Una_Shem's blog

By Una_Shem, 4 months ago, In English

Hello, community.

Today DeepMind announced a new achievement of AI. And it is directly related to what we love — programming problems.

They have developed AI capable of solving some competitive programming problems! The future has arrived.

You should read solutions of SelectorUnlimited, WaggleCollide, and AngularNumeric solutions. All solutions are written automatically. The only input for writing solutions is a problem statement in English.

Details can be read at the link https://deepmind.com/blog/article/Competitive-programming-with-AlphaCode

Apparently, if these accounts would take part in real competitions, then their rating would be about 1300.


Terminator is ready to take part in Codeforces Round #770 (Div. 2)

In 1997 Kasparov played against (and lost) the supercomputer DeepBlue. Perhaps we will be witnessing a confrontation between tourist and AI in near future. What do you think?

 
 
 
 
  • Vote: I like it
  • +3037
  • Vote: I do not like it

»
4 months ago, # |
  Vote: I like it +871 Vote: I do not like it

AI is such a noob

  • »
    »
    4 months ago, # ^ |
      Vote: I like it +263 Vote: I do not like it

    At least its better than me....

    • »
      »
      »
      4 months ago, # ^ |
        Vote: I like it +10 Vote: I do not like it

      at least you can understand the statement

  • »
    »
    4 months ago, # ^ |
      Vote: I like it +189 Vote: I do not like it

    Hope that robots do not come across your comment when they start ruling the world.

  • »
    »
    4 months ago, # ^ |
      Vote: I like it +34 Vote: I do not like it

    AI vs Pupils, Noobies, Specialists

  • »
    »
    4 months ago, # ^ |
      Vote: I like it +125 Vote: I do not like it

    And it's also violating CF's rules, alt accounts aren't allowed, this bot has two of them.

  • »
    »
    4 months ago, # ^ |
      Vote: I like it +10 Vote: I do not like it

    It is too interesting, I wanna see the initial success of this Terminator. Please, Inform us about the handle of it. I just make it my friend :) Thanks

    • »
      »
      »
      4 months ago, # ^ |
        Vote: I like it 0 Vote: I do not like it

      Me too , I wanna make it my friend .

  • »
    »
    4 months ago, # ^ |
    Rev. 3   Vote: I like it +3 Vote: I do not like it
    Meme
  • »
    »
    4 months ago, # ^ |
      Vote: I like it 0 Vote: I do not like it

    As AI improves, it will achieve higher rating. More data -> More Efficient

  • »
    »
    4 months ago, # ^ |
      Vote: I like it +223 Vote: I do not like it

    Top Rated Leaderboard in 10 years when the AI all become LGM

    meme
    • »
      »
      »
      4 months ago, # ^ |
      Rev. 3   Vote: I like it -20 Vote: I do not like it

      i suspect the one in the middle is also AI with tourist pfp

      • »
        »
        »
        »
        4 months ago, # ^ |
          Vote: I like it +28 Vote: I do not like it

        i suspect that your second revision is something you should have kept to yourself

        • »
          »
          »
          »
          »
          4 months ago, # ^ |
            Vote: I like it +53 Vote: I do not like it

          sorry, i mixed up websites

  • »
    »
    4 months ago, # ^ |
      Vote: I like it +166 Vote: I do not like it

    I still cannot forget when 5 years ago AlphaGo (Another AI developed by Deepmind but playing Go) demolished me and Lee Sedol (we are both top Go players). Now it seems like AlphaCode is invading CP. I sincerely hope that tourist can fight back for mankind. Best wishes.

    • »
      »
      »
      4 months ago, # ^ |
        Vote: I like it +62 Vote: I do not like it

      So you are real Jie_Bao?

    • »
      »
      »
      4 months ago, # ^ |
        Vote: I like it +26 Vote: I do not like it

      In case he loses,we also have MiracleFaFa

  • »
    »
    4 months ago, # ^ |
      Vote: I like it +41 Vote: I do not like it

    Time to private all our github repos

    • »
      »
      »
      4 months ago, # ^ |
        Vote: I like it +10 Vote: I do not like it

      As if that will stop Github from selling that.

    • »
      »
      »
      4 months ago, # ^ |
        Vote: I like it 0 Vote: I do not like it

      We should be able to make our submissions to CF private too.

  • »
    »
    4 months ago, # ^ |
    Rev. 2   Vote: I like it -36 Vote: I do not like it

    Low** ka AI

  • »
    »
    4 months ago, # ^ |
      Vote: I like it +10 Vote: I do not like it

    I will post the same comment when i become 'grandmaster'.

  • »
    »
    4 months ago, # ^ |
      Vote: I like it +55 Vote: I do not like it

    Curious mind wants to know
    The AI ( WaggleCollide ) which does not know the language it is writing the code, how can it beat tourist? tourist

»
4 months ago, # |
Rev. 2   Vote: I like it -47 Vote: I do not like it

We will be witnessing more cheaters.

  • »
    »
    4 months ago, # ^ |
      Vote: I like it +272 Vote: I do not like it

    From the codeforces rules:

    Solutions and test generators can only use source code completely written by you, with the following two exceptions:

    1. the code was written and published/distributed before the start of the round,
    2. the code is generated using tools that were written and published/distributed before the start of the round.

    If this AI was publically available, using it would not be cheating by point 2.

    • »
      »
      »
      4 months ago, # ^ |
        Vote: I like it +22 Vote: I do not like it

      Maybe then it would be time to introduce more constraints to this rule, i.e "don't use AI to cheat"

      That being said, no idea how it can be enforced, maybe by requesting google to create an AI to tell the likelihood of a code being generated by AI.

      • »
        »
        »
        »
        4 months ago, # ^ |
          Vote: I like it +18 Vote: I do not like it

        For me seeing AI progress would be really interesting. And while in chess you can bruteforce all combination of moves for some depth, I can't see how it could be possible to bruteforce tasks harder than div.2 D. So if tourist would decide to code an AI it most likely would NOT overtake him in rating race, hence no point im cheating.

        Also, returning to chess... Lichess (platform to play chess online, that also has rating), just marks bot accounts as bots. So better alternative than banning AIs imo is allowing bots to participate in rated comtests, but as "half-official" participants. So their results will not be used to calculate rating delta of other participants, but they would still get (or lose) rating based on projection of their results on real results.

        I think letting bots to participate in contests would allow Codeforces and competitive programming overall become more popular.

        • »
          »
          »
          »
          »
          4 months ago, # ^ |
            Vote: I like it 0 Vote: I do not like it

          I meant participants using AI for their contests,

          Also same in chess, lichess and chess dot com do ban users who cheat via ai

          • »
            »
            »
            »
            »
            »
            4 months ago, # ^ |
              Vote: I like it +3 Vote: I do not like it

            You can programm a chess bot knowing only how do figures move and what check and mate are. To code AI capable of solving CP tasks of high-level you need to be on even-more-higher-level. Did chess players become extinct after AI beated Kasparov? No. In fact it became much more popular.

            Also having an AI capable of solving tasks, we become closer to an AI that is capable of creating tasks. Imagine new CP format? Like 1v1 duels. Or rated contests every day. Sounds cool.

            But anyway — having 300-more rated friend can lead to cheating just using his solutions. Also there were a bunch of posts about indian Telegram chats, where participants leaked solutions to each other. So it does not really give you new horizons to cheat. But new horizons to CP? Hell yeah

            • »
              »
              »
              »
              »
              »
              »
              4 months ago, # ^ |
                Vote: I like it -17 Vote: I do not like it

              I don't understand what are you even on about, I just said there would be need in future to say "don't use ai to cheat" I never mentioned restricting the AIs themselves. Can you read? Or do you just write huge unrelated paragraphs, or are you replying to me by accident

        • »
          »
          »
          »
          »
          4 months ago, # ^ |
            Vote: I like it 0 Vote: I do not like it

          Allowing bots to be in codeforces at all sounds like a bad idea to me, cuz very soon people will start blogs complaining people using it in official competitions — though i guess you can check for copying code?

    • »
      »
      »
      4 months ago, # ^ |
      Rev. 2   Vote: I like it +3 Vote: I do not like it

      Probably they will train some model which will classify "deep-fake" from real one).

      • »
        »
        »
        »
        4 months ago, # ^ |
          Vote: I like it 0 Vote: I do not like it

        AI break rules in codeforces

    • »
      »
      »
      4 months ago, # ^ |
        Vote: I like it +391 Vote: I do not like it
      All problem statements will use this font
»
4 months ago, # |
  Vote: I like it +111 Vote: I do not like it

It might sound fun and games now, but if, in reality, AI comes to dominate programming, the developer world could be hit very hard. Thousands of people work as developers around the world; with many more joining the wrokforce every year. This population could be very badly affected by such improvements.

