Flash Report: ChatGPT-4o Takes on Competitive Programming [English Subtitles]

Рет қаралды 7,942

Ай бұрын

We had ChatGPT equipped with GPT-4o solve problems from AtCoder Beginner Contest 353.
Problems: atcoder.jp/contests/abc353/tasks
Code
A: atcoder.jp/contests/abc353/su...
B: atcoder.jp/contests/abc353/su...
C: atcoder.jp/contests/abc353/su...
Problem Explanation: • 4分でAtCoder Beginner Co...
X: / evima0

Пікірлер: 17

@yasaka-jp Ай бұрын

GPT-4 から GPT-4o になってコード生成系のタスクの質が悪くなった、というのを話題にしているツイートを見た

@hitsuki_karasuyama Ай бұрын

4oは動画や漫画からも学習しているから、サブカル知識とかは増えたけどプログラミングとかは大して改善していない本番はGPT5

@user-jy8ev9th7n Ай бұрын

なるほどぉ　ユーザーとしては、汎用性が高いのもありがたいけど、特化型つくってくれたらそれもありがたい

@zheng166 Ай бұрын

I recognize that the advantages of gpt-4o are that the speed and price are half of what they used to be, and now it has visual capabilities, and the latency of voice conversations has been reduced through webrtc technology

@zheng166 Ай бұрын

They did not advertise any progress in understanding and reasoning in the model, maybe we need a contest programming arena for models.

@tomanification_music_monke983 Ай бұрын

生成の試行プロセスを減らす機構が入ったとかなんだろうな。たぶん演算しないで類似の回答を引っ張り出してくること優先とかになってコードの修正しようとしてもコードの演算させてくれない問題になってきてる気がする

@AngrySadEight Ай бұрын

迅速な投稿、お疲れ様です。 ......ある意味衝撃の結末でしたね。

@masuo64 Ай бұрын

サムネかわいいですね。

@evimalab Ай бұрын

ありがとうございます。（容易に察せられる通り ChatGPT 4o に描いてもらいました。）

@asamas9648 Ай бұрын

タイムリー

@rupanugapmishra7128 Ай бұрын

thank god agi is still far if it cant even do 800 elo atcoder question... tho its probably 1400 in codeforces... why not try codeforces next? see till which elo it can solve in codeforces?

@evimalab Ай бұрын

They have a long way to go, I guess. Making a more extensive version of this video is probably an option, but I'm not sure whether I should do it as one of those who prepare these contests. Although this time GPT-4o failed, sometimes AI can do quite well. For example, in ABC349, GPT-4 Turbo was able to solve A to E quickly, achieving a performance rating of about 2100. I don't think it's actually possible to "ban AI" during contests (how would you detect it, and what exactly qualifies as AI?), but I don't want to promote using AI either.

@rupanugapmishra7128 Ай бұрын

@@evimalab gpt 4 was able to do e in ABC349? pretty sure its a standard recursion question and gpt must have been trained a lot in those recursive tic tac toe algorithms... thats probably it was able to do it since it was in the training data right? like it has been ingrained in its neural net....