One might argue that cheating is a risk already given that it’s possible to discuss problems with others without being caught. This means that even in the best case, access to AI would give cheaters an advantage comparable to access to an editorial, which is clearly a game-breaking edge. Even if we assume for the sake of argument that the hypothetical AI writes code that is watermarked in a way that can’t be removed through minor modifications, it seems reasonable to assume given the capability of ChatGPT to provide mostly reasonable explanations for the code it generates that an AI capable of solving any competitive programming problem will be able to explain the solution idea in words comparably well to an editorial (or, even if this is not the case at launch, it seems nearly certain that updated versions of the model will eventually develop this capability). Neither of these mitigating factors apply to CF. The relevant differences between chess and Codeforces are that in chess, my understanding (as someone who doesn’t play chess…) is that (a) online chess platforms are pretty good at detecting blatant cheating, so only those who cheat in relatively subtle ways can go undetected, and (b) OTB tournaments are frequent enough to serve as the main form of serious competition, meaning that (1) players who consistently perform much better online than OTB may attract suspicion and (2) cheating online is not as problematic because OTB rating is considered the primary metric of skill. Maybe I’m wrong, but it’s hard to imagine a situation where a public AI model becomes stronger than the best competitive programmers but online programming contests don’t die out. I respect your point of view, and even though we might think differently, it's cool to exchange ideas. This enables me to provide relevant and helpful information when answering questions or discussing topics. My training data includes a vast amount of text from various sources, such as websites, books, and code repositories. Instead, I generate responses based on the patterns and knowledge I have learned during my training. At the same time, it's not accurate to label it as mere copying and pasting of existing code.Īs an AI language model, I do not directly copy and paste from existing code or text. When programmers use "int" or name a variable "tmp," are they copying from one another? Or are they simply utilizing existing knowledge? When someone develops a solution, it can't be considered wholly original, as it is a combination of pre-existing knowledge. I believe our disagreement stems from the choice of words, particularly the term "copy-paste." Setting aside AI models for a moment, let's consider writing code in C++. Therefore, this contest will be shock the whole world again, end the old era and open the new era, and we contestants will all witness the history. We can also say, if AI beat tourist and rank No.1 in a contest, it means AI beat the whole education system, make all online exams and contests cheatable. However, think about it, why alphago vs Sedol lee or Jie Ke is most famous and shock the whole world? Because go is the hardest checkerboard game AI can beat, beat most famous go player means beat all checkerboard. The reason why I post this, is that many people, including me, worries that chat-GPT will destroy the competitive programming, like alphago destroy the go. Therefore, I am proud to say, competitive programming, like us codeforces, is one of the hardest obstacles chat-GPT will overcome. Codeforces is the only one that GPT behave below 5%. We can see from this picture that latest Chat-GPT have very outstanding performance at all kinds of exams GRE (beat me hard in GRE verbal), SAT, LSAT.īut for codeforces, original GPT only have a rating of 260, the GPT-4 is slightly better, with a rating of 392. Now the newest GPT have published its academic performance.
0 Comments
Leave a Reply. |
AuthorWrite something about yourself. No need to be fancy, just an overview. ArchivesCategories |