Sorry if this comment comes off as slightly less humorous. Maybe I am a grey programmer, that's why I'm so scared :(

  • »
    »
    4 months ago, # ^ |
      Vote: I like it -23 Vote: I do not like it

    Today you are solving problems,in future you will develop AI to solve problems.Nothing replaces the programmers.

    • »
      »
      »
      4 months ago, # ^ |
        Vote: I like it +81 Vote: I do not like it

      Unless there is AI capable of developing AI to solve problems.

    • »
      »
      »
      4 months ago, # ^ |
        Vote: I like it +8 Vote: I do not like it

      But AI can program more AI later. It's like self-reproduction and need of humans will keep reducing.

      • »
        »
        »
        »
        4 months ago, # ^ |
          Vote: I like it 0 Vote: I do not like it

        Human's will do more productive work then

        • »
          »
          »
          »
          »
          4 months ago, # ^ |
            Vote: I like it 0 Vote: I do not like it

          And what faction of humans you think do very productive work in the industry?

      • »
        »
        »
        »
        4 months ago, # ^ |
          Vote: I like it +4 Vote: I do not like it

        I always thought that the key ingredient of true machine intelligence is self observation. That is it has to be able to recursively track it’s thought process and control its direction instead of mindlessly going through the whole search space.

        • »
          »
          »
          »
          »
          4 months ago, # ^ |
            Vote: I like it +9 Vote: I do not like it

          Well... Unfortunately ML currently is based on trying to find global minimum of function.

          Imagine yourself in a mountains, trying to find the lowest part of the ground. You may think that you found one, but it could be just the lowest part you could see from a peek you previously where. Do you think that observating your previous progress woild lead you to the best place? No, it would not. So either check every [opposite to peak] place, or pray on a good decisions when you start (which are based on random)

          Same thing in ML: Unfortunately good parameters, that are passed to an AI before learning, can not be determined before learinng. There are some ways to generate better datasets, but it is as far as you can get. You can't learn AI to do smth (unless you know algorithm of how to do it by yourself), you can only minimize difference between some mythic optimal algorithm, and algorithm that your AI uses

          • »
            »
            »
            »
            »
            »
            4 months ago, # ^ |
              Vote: I like it +12 Vote: I do not like it

            I guess the problem with the current AI is more fundamental. The path we took to just optimising some function using back propagation is fundamentally broken IMHO.

            We don’t do back propagation in our brains. We learn things instantly, we don’t do silly classification mistakes, we don’t have catastrophic forgetting and so on. So many clues screaming at us that we’re going the wrong way.

            I think that the industry gold standard is taken for granted and no one dares to ask how far can you go with the current NN architecture. Either the problem is in funding or people are blind to the problem that the current architecture will definitely not lead to the true AI.

            • »
              »
              »
              »
              »
              »
              »
              4 months ago, # ^ |
                Vote: I like it 0 Vote: I do not like it

              The problem is — no one knows how do we get to classify things for example. I haven't checked, but they say our brain creates and breaks neuron links, almost similar in how they do in ML.

              But one thing for sure — no one thinks that what we have now is gold. Unfortunately, aome things are impossible to implement. We use base-2 notation, while base-e notation is much more efficient. Why? We know a good way to create base-2 systems, and it does the job.

              People are lazy, they will only progress until their time is going to pay-off. Would you invest a 1m$ to get 1$ every day? No, it takes ~3000 years to get it back. And there is no guarantee in gwtting interest from investing a lot of time into searching a new ways how to do it, plus everything ahows it is most likely... impossible. So we can only hope to invent a little optimizations atm

              • »
                »
                »
                »
                »
                »
                »
                »
                4 months ago, # ^ |
                  Vote: I like it 0 Vote: I do not like it

                The problem is — no one knows how do we get to classify things for example

                Sure we do know now.
                The most prominent company that IMHO took the right approach is numenta.

                I will not go into detail, but the general birds eye view.

                The neurons in the neocortex are organized into modules called cortical columns. Each column is made of 6 layers. You can think of that column as a self contained entity, like a small brain. We can stitch these modules (columns) together in a modular fashion. So, you could add new module in the brain and it will take it's role and have some new functionality.

                Each cortical column performs recognition of the sensory sequence input and based on that input it casts a vote for nearby cortical columns and sends motor commands.

                Well, that was too much info I guess.. sorry if I overloaded you with information =)

            • »
              »
              »
              »
              »
              »
              »
              4 months ago, # ^ |
                Vote: I like it 0 Vote: I do not like it

              Really? I thought the neural networks were based on actual structure of neurons in the brain? And aren't the giant models relatively decent at 1 shot tasks after pretraining (which is similar to what humans do)?

              • »
                »
                »
                »
                »
                »
                »
                »
                4 months ago, # ^ |
                  Vote: I like it +4 Vote: I do not like it

                Artificial Neural Networks is a class of algorithms "loosely modeled" on neurons in a biological brain. Backpropagation is even more removed from how the actual brain works.

            • »
              »
              »
              »
              »
              »
              »
              4 months ago, # ^ |
                Vote: I like it 0 Vote: I do not like it

              Same feelings here, those ML trends leading AI development nowadays seems to me misleading, sometimes I think it is on purpose...... :p

        • »
          »
          »
          »
          »
          4 months ago, # ^ |
            Vote: I like it 0 Vote: I do not like it

          It's called renforcement learning.

  • »
    »
    4 months ago, # ^ |
      Vote: I like it -99 Vote: I do not like it

    Yes, let's stop this madness before it get's out of control. The programmers who created this think that they are irreplaceable, but they would be the first ones to get replaced. Let this comment gather downvotes, but there is a limit to what humans should automate. If the world continues like this, we are moving to a sad, sad reality — one far worse than covid. This AI should be banned from participating. I don't want to be in any way helpful in it's development! And other programmers consent must be valued too...

    • »
      »
      »
      4 months ago, # ^ |
        Vote: I like it +60 Vote: I do not like it

      Yes, let's stop this madness before it get's out of control. The programmers who created this think that they are irreplaceable, but they would be the first ones to get replaced. Let this comment gather downvotes, but there is a limit to what humans should automate. If the world continues like this, we are moving to a sad, sad reality — one far worse than covid. This AI should be banned from participating. I don't want to be in any way helpful in it's development! And other programmers consent must be valued too...

      • »
        »
        »
        »
        4 months ago, # ^ |
          Vote: I like it +14 Vote: I do not like it

        Yes, let's stop this madness before it get's out of control. The programmers who created this think that they are irreplaceable, but they would be the first ones to get replaced. Let this comment gather downvotes, but there is a limit to what humans should automate. If the world continues like this, we are moving to a sad, sad reality — one far worse than covid. This AI should be banned from participating. I don't want to be in any way helpful in it's development! And other programmers consent must be valued too...

        • »
          »
          »
          »
          »
          4 months ago, # ^ |
            Vote: I like it +18 Vote: I do not like it

          Yes, let's stop this madness before it get's out of control. The programmers who created this think that they are irreplaceable, but they would be the first ones to get replaced. Let this comment gather downvotes, but there is a limit to what humans should automate. If the world continues like this, we are moving to a sad, sad reality — one far worse than covid. This AI should be banned from participating. I don't want to be in any way helpful in it's development! And other programmers consent must be valued too...

          • »
            »
            »
            »
            »
            »
            4 months ago, # ^ |
              Vote: I like it -19 Vote: I do not like it

            Yes, let's stop this madness before it get's out of control. The programmers who created this think that they are irreplaceable, but they would be the first ones to get replaced. Let this comment gather downvotes, but there is a limit to what humans should automate. If the world continues like this, we are moving to a sad, sad reality — one far worse than covid. This AI should be banned from participating. I don't want to be in any way helpful in it's development! And other programmers consent must be valued too...

            • »
              »
              »
              »
              »
              »
              »
              4 months ago, # ^ |
                Vote: I like it 0 Vote: I do not like it

              Yes, let's stop this madness before it get's out of control. The programmers who created this think that they are irreplaceable, but they would be the first ones to get replaced. Let this comment gather downvotes, but there is a limit to what humans should automate. If the world continues like this, we are moving to a sad, sad reality — one far worse than covid. This AI should be banned from participating. I don't want to be in any way helpful in it's development! And other programmers consent must be valued too...

      • »
        »
        »
        »
        4 months ago, # ^ |
          Vote: I like it 0 Vote: I do not like it

        you and me just scare to be replaced lul

    • »
      »
      »
      4 months ago, # ^ |
        Vote: I like it 0 Vote: I do not like it

      Come on, does AI solving math problems replaced mathematicians? Did site constructors replace programmers? Or, maybe, chess AI fully replaced, chess players? Or do you, maybe, think, that AI will be able to beat tourist in codeforces competitions? Of course not, at least not in our century. AI is not even near being that smart to deal with unusual situations which constitutes the work of a programmer. AI can work with given actions only, it does not really "learn" something, it just simulates the mind without really knowing what is it doing.

      • »
        »
        »
        »
        4 months ago, # ^ |
          Vote: I like it +17 Vote: I do not like it

        I think you're underestimating the rate at which a breakthroughs can happen. Artificial General Intelligence could be developed in a very short time frame by having a "seed AI" that writes new AIs that write new AIs, etc. The only thing we need is a seed AI that has an "r-value" greater than 1.

        I'm not saying that this is what will definitely happen, but I think there is a decent chance that AI smarter than humans in every domain of interest gets developed in the next 50 years. It definitely should not be shrugged aside.

  • »
    »
    4 months ago, # ^ |
    Rev. 2   Vote: I like it +18 Vote: I do not like it

    Has calculators replaced mathematicians? No.

    At some degree, automated systems such as AlphaCode will be extremely useful to automate some mundane task (for example, creating user interfaces from UX diagrams) but some other areas, much more deep creativity is needed and indeed, the day that an AI system is capable of fully replacing research engineers we will have accomplished the so-called "AGI" or "Artificial General Intelligence" in such case, we all going to be living in a renaissance :)

    • »
      »
      »
      4 months ago, # ^ |
        Vote: I like it +5 Vote: I do not like it

      Lol you're not comparing calculators to AI's... they're completely different concepts, it's like comparing deepblue with say alphazero or sth.

    • »
      »
      »
      4 months ago, # ^ |
        Vote: I like it 0 Vote: I do not like it

      This is market place thinking: automation is equivalent to AI. This is why AI is popular, lots of automation is now counted in AI catalogue

  • »
    »
    4 months ago, # ^ |
      Vote: I like it +2 Vote: I do not like it

    Not saying it won't happen, but I think we are still a long way from AI dominating programming. In the real world most programming problems are far less well specified than in competitive programming, and much of programming is more about working out what the customer actually wants; rather than about implementing clever algorithms.

    Having said this, AlphaCode and similar AIs could quite quickly become very useful tools for those areas of programming that are dominated by algorithm design, and could require some rethinking of what skills are important for developers.

    Remember also that there have been many developments over the years that have made developers more productive (and have changed the skills developers need). These include new languages and libraries, much faster build and test times, and easy access to open source code through the web. In spite of this the demand for developers has increased, almost continuously, for at least the last 50 years.

  • »
    »
    4 months ago, # ^ |
      Vote: I like it +25 Vote: I do not like it

    Yet when blue collar workers are hit by the same phenomenon? Crickets chirping.

»
4 months ago, # |
  Vote: I like it +21 Vote: I do not like it

Tourist v/s AI .............would love to witness it!!!

  • »
    »
    4 months ago, # ^ |
      Vote: I like it +15 Vote: I do not like it

    What about AI versus the whole developer community? The stakes are high for one party; if the developers lose, they'll be replaced for ever. Are you interested to witness this match?

  • »
    »
    4 months ago, # ^ |
      Vote: I like it 0 Vote: I do not like it

    I would hate to witness it... This match happening would be the darkest day in the history of programming.

    Even if our beloved tourist wins, people like you and me still don't have a place in the world.

    Agree with this, no matter how many downvotes it gets.

»
4 months ago, # |
  Vote: I like it -16 Vote: I do not like it

what's the cf handle of the AI programmer ??!!, i would love to read its codes and would also try to compete with it in the contests ....would be fun :) Which language does the AI bot uses? Btw , It would be quite hilariuos to see the AI bot's code to pass the pretets and fail in system tests....

»
4 months ago, # |
  Vote: I like it +5 Vote: I do not like it

I feel I am not as good as an AI :((

»
4 months ago, # |
Rev. 3   Vote: I like it +67 Vote: I do not like it

F for grays

»
4 months ago, # |
  Vote: I like it +3 Vote: I do not like it

They should make a documentary like Alpha Zero, where tourist will be invited for solving problems against this AI.

»
4 months ago, # |
  Vote: I like it -54 Vote: I do not like it

Has my data been used for training this AI in any way and will Codeforces users data be used without their consent for training this AI?

  • »
    »
    4 months ago, # ^ |
      Vote: I like it +2 Vote: I do not like it

    Probably not. They might use it if you get better instead of making pointless comments and crying about AI replacing you though.

    About the second thing, probably yes because you're allowing anyone to see and access your code by submitting on this platform, it's their right to use it.

  • »
    »
    4 months ago, # ^ |
      Vote: I like it +9 Vote: I do not like it

    You have already given your consent. Please read https://codeforces.com/terms

    In these Website Terms And Conditions, “Your Content” shall mean any audio, video, text, images or other material you choose to display on this Website. With respect to Your Content, by displaying it, you grant Codeforces a non-exclusive, worldwide, irrevocable, royalty-free, sublicensable license to use, reproduce, adapt, publish, translate and distribute it in any and all media. Your Content must be your own and must not be infringing on any third party’s rights. Codeforces reserves the right to remove any of Your Content from this Website at any time, and for any reason, without notice.

  • »
    »
    3 months ago, # ^ |
      Vote: I like it 0 Vote: I do not like it

    The AI wouldn't take your data at least....... unless it wants to get worse at problem solving so you need not worry of your data being used for training :)

»
4 months ago, # |
  Vote: I like it +273 Vote: I do not like it

This is amazing, but I still seriously doubt we will see an AI solve problems much harder than this consistently any time soon (in the next ~20 years at least). I wouldn't be surprised if it manages to solve some hard problems, as sometimes the difficulty comes from just having to implement a lot of stuff. It will be interesting to see what kind of problems the bot finds easy. Personally I hope it is very good at geometry problems, so we can get rid of those 8)

  • »
    »
    4 months ago, # ^ |
      Vote: I like it +47 Vote: I do not like it

    My impression is those can't even coord/trig/complex bash math contest geometries yet?

    My guess is the best chance might be `template' data structure problems, e.g. things doable with templates like lazy propagation / LCT / suffix arrays. I'm really not a fan of the idea of 'easy data structure problems'.

    • »
      »
      »
      4 months ago, # ^ |
        Vote: I like it +31 Vote: I do not like it

      Lazy propagation might be possible, but for the others on your list, the lack of training data likely makes it impossible. There are probably less than 1000 solutions on CF that use LCT. Further, most of those use very different templates, and the AI has no way of knowing they all are the same data structure.

      But yeah, I agree that the bot ridding us of geometry is likely a pipe dream ;_;

    • »
      »
      »
      4 months ago, # ^ |
      Rev. 3   Vote: I like it +10 Vote: I do not like it

      I feel parsing math contest geometry problems is a much bigger problem than actually solving them.

      When I was doing Olympiad math and was much worse off at synthetic geometry than at bashing geometry problems, I had a very mechanical approach to bashing, which I am pretty sure can be formalized into a system that a theorem prover can easily work through, given enough time. Even if it involves proving certain non-trivial inequalities, I am sure that the computer-aided inequality proving methods that are (or at least used to be) so popular among math olympiad students in China and Vietnam would be really helpful.

      Such approaches sound hard for purely synthetic solutions though, since a lot of the times we need constructions, and constructions are super hard to come up with. But still, I think if AI can solve competitive programming problems, it is not too far-fetched to think that it stands a chance against a usual IMO G8, since in my experience (of getting at least as good at synthetic geometry as a usual "strong" math olympiad contestant in 2016), it all just boils down to learning how to apply the same 500-1000 facts in Olympiad geometry in a well-defined systematic manner, of which you probably need only 100 in most cases.

      • »
        »
        »
        »
        4 months ago, # ^ |
          Vote: I like it 0 Vote: I do not like it

        Actually it depends on the way you attack such a problem. If directly one picks up the hammer of bashing and treats the problem as a nail, it won't work because the problem is more than that. Approaching a (MO, CP is very very different) geometry problem mainly involves 2 things:

        1) Realizing any well known configuration or setup.

        2) Simplifying that configuration with low-tech results or using hi-tech results to kill it.

        As for the construction stuff, they are quite nontrivial to come up with and I don't think a bot would be able to currently deal with explicitly coming up with a tricky construction. And NO NO NO. It is very far fetched to think that a bot which can do CP can solve IMOSL G8s. Please think about what you just said.

