It's a misunderstanding, it's a misunderstanding. Xie Saining, an alumnus of Shanghai Jiao Tong University mistakenly identified as "one of the authors of Sora," urgently clarifies this matter.
His statement was posted on his Moments (a feature in Chinese social media platforms similar to Facebook timeline) as well as the "交大校友荟" (Jiao Tong University Alumni Association) account on Jiao Tong University's official platform. He mainly addressed several points:
Sora is entirely the work of OpenAI, with no involvement from Xie Saining himself.
However, he will be conducting related research at New York University in the future.
To develop achievements like Sora, talent comes first, data comes second, and computing power comes third.
Some stories in science fiction works like "Black Mirror" may soon become reality, so it's important to be prepared with corresponding oversight systems.
The original article spreading false information has also been taken down by the platform after everyone's reports.
This misunderstanding finally seems to have come to an end.
However, there are still some details worth noting in the information revealed by Thessain. Bill Peebles, the familiar Sora development lead, mentioned that the team has been working at high intensity, almost without sleep, for nearly a year.
Bill Peebles joined OpenAI in March 2023, which is just under a year ago.
This also indirectly refutes the previous rumor that "Sora has been developed long ago but just hasn't been released."
As for why this outrageous misunderstanding arose, it all stems from a paper co-authored by Sheshaing and Peebles.
What caused this misunderstanding?
The DiT model referenced in the Sora technical report is precisely the result of collaboration between Sheshaing and Peebles.
At the time, Dr. Peebles was interning at Meta during his studies, while Sheshaing had not yet fully departed from FAIR.
DiT was the first to integrate Transformer with diffusion models, and the related paper was accepted as an Oral presentation at ICCV 2023.
Although the specific architecture details of the Sora model haven't been disclosed, Xie's analysis from the Sora technical report suggests that it's likely built upon the foundation of DiT.
In addition, Sora also drew inspiration from many other academic achievements. For instance, the method of encoding training data into patches was referenced, including Google DeepMind's Patch n’ Pack, among others.
By drawing on various strengths and firmly adhering to the guidance of Scaling Law, the remarkable achievements of Sora have been accomplished.
Indeed, greatness recognizes greatness.
To dispel any misunderstanding, Yann LeCun is indeed a luminary in the field of AI research, currently serving as an Assistant Professor at New York University.
He graduated from the ACM class of Shanghai Jiao Tong University for his undergraduate studies, and earned his PhD from the University of California, San Diego in 2018.
After graduating, Xie Saining joined Facebook AI Research (FAIR), where he collaborated with Kaiming He on proposing ResNeXt.
Currently, his total citations on Google Scholar have reached 45,000, with an impressive h-index of 31.
Just before Sora's release not long ago, he led a team at New York University to give the DiT model a major upgrade, introducing a new model called SiT.
SiT shares the same backbone network as DiT, but offers improved quality, speed, and flexibility.
Xie Saining believes that engaging in research related to Sora is one of the significant responsibilities that academia can shoulder in this era.
Let's look forward to his new achievements together.
Reference link: 1
