Armenian spinning top Quiz
Let's get dangerous! (c)
I'll try to send (or give, or whatever) an Armenian spinning top to the person who gives the best explanation of the difference between the left and right pictures.
I solemnly swear that both pictures were done with the same camera, at the same place, the same spinning top.
Let's set deadline for explanations as 13.00 MSK, 13 Jul 2000 + 2*13.
Let's get dangerous! (c)
I'll try to send (or give, or whatever) an Armenian spinning top to the person who gives the best explanation of the difference between the left and right pictures.
I solemnly swear that both pictures were done with the same camera, at the same place, the same spinning top.
Let's set deadline for explanations as 13.00 MSK, 13 Jul 2000 + 2*13.
❤2🤔1
Water jets - navigation
A thought here, a physics puzzle there... One by one, I collected quite a nice series of posts about fluid dynamics.
It flows from the law of conservation of energy to a DIY tower printed on a 3D printer.
Enjoy your ride down the torrent!
If you want to start quickly
Main line of the story.
106 - Water jets | tg / web
107 - Pouring-out velocity | tg / web
108 - Not exponential. This time. | tg / web
109 - New school level physics problem | tg / web
258 - The Water Tower | tg / web
272 - Jets are approximated | tg / web
273 - Stream model | tg / web
Full list
The device
106 - Water jets | tg / web
258 - The Water Tower | tg / web
270 - Experiment video and photo | tg / web
Physics of the jet
107 - Pouring-out velocity | tg / web
108 - Not exponential. This time. | tg / web
109 - New school level physics problem | tg / web
Experiment and model
272 - Jets are approximated | tg / web
273 - Stream model | tg / web
A thought here, a physics puzzle there... One by one, I collected quite a nice series of posts about fluid dynamics.
It flows from the law of conservation of energy to a DIY tower printed on a 3D printer.
Enjoy your ride down the torrent!
If you want to start quickly
Main line of the story.
106 - Water jets | tg / web
107 - Pouring-out velocity | tg / web
108 - Not exponential. This time. | tg / web
109 - New school level physics problem | tg / web
258 - The Water Tower | tg / web
272 - Jets are approximated | tg / web
273 - Stream model | tg / web
Full list
The device
106 - Water jets | tg / web
258 - The Water Tower | tg / web
270 - Experiment video and photo | tg / web
Physics of the jet
107 - Pouring-out velocity | tg / web
108 - Not exponential. This time. | tg / web
109 - New school level physics problem | tg / web
Experiment and model
272 - Jets are approximated | tg / web
273 - Stream model | tg / web
👍4❤1
This media is not supported in your browser
VIEW IN TELEGRAM
The troll farm.
My colleagues reminded me about a beautiful platform. It is a nice way to study programming languages and algorithms. Unlike leetcode, where the feedback is basically "your solution didn't pass test 573", in Codingame you can actually see how your program controls some characters, trolls this time.
The idea of the spring contest is simple. There are resources on the map: trees, water, iron, and so on. The map is symmetrical. You have a tent, and your opponent has a tent. Your program can issue simple commands: move your troll, chop, pick up, put down. The goal is to gather as much fruit and wood as possible.
Of course, I decided to participate using agents and a programming language (Rust) in which I have absolutely no experience.
Vibe programming quickly pushed me through the first two leagues and then I got stuck in the Gold league. It happened exactly when Fable became available again. It gave me a serious boost in the rankings, but not enough to beat Boss 5 and reach the next league.
Then ChatGPT 5.6 appeared and I reconstructed a bot from the next league. Unfortunately, Codingame decided to go offline, so I can not share my current position.
To have something entertaining to read during the flight, I rendered a document with small introduction to Rust and a description of bot's logic. It turned out to be really good reading. I thoroughly enjoyed it on the plane and even solved a few simple problems in Rust. I think I will share these funny little programs in the next posts.
My colleagues reminded me about a beautiful platform. It is a nice way to study programming languages and algorithms. Unlike leetcode, where the feedback is basically "your solution didn't pass test 573", in Codingame you can actually see how your program controls some characters, trolls this time.
The idea of the spring contest is simple. There are resources on the map: trees, water, iron, and so on. The map is symmetrical. You have a tent, and your opponent has a tent. Your program can issue simple commands: move your troll, chop, pick up, put down. The goal is to gather as much fruit and wood as possible.