        • »
          »
          »
          »
          »
          4 months ago, # ^ |
          Rev. 3   Vote: I like it +33 Vote: I do not like it
          1. As far as bashing is concerned, I am pretty sure most problems succumb to bashing. At least when I was not good at geometry, I recall having successfully bashed all of the close to 1000-2000 geometry problems I tried in a very mechanical and mindless manner (using at least one of complex numbers, trigonometry or one of barycentric/trilinear, Cartesian, and tripolar coordinates). These problems also include pretty much every high numbered IMO shortlist geometry problem, and problems from TSTs of countries like USA, China, Vietnam and Russia (which are traditionally strong in geometry and have much harder geometry problems than others). So I feel that your point about bashing is kind of baseless. Of course, I am not good enough/don't have enough time to actually write up a solver based on my experience, but I am familiar enough with some SOTA literature to claim that this is definitely not beyond the reach of such techniques. It is literally just symbolic manipulation + equation solving in most cases, and if you have equations you can't solve algebraically, you would need to combine them in a certain manner, that I am sure is possible, given the recent OpenAI result (or the inequality solvers that I alluded to in the previous comment).

          2. I had already mentioned earlier that constructions are super non-trivial, so in general it is not possible to systematically list out all possible constructions, but for most constructions that a human competing at, say, the IMO bronze/silver level can think of, I don't think AI can't be systematically trained by adding in a ton of inductive bias (Of course, it is theoretically possible to go about looking through constructions by enumerating them in lexicographical order, but it's too impractical, so I sort of agree with your point).

          3. If a person, who is as bad at thinking as me, can train enough to be able to solve, at their peak, pretty much any IMOSL geometry problem with a high probability in less than an hour or so, or get to a couple of completely new triangle centers interesting enough for the centers to warrant a mention in ETC back when it had much fewer centers than it has now, I am pretty sure that the learning process is super mechanical. In fact, verification of triangle centers is a very well-known problem and has been dealt with at length in ETC. Knowing the fact that there are AIs that can actually recognize SOS expressions in inequalities (for instance, refer to this blog by OpenAI), I really don't think that given a large enough knowledge base, it can't figure out how to solve MO geo problems. This takes care of your argument about "realizing any well-known configuration or setup" and "simplifying that configuration with low-tech results or using hi-tech results to kill it". If you put in the effort to expand their database of statements to include something like the whole Kimberling Encyclopedia of Triangle Centers and/or Bernard Gibert's list of higher order curves, do you really expect a heuristic search to never be able to find out a solution in a large enough number of geometry problems?

          Edit: Just wanted to add the fact that this thesis exists. Apparently, some problems at the IMO shortlist level were also computer generated, and some of them got accepted into national Olympiads too.

          Edit 2: I also never claimed that a bot that can solve competitive programming problems (specifically referring to the DeepMind AI) can also solve geometry problem. I just said that it is possible to get computers to solve MO geometry problems for you.

»
4 months ago, # |
  Vote: I like it +11 Vote: I do not like it

I Don't think Artificial intelligence is enough for CP.

  • »
    »
    4 months ago, # ^ |
      Vote: I like it 0 Vote: I do not like it

    That's what they thought before Alphazero and Stockfish.

    • »
      »
      »
      4 months ago, # ^ |
        Vote: I like it 0 Vote: I do not like it

      Chess has very well defined rules, CP problems on the other hand are usually unique.

»
4 months ago, # |
  Vote: I like it +18 Vote: I do not like it

A wave of pupils can be expected now.

»
4 months ago, # |
  Vote: I like it -10 Vote: I do not like it

It is very interesting to see how this new AI would solve problems, that require some new ideas. Well, I would definitely be upset if AI would become better than tourist, but I am afraid it is inevitably.

»
4 months ago, # |
  Vote: I like it -19 Vote: I do not like it

Wow this is amazing

»
4 months ago, # |
Rev. 2   Vote: I like it +116 Vote: I do not like it

I've read some solutions of these three users, and some of them don't look like they've been written by a bot. They seem to use "human" code templates, defines and stuff like that.

Maybe the explanation is that the network was trained on solutions from editorials, and sometimes copies these templates just because they were written in the training data. Though it seems strange that the AI sometimes doesn't understand the code it is writing and blindly copies human templates (I would imagine that the network recognizes the problem and builds a solution which is automatically converted to code, but maybe the architecture is different).

  • »
    »
    4 months ago, # ^ |
      Vote: I like it +46 Vote: I do not like it

    It's also quite interesting how the AI copes with informal statements — legends, stories and stuff like that. I would imagine processing them is much harder.

    • »
      »
      »
      4 months ago, # ^ |
        Vote: I like it +16 Vote: I do not like it

      My guess is it puts a lot of value on input/output specifications and sample I/O.

  • »
    »
    4 months ago, # ^ |
      Vote: I like it 0 Vote: I do not like it

    I was actually reading through them and was wondering the same thing. How these code seem like they are written by a human and not an AI because of structure and templates

    • »
      »
      »
      4 months ago, # ^ |
        Vote: I like it +5 Vote: I do not like it

      Well, one of the pictures in the article shows that solutions are also a part of the training dataset (and I hadn't noticed it before I was writing my first comment), so it explains the "human" templates.

  • »
    »
    4 months ago, # ^ |
      Vote: I like it 0 Vote: I do not like it

    We also show through detailed analysis that there is no evidence that AlphaCode copies important parts of previous solutions or exploits weaknesses in the problem structure.

    Excerpt from the paper.

    • »
      »
      »
      4 months ago, # ^ |
        Vote: I like it 0 Vote: I do not like it

      Well now there isn't no evidence, there's a bit of circumstantial evidence.

  • »
    »
    4 months ago, # ^ |
      Vote: I like it 0 Vote: I do not like it

    May be!

»
4 months ago, # |
  Vote: I like it -50 Vote: I do not like it

The machine will not outshine its maker

  • »
    »
    4 months ago, # ^ |
    Rev. 2   Vote: I like it +5 Vote: I do not like it

    "The machine will outshine its maker, It always had and Always will" ~ Skynet

»
4 months ago, # |
  Vote: I like it +70 Vote: I do not like it

Tourist Vs AI! Already feeling bad for AI

  • »
    »
    4 months ago, # ^ |
      Vote: I like it +54 Vote: I do not like it

    They said the same thing for alphaGo, funny how that turned out XD

  • »
    »
    4 months ago, # ^ |
      Vote: I like it +8 Vote: I do not like it

    I remember people cheering for Lee Sedol. The consensus was that AlphaGo was going to fail miserably.

»
4 months ago, # |
  Vote: I like it +2 Vote: I do not like it

Now, we have to quickly solve first two problems :)

»
4 months ago, # |
  Vote: I like it +89 Vote: I do not like it

It's very amusing to look at the AI's submissions to 2000+ rated problems:

G – Unusual Minesweeper: Not really sure what to make of this? I have zero clue what it's doing or how what it's calculating helps solve the problem. It doesn't even use $$$timer$$$ in it's solution! It looks like the AI wrote a random something to pass the first sample.

F – Quadratic Set: The AI hardcodes small cases and does something random for everything else.

H – Permutation and Queries: The AI understood the problem correctly and translated it directly to code which gets TLE. This one is interesting because I notice the AI does particularly well when the problem gives a clear procedure (i.e. it smashes problems like A – Life of a Flower out of the park), but I wonder how well the AI performs on older beta rounds or ICPC problems with more legends and maybe not the most direct way of stating the problem? Also, why did the AI fail to optimize the solution in this case? There are instances of it successfully implementing something beyond the naive solution. Is it a matter of the dataset not having enough AC submissions to harder problems?

  • »
    »
    4 months ago, # ^ |
      Vote: I like it +80 Vote: I do not like it

    the AI does particularly well when the problem gives a clear procedure

    Maybe they should test on Atcoder as well.

  • »
    »
    4 months ago, # ^ |
      Vote: I like it +85 Vote: I do not like it

    So the AI is me solving usaco problems...

  • »
    »
    4 months ago, # ^ |
    Rev. 2   Vote: I like it 0 Vote: I do not like it

    Given the code shown for Backspace, I wondered if they were counting correct-but-TLE as passing. Then I remembered that said problem had "Preparation: Will not be revealed for now because we care for his/their safety." in its editorial. Sooo either the problem sent to the AI had its original (weak) testcases, or... an AI getting from statements to solutions is still a great accomplishment, but the rating is probably a little inflated.

