Sora built a social network
OpenAI launched Sora 2 as a TikTok-style app where every video is generated. The model is impressive. I'm less sure a feed of generations gives people a reason to keep watching.
OpenAI released Sora 2 last Tuesday, and alongside it a new iPhone app called Sora. It’s invite-only in the US and Canada, and it went to the top of the App Store within a few days. The app is a vertical feed, like TikTok, but every video in it is generated. You make a video from a prompt, or remix someone else’s, and there’s a feature called Cameos: you record a short verification video of yourself, and then you, and friends you give permission to, can put you into generated videos.
A personal note: I recently moved into a new role at YouTube working on generative video. This post covers the public product and my own reaction to it.
The model is a real step forward. The physics is noticeably better than the original Sora from February last year, which I wrote about for its broken glasses and multiplying wolf pups. Sora 2’s clips have synchronized dialogue and sound, like Veo 3, and objects behave more like objects. OpenAI’s launch post shows a missed basketball shot bouncing off the backboard instead of teleporting into the hoop, and frames that as the model being willing to let things fail realistically. I think that’s the right thing to highlight.
The app is the more interesting decision. OpenAI could have shipped Sora 2 as a creation tool, the way most video models are offered. Instead they built a social network, with feeds, follows, likes, remixes and a recommendation system. The Cameo feature makes it personal: the most popular content in the first week has been people putting themselves and their friends into absurd situations, and a lot of videos of Sam Altman, who made his cameo public.
My view, which I hold with less confidence than usual: a feed of generations is a great launch week and a hard long-term product.
What makes people come back to TikTok or YouTube isn’t that the videos are well made. It’s people: creators with a point of view, whose lives, skills or humor you follow over time. Generated content lowers the cost of making things to near zero, which means the supply is infinite, and infinite supply makes each piece worth less attention. The novelty of seeing yourself in a generated video is powerful but wears off. What’s left after that has to be a reason to watch other people’s generations, and I don’t think the model alone provides one.
There’s also the IP question, which OpenAI is visibly struggling with this week. The feed quickly filled with copyrighted characters, and they’ve announced changes to give rights holders more control.
I think where generation is most useful is inside existing social and creator platforms, as a tool that helps real creators make things they couldn’t make before, rather than as a separate world of generated content. But the week has been a great real-world experiment, and I’ll be watching the retention numbers, if they ever become public, more closely than the downloads.