I’ve spent the last week or so experimenting with Sora 2. Here’s what I’ve learned.
Bias towards cameos:
The big breakthrough of Sora is how quickly and naturally it can make “deep fakes” of its users. When you open the app, you record yourself saying 3 different numbers, turn your head in 2 directions, and bam: they’ve got you. Now you can reference your “cameo” in any clip to start directing your digital doppelganger to do whatever you want, and you can reference other users’ cameos as long as they’ve given you permission.
While Sora is technically able to accept prompts without cameos, it’s become clear that the system wants cameos. The “top” tab is full of familiar faces, from Sam Altman to Jake Paul to (increasingly) historical figures like MLK and JFK.
Regarding why OpenAI would invite all of this copyright vulnerability, I think they are placing a small bet (for them at least) that Sora takes off as its own social platform, and having a popular cameo becomes a valuable brand extension channel. Seeing Mark Cuban and Sam Altman 50+ times a day has to be good for Sam’s name recognition. Now extrapolate that to the next influencer, marvel character, brand mascot.
Evidence: Look how easily Sora casts me as a fashion influencer in a “Get Ready With Me” video:
Slot Machine for Memes
Speaking of the “top” tab, what you find browsing Sora is not so much “unfettered creativity” as “variations on a theme.” It’s hard to create a clip from nothing that is engaging and not immediately recognizable as slop.
When creating videos, it’s a bit like playing a slot machine that takes 5-10 minutes to process. The results can be batshit crazy, miss the brief entirely, and oftentimes just kind of boring. Then you have another roll of the dice by posting the thing, and waiting to see if anyone engages with it.
When a clip does break through, other users “remix” it, which is a way for them to alter the prompt while keeping the original video as reference material. This is basically a new version of those old meme generator sites where you could change the text on top of familiar memes.
1 idea per clip
The system can’t really deal with a prompt that has more than one central character or idea. If your prompt has more than one “and then” Sora struggles with complex sequential prompts.
Surrealism > naturalism
This might just be a personal preference, but I prefer my AI videos absurd, dreamlike, knowingly unrealistic as opposed to deep fakes of already mid content formats. And that’s also playing out in the top tab, which has everything from a pile of mac & cheese getting arrested to teenagers ding dong ditching while levitating.
My challenge: Make a 30 second ad for Runpoint
Based on all of those lessons learned, I set out to see if I could make an ad for Runpoint based on this value prop Sam and I have been batting around based on the idea that so many AI initiatives get stuck in the pilot phase: “Escape Pilot Purgatory”
The Surrealism > Realism rules pointed me towards an idea inspired by Beetlejuice, where AI pilots were stuck in a bureaucratic nightmare of a waiting room.
Bias towards cameos meant I’d have to cast myself and Sam. Every other original character idea flopped, especially after trying to persist that character in multiple clips.
The Slot machine nature of the platform meant that the end result came out completely unhinged and creepy, but I gotta say I still kind of like it.
So without further ado, I give you Pilot Purgatory: