On August 26th, Eliezer tweeted
(https://twitter.com/ESYudkowsky/status/1563282607315382273):
In 2-4 years, if we're still alive, anytime you see a video this beautiful, your first thought will be to wonder whether it's real or if the AI's prompt was "beautiful video of 15 different moth species flapping their wings, professional photography, 8k, trending on Twitter".
Will this tweet hold up? (The part about AI video generation, not about whether we'll all be dead in 2-4 years.) Giving max date range to be generous.
This market resolves YES if at close (end of 2026) my subjective perception is that this was a good take--e.g., AI-generated video really is that good--and NO if it seems like Eliezer was importantly wrong about something, e.g., AI-generated video still sucks, or still couldn't be the cause for serious doubt about whether some random moth footage was made with a camera or not.
I reserve the right to resolve to an early YES if it turns out Eliezer was obviously correct before the close date. I won't dock points if he ends up having been too conservative, e.g., a new model comes out in 6 months with perfect video generation capabilities.
I guess this market resolves N/A if we all die, but, well, y'know.
Betting policy: I will not bet in this market (any more than I already have, and I've long sold all my shares).
1,000People are also trading
This question is way too high because of a resolution detail people seem to be overlooking. I pointed it out around a year ago:
As I understand it, this market only resolves YES if there's realistic video generation AND no mechanism in place to reliably tell apart AI-generated from real video on Twitter.
I think that conditioning on realistic video, there's like at least a 20% chance civilization is forced to find a way to mark videos on Twitter as either real or AI-generated, so this question should be trading at most around 75%.
This question is not directly measuring AI video generation. Instead, it's measuring how much doubt people will feel when they look at a video on twitter. And there are ways to make the amount of doubt low!
If this is not true, @journcy or @mods should elaborate.
@Nikola It's crazy how much free mana from AI over-optimists there is constantly, including from just not properly reading the resolution criteria. And I say this as someone with fairly short timelines. But I can't invest in correctional pessimism properly, because I'm trying to free up mana to bet in other over-optimistic markets. XD
@Nikola I suspect this is not true for two reasons:
This did not happen with photoshop, even when sites like Pinterest were flooded with photos that claimed to be real, and were not.
In order for it to impact this market, the indicator would have to be crystal clear -- it would have to proceed/accompany your "first thought" about the video. I'm not sure that Community Notes clears this bar, and we are not reliably seeing even Community Notes on AI produced media.
This +2 years pretty much clinches it for me
Even if AI can make something like that moth video this shouldn't resolve YES... there are many, many types of videos that are really popular online that AI can't make. Could use some clarification from @journcy on how they're thinking about it (i.e. is it more about copying the moth video or copying "most" types of online video) but if we take their criteria seriously YES is way overvalued
@ChrisRigas I think their criteria is very clear:
> This market resolves YES if at close (end of 2026) my subjective perception is that this was a good take--e.g., AI-generated video really is that good--and NO if it seems like Eliezer was importantly wrong about something, e.g., AI-generated video still sucks, or still couldn't be the cause for serious doubt about whether some random moth footage was made with a camera or not.
It's clearly not about copying the moth video specifically, and not about being able to copy literally every kind of video you'd find online, but about AI being able to create short videos of the same category as the moth video, in a way that is hard to distinguish from real videos even with a closer look, such that there's a reasonable doubt about whether any particular video like that is AI generated.
The key phrase here is "AI generated video really is that good". If that condition is met, then this question should resolve yes, in my interpretation. And I think Veo 2 is there - I'm confident that once Veo 2 is released to the public, there will be no doubt that this question will resolve YES.
@MetallicDragon From their comment:
To me, this tweet will have stood the test of time if it is in some sense embedded in the zeitgeist that a level of skepticism is warranted for essentially any video you see online. (With obvious caveats about things you know existed prior to the 2020s, etc.)
how do you get from there to "clearly not about being able to copy literally every kind of video you'd find online"?
it's not about there being a couple edge cases, but like the majority of types of online video right now could not remotely be made with AI
@ChrisRigas That quote does seem to conflict with the text I quoted. I hold the text of original question as holding more weight than that comment. There is zero indication in the original question that AI would need to be able to generate any arbitrary internet video. Rather, the way I read it, it only needs to apply to videos in roughly the same category as the moth video, in a way that causes doubt in a significant portion of the population when they see such a video.
I think most people would read the question the same way. If the asker really did mean that it would need to apply to any arbitrary internet video (including long-form coherent scripted content with matching audio), that's a hugely important detail to be left out of the original question and only mention it in a comment.
@Shai agreed. there is now a region of substantial uncertainty where many Veo 2 videos seem like they could be real; however, the most high-quality real videos such as this one are still beyond the capabilities of SotA video generation
@Shai I think you know too much about moths? Look at the flailing legs, the floofy bodies, the weird yellow one doing a faceplant. This does look a lot like a series of AI generated clips. It doesn't look much like what most people think moths look like.
@Duncn Looking closely at it, the wing beats don't particularly map to changes in acceleration. Clearly not made by a entity that has a strong grasp of physics.
I have changed my mind on this market after holding out for a long time. Mostly in response to veo 2.
There's a bunch of subjectivity in who is "you" and how strictly to interpret "anytime". But we're getting to the point where suitably curated AI generated videos pass a casual inspection. And the kind of thing you're liable to come across on the internet will be such curated videos almost by definition.
Gotta hand it to Eliezer, in his tweet he predicted very well the kind of thing that AI has turned out to be good at, namely short form visual bling with relatively low consistency / coherency requirements.
I still think that generating high-quality videos to an arbitrary prompt is very hard. Especially if (1) the subject is specific, (2) the result is one-shot rather than curated, (3) the time is longer than a few seconds.
I still think that generating these videos will be expensive for the foreseeable future. Sora's current high pricing and low availability* points at something fundamental. I think veo 2 will be roughly similar.
* for future reference, Sora currently gives you 50 low-res videos per month for $20, with restricted sign ups.
This is the best I could do on Sora for "a highlight of an NFL play where an interception is returned for a touchdown".
Again, don't think we're anywhere close to realism on this.
https://sora.com/g/gen_01jf09c5atee8t7h2s7641nvph
Not sure I see how people are so confident on this.
From @journcy
To me, this tweet will have stood the test of time if it is in some sense embedded in the zeitgeist that a level of skepticism is warranted for essentially any video you see online. (With obvious caveats about things you know existed prior to the 2020s, etc.)
I don't think Sora are others are close to generating many types of common video with any type of realism. Take this example: "a highlight of an NFL play where an interception is returned for a touchdown". I don't have access to Sora but I really doubt it can get the motions of all 22 players right, the physics of the ball being thrown, the quick changes in direction, etc. with any level of detail right. Possible it will get there but it seems to me that would be part of the resolution criteria and we're not particular close...
I'm already at the point where i double check to make sure short political vidoes arent AI
(How do I include my bet as part of my comment?)