Sora背后团队：应届生领衔，00后加入，还特招艺术生

文章转载来源：AIcore

文章来源：量子比特

图片来源：无界AI生成

目前全球最受关注的技术团队是哪支？

Sora 团队成为了人们关注的焦点。

项目负责人的评论区不仅人满为患，还成为最受欢迎的“景点”。

才华横溢的成员们的履历也持续受到关注。

△来自微博博主@木雅

大家发现这支队伍相当年轻：两位带头人去年（2023年）刚刚博士毕业，队伍里甚至还有2000年后的球员……

但它也确实很棒：

Tim Brooks，DALL-E 3 的作者之一，GitHub 5.7k 项目 InstructPix2Pix 的作者，2021 年至 2022 年在 NVIDIA 实习时是视频生成研究的项目负责人。

William (Bill) Peebles 与 Xie Senin 合作开发了 Sora 的技术基础之一——DiT（扩散变压器）。该论文还入围了 CVPR 2022 最佳论文候选名单。

……

今天我们就来详细聊聊这支球队的来历。

由应届毕业生带领的团队

包括Tim和Bill在内，Sora的主要负责人有3人（以下排名不分先后）。

Tim Brooks 也是 DALL-E 3 的作者，去年 1 月刚从加州大学伯克利分校获得博士学位。

Tim 就读于卡内基梅隆大学，主修逻辑和计算，辅修计算机科学。在此期间，他在Facebook的软件工程部门实习了四个月。

2017年，蒂姆本科毕业后，先是在谷歌工作了近两年，在Pixel手机部门研究AI摄像头，随后前往伯克利AI实验室攻读博士学位。

在伯克利攻读博士学位期间，Tim 的主要研究方向是图像和视频生成。他还在 NVIDIA 实习并领导了​​一项视频生成研究。

回到校园后，Tim 与导师 Alexei Efros 教授和博士后 Aleksander Holynski（现就职于 Google）共同开发了 AI 图像编辑工具 InstructPix2Pix，并入选 CVPR 2023 亮点。

去年1月，Tim顺利博士毕业，加入OpenAI，并先后参与了DALL-E 3和Sora的工作。

值得一提的是，蒂姆不仅在专业领域拥有很高的技术水平，而且还是一位多才多艺的人。

据蒂姆本人介绍，他还喜欢摄影和音乐。他高中时拍摄的照片获得了国家地理杂志的奖项。他曾在百老汇演出并荣获 B-box 国际奖项...

与蒂姆就读同一所学校并于四个月后毕业的威廉·皮布尔斯也是索拉的另一位负责人。

(Peebles uses the nickname Bill on ? and the full name William on Linkedin and when signing papers. He will be referred to as Bill below.)

Bill studied at MIT as an undergraduate, majoring in computer science. He participated in research on GAN and text2video, and also interned in NVIDIA's deep learning and autonomous driving team, studying computer vision.

After graduation and before officially starting his Ph.D., he also participated in a summer internship at Adobe, and his research was still on GAN. This project and (then) Chinese scholar Zhu Junyan of Carnegie Mellon University (also a student of Professor Efros, now at MIT) The group has cooperated and became a candidate for the best paper in CVPR 2022.

After that, at the beginning of the semester, Bill went to Berkeley to study for a doctoral degree in Professor Efros's research group. His research results were selected into academic conferences such as SIGGRAPH, ICCV, and CVPR for many times.

In May 2022, Bill went to Meta for a half-year internship, and collaborated with Xie Saining (Bill had not left Meta when he started his internship) to publish the DiT model, which combined the Transformer and the diffusion model for the first time.

This result was accepted as an Oral paper in ICCV 2023. It is worth mentioning that Sora, released by OpenAI this time, is believed to be built based on DiT.

In May last year, Bill also graduated from Berkeley and joined OpenAI.

In addition to these two researchers who joined last year, Aditya Ramesh, another leader of the Sora team, is the "old man" of OpenAI.

Aditya is the creator of DALL-E and has led the research on three generations of DALL-E. He is a co-author on all three versions of the paper.

And such a master who led three generations of DALL-E and now leads the Sora team only has a bachelor's degree.

According to LeCun, Aditya studied at New York University as an undergraduate and participated in some projects in his laboratory.

In the meantime, Aditya was already studying generative models and published a paper with LeCun.

After graduation, Aditya wanted to continue his studies, but was retained during OpenAI's summer internship and became a formal researcher.

Joined after 00

Aditya Ramesh is not the only undergraduate student on Sora’s team.

As mentioned earlier, there is a "post-00s" Will DePue in this team, who just graduated from the Department of Computer Science at the University of Michigan in 2022.

When he was a senior in college, he started a business, a market consulting company called Deep Research, which was later acquired by Commsor.

In July 2023, I joined OpenAI. According to his LinkedIn information, he just joined the Sora project team in January this year.

In addition, neither David Schnurr nor Joe Taylor has a Ph.D. The former graduated from the University of California, Santa Barbara, and the latter graduated from the Art University of San Francisco.

And as Aditya Ramesh himself said, many members of the Sora team are the authors of DALL-E 3.

Including two Chinese, Li Jing and Yufei Guo.

Li Jing is the co-author of DALL-E 3. He graduated from the Department of Physics of Peking University in 2014 and received his PhD in Physics from MIT in 2019. After working as a postdoc at Meta for more than 2 years, Li Jing joined OpenAI in 2022.

Among the Chinese authors is Ricky Wang, who just switched from Meta/Instagram to OpenAI in January this year. The other two, Yufei Guo and Clarence Ng, do not have much public information.

Also new to the job is Conner Holmes. When he was working at Microsoft, he participated in the inference optimization work of DALL·E 3 as a foreign aid, and later joined OpenAI.

Finally, take a look at the full list of authors:

Judging from the team's formation and research foundation, Sora should be OpenAI's latest achievement in the past six months, rather than "it has been around for a long time but has been held back" as reported online.

However, the explosion of Sora and the continued gathering of top talents shocked everyone to reconsider OpenAI's technological leadership.

Just today, the author released Sora's new work, which even includes multi-camera videos of "the same scene".

Netizens’ mood be like:

Now, it’s video generation, what’s next?

Reference links: [1]https://www.wpeebles.com/ [2]https://www.timothybrooks.com/about/ [3]http://adityaramesh.com/about.html

热点：比特币背后 加入nft 元宇宙加入 加入币圈 SOL官网