Sora App
Today we’re releasing Sora 2, our flagship video and audio generation model.
The original Sora model from February 2024 was in many ways the GPT‑1 moment for video—the first time video generation started to seem like it was working, and simple behaviors like object permanence emerged from scaling up pre-training compute. Since then, the Sora team has been focused on training models with more advanced world simulation capabilities. We believe such systems will be critical for training AI models that deeply understand the physical world. A major milestone for this is mastering pre-training and post-training on large-scale video data, which are in their infancy compared to language.
With Sora 2, we are jumping straight to what we think may be the GPT‑3.5 moment for video. Sora 2 can do things that are exceptionally difficult—and in some instances outright impossible—for prior video generation models: Olympic gymnastics routines, backflips on a paddleboard that accurately model the dynamics of buoyancy and rigidity, and triple axels while a cat holds on for dear life.
Prior video models are overoptimistic—they will morph objects and deform reality to successfully execute upon a text prompt. For example, if a basketball player misses a shot, the ball may spontaneously teleport to the hoop. In Sora 2, if a basketball player misses a shot, it will rebound off the backboard.
The Sora app for iOS is available to download now, and it can be used in the United States and Canada. Those invited to the app will be able to use Sora 2 on the Sora website.
People on Sora who generate videos of Altman are especially getting a kick out of how blatantly OpenAI appears to be violating copyright laws. (Sora will reportedly require copyright holders to opt out of their content’s use — reversing the typical approach where creators must explicitly agree to such use — the legality of which is debatable.)
[…]
Aside from its algorithmic feed and profiles, Sora’s defining feature is that it is basically a deepfake generator — that’s how we got so many videos of Altman. In the app, you can create what OpenAI calls a “cameo” of yourself by uploading biometric data. When you first join the app, you’re immediately prompted to create your optional cameo through a quick process where you record yourself reading off some numbers, then turning your head from side to side.
Each Sora user can control who is allowed to generate videos using their cameo. You can adjust this setting between four options: “only me,” “people I approve,” “mutuals,” and “everyone.”
It’s been a long time since I’ve been this sucked into an app. Such was the situation I found myself in last night with OpenAI’s new version of Sora. Once I got access, I found it nearly impossible to put it down. I just kept wanting to remix everything I scrolled past. Yes, it was incredibly dumb. Yet highly amusing! And technically, very interesting, albeit in mildly troubling ways. But everyone else will write about that aspect, and rightfully so. My angle here is simply that OpenAI remains so good at creating these types of viral products. Underlying tech aside, that team continues to seem to know how to productize better than anyone else in the space.
Case in point: Meta launched a similar foray just days before in the form of “Vibes”. Now, did they rush it out the door to get ahead of this Sora 2 launch? Hard to say for sure, but it sure feels that way. The product, if you even want to call it that, is so half-baked and obtuse to use that it’s more like an employment quiz.
Started using the Sora app and it’s like TikTok for AI generated videos.
I used to think it would take a year or two for AI videos to become as popular as influencer content on social media but I can see this app causing that to happen by the end of the year.
Sora, though invitation-only at the moment, is currently #3 in the U.S. App Store.
[…]
Also, I’m sure Sora will eventually come to Android. But, to play with it now, you need an iPhone.
Previously:
