Jim Fan 是 OpenAI 的第一位实习生，现任英伟达资深研究科学家兼 AI 代理主管。!!!!!

首先Jim Fan 的两个发文是同一天！！！！

人家是解释了 OpenAI 没有那么低级直接 for(i=0;i<n;i++){UE;}

何谈辟谣？

既然很多朋友都热衷讨论，sora的训练集到底是什么？

那么我们就来看看研究下

拉尔夫布鲁克斯 AT&T 数据科学家

I do agree that it was likely trained with Unreal Engine 5 or something quite similar. Having watched quite a lot of UE5 demos around its launch I found the Sora examples, the first on in particular, to have quite an UE5 look to it.

And honestly it makes perfect sense to use game engines as a data source. Not only can you produce enormous amounts of near photo-realistic footage in various scenarios using it, you also get far more than just visual data. You get detailed depth data, motion data, and so many other things which you can feed into the AI to teach it about the world.