Sora 2 Bugs
I’ve been having an absolute blast with the recently released Sora 2, pushing its limits, testing edge cases, and generally seeing what kind of creative chaos I can conjure.
Yesterday, I had a fun idea to create a spoof video of @searls and myself presenting at a fictitious conference. You know, the kind of thing that would get a chuckle at the next tech meetup.
The Future of “JS”
Original Prompt: dmosher and searls are panelists at a tech conference on the future of JS. Searls thinks the conference is about him because his initials are JS but it’s actually about the future of JavaScript.
It turned out better than I expected! I just wish they would have let the sentence continue so Justin could get to the punch line about getting to his autobiography, alas… perhaps in the next remix. 😂
A Quick Primer on Sora’s Onboarding
For those who haven’t been through it yet, when you onboard to Sora, it has you perform a rather sci-fi ritual. You read a sequence of numbers out loud while it captures an audio fingerprint—essentially training the model to understand the unique characteristics of your voice so it can synthesize it in generated videos. It does the same dance with your face, capturing a “likeness fingerprint.”
So, What Happened?
I was attempting to remix my own video and hit an odd error after a few attempts. The resulting draft played normally at first—same conference setting, same general vibe. But then Justin’s voice came through reading numbers. Those numbers. The exact sequence from what I think was likely his onboarding session.
Sora had somehow embedded what I think was Justin’s audio fingerprint into my remix. I immediately tried to export the video for documentation, but the publish button was grayed out, share options threw errors, and the download link was mysteriously absent. With the evidence literally vanishing before my eyes, I did the only thing I could—fired up PhotoBooth.app
to capture this janky screen recording:
Shortly after I captured this, the draft mysteriously vanished from my dashboard entirely.
Remix Prompt: “Later in the panel discussion the two are talking about famed techno-prophet Gary Bernhardt’s prediction on the birth and death of JS (JavaScript) and @searls again has to be reminded the topic is not about him being born or dying”
Speculation
I’m only speculating here, but after spending some time thinking about the architecture, here’s my best guess at what went wrong:
Sora appears to be storing user ‘fingerprints’ (the audio and visual data from onboarding) in some kind of shared embedding space. When you remix videos multiple times in quick succession, especially when referencing other users, there seems to be a race condition or caching issue that can cause these embeddings to bleed into the generated content.
The fact that Sora immediately prevented any form of export suggests there’s some level of awareness built into the system about what constitutes protected data. But it seems like the safeguards kicked in after generation, not before—classic case of closing the barn door after the horse has bolted.
I’ll keep experimenting with Sora 2, because honestly, it’s incredible. Though next time I try to spoof a conference talk, I’ll make sure the only thing that gets leaked is Justin’s actual autobiography. Found any interesting edge cases yourself?