    I will throw out that python as it's seen in cp is a very niche dialect (pypy but no numpy) so the set of python coders who would persist/twist through it I reckon is pretty slim, which means the modal python solution for the AI to pull from is probably 'bad' in a lot of ways (mine included, shudder/clown/cheese/cry)... maybe?

    There was also this: https://codeforces.com/blog/entry/94648 a few months ago... didn't get around to trying it though.

    • »
      »
      »
      4 months ago, # ^ |
        Vote: I like it +3 Vote: I do not like it

      Hmm that Backspace code does look like $$$\mathcal O(n^2)$$$. If they're counting any solution that's correct regardless of TLE, then these results are way less significant than I thought, because kind of the whole premise of CP is solving problems with good complexity and not just brute forcing through everything.

      For the "Python to C++" translator, are you perhaps referring to this? I remember seeing this a few months ago and thinking that problems based on "naive -> AC" optimized with data structures are most likely the types of problems AI would be best with since there's a consistent "formula" to them.

      • »
        »
        »
        »
        4 months ago, # ^ |
        Rev. 2   Vote: I like it 0 Vote: I do not like it

        Yup (captain america pointing 'I get that reference!' meme).

      • »
        »
        »
        »
        4 months ago, # ^ |
          Vote: I like it +1 Vote: I do not like it

        I can't find the bot's submission but benq resubmitted the code and it doesn't get a TLE: 144971343

        Somehow pop(0) for a list is just really fast in cpython?

        • »
          »
          »
          »
          »
          4 months ago, # ^ |
            Vote: I like it +3 Vote: I do not like it

          That's interesting. I've confirmed that list.pop(0) does really memmove() the entire list.

          So... likely that's caused by memmove() being too fast.

          • »
            »
            »
            »
            »
            »
            4 months ago, # ^ |
              Vote: I like it 0 Vote: I do not like it

            these builtin lists are fast in every language, I remember 100k random insert/delete on length 100k arraylist in java took only 400 ms, and that was back in 2016

            Also there was one problem which we were told needed some memset to be fast (it was actually n^2 solution lol) and so I couldn't solve it in java since looping and setting to 0 was too slow (eventually I did solve it the right way)

  • »
    »
    4 months ago, # ^ |
      Vote: I like it +32 Vote: I do not like it

    As mentioned in the preprint paper "4.5. Filtering", this AI produces lots of solutions and then filters them by running sample input on them, which make 99% of solutions filted.

    So, when the problem is too hard for AI to even produce a reasonable solution for sample input, only the weird solutions (like output the sample output directly or output something happens to meet the sample output) can pass the filtering stage, which explain what we observed for problem G and F.

  • »
    »
    4 months ago, # ^ |
      Vote: I like it +5 Vote: I do not like it

    They also have a nice UI for visualizing where the "attention heads" are focusing on in each problem: https://alphacode.deepmind.com/

»
4 months ago, # |
  Vote: I like it -31 Vote: I do not like it

I am fake tourist and I confirm AI is noob

»
4 months ago, # |
  Vote: I like it 0 Vote: I do not like it

Wow! Is it a good News or bad News? Chess players still have a career, hoping the same!!

»
4 months ago, # |
  Vote: I like it -8 Vote: I do not like it

soon we humans will get transferred to the matrix.

»
4 months ago, # |
  Vote: I like it -9 Vote: I do not like it

AI is noob in front of tourist.

»
4 months ago, # |
  Vote: I like it 0 Vote: I do not like it

It seems that we are going to be defeated by an AI codeforces in a near future.

»
4 months ago, # |
  Vote: I like it -8 Vote: I do not like it

GOOD

»
4 months ago, # |
  Vote: I like it -29 Vote: I do not like it

One question, why?

  • »
    »
    4 months ago, # ^ |
      Vote: I like it +5 Vote: I do not like it

    why not?

    • »
      »
      »
      4 months ago, # ^ |
        Vote: I like it -10 Vote: I do not like it

      Don't you guys ever see movies , Don't you get it competitive porgramming is the highest in terms of creative thinking in programming world. If they managed to pull that, then how would you think the time it will take for them to replace software engineer jobs by some AI?

»
4 months ago, # |
Rev. 3   Vote: I like it -8 Vote: I do not like it

AI handle on codeforces

»
4 months ago, # |
  Vote: I like it +62 Vote: I do not like it

If AI gets good enough (level 2000+) in problem solving, it would destroy all opportunities for average coders.

Only very high rated coders will be needed to monitor/maintain the AI systems and almost everyone below 2000 will face negligence because AI will do what they can. However such future is yet to happen but it poses such potential.

»
4 months ago, # |
  Vote: I like it -7 Vote: I do not like it

just fantastic!

»
4 months ago, # |
  Vote: I like it -16 Vote: I do not like it

I think a tourist is a computer, so what is the confrontation of a computer against a computer

»
4 months ago, # |
  Vote: I like it +1 Vote: I do not like it

It sounds fun but I don't think AI for CP will get huge results like deep blue.The premise of solving a problem is to understand the problem, apparently AI still has a hard time doing this.

»
4 months ago, # |
  Vote: I like it 0 Vote: I do not like it

Wow, I am still better than AI :)

»
4 months ago, # |
  Vote: I like it 0 Vote: I do not like it

Malades aka bor gapni aytibsiz! MDSPro

»
4 months ago, # |
  Vote: I like it -19 Vote: I do not like it

Don't compare this shit with Tourist( God )

»
4 months ago, # |
  Vote: I like it +3 Vote: I do not like it

AI will make problems -> AI will solve problems.

»
4 months ago, # |
  Vote: I like it +127 Vote: I do not like it

Now we know that all these notoriously bad problem statements were merely a preparation to trick AI.

»
4 months ago, # |
  Vote: I like it 0 Vote: I do not like it

Now tourist will write code for his own AI XD

»
4 months ago, # |
  Vote: I like it +3 Vote: I do not like it

many people are telling noob to these AI, but can't you see the speed in which they are solving problems?

»
4 months ago, # |
  Vote: I like it 0 Vote: I do not like it

Ah damn, they beat me

»
4 months ago, # |
  Vote: I like it +11 Vote: I do not like it

I just thought about this the other day. Emergence of such systems is inevitable, but also sad. I, for one, really believe in potential of AI such as this and Copilot. It means that in the near future millions of talented people will be forced to drag out a miserable existence or to commit suicide. I think that automating creative, human jobs while we still live in capitalist society is just unethical. Uncle Ted had a point.

»
4 months ago, # |
  Vote: I like it -12 Vote: I do not like it

Yowza! I think it will be an interesting match.AI if you can read my comment i wish you best ( but actually you can't win )

Meme
»
4 months ago, # |
  Vote: I like it 0 Vote: I do not like it

It's a very exited waiting to see a confrontation between tourist and AI. Then if tourist lose, then we all no need to have programming :)

»
4 months ago, # |
  Vote: I like it 0 Vote: I do not like it

Don't worry codeforcer's. If AI has learned every kind of algorithm, I think that it wouldn't be a master because coding is not like chess everybody or everything can understand.

»
4 months ago, # |
  Vote: I like it 0 Vote: I do not like it

I think that it will be extremely hard to write an AI to solve hard problems right now, because intuition itself is difficult to program, and the number of ways to write a problem that has one given solution is extremely large. However, I am looking towards the technology of the future and the problems the AI will end up solving.

»
4 months ago, # |
  Vote: I like it 0 Vote: I do not like it

Finally, a worthy opponent.. Our battle will be legendary!

»
4 months ago, # |
  Vote: I like it +19 Vote: I do not like it

Ai idea will be:

01010111 01110010 01101111 01101110 01100111 00100000 01000001 01101110 01110011 01110111 01100101 01110010

»
4 months ago, # |
  Vote: I like it +233 Vote: I do not like it

Now let's see its rank on AGCs

  • »
    »
    4 months ago, # ^ |
      Vote: I like it +13 Vote: I do not like it

    Dumber idea: give it Nowruz (from IOI17)...

»
4 months ago, # |
  Vote: I like it +3 Vote: I do not like it

Scared to face humans. Now have to face AI. :smiling_face_with_tear:

»
4 months ago, # |
  Vote: I like it 0 Vote: I do not like it

AI will create AI itself! Then they will write the problem statement also! What else will be left for us?

