Can OpenAI build a social network?

What does Sora tell us about the future of A.I.?

Greetings from Read Max HQ! In this week’s newsletter, a consideration of OpenAI’s new video-generation app “Sora,” a subject I discussed on KQED on Wednesday in a segment you can listen to here.

On Monday, I’ll be discussing “The Aesthetic Allure of Fascism” with Juliet Jacques and Nadia Meeran in a Zoom conversation hosted by the Future Narratives Lab. The event starts at 1:30 p.m. ET; you can register to attend here.

A reminder: This newsletter is my full-time job, and, due to the basic structure of subscription businesses and the reality of “churn,” I need to sign up new paying subscribers every week to keep going. If you’ve been enjoying the work (and it is work, and also procrastination) that goes into this newsletter, please considering rewarding it with a paying subscription, for effectively the same price as buying me one (1) (cheap) beer at a local bar.

Over the past week, thanks to OpenAI’s new generative-A.I. video app Sora, amateurs and non-enthusiasts have had free, push-button access to state-of-the-art text-to-video generation, in many cases for the first time. Now you, too, just like any expert deep-faker or slop magnate, can instantly generate shockingly realistic short videos based on prompts like, e.g., “Sam Altman giving a product demonstration on stage. He says ‘one more thing’ and then begins to show symptoms of c diff, doubling over and making a huge mess on stage”:

What you are seeing here, besides an A.I.-generated video of OpenAI C.E.O. Sam Altman blasting diarrhea on stage, is a pivotal moment in the A.I. boom. I don’t quite mean the technology (video-generation software of this polish and reliability has already been available), or the sudden, extreme public reckoning with the quality of frontier models (even if this is a free, frictionless, consumer-marketed deepfakes product), though both aspects are important.

No, what I mean is: it has a feed. After years of establishing “the chatbot” as the dominant paradigm for interacting with generative A.I. models, OpenAI is trying something else--something very familiar. Sora is a social media app, in the same sense as TikTok or Instagram, and its design and existence suggests some interesting things about the future of OpenAI--and of the A.I. boom more broadly.

As an app, Sora feels remarkably like TikTok. You open it to a feed of videos, which can be set to show only videos from people you follow, or “latest,” or “For You”; any video you post to your profile page will also appear in others’ feeds. The “A.I.” component is that all the videos you see are generated with the Sora model by typing a prompt into a text box and waiting 5-10 minutes (or more, if you need to tweak the prompt and regenerate.)

In many ways this design is quite similar to Facebook’s new app “Vibes,” another video-generating app similarly organized around a feed through which you can consume A.I. videos made by others. What distinguishes Sora in particular is “cameos”: You can upload a short video of yourself to the app, and suddenly turn your likeness into a prompt-able, generate-able character--not just for yourself, but for any friends to whom you extend permission. If you’re followed by, say, New York Times ace tech reporter Mike Isaac, you can make videos of him “tearfully admitting to two police detectives that he gets all his ides from the newsletter ‘read max’ and that he’ll never live up to it”:

Or if I follow you, as I do New York magazine columnist John Herrman, you can prompt a video of me “making a series of exaggerated dreamworks/pixar/disney facial expressions for a TikTok video, influencer style, dancing”:

You can even (as Altman has done) allow anyone on the app to generate videos featuring your likeness--hence the video at the top of this post.

These cameos help explain Sora’s immediate viral success, especially when compared to Vibes: They’ve managed to enable the kind of simultaneous narcissism and sociality that once defined (and has since largely become absent from) social media. The feed is engaging because it shows you images of yourself and your friends. (Or, sometimes, of influencers who occupy friend-like spaces in one’s psychic universe.)

The other thing that explains Sora’s viral success, of course, is OpenAI’s longtime secret weapon: Loose guardrails and half-thought-through safety measures. The relative freedom that the Sora model offers users is part of what makes it so fun for small-scale socially contextualized videos--but also what makes it seem so alarming as a new burden on the eroding ecosystem of trust and verification. Sora won’t let you generate video of living humans or sexual scenarios, but it’s very easy to make videos of deceased public figures like Martin Luther King, Jr., or simply of wholly made up politically charged scenarios:

@drewharwellThe Sora AI disinfo nightmare is here Fake videos of real people are going viral and the piddly little watermark doesn't work #sora #sora2 #openai #aivideo #disinformation

Tiktok failed to load.