Of course, I decided to participate using agents and a programming language (Rust) in which I have absolutely no experience.
Vibe programming quickly pushed me through the first two leagues and then I got stuck in the Gold league. It happened exactly when Fable became available again. It gave me a serious boost in the rankings, but not enough to beat Boss 5 and reach the next league.
Then ChatGPT 5.6 appeared and I reconstructed a bot from the next league. Unfortunately, Codingame decided to go offline, so I can not share my current position.
To have something entertaining to read during the flight, I rendered a document with small introduction to Rust and a description of bot's logic. It turned out to be really good reading. I thoroughly enjoyed it on the plane and even solved a few simple problems in Rust. I think I will share these funny little programs in the next posts.
👍1🔥1
Please open Telegram to view this post
VIEW IN TELEGRAM
🤯5🌚1
Vibe trolling
I have already started talking about Trolls Farm. There are two camps on a symmetric map, and your task is to collect more resources than your opponent. As always in strategy games, you have to balance strategy and tactics. You can plant a lot of trees, collect a lot of fruit, and train a super power troll. But by the time you do that the game may already be ending, and your mighty troll will have no time left to collect resources.
Let us make this a question-and-answer session.
Why vibecoding? Because, unfortunately, I do not have enough time to write all the code myself. At the same time, optimal system management is a very interesting topic to me.
Why spend time on this at all? It is an interesting task at the edge of contemporary agentic programming abilities. It is also a good opportunity to compare OpenAi Sol with Fable.
And what? Fable stopped at 135th place, Sol super max brought me to 6th place in the global rating.
How should you approach a task like this? Give the agent tools for working with the platform: compile the code, submit it to the arena, and read the current position in the leaderboard. You do not need to run an MCP server. All the models can figure out the API and use it. Codex does this faster. Claude usually needs to be explicitly told to investigate the API and write the tools down.
Just run agent and wait? Not quite. It can give you the first prototype, but after that you have to keep pushing the development toward higher places.
What works? Analysis of recent battles. Looking for recurring weak points across a series of games. Numbers, statistics. Idiotically simple, but deep models of the world, like "gain per turn" for resources, which allows to write down a simple expression like tree_gain / round_trip_length and select goal according to this expression.
What doesn't work? Manual watching games and making guesses: "Plant more trees near the camp, use seeds to train powerful trolls!" It gives you +5 wins and +15 loses. Thinking about super complex algorithmic approaches and attempts to implement something, which "automatically finds the best moves just out of the game rules"
Do you feel satisfied? Not quite. Now I understand programming multiagent systems slightly better, but now I want to make a close analysis of the program, main ideas, moving parts. So stay tuned, the case is not closed yet.
I have already started talking about Trolls Farm. There are two camps on a symmetric map, and your task is to collect more resources than your opponent. As always in strategy games, you have to balance strategy and tactics. You can plant a lot of trees, collect a lot of fruit, and train a super power troll. But by the time you do that the game may already be ending, and your mighty troll will have no time left to collect resources.
Let us make this a question-and-answer session.
Why vibecoding? Because, unfortunately, I do not have enough time to write all the code myself. At the same time, optimal system management is a very interesting topic to me.
Why spend time on this at all? It is an interesting task at the edge of contemporary agentic programming abilities. It is also a good opportunity to compare OpenAi Sol with Fable.
And what? Fable stopped at 135th place, Sol super max brought me to 6th place in the global rating.
How should you approach a task like this? Give the agent tools for working with the platform: compile the code, submit it to the arena, and read the current position in the leaderboard. You do not need to run an MCP server. All the models can figure out the API and use it. Codex does this faster. Claude usually needs to be explicitly told to investigate the API and write the tools down.
Just run agent and wait? Not quite. It can give you the first prototype, but after that you have to keep pushing the development toward higher places.
What works? Analysis of recent battles. Looking for recurring weak points across a series of games. Numbers, statistics. Idiotically simple, but deep models of the world, like "gain per turn" for resources, which allows to write down a simple expression like tree_gain / round_trip_length and select goal according to this expression.
What doesn't work? Manual watching games and making guesses: "Plant more trees near the camp, use seeds to train powerful trolls!" It gives you +5 wins and +15 loses. Thinking about super complex algorithmic approaches and attempts to implement something, which "automatically finds the best moves just out of the game rules"
Do you feel satisfied? Not quite. Now I understand programming multiagent systems slightly better, but now I want to make a close analysis of the program, main ideas, moving parts. So stay tuned, the case is not closed yet.
🔥5
The jumping wheel
About ten years ago, my friend told me a story. At MIPT, a professor of theoretical mechanics promised to give a free "excellent" mark for his course to anyone who solved a problem and defended the solution.
The problem is seemingly simple. There is a wheel of mass M with a small but heavy point mass m at the top of it. You are to find the ratio m/M at which the wheel jumps for the first time.
I attacked this problem a few times, when I had several hours of free time. The last time was during my vacation in Montenegro. To be honest, I felt extremely stupid, when I gave this problem to Claude. It produced a solution that looked viable, but I couldn't understand it.
So, let's review the state of SOTA neural networks. Claude: solved the problem and could discuss the solution. ChatGPT: knows the answer, heavily hallucinates. Me: almost totally understand Claude's approach.
As a small bonus: when m >> M, the wheel jumps when the point accelerates upward. This happens when the dot is in the rightmost position (assuming that the wheel moves to the right and rotates clockwise).
The post is already extremely long. Interesting links, the history of this problem, and recent articles - next time.
About ten years ago, my friend told me a story. At MIPT, a professor of theoretical mechanics promised to give a free "excellent" mark for his course to anyone who solved a problem and defended the solution.
The problem is seemingly simple. There is a wheel of mass M with a small but heavy point mass m at the top of it. You are to find the ratio m/M at which the wheel jumps for the first time.
I attacked this problem a few times, when I had several hours of free time. The last time was during my vacation in Montenegro. To be honest, I felt extremely stupid, when I gave this problem to Claude. It produced a solution that looked viable, but I couldn't understand it.
So, let's review the state of SOTA neural networks. Claude: solved the problem and could discuss the solution. ChatGPT: knows the answer, heavily hallucinates. Me: almost totally understand Claude's approach.
As a small bonus: when m >> M, the wheel jumps when the point accelerates upward. This happens when the dot is in the rightmost position (assuming that the wheel moves to the right and rotates clockwise).
The post is already extremely long. Interesting links, the history of this problem, and recent articles - next time.
👍4
Hopping hoop dynamics
In the series about stochastic AUROC, I already stated that the bleeding edge of science is not as distant as one might imagine. Let's again check that a bunch of expressions one can write down without breaking a sweat, can be the subject of a contemporary publication.
The article I used as the title picture for this post was published in the May 2026 issue of the American Journal of Physics 94(5):359–362. DOI:10.1119/5.0291968
Researchgate provides a downloadable pdf. It contains a rigorous analysis of the problem. But let's find some interesting facts here. First of all, the age of this problem is about one century. We can trace it through J.E. Littlewood's book "A mathematician's miscellany". The problem is attributed to H. A. Webb, who lived in 1862–1961. I think we can safely assume that he stated this problem around 1900 and it's like 130 years old.
The next thing which intrigues me, is the following phrase: "Still, these studies have not focused on the ideal, absolutely rigid, massless, and non-slipping hoop originally proposed by Littlewood." I hope that its true. And, if it is, it is wonderful. You always are to check all your assumptions. Like you know that there is a problem, which your professor gives to their students for years. One can assume that this problem and its solution is well-known. Another can run deepresearch and get the result: it is an unpublished result. And make one more publication.
I totally can't get this statement: "We will show that hopping is indeed possible, but if the hoop satisfies the conditions for hopping it will hop at the first instant of release." For me it's totally insane. The reason for the hop is the motion of this hoop and there is no motion at all at the very beginning.
Next time I want to share links on experiments with Littlewood's hopping hoop.
In the series about stochastic AUROC, I already stated that the bleeding edge of science is not as distant as one might imagine. Let's again check that a bunch of expressions one can write down without breaking a sweat, can be the subject of a contemporary publication.
The article I used as the title picture for this post was published in the May 2026 issue of the American Journal of Physics 94(5):359–362. DOI:10.1119/5.0291968
Researchgate provides a downloadable pdf. It contains a rigorous analysis of the problem. But let's find some interesting facts here. First of all, the age of this problem is about one century. We can trace it through J.E. Littlewood's book "A mathematician's miscellany". The problem is attributed to H. A. Webb, who lived in 1862–1961. I think we can safely assume that he stated this problem around 1900 and it's like 130 years old.