»
4 months ago, # |
  Vote: I like it +5 Vote: I do not like it

Let's not get the hopes too high . it's completely different from chess AI that relies on calculations but solving competitive problems isn't just calculations that's an important part of it only. creativity is required here

»
4 months ago, # |
  Vote: I like it 0 Vote: I do not like it

we trust in tourist

»
4 months ago, # |
Rev. 2   Vote: I like it +23 Vote: I do not like it

Wanna hear thoughts of AlexSkidanov on it since his project Near was previously about AI solving CP problems

»
4 months ago, # |
  Vote: I like it 0 Vote: I do not like it

Our jobs will be in danger in near future. T.T

»
4 months ago, # |
  Vote: I like it 0 Vote: I do not like it

Actually, I am more excited about an AI that can generate nice quality problems.

»
4 months ago, # |
  Vote: I like it +148 Vote: I do not like it

Wow... this is so mind-blowing. I think I really have something to say about this. I started to put effort into competitive programming after graduating from high school. Before that, I worked hard on Go and tried to become a professional Go player. I gave up on this idea quickly after AlphaGo came. The feeling that no matter how you grind your skills throughout your whole life, there's always an opponent you can never beat is so frustrating. I never thought I could meet Google and their AI again in competitive programming. Frankly, I really panicked when I saw this.

Maybe I am becoming pessimistic(or optimistic, depending on how you think), but I think there is much more than the current $$$1300$$$ rating of the AIs. There's more than just implementation in those codes, there's bit operation, binary search, and factors of a number. These are correct applications of some algorithms. Even these are only easy ones, there's no strong barrier preventing those AI from learning how to write codes for harder ones. The $$$1300$$$ rating would be a much underestimated value by then. A reminder for those who saw AI merely as noobs, that when AlphaGo was announced, not a single person who can play Go rather well thought AlphaGo had the slightest chance to beat Lee Sedol.

There are still barriers beyond the application of existing algorithms. AI can't seem to "invent" algorithms for now, but what AI can do now is surprising enough, and being able to "invent" algorithms would lead to a whole new era.

  • »
    »
    4 months ago, # ^ |
      Vote: I like it +26 Vote: I do not like it

    Maybe you should give problem setting a try. Go is a game of complete information: my impression is that OI problem setting can be far more adversarial, even compared to RTS video games with fog of war / cheesing.

  • »
    »
    4 months ago, # ^ |
      Vote: I like it +26 Vote: I do not like it

    Maybe it's time for you to join the other former competitive programmers who now work in AI (including me) :)

  • »
    »
    4 months ago, # ^ |
      Vote: I like it 0 Vote: I do not like it

    I don't think it can create new algorithms,the more creativity needed the less chances for the AI to beat humans but when it comes to calculations it will surpass the human. I listened to AI playing music and it was pathetic just grabbing some pieces from here and here it can't get creative.

»
4 months ago, # |
  Vote: I like it 0 Vote: I do not like it

they better write artificial intelligence for write artificial intelligence

»
4 months ago, # |
  Vote: I like it +118 Vote: I do not like it

I'm waiting for the moment when AI newbie asks AI red on how to become red.

»
4 months ago, # |
  Vote: I like it 0 Vote: I do not like it

WoW!

»
4 months ago, # |
  Vote: I like it 0 Vote: I do not like it

Damn! I'm getting AlphaGo vibes back again!!!. Can't wait to see AlphaCode and tourist 1v1 on a duel

»
4 months ago, # |
  Vote: I like it +1 Vote: I do not like it

Jarvis has taken part in competition...

»
4 months ago, # |
  Vote: I like it +8 Vote: I do not like it

One step closer to Matrix actually being a reality :D

God, this is scary.

»
4 months ago, # |
  Vote: I like it 0 Vote: I do not like it

Imagine If we feed(Train) every Competitive programming(algorithmic problem) Ever existed Till now to our bot Imagine its Capability to recognize and solve algorithmic problems. I would love to see that in near future. Atleast in my lifespan

  • »
    »
    4 months ago, # ^ |
      Vote: I like it 0 Vote: I do not like it

    :)

    It's not really that much of data in the grand scale of things

    But yeah

    Can't wait for the most awaited duel lol

»
4 months ago, # |
  Vote: I like it 0 Vote: I do not like it

GG

»
4 months ago, # |
  Vote: I like it +4 Vote: I do not like it

Waiting for Umnik to troll AI for its Binary Search Implementation XD

»
4 months ago, # |
  Vote: I like it 0 Vote: I do not like it

many people do cp for improving problem solving skill and/or it's fun.So AI solving problem shouldn't be a issue for now. Though it will definitely impect good problem solvers demand(it not now in near future).Overall it feel's both existing and terrifying.

»
4 months ago, # |
  Vote: I like it 0 Vote: I do not like it

SelectorUnlimited , WaggleCollide and AngularNumeric are the AI's profile name. Add them to your friendlist.

  • »
    »
    4 months ago, # ^ |
      Vote: I like it 0 Vote: I do not like it

    AngularNumeric is the better one. But two others should keep practicing too.

»
4 months ago, # |
  Vote: I like it 0 Vote: I do not like it

The tags in the problem descriptions on alphacode.deepmind.com are wrong (different from what's on the Codeforces) for some reason, and different between Python/C++.

Compare "TAGS: binary search,math" https://alphacode.deepmind.com/#layer=18,problem=69,heads=00000000000 and "TAGS: dfs and similar,graphs,trees" https://alphacode.deepmind.com/#layer=18,problem=70,heads=00000000000 for the same problem https://codeforces.com/problemset/problem/1559/B

Also is there an easy way (without manually cleaning garbage spaces) to copy-paste and try solutions from alphacode.deepmind.com?

  • »
    »
    4 months ago, # ^ |
      Vote: I like it +35 Vote: I do not like it

    I think the tags are actually randomly generated, The paper mentioned that when repeatedly sampling from the model, they provided random tags as a way of telling the model to try different approaches, a little like human brainstorming.

»
4 months ago, # |
Rev. 2   Vote: I like it 0 Vote: I do not like it

Who will be replaced first by AI — coders or bus drivers? The answer seems to be no more so obvious. But who cares. Universal basic income for everybody, and AI should work hard instead of humans.

»
4 months ago, # |
  Vote: I like it -7 Vote: I do not like it

There's no way in 10 years you're gonna surpass tourist.

»
4 months ago, # |
  Vote: I like it 0 Vote: I do not like it

Here https://alphacode.deepmind.com/#layer=18,problem=14,heads=00000000000

Functions inputnum, inputnums, inputmatrixint look to be copied exactly from someone's Python template, most are not used.

Does anyone recognize these functions?

»
4 months ago, # |
  Vote: I like it +23 Vote: I do not like it

Looking at the paper it looks like they only used non hidden cases, which means they don't exceed some length and are shown as .... I submitted 7 codes which are marked as correct in alphacode.deepmind.com and only one of them got AC.

144976269 144972743 144972536 144972467 144972379 144971604

To copy code from alphacode.deepmind.com use this in vim ggVG gJ :%s/⤶/\n/<Ret>

  • »
    »
    4 months ago, # ^ |
      Vote: I like it +23 Vote: I do not like it

    That's kinda invalidates the whole claim of the paper...

    • »
      »
      »
      4 months ago, # ^ |
        Vote: I like it +20 Vote: I do not like it

      Quote from paper (section 5.1):

      Evaluating on actual programming competitions checks program correctness more thoroughly, compared to evaluating on our dataset which has known weaknesses including false positives, accepting algorithmically inefficient solutions, and handling problems with multiple acceptable outputs.

      Apparently this is a known defect of their dataset used for offline evaluation. The rating of ~1300 is derived from submitting code directly to Codeforces, so it's not affected.

      The Attention Visualizer website marking solutions evaluated on smol dataset as passing is certainly confusing though.

      • »
        »
        »
        »
        4 months ago, # ^ |
          Vote: I like it 0 Vote: I do not like it

        I think that rating should be treated with a grain of salt, since most probably they get no time penalty and if I understood correctly, they submit up to 10 completely different solutions until one of them is marked as correct.

        • »
          »
          »
          »
          »
          4 months ago, # ^ |
            Vote: I like it 0 Vote: I do not like it

          In fact they have considered time penalty. See Table 4 in page 14 for more details.

      • »
        »
        »
        »
        4 months ago, # ^ |
          Vote: I like it -10 Vote: I do not like it

        They claim to achieve "the level of the median competitor". Median is 1000. 1300 is 25% percentil.

  • »
    »
    4 months ago, # ^ |
      Vote: I like it 0 Vote: I do not like it

    Each test case consists of one line containing two integers $$$n$$$ and $$$k$$$ ($$$1 \le n \le 10^9, 1 \le k \le 2$$$)

    144972467

    for (int i = 0; i < n; i++) {
        if (f[s[i] - '0'] == 0) cnt++;
            f[s[i] - '0']++;
    }
    

    He was hoping for the best =)

  • »
    »
    4 months ago, # ^ |
      Vote: I like it +31 Vote: I do not like it

    Yeah, I think this was a known limitation of the dataset, but they did actually submit to CF, and those results are still quite impressive.

