# Building OpenAI o1

## Метаданные

- **Канал:** OpenAI
- **YouTube:** https://www.youtube.com/watch?v=3k89FMJhZ00
- **Дата:** 12.09.2024
- **Длительность:** 3:16
- **Просмотры:** 241,763
- **Источник:** https://ekstraktznaniy.ru/video/11478

## Описание

Top row (left to right): Mark Chen, Giambattista Parascandolo, Trapit Bansal, Łukasz Kaiser, Hunter Lightman, Karl Cobbe, Łukasz Kondraciuk, Szymon Sidor, Noam Brown, Hongyu Ren, Liam Fedus, Hyung Won Chung

Bottom row (left to right): Ilge Akkaya, Jakub Pachocki, Shengjia Zhao, Jason Wei, Wojciech Zaremba, Jerry Tworek

Host: Bob McGrew

More here: www.openai.com/o1
Full list of contributors: https://openai.com/openai-o1-contributions/

## Транскрипт

### Segment 1 (00:00 - 03:00) []

We're starting a series of new models with  the new name o1 and this is to highlight the   fact that you might feel different when you use  o1 as a compared to previous models such as GPT-4o   so as others will explain later o1 is a reasoning  model so it will think more before answering your   question. We are releasing two models o1-preview  which is to preview what's coming for o1 and o1   mini which is a faster slow smaller and faster  model that is trained with a similar framework as   o1 so we hope you like our new naming scheme o1. So  what is reasoning anyway so one way of thinking of   reasoning is that there are times where we ask  questions and we need answers immediately because   they're simple questions. For example if you ask  what's the capital of Italy you know the answer   and you don't really have to think about  it much but if you wonder about a complex   puzzle or you want to write a really good business  plan, the novel, you probably   want to think about it for a while and the more  you think about it the better the outcome so   reasoning is the ability of turning thinking time  into better outcomes whatever the task you're   doing. It's been going on for a long time but I  think what's really cool about research is there's   that aha moment there's that particular point  in time where something surprising happens and   things really click together. Are there any times  for you all when there was you had that aha moment? There was the first moment when the moment was hot  of the press we started talking to the model and   people were like wow this model is really  great and starting doing something like that   and I think that there was a certain moment in our training process where we trained like put   more compute in RL than before and train first  all generating coherent chains of thought and we   so wow this looks like something meaningfully  different than before and I think for me   this is the moment. I think related to that  when we think about like training a model for   reasoning one thing that immediately jumps to mind  is you could have humans write out their thought   process and train on that. An aha moment for me  was like when we saw that if you train the model   using RL to generate and hone its own chain  of thoughts it can do even better than having   humans write chain of thought for it. And that was  in aha moment that you could really scale this and explore models reasoning that way. For a lot of  the time that I've been here we've been trying to   make the models better at solving math problems as  an example and we've put a lot of work into this   and we've come with a lot of different methods  but one thing that I kept like every time I would   read these outputs from the models I'd always be  so frustrated that the model just would never   seem to question what was wrong or when it was  making mistakes or things like that but one of   these early uh o1 models when we trained it and we  actually started talking to it we started asking   it these questions and it was scoring higher on  these math tests we were giving it we could look   at how it was reasoning and you could just see  that it started to question itself and have really   interesting reflection and that was a moment for  me where I was like wow like we we've uncovered   something different this is going to be something  new and it was just like one of these coming   together moments that that was really  powerful. Thank you and congrats on releasing this.