I've been watching these videos for about an hour now.
I really want to call this the "Suno moment" for AI video.
Prior to Sora 2, you had to prompt a lot of clips which you then edited together. You had to create a starting frame, maybe do some editing. Roll the dice a lot.
Veo 3 gave us the first glimpse of a complex ensemble clip with multiple actors talking in a typically social media or standup comedy fashion. But it was still just an ingredient for some larger composition, and it was missing a lot of the soul that a story with a beginning-middle-end structure has.
Sora 2 has some internal storytelling mechanic. I'm not sure what they did, but it understands narrative structure and puts videos into an arc. You see the characters change over the course of the video. They're not just animated Harry Potter portraits. They're alive. And they do things that change the world they're in.
Furthermore, Sora 2 has really good "taste" and "aesthetic", if that makes sense. It has good understanding of shot types, good compositions, good editing, good audio. It does music. It brings together so much complexity in choice and arranges them into a very good final output.
I'm actually quite blown away by this.
Just like Suno made AI music simple and easy - it handled lyrics, chorus, beat, medley, etc. - this model handles all of the ingredients of a 10 second video. It's stunning.
Sora 2 isn't the highest quality video model. It doesn't have the best animation. But it's the best content machine I've ever seen.
I can see this, it's extremely impressive from a technological standpoint, and I've already been caught by the first convincing fakes on Reddit (an army person giving an anti-Trump speech). But I'm also worried, as it's a super easy channel to create convincing fakes, mass produced 'content' for mass consumption, etc.
Now these things aren't new, fake videos / images go back decades if not a century. But they took some effort to make, whereas this technology makes it possible for it to take less effort than it took for me to write this comment.
Of course, it's always my choice; if I stop visiting Reddit and touch grass instead it really won't affect me directly.
Maybe some MAY end up in compiliations in ten years, much like Vines do today. But there will be a million times more tiktoks and a billion times more AI generated videos than there were vines, so if 0.01% of vines became memetic, the amount of AI generated ones will be infintesimal.
Content is all ephemeral on some time scale, but you can cache the near-term content to maximize the views and cut back on compute costs. Some model or human made it (the cost), it's trending (the value), so keep it around for a bit.
Everything has a relevancy and penetration decay curve.
The funny thing is, I think this law applied in the classical era (1950's, 1990's, etc.), we just weren't creating at scale to realize it.
Maybe it's just one dominant variable: novelty. I'd be curious to see how we might model this.
I've been watching these videos for about an hour now.
I really want to call this the "Suno moment" for AI video.
Prior to Sora 2, you had to prompt a lot of clips which you then edited together. You had to create a starting frame, maybe do some editing. Roll the dice a lot.
Veo 3 gave us the first glimpse of a complex ensemble clip with multiple actors talking in a typically social media or standup comedy fashion. But it was still just an ingredient for some larger composition, and it was missing a lot of the soul that a story with a beginning-middle-end structure has.
Sora 2 has some internal storytelling mechanic. I'm not sure what they did, but it understands narrative structure and puts videos into an arc. You see the characters change over the course of the video. They're not just animated Harry Potter portraits. They're alive. And they do things that change the world they're in.
Furthermore, Sora 2 has really good "taste" and "aesthetic", if that makes sense. It has good understanding of shot types, good compositions, good editing, good audio. It does music. It brings together so much complexity in choice and arranges them into a very good final output.
I'm actually quite blown away by this.
Just like Suno made AI music simple and easy - it handled lyrics, chorus, beat, medley, etc. - this model handles all of the ingredients of a 10 second video. It's stunning.
Sora 2 isn't the highest quality video model. It doesn't have the best animation. But it's the best content machine I've ever seen.
I can see this, it's extremely impressive from a technological standpoint, and I've already been caught by the first convincing fakes on Reddit (an army person giving an anti-Trump speech). But I'm also worried, as it's a super easy channel to create convincing fakes, mass produced 'content' for mass consumption, etc.
Now these things aren't new, fake videos / images go back decades if not a century. But they took some effort to make, whereas this technology makes it possible for it to take less effort than it took for me to write this comment.
Of course, it's always my choice; if I stop visiting Reddit and touch grass instead it really won't affect me directly.
Maybe some MAY end up in compiliations in ten years, much like Vines do today. But there will be a million times more tiktoks and a billion times more AI generated videos than there were vines, so if 0.01% of vines became memetic, the amount of AI generated ones will be infintesimal.
Content is all ephemeral on some time scale, but you can cache the near-term content to maximize the views and cut back on compute costs. Some model or human made it (the cost), it's trending (the value), so keep it around for a bit.
Everything has a relevancy and penetration decay curve.
The funny thing is, I think this law applied in the classical era (1950's, 1990's, etc.), we just weren't creating at scale to realize it.
Maybe it's just one dominant variable: novelty. I'd be curious to see how we might model this.