ICFPC2026. Postmortem. AI, LLM, Agents. Unagi team

Instead of posts "I ran one-shot prompt, it ran for 10 hours and gave me Counter Strike" I prefer to write about what agents didn't manage to solve. Why?

First of all, because it's boring. Nowadays all channels are about "Agents and LLM can do everything". It is not true. And it is important to understand, what they can't.

For example, it is not enough just to use LLM or agent to have good results in contests. One year ago I totally forgot about the contest and my participation was reduced to copy-pasting code and text from chatgpt to contest platform forth and back. In the tablet. About three hours during the taxi ride. And the result wasn't bad at all. 30th place.

This time two 100$ subscriptions, tough juggling with 5 agents, collaboration with friends gave 40th place.

Let's check what you should have to be at the top. I checked repository of Unagi team.

🏋️ Six participants with roles: autonomous agent system; iteration over solutions; devops; tools; hand-crafted solutions; holistic solutions; catalogue + room compiler
🏋️ Infrastructure prepared in advance: repository, CPU computational farm, AI agent collaboration tool
🏋️ Careful review and "lessons learned" from previous contests

I'm going to dive deeper into their machinery.

#friday #shitposting

Be yourself.

# We are not the center of the digital universe

Soviet ideologists were geniuses. History does go in a spiral. Back in the eighties a programmer was sitting in a dark room and typed something in dark terminal windows. Now, a quarter of the way into the 21st century we are sitting again in claudes, codexes, geminis in these dark konsoles.

I sit there too - for work and, you know, for trolls. But for study I use claude code in Visual Studio Code. It's hard to study when the whole program is written for you. So I need to tell my code assistant "You're my study assistant. You give me short answers. You don't give me advice or write a program until I ask you" You know. It's a prompt for an agent.

But then there is no use in this exercise if I just start asking the bot "write me this, write me that..." I need to think. I need to find a precise question, that covers the piece of the puzzle which eludes me. If it happens again and again, I need to find the pattern and ask a slightly wider question. You know. It's a prompt for me.

At this moment it's silly to think that we are ther center of the universe and continue to command servants who live in your computers. Claude, codex and we, humans, are agents and neural networks who are distilling each other trying to grasp grains of knowledge and to optimize our metrics. Copernikus, Galileo... You are right. We, humans, are not centers of the universe. Bring your pitchforks and torches. I'm here. I'm ready.

A solar symbol on a manhole cover in Armenia. Have a nice Friday!

P.S. Feel free to share something interesting or funny in the comments!

Precision and accuracy, bias and variance

Eleven years ago, my boss at Yandex told me that reading machine learning books in Russian was a bad idea. If you did, you wouldn't know the right terms to search for online. So, for me, the language of machine learning is English. It is more or less OK. But not always. And I'm not talking (yet) about the beautiful soup of binary classification metrics. Today, let's talk about two very simple things. Precision and accuracy.

Intuitively both are about not being wrong. For a long time, that was enough for me. I memorized that accuracy is the fraction of correct predictions. Precision is the fraction of actually positive objects out of all we predicted as the positive. Suddenly my son showed me his schoolbook on statistics, where everything was explained. Precision is how close to each other your predictions are. Pictures in the book explained precision as distance between your shots and accuracy - how close your shots are to the bull's eye.

Codex is not happy with my vague approach. If you squint, you can see that this precision-accuracy pair looks resembles two quite challenging ML concepts: variance and bias. To be more precise, the relationships run in opposite directions: low precision corresponds to high variance, while high bias tends to reduce accuracy.