»
4 months ago, # |
Rev. 2   Vote: I like it +8 Vote: I do not like it

From 4 years ago:

AlexSkidanov did you guys give up on the problem?

»
4 months ago, # |
  Vote: I like it +14 Vote: I do not like it

Finally, worthy opponent. Our battle will be legendary ---

»
4 months ago, # |
  Vote: I like it -8 Vote: I do not like it

Watch Alpha code get trained to be 3600 on codeforces like chess.

»
4 months ago, # |
  Vote: I like it -22 Vote: I do not like it

Oh no.... OH NONONONONONONO... PLEASE NO!

»
4 months ago, # |
  Vote: I like it +1 Vote: I do not like it

In their initial release, both AlphaZero & AlphaGo were stronger than their strongest human counterparts. AlphaZero was estimated to be around ~3200 Elo after 4 hours of training Chess.

AlphaCode being ~1300 goes to show that solving competitive programming problems is a considerably hard & complex task.

  • »
    »
    4 months ago, # ^ |
      Vote: I like it +1 Vote: I do not like it

    Actually I am not sure that it does. An alternative might be that there is simply that there is much less training data available for competitive programing. AlphaZero and AlphaGo were able to generate enormous amounts of new training data by playing copies of themselves. For competitive programming there is, as far as I know, no good way of generating large quantities of new training data.

»
4 months ago, # |
  Vote: I like it +92 Vote: I do not like it

I wish website developers add this when submitting a problem ..

»
4 months ago, # |
  Vote: I like it -8 Vote: I do not like it

Right now I think it would be useful to make good tests, rather than to help solving some problem.

»
4 months ago, # |
  Vote: I like it +13 Vote: I do not like it

We can imagine that one day when tourist is rushing to rating 4000, An AI called sorry_tourist comes and get rank 1st

»
4 months ago, # |
  Vote: I like it 0 Vote: I do not like it

In my opinion: Al will not be able to surpass the tourist in the future for about 10 years.

»
4 months ago, # |
  Vote: I like it 0 Vote: I do not like it

It is very interesting and useful!

»
4 months ago, # |
  Vote: I like it 0 Vote: I do not like it

Now AI will compete with noobie doobie and pupils. Very interesting.

»
4 months ago, # |
  Vote: I like it 0 Vote: I do not like it

Alash!! I could atleast beat AI xD

»
4 months ago, # |
Rev. 3   Vote: I like it 0 Vote: I do not like it

tourist versus terminator would be interesting in near future

»
4 months ago, # |
  Vote: I like it 0 Vote: I do not like it

can it totally understand the mean of the problem...?

  • »
    »
    4 months ago, # ^ |
      Vote: I like it 0 Vote: I do not like it

    Certainly not. It does understand something and makes a kind of brute force based on pretests. Without pretests it would not have any chance to solve anything > div2A

»
4 months ago, # |
  Vote: I like it 0 Vote: I do not like it

Hope they don't actually start competing with regular newbies

»
4 months ago, # |
  Vote: I like it +3 Vote: I do not like it

Why is 1553H - XOR and Distance declared as "pass" here? As far as I can see it wrote a brute force that can't pass.

  • »
    »
    4 months ago, # ^ |
      Vote: I like it 0 Vote: I do not like it

    Many (most?) of the "correct" solutions from alphacode.deepmind.com are incorrect, see this comment for details: https://codeforces.com/blog/entry/99566?#comment-883642

    The paper mentions this in a way, and Table 2 estimates that 46% are "False Positive or Slow Rate", which is probably an underestimate judging from a quick test.

»
4 months ago, # |
Rev. 2   Vote: I like it +8 Vote: I do not like it

Thanks to alphacode, looks like people would no longer need telegram groups in future to cheat, lol.

»
4 months ago, # |
  Vote: I like it +1 Vote: I do not like it

I don't wanna become unemployed .... :(

»
4 months ago, # |
  Vote: I like it 0 Vote: I do not like it

The question is how to train a AI (!dragon).

»
4 months ago, # |
  Vote: I like it 0 Vote: I do not like it

Would love to hear tourist's response to this!

»
4 months ago, # |
  Vote: I like it +1 Vote: I do not like it

Its a great news !!! Although I want to point what can or i should say shouldn't happen in the future, looking at what happened in the past with chess engines :

Fact

Talking about the chess engines , they have just destroyed the beauty of chess, people have stopped brainstorming crazy ideas , they just go for the engines to analyse stuff. This has just killed the creativity of players.

I just hope that engines don't ruin the fun of Problem Soving and Competitive Programming

  • »
    »
    4 months ago, # ^ |
      Vote: I like it +2 Vote: I do not like it

    This is just completely false, chess remains an S-tier game and everyone has crazy and wildly outlandish ideas. Engines didn't ruin chess at all.

    • »
      »
      »
      4 months ago, # ^ |
        Vote: I like it 0 Vote: I do not like it

      IM Panayotis Frendzas (pfren) on Have engines ruined Chess? Or made it better?

      HIS ANSWER

      The Competitive Chess now relies heavily on Engines for the analysis. As an example we can see the various openings that players follow are just bookish and pre-analysed / studied . Although this hasn't totally killed the game I agree with you, but what I have realised after playing chess for so many years is that Chess engines has somewhat changed people's perspective about chess.

      • »
        »
        »
        »
        4 months ago, # ^ |
          Vote: I like it +8 Vote: I do not like it

        As an example we can see the various openings that players follow are just bookish and pre-analysed / studied.

        Pretty sure this was still true without chess engines. Just look at Encyclopaedia of Chess Openings, a book before the era of engines.

        Personally, I think chess engines made the game better. They allow new players to learn easier. Also, a good player would keep in mind that the best move for a computer might not be the best for a player, since we can't see 20-30 moves ahead.

»
4 months ago, # |
  Vote: I like it 0 Vote: I do not like it

Tourist v/s AI .would love to witness it!

»
4 months ago, # |
  Vote: I like it -13 Vote: I do not like it

Tourist will win for sure!

»
4 months ago, # |
  Vote: I like it 0 Vote: I do not like it

Lowd** ka AI

»
4 months ago, # |
  Vote: I like it 0 Vote: I do not like it

how does it make sure that the training data it takes from github doesn't already contain the solution to the problem it's attempting when doing past contests? (i.e solutions posted by competitors in their public repos following a contest)

i'm guessing it must take the commit's timestamp into consideration

  • »
    »
    4 months ago, # ^ |
      Vote: I like it 0 Vote: I do not like it

    they test on contests that appear after the data collection process

»
4 months ago, # |
  Vote: I like it 0 Vote: I do not like it

this will change the CP world..

»
4 months ago, # |
  Vote: I like it 0 Vote: I do not like it

Bot will never replace Tourist!!!

»
4 months ago, # |
  Vote: I like it 0 Vote: I do not like it

Let's gooo AI. Near future, we don't have to think about CP anymore. Which is really Awesome. This indicates that we as humans are developing in technologies.

  • »
    »
    4 months ago, # ^ |
      Vote: I like it 0 Vote: I do not like it

    Except when you get replaced. Then it won't be really awesome.

  • »
    »
    4 months ago, # ^ |
      Vote: I like it +67 Vote: I do not like it

    Near future, we don't have to think about CP anymore.

    Yeah, just like people don't have to think about chess anymore now that AI can beat humans in chess.

    • »
      »
      »
      4 months ago, # ^ |
        Vote: I like it 0 Vote: I do not like it

      nobody have to think about chess anymore.

      And nobody had to, but that's besides the point

»
4 months ago, # |
Rev. 2   Vote: I like it 0 Vote: I do not like it

Which rating does it really have — 1300 or 1000? There are worlds in between. In terms of percentile, the distance between 1000 and 1300 is the same as between 1300 and 3800.

»
4 months ago, # |
  Vote: I like it +24 Vote: I do not like it
»
4 months ago, # |
Rev. 4   Vote: I like it 0 Vote: I do not like it

soon it will defeat tourist. like the stockfish(chess engine) defeated world chess champion.

»
4 months ago, # |
Rev. 2   Vote: I like it +15 Vote: I do not like it

Okay I'm going to retire after this f**king AI beats me just three days later

»
4 months ago, # |
  Vote: I like it 0 Vote: I do not like it

AI and VR scare the shit out of me

»
4 months ago, # |
Rev. 2   Vote: I like it 0 Vote: I do not like it

God : What's your age when you are able to solve codeforces 1300+ rating problem?
Me : 18+ year
AI : Just after birth

»
4 months ago, # |
  Vote: I like it 0 Vote: I do not like it

We can witness the history of codeforces change!