Enable 3rd party cookies or use another browser

The downside potential of this kind of generative-A.I. technology is such well-trod territory at this point that I almost hesitate to elaborate it for fear of boring readers: The ability of nearly anyone on the planet to create believable video of anything, on demand, for free, is alarming not simply because it might be misleading to naive viewers, but, worse, because it makes cynical viewers of us all, suspicious of any video and disbelieving of what was once gold-standard evidence of “what happened.” What can be done about this is difficult to say: Even if it were possible to remove Sora from the app store, similarly high-quality, freely accessible models already exist, and still have room to improve.

As it grows, too, Sora opens up the possibility of strange and unpredictable second-order effects as well: In the week since it’s launched, Sora’s main feed has increasingly become dominated by videos of the YouTuber/boxer/troll Jake Paul, who has freely allowed his likeness to be used by anyone on the platform. For a certain kind of influencer business model, it seems--one orientated around maximum exposure and total shamelessness--there is no reason not to allow yourself to become “victim” to thousands of deepfakes. You’re effectively outsourcing, for free, your main job of creating aggravating and polarizing content featuring you!

Of course, much of what I’m saying here about world with cheap and easy access to deepfake generation could and would be said if Sora were, like ChatGPT, or most of its video-generation rivals, solely a chatbot--a blank box into which you typed text and a screen on which output appears, all of it (supposedly) private, the app situated somewhere between a toy and a tool. What stands out here is that OpenAI has decided to release Sora packaged as something else: A social media platform.

Since its inception (or since its reconstitution as an unwieldy matryoshka of non-profit and P.B.C.) OpenAI has been spending an enormous amount of cash, and making an enormous number of commitments, without ever turning a profit. (Altman said this week that profitability is “not in my top 10 concerns.”) This is par for the course for Silicon Valley, where start-ups are given long runways to grow and gain users or customers before investors start clamoring for profits (or, at the very least, for an exit). And anyway, OpenAI has a fairly straightforward business proposition to look forward to: it can sell its A.I. services to businesses and consumers, a strategy that will make the company about $12.7 billion in revenue when all is said and done this year.

This is a fantastic amount of money, and in other industries, or for other companies, the quickly growing revenue and falling costs might be enough. But this is Silicon Valley, and ChatGPT is the fastest-growing app of all time. This week Altman said that it has 800 million weekly active users, which probably puts it in spitting distance of Facebook for size--and yet Meta brought in $164 billion last year.

In a strange sense, had ChatGPT grown more slowly, OpenAI might have trundled along as a straightforward subscription business with the expectation that eventually the revenue from its subscriptions would outpace the cost of providing its services. But in our universe, where Altman and his investors are seeing 800 million people hit the website every week--99.75 percent of whom aren’t paying for a subscription--different business models become attractive as well. Like, say, Meta’s: Why not build your model into a social feed, against which you can sell advertisements? (Certainly, Meta itself has had this same idea, hence “Vibes.”)

Earlier in this A.I. boom, L.L.M. chatbots were sometimes held up as marking a generational shift in Silicon Valley--a step beyond the megaplatforms that had come to define the software industry. Here at Read Max, we have long felt that large language models fit much more seamlessly into the lowest-common-denominator engagement-hunting that marks the social-media platform’s business model than is often admitted. At some point, watching the amount of engagement-seeking slop being created on ChatGPT for posting elsewhere, OpenAI must have begun to wonder: Why don’t we just make a place for them to post it here, so we can get a piece of that action?

A.I. as normal technology (derogatory)

If it works, then we might look back on Sora as the moment OpenAI settled in and allowed itself to be fully annexed by the social-platform sector--the A.I. boom ultimately less a regime change than the minor origin story for the latest entrant into small club of mega-platforms minting money from targeted advertising.

But… will it “work”? What makes social-media platforms like Instagram and TikTok so wildly successful is the extremely favorable unit economics: The marginal costs of hosting and serving videos is negligible, while the cost of creating the content is off-laid to users, so the platforms can grow effectively unimpeded. Sora, by contrast, internalizes the costs of creation, which are still significant (if improving): Every new video generated by a user costs the company more money.

There’s no easy way out of this dilemma. We’re a long way off from inference costs dropping so low that the marginal cost of generating videos is as insignificant as that of hosting or serving them. And, as the viral success of Sora over the more consumption-oriented “Vibes” suggests, the real attraction of this type of A.I.-video app seems to lie in the quality and versatility of what it can generate, not in the experience of raw-dogging a feed of slop.

In which case we might look back on Sora as pivotal in another direction--a failed gambit to turn OpenAI into the next Meta, rebuffed by the economic logic of the tech itself.