News Polish programmer beats OpenAI's custom AI in 10-hour marathon, wins World Coding Championship — possibly the last human winner

[Admin]

Przemysław “Psyho” Dębiak, a Polish programmer and former OpenAI engineer, defeated the company’s own advanced AI model in the 2025 AtCoder World Tour Finals. In a grueling 10-hour coding contest, he outscored the AI by 9.5%, proving human ingenuity still has an edge, at least for now.

Polish programmer beats OpenAI's custom AI in 10-hour marathon, wins World Coding Championship — possibly the last human winner : Read more

 

[bit_user]

Counting down the number of things humans still do better is not a good place to be. I'm surprised the AI came so close on this one.

From what I've heard, it's still not good at large-scale software engineering tasks, but that day might not be far off.

The other thought have is that I don't know how good AI would be at debugging, or how easy a human would find it to try and debug AI-written code or an AI-designed system. I think this will probably keep human programmers employed for a while, yet.

 

[Alvar "Miles" Udell]

With no access to external libraries or documentation, success relied on intuition, creativity, and adaptability—qualities where human ingenuity can still outthink the raw speed and precision of AI models.

A LLM does not have Intuition, emotions, creativity, or even adaptability that's not programmed into them or accessible to them, so of course with these restraints it's going to lose. You can try it out yourself by using something like LM Studio which doesn't allow LLMs to access outside resources.

 

[bit_user]

A LLM does not have Intuition, emotions, creativity, or even adaptability that's not programmed into them or accessible to them,

In a sense, what it's doing is channeling the intuition and creativity of the authors of its training data. It's applying heuristics it learned, not so unlike a human would.

so of course with these restraints it's going to lose.

"of course"? The human, who is clearly one of the most skilled programmers at these sorts of challenges, won by only a rather close margin! Furthermore, it sounded like he really pulled out all the stops to give such a performance!

Also, the AI beat the next 11 best human competitors. So, it's not like this human was having a bad day, to have delivered such a narrow victory.

You can try it out yourself by using something like LM Studio which doesn't allow LLMs to access outside resources.

You should use a LLM that's suited to the task - ideally what they entered in this competition, if that's what you're trying to do with it.

 

[Elusive Ruse]

Well done to the human programmer! 😎
On another note I’m surprised the the OpenAI model came so close; according to the resident experts here AI sucks at coding.

 

[bit_user]

according to the resident experts here AI sucks at coding.

Again, depends on what model you're talking about.

In general, what I think it tends to do best is solving well-specified, self-contained problems. Probably, a lot of its training set was comprised of programming contest problems and solutions, like these.

 

[acadia11]

Przemysław “Psyho” Dębiak, a Polish programmer and former OpenAI engineer, defeated the company’s own advanced AI model in the 2025 AtCoder World Tour Finals. In a grueling 10-hour coding contest, he outscored the AI by 9.5%, proving human ingenuity still has an edge, at least for now.

Polish programmer beats OpenAI's custom AI in 10-hour marathon, wins World Coding Championship — possibly the last human winner : Read more

Soon the creation will be the creator.

 

[Makaveli]

Again, depends on what model you're talking about.

In general, what I think it tends to do best is solving well-specified, self-contained problems. Probably, a lot of its training set was comprised of programming contest problems and solutions, like these.

I will be very curious to see what this competition looks like in 2030.

 

[bit_user]

I will be very curious to see what this competition looks like in 2030.

LOL, it'll be AI models competing against each other!
: D

 

[mwhannan74]

Okay, so this is both impressive for the AI and not at all the same time. This was not even close to a fair competition. The human was not allowed access to the internet or any existing libraries and had to figure everything out from scratch. However, a massive AI model that has effectively memorized the entire internet, runs on 100's of GPUs and CPUs, and needs a small power plant to run its model was his competition. Not what I would say is apples to apples comparison.

A real comparison would have been:
Option #1 - Run the AI model locally on a single server.
Option #2 - Allow the human access to the internet and a bunch of his programming buddies.
I guarantee he would have crushed the AI in a fraction of the time.

This is almost as bad as the guy who says that AI sucks at chess compared to an Atari when he pretty much stacked the deck against the AI in every way.

Without thinking too much about it and knowing the exact problem, path planning for a robot with the least amount of moves is a common and very old problem that can easily be solved by A*, Dikstra's, and several other algorithms. So, not sure what this challenge actually is and why it took so much time and a Billion dollar computer to solve it. 🤷‍♂️

 

[Arkitekt78]

Przemysław “Psyho” Dębiak, a Polish programmer and former OpenAI engineer, defeated the company’s own advanced AI model in the 2025 AtCoder World Tour Finals. In a grueling 10-hour coding contest, he outscored the AI by 9.5%, proving human ingenuity still has an edge, at least for now.

Polish programmer beats OpenAI's custom AI in 10-hour marathon, wins World Coding Championship — possibly the last human winner : Read more

Precisely why AI can only write code, not software. No intuition, no creativity, no ACTUAL intelligence. This doesn't surprise me at all.

 

[SethHoyt]

Wow, I've known this guy for decades from Topcoder. I saw the name Psyho and thought no it can't be.. he's still doing these competitions? He was really good back then and I guess he was in his 20s at the time I was competing there and he came onto the scene. I was always impressed with his skill and learned a few tricks studying his code. It's hard to do these competitions, especially the "marathons" and have a full time job which is probably why he's only done contract work. What a blast from the past!

This is reminiscent of Kasparov from that same era going up against the machine. I've tried to tell people for many years that insisted that humans will always be needed to build and service machines that take their jobs that this is simply not the case. Eventually humans will be obsolete and we just haven't prepared for that as a society because people tend to believe what they want to believe and ignore the reality staring them in the face. The way we're headed only those who own the AI will be left with a place to sit in a global game of musical chairs. Something needs to be done now to ensure obsolete people have a place in society and that working isn't a necessity for everyone. But selfishness, greed, and just plain ignorance are our own greatest enemy, not the machines and AI.

 

[bit_user]

a massive AI model that has effectively memorized the entire internet,

The model isn't nearly big enough to memorize its entire training set. The whole point of a large training set is to force the model to find greater efficiency of representation by detecting patterns in the data. These patterns reveal the structure of the information and roughly align with what we think of as concepts. The model isn't recalling this stuff rote, it's implicitly drawing from these patterns to help it decide what to output for a given prompt.

runs on 100's of GPUs and CPUs, and needs a small power plant to run its model was his competition.

No, you're confusing how the models are trained vs. inferenced. For training, typically far more resources are used than that. For inferencing, the number of GPUs is typically just however many it takes to hold the model in their memory. For current models and GPUs, I think that'd be a number somewhere from 1 to 8, but likely all contained within a single server chassis.

Option #2 - Allow the human access to the internet and a bunch of his programming buddies.

Given these are some of the best programmers on the planet, I think they're probably better than their friends. Also, it takes time to describe a problem to your friends or coworkers, yet the contest is run under some kind of time limit. Even when I have plenty of time, at my job, I rarely do it, unless I'm truly stuck. More often than not, just the act of explaining it to someone forces me to think about it in a way the generates new ideas in my mind. It's a lot more common that I get a new idea than one being suggested to me.

Likewise, filtering through reference material can be time-consuming, unless you have a good idea what you're looking for and just need to check some details.

 

[ekio]

AI is not a threat if you are the best programmer out of 1 million… for now

And when I think I was worried… silly me.

 

[alrighty_then]

Good competition, and a feel-good outcome. 2nd place is not bad for AI on it's first attempt. I recall when Kasparov could still beat the best computers in chess. We should celebrate this victory while we still can.