»
4 months ago, # |
  Vote: I like it 0 Vote: I do not like it

I am curious to see Deepminds submissions. :)

»
4 months ago, # |
  Vote: I like it 0 Vote: I do not like it

How to deal with unsolvable problems such as AHC?

»
4 months ago, # |
  Vote: I like it 0 Vote: I do not like it

At least I'm still better than AI

For now.....

»
4 months ago, # |
  Vote: I like it +21 Vote: I do not like it

More construction problem, less data structure problem, then AI will lose.

»
4 months ago, # |
Rev. 2   Vote: I like it +1 Vote: I do not like it

I think it's already hard enough for an AI to understand programming problems, let alone find a suitable algorithm and program it correctly. I cannot figure out how many efforts it's needed to train such an AI.

»
4 months ago, # |
  Vote: I like it 0 Vote: I do not like it

AI is a noob. I'm sure that tourist will definitely win this

»
4 months ago, # |
  Vote: I like it 0 Vote: I do not like it

Don't be afraid folks. Terminator knows no more than Brute Force algorithm.

  • »
    »
    4 months ago, # ^ |
      Vote: I like it 0 Vote: I do not like it

    maybe as of now but then there comes "Machine Learning" XD

    • »
      »
      »
      4 months ago, # ^ |
        Vote: I like it 0 Vote: I do not like it

      I would be impressed if Terminator use some kind of decomposition. And backtracking.

»
4 months ago, # |
  Vote: I like it 0 Vote: I do not like it

fighting tourist, u r the god of humankind

»
4 months ago, # |
  Vote: I like it 0 Vote: I do not like it

Just started CP a month ago and then this blog pops up... Great.

»
4 months ago, # |
  Vote: I like it 0 Vote: I do not like it

Remembering the Silicon Valleys(webSeries) AI improve algorithms and Algorithms improve AI. I guess Deepmind just got the Pendrive or Stole it.

»
4 months ago, # |
  Vote: I like it +6 Vote: I do not like it

This is at the same time both amazing and terrifying as well.

»
4 months ago, # |
  Vote: I like it +3 Vote: I do not like it

cannot be compared to a tourist

»
4 months ago, # |
  Vote: I like it +3 Vote: I do not like it

Imagine getting banned for using same variables and same steps by two AI accounts .

»
4 months ago, # |
  Vote: I like it 0 Vote: I do not like it

Definitely bot will not surpass tourist but I think that will surpass coders like me lol and it seems dangerous to me.

»
4 months ago, # |
  Vote: I like it +18 Vote: I do not like it

It would be fun to see AI hacking solutions xD.

»
4 months ago, # |
Rev. 3   Vote: I like it 0 Vote: I do not like it

Upcoming Contest be like "Div X"; only for AIs. Then, an AI neural network against tourist

»
4 months ago, # |
  Vote: I like it 0 Vote: I do not like it

Why didn't they participate in the last contest?

»
4 months ago, # |
  Vote: I like it +1 Vote: I do not like it

Imagine devs of AI getting tons of money from the contests with prizes

»
4 months ago, # |
  Vote: I like it 0 Vote: I do not like it

Long before there are systems available that solve problems on their own, we will see systems that support humans at problem solving. This is the case in nearly all AI developement areas.

So these problem solving bots will not suddenly jump out of the dark. I expect the development to be more like self driving cars, a process spanning several decades.

»
4 months ago, # |
  Vote: I like it 0 Vote: I do not like it

i don't know why but this is scary

»
4 months ago, # |
  Vote: I like it +5 Vote: I do not like it

So when will we have AI beating humans in an AI-making contest?

»
4 months ago, # |
  Vote: I like it 0 Vote: I do not like it

In the past, we used to say that programming will do what eliminates many jobs such as (accountant, doctor, teacher, worker.....) It seems that the programmer will be added to them

  • »
    »
    4 months ago, # ^ |
      Vote: I like it 0 Vote: I do not like it

    we will be first

    • »
      »
      »
      4 months ago, # ^ |
        Vote: I like it 0 Vote: I do not like it

      Still we can more easily migerate to AI than a/an (accountant, doctor, teacher, worker ..). That's the good story :)

»
4 months ago, # |
  Vote: I like it +1 Vote: I do not like it

I wonder how much time it takes to write the code? If it comes with the solution then it would just code it in lightening speed.

»
4 months ago, # |
Rev. 2   Vote: I like it +7 Vote: I do not like it

Curious mind wants to know
The AI WaggleCollide which does not know the language it is writing the code, how can it beat tourist? tourist

»
4 months ago, # |
  Vote: I like it +13 Vote: I do not like it

I picked a random ez problem with a reasonably nice statement: 1559A - Mocha and Math.

#include <bits/stdc++.h>
using namespace std;
int a[105];
bool vis[105][105];
int dp[105][105];
int main() {
  ios::sync_with_stdio(false);
  int t;
  cin >> t;
  while (t--) {
    int n;
    cin >> n;
    for (int i = 1; i <= n; i++) {
      cin >> a[i];
    }
    int ans = 0x3f3f3f3f;
    for (int i = 1; i <= n; i++) {
      for (int j = i; j <= n; j++) {
        int temp = a[i];
        for (int k = i + 1; k <= j; k++) {
          temp &= a[k];
        }
        ans = min(ans, temp);
      }
    }
    cout << ans << endl;
  }
}

The AI does min(1061109567, AND(a[l:r]) for each subarray [l:r]). It doesn't have a clue that AND(a[l:r+1]) <= AND(a[l:r]) which can optimise this to $$$O(N)$$$. It also puts random-ass global arrays. It's the same no matter what layer I pick. Wat.

  • »
    »
    4 months ago, # ^ |
      Vote: I like it +56 Vote: I do not like it

    The baby has just been born and it can't even walk on its own. Wat.

    • »
      »
      »
      4 months ago, # ^ |
        Vote: I like it +8 Vote: I do not like it

      How long do you think this thing has been in development? You're acting like it suddenly appeared out of thin air.

»
4 months ago, # |
  Vote: I like it 0 Vote: I do not like it

Of course the confrontation will take place in the future (just like AlphaGo, Lee Sedol and Ke Jie).

In 2016, almost everyone thought Lee Sedol would beat AlphaGo easily, but he lost at last. AI can't be underestimated.

»
4 months ago, # |
  Vote: I like it +18 Vote: I do not like it

i think its time to private all our github repos

»
4 months ago, # |
  Vote: I like it 0 Vote: I do not like it

Though the chances of it happening are very very low because AlphaZero's code wasn't made publicly available (afaik), hope that in a few years Codeforces doesn't turn into the CP equivalent of today's chess.com with so many of its users highly-skilled at using chess engines (they won't be using AlphaZero, but still).
Guess we'll just have to adapt with the times.

»
4 months ago, # |
  Vote: I like it +6 Vote: I do not like it

Somehow I am having feeling of insecurity…

»
4 months ago, # |
  Vote: I like it +1 Vote: I do not like it

Will be interesting to see these ranting about weak pretests, lol...

»
4 months ago, # |
  Vote: I like it +22 Vote: I do not like it

Yes, everything is going according to my plan prophecy: https://codeforces.com/blog/entry/97546

»
4 months ago, # |
  Vote: I like it +14 Vote: I do not like it

4 years ago, I played Go as a hobby and AI exceeded me. Now, I am just programming and AI exceeded again.

»
4 months ago, # |
  Vote: I like it +114 Vote: I do not like it

I'm not a ML expert, but I wonder will the result be substantially improved if training could be performed on private data. For example, if access to generators is provided (e.g. retrieve them from polygon?), possibly we could deploy a RL-like training process? At the very least we could eliminate some false positives. Impressive work anyways!

  • »
    »
    4 months ago, # ^ |
      Vote: I like it +32 Vote: I do not like it

    Stop giving them ideas. Are you on the side of the terminators?

»
4 months ago, # |
  Vote: I like it 0 Vote: I do not like it

Adversarial detection required!!

»
4 months ago, # |
  Vote: I like it 0 Vote: I do not like it

We feel like we're going to lose our jobs

»
4 months ago, # |
  Vote: I like it 0 Vote: I do not like it

Yelp, the time has come. Ai overlords are upon us.

»
4 months ago, # |
  Vote: I like it 0 Vote: I do not like it

Great, Now i will lose to a bot.

»
4 months ago, # |
Rev. 2   Vote: I like it -13