Репост из: Data, Stories and Languages
Kaggle Nemotron Competition Writeup for the 1st place: codex написал весь код, но думать он не смог
https://www.kaggle.com/competitions/nvidia-nemotron-model-reasoning-challenge/writeups/1st-place-solution
Закончилось очередное соревнование на каггле. Участник победившей команды поделился опытом:
"All code was written with Codex. We did not write a single line of code directly. However, the ideas, analysis direction, and trace design decisions that improved the score almost never came from Codex."
Но есть нюанс:
"I don't know for sure, but other than Gemini 3.1 Pro Extended, almost all other AIs were giving conflicting ideas, and most of those ideas were just bad. Specially Claude, like it was deliberately trying to emphasize on an already wrong proven hypothesis. One thing this competition taught me is, AI is still far way from working without intervention."
Всё, решено - далеко нам до AGI.
#ai #kaggle
https://www.kaggle.com/competitions/nvidia-nemotron-model-reasoning-challenge/writeups/1st-place-solution
Закончилось очередное соревнование на каггле. Участник победившей команды поделился опытом:
"All code was written with Codex. We did not write a single line of code directly. However, the ideas, analysis direction, and trace design decisions that improved the score almost never came from Codex."
Но есть нюанс:
"I don't know for sure, but other than Gemini 3.1 Pro Extended, almost all other AIs were giving conflicting ideas, and most of those ideas were just bad. Specially Claude, like it was deliberately trying to emphasize on an already wrong proven hypothesis. One thing this competition taught me is, AI is still far way from working without intervention."
Всё, решено - далеко нам до AGI.
#ai #kaggle