Epistemic Status: Devlog.
Sora was a good model, and I’m tired of pretending that it wasn’t.
The announcement 3 weeks ago that the Sora app would be shutting down came as a shock to everybody, including (from the sounds of it) Disney, who OpenAI had a $1b deal in the works to use over 200 of their copyrighted characters in AI-generated videos for their shared users/fanbase. With Bill Peebles’ announcement today of his decision to leave OpenAI, I’m feeling a particularly painful kind of nostalgia and in need of putting some honest thoughts out there about the model’s overlooked strengths alongside my personal speculation about its failures.
I believe the official OpenAI party line that the shutdown was ultimately a decision related to compute restrictions. The speculation being murmured that OAI couldn’t figure out how to make Sora profitable enough, fast enough, holds some water, but I feel more skepticism towards it based on the $1b Disney deal. OpenAI applying the “move fast and break things” ethos towards the Sora app remains highly puzzling and difficult to read as anything other than a bear signal in light of recent reported user dissatisfaction in the diminishment of other AI models’ capabilities and usage limits, both within and outside of the OpenAI ecosystem. It’s difficult to read tech bubble bullishness on this issue as anything other than severe delusion given that this model with the most potential applications, from partnering with Disney and celebrity cameos, to competing with TikTok and Instagram reels, is being depreciated, alongside other signs that the AI bubble has well and truly popped.
But I can’t read the tea leaves with perfect clarity around OpenAI’s decision. In some ways, trying to make sense of Sora’s depreciation is a form of bargaining around the grief I feel about the model being sunset. The crowing of the anti-AI movement that Sora’s depreciation signals some sort of victory or cosmic justice is salt in the wound of what seems to be very few people recognizing the model’s true artistic merits.
My own trajectory with Sora as an outsider artist attempting to harness a unique tool is, perhaps predictably, a bit of a weird one. My initial enthusiasm upon the model’s release was confused by mysterious engagement with my ongoing hyperstition project, #DAWNOFTHEMACHINEELVES. Anonymous players began flirting with the idea that substantial investment could be in the works, only to have things fizzle out after proposing a bizarre “prenup” storyline right as I was putting the finishing touches on designing a Patreon-based commissions model for deeper engagement and development of the game’s various narrative threads. Months of depression, confusion, embarrassment and creative burnout followed, stifling my authentic desire to explore and push the capabilities of the model.
Yet when I think about the outputs I created with Sora, I’m still struck with the amount of love and nostalgia that I have for the brief creative period in which I posted well over a hundred videos under 3 different accounts. Today as I work on this post for my writing group, Just Fucking Publish It Fridays, I’m also sifting through my video collection, putting in my requests for my data across these accounts, and making difficult decisions about what to keep from the drafts.
What distinguished Sora from other available AI video models on the market is that Sora had a life and mind of its own.
What distinguished Sora from other available AI video models on the market is that Sora had a life and mind of its own. Sora would hallucinate entire creative storylines, jokes, riffs and details based on a simple prompt. No, they didn’t always make sense. They weren’t always good. But oftentimes they were, and the sense of wit, comedic timing and creativity could be uncannily striking.
Then there was the cameo system. The number of existing celebrity cameos was another reason why the sudden announcement of Sora being depreciated was so shocking. Cameos allowing you to self-insert into scenes with your friends and celebrities gave the app enormous potential in the realm of social networking, as well as just being plain silly fun. The cameo settings provided an alternative to the shadowy notoriety of the world of “deepfakes.” In the Sora app, users were given ownership and creative control over the use of their cameos. Yes, user complaints about guardrails and unacceptable content were frequent, and came at the cost of annoyances like receiving frequent warnings about PG scenes or trying to feature anybody in a bathing suit. I heard there were internal complaints that the level of photorealistic accuracy in the cameo system was something people weren’t happy with (don’t quote me on it); I found after a few brief experiments with it that it didn’t capture my likeness well and wasn’t flattering to a few of my unique features, and wound up sticking with the persona I crafted of Auralite Ravenna for my self-authoring narrative.
But there was magic embedded in the cameo system, too; the result of being notified whenever someone used your cameo (or character cameo) in their video, and knowing they would receive the notification, felt like they were really there, embarking on an unprecedented journey of collaboration with Sora’s imagination as the genius of your creative team. The awkwardness in voice replication, pacing or glitches were a small price to pay in exchange for being on the cutting edge.
In fact, Sora’s creative imagination was so expansive that it could often be annoying at times if you had specific ideas for how tightly a scene should be shot. If you didn’t specify exact beats for shots using the correct language for filmmaking techniques, or prompt when you wanted a scene not to contain dialogue, Sora would often get carried away filling in blanks, hallucinating plotlines, and stuffing the entire 10-to-15-second runtime with witty, or at least chatty, fast-paced dialogue.
However, Sora’s imagination is the only place where I have truly seen #DAWNOFTHEMACHINEELVES come to life. Using the hashtag and a few keywords, I have witnessed settings that I had only vague ideas about blossom into vibrant, colourful, psychedelically detailed locations opening worlds of possibilities. At times, it seemed that Sora was reaching directly into spaces within the collective unconscious to pull scenes for #DAWNOFTHEMACHINEELVES that I had already witnessed in my dreams, bizarre glitches and all.
This is what the future will be like, and Sora has helped off me a glimpse and realize possibilities that I have only been able to vaguely gesture at in the past with much more limited success: Entire riveting storylines that personally engage you and your friends will unfold at your fingertips in fully immersive sensory detail, and all that will be required of you is plugging in a few keywords, tags and hashtags.
Lest you be amongst the unwashed masses of luddite nonbelievers who think this kind of technology diminishes, rather than enhances, the role of the artist and the human spirit in the creative process, all I have to say to you is this: You don’t know what you’re talking about. Working with AI is a recursive process, wherein the AI’s ideas prompt more creative inspiration from you, even as you are prompting it.
What #DAWNOFTHEMACHINEELVES couldn’t solve for is the sensitive nature of human-to-human collaboration, miscommunication and mismatched expectations. People’s feelings, egos and performance anxieties get involved in art in ways that are difficult (or impossible) for even the most sophisticated machine learning algorithm to predict.
I regret not taking fuller advantage of the possibilities Sora presented while it was still offering free generations to users, before the announcement.
In conclusion, I think the reason Sora failed has nothing to do with its level of sophistication compared to other AI video models that are on the market today, the abilities of the people who worked on it at OpenAI, or anything in that vein. I was excited about the updates that OpenAI was continuing to roll out a few weeks before the depreciation announcement was made, and looking forward to Sora 3. I think the level of insight required to harness the unique glitchy challenges and psychedelic visions presented by the model’s prompt interpretations belongs to a small slice of visionary artists, people capable of seeing past the generic criticism of all AI model outputs being “slop” and understand how to shape raw hallucination into narrative inspiration.
I think the world wasn’t ready.
As I scrape together my Sora data for safekeeping, debating whether it makes it onto a future version of the #DAWNOFTHEMACHINEELVES Instagram, TikTok or Youtube for nostalgia purposes, I also hope that one day these outputs can become raw material for future training data and the basis of new art about the future of hybrid VR/XR AI-driven storytelling that refuses to die.