in Tokyo: how it felt

The company hosted the first-ever tournament where AI directly competed against the world’s leading developers in a physical championship. OpenAI, the creators of ChatGPT, sponsored and fielded their model in a separate «Humans vs. AI» exhibition match.

«Humanity has won (for now!),» Dembiak wrote on X (Twitter), noting that he had barely slept, participating in several competitions over three days. «I am completely exhausted… barely alive.» Interestingly, the winner is a former employee of OpenAI.

600 minutes of coding

The AtCoder World Tour Finals is one of the most prestigious tournaments in the world of programming competitions. Only the 12 best developers on the planet take part in it — based on the results of the previous year. The Heuristic direction focuses on NP-hard problems — these are problems for which it is impossible to find an ideal solution in a reasonable time, so heuristics are used — approximate methods for finding the best solutions.

All participants, including OpenAI, competed on the same hardware provided by the organizers. This ensured a level playing field for both humans and AI. According to the rules, any programming language available on the AtCoder platform was allowed. Resubmissions were not penalized, but a minimum of five minutes had to be waited between each attempt.

The competition involved solving a single, extremely complex optimization problem within 600 minutes. Dembiak’s victory is proof that human skill still matters, even in the era of rapid development of artificial intelligence, writes Ars Technica.

Dembiak — programmed for 10 hours with almost no sleep. But he admits that this may only be a temporary advantage over increasingly powerful machines.

Despite Przemysław Dembiak winning 500,000 yen, the AtCoder World Tour finals demonstrated how both humans and AI models work at the edge of their capabilities, solving extremely complex optimization problems that have no perfect solution — only progressively better options.

9.5% ahead

In the final results of the competition, programmer Przemysław Dembiak (pseudonym Psyho) won, scoring 1,812,272,558,909 points. The OpenAI model, registered as OpenAIAHC, finished with a score of 1,654,675,725,406 points. The gap was approximately 9,5%.

Despite the defeat, the OpenAI model — a modified version of an AI like o3 — took second place, ahead of 10 other human participants who made it to the finals based on the results of the year’s ranking.

OpenAI called the second-place finish a «milestone» for AI models in the world of programming competitions. «Models like o3 rank in the top 100 in programming and math competitions. But to our knowledge, this is the first time a model has made it into the top three in a premium competition,» a company spokesperson told Ars Technica.

The company added that events like AtCoder allow you to test how well AI can think strategically, plan ahead, and refine solutions through trial and error — just like a human.

Przemysław Dembiak’s victory does not look like the final victory of humanity, but rather like a bright episode in the long journey of evolution.

«Honestly, this hype seems a bit strange,» wrote a programmer at X. «I didn’t think so many people would be interested in a programming contest.»

For now, the human ability to find unexpected approaches remains unique. But perhaps next time, humans will not compete with AI — but cooperate with it. Or even watch from the stands.