The next thing which intrigues me, is the following phrase: "Still, these studies have not focused on the ideal, absolutely rigid, massless, and non-slipping hoop originally proposed by Littlewood." I hope that its true. And, if it is, it is wonderful. You always are to check all your assumptions. Like you know that there is a problem, which your professor gives to their students for years. One can assume that this problem and its solution is well-known. Another can run deepresearch and get the result: it is an unpublished result. And make one more publication.
I totally can't get this statement: "We will show that hopping is indeed possible, but if the hoop satisfies the conditions for hopping it will hop at the first instant of release." For me it's totally insane. The reason for the hop is the motion of this hoop and there is no motion at all at the very beginning.
Next time I want to share links on experiments with Littlewood's hopping hoop.
👍2
Please open Telegram to view this post
VIEW IN TELEGRAM
❤7🤔2
ICFPC 2026
Sorry for jumping from topic to topic. One day I should collect all plots I started and finalize them somehow. But. It's the event I have been waiting for the whole year.
It's ICFPC 2026.
I have been taking part in it since 2008. It is not about a prize or anything. It's about learning something new and communicating with friends. Because of ICFPC I know about orbital motion and Sundman transformation. I like the idea of the Endo contest very much. The authors created a special rules of computing based on substring search and substitution. You had to write a highly optimized DNA executor and then solve tons of quests. Sand computers is also "a quest inside a quest".
It's very interesting to see what it will be this year.
It's already announced that there will be computation performed by little men. So, probably, the first task will be to write a highly optimized program, as in Endo and Sand computers. We will see.
Sorry for jumping from topic to topic. One day I should collect all plots I started and finalize them somehow. But. It's the event I have been waiting for the whole year.
It's ICFPC 2026.
I have been taking part in it since 2008. It is not about a prize or anything. It's about learning something new and communicating with friends. Because of ICFPC I know about orbital motion and Sundman transformation. I like the idea of the Endo contest very much. The authors created a special rules of computing based on substring search and substitution. You had to write a highly optimized DNA executor and then solve tons of quests. Sand computers is also "a quest inside a quest".
It's very interesting to see what it will be this year.
It's already announced that there will be computation performed by little men. So, probably, the first task will be to write a highly optimized program, as in Endo and Sand computers. We will see.
🔥1
ICFPC 2026. Lightning round.
This year, the contest authors revived an idea from one of the previous contests: they came up with a special 2D graph-based programming language and created a kind of LeetCode for it.
The contest lasts three days, with a lightning leaderboard at the end of the first one. As you can see, our team made it onto the first page.
Of course, I'm doing it with agents, but who isn't? The funniest thing I've heard in the Discord channel so far was: "My friends decided to leave the contest because they ran out of tokens."
I'm very proud that I wrote the first submission entirely by hand. Then came a lot of infrastructure work: downloaders for problem statements, program parsers, submitters for our solutions, and a simulator of the little men's computation process.
By the end of the day, 9 problems out of 12 had been solved. I went to bed and left Codex and Claude with the goals: "submit all unsubmitted problems" and "improve the scores of submitted ones."
To my surprise, they managed to achieve both goals in six hours of work.
This year, the contest authors revived an idea from one of the previous contests: they came up with a special 2D graph-based programming language and created a kind of LeetCode for it.
The contest lasts three days, with a lightning leaderboard at the end of the first one. As you can see, our team made it onto the first page.
Of course, I'm doing it with agents, but who isn't? The funniest thing I've heard in the Discord channel so far was: "My friends decided to leave the contest because they ran out of tokens."
I'm very proud that I wrote the first submission entirely by hand. Then came a lot of infrastructure work: downloaders for problem statements, program parsers, submitters for our solutions, and a simulator of the little men's computation process.
By the end of the day, 9 problems out of 12 had been solved. I went to bed and left Codex and Claude with the goals: "submit all unsubmitted problems" and "improve the scores of submitted ones."
To my surprise, they managed to achieve both goals in six hours of work.
Please open Telegram to view this post
VIEW IN TELEGRAM
❤1👍1
Please open Telegram to view this post
VIEW IN TELEGRAM
Please open Telegram to view this post
VIEW IN TELEGRAM
Please open Telegram to view this post
VIEW IN TELEGRAM
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.
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.
🔥3
Please open Telegram to view this post
VIEW IN TELEGRAM
👍1
Please open Telegram to view this post
VIEW IN TELEGRAM
❤3