Even if AI can generate convincing video like this, I predict that AI detection will also advance at a similar (if not more advanced) rate so that the question "Is this real" will not be one that we would ask in 2-4 years.

@DanielParker any basis or intuitions for that claim? The AI detection for AI generated text has clearly not kept up

@firstuserhere It's true that detection of AI generated content has lagged behind the generation of AI content, but it's a very new and important field that a lot of people are working on, so I expect it to catch up. At the absolute least, I expect that watermarks will be included by most commonly used video generating algorithms and that Twitter/YouTube (by 2026) will automatically flag such videos in a way that will be easy for users to recognize. Black-market/custom software will be able to generate videos without watermarks, of course, and it's plausible/likely that humans would have trouble distinguishing AI generated content from real videos, but even if watermarks are the only protection, I expect that actual instances of confusion will be relatively rare in 2026 so that most people will not think to wonder "was this generated by AI?"

@DanielParker while we're at it, did you notice that dalle-3 images are no longer watermarked the way dalle or dalle-2 images were?

@firstuserhere Just because you can't see the watermark doesn't mean that it's not there. Bing Image Creator (which uses DALL-E 3) definitely does include an invisible-to-humans digital watermark. I haven't looked into the details, but my guess is that this is a security feature which is inherent to DALL-E 3, not something that was added by Bing. It would probably be relatively easy to run an image (or a video) through a filter to remove such a watermark, but people posting videos would not do that unless they were actively trying to deceive their audience (i.e. this would happen only in a small minority of cases). As long as the watermark is there by default, Twitter/YouTube should be able to trivially flag most AI generated content.

Sadly, there will be edge cases like deliberate disinformation campaigns or people who attempt to fraudulently pass of AI content as their own work, but those should be rare. Hopefully, the white hats will be able to develop techniques to detect even those cases, but even if they don't, the provenance of the vast majority of content should be identifiable at a glance by 2026.

@DanielParker /Ernie/openai-images-have-a-useful-and-har

/Ernie/midjourney-images-can-be-effectivel

@Ernie For the purposes of my argument, an easy to remove watermark (even just metadata associated with the image) is fine. As long as most AI generated content has the watermark, social media can be easily designed to prevent most people from being confused by its provenance in most cases.

The edge cases where people could still be confused by malicious actors are important and smart people are working on the problem, but that's not what Yudkowsky's tweet was about. If you see a pretty video about moths on social media, there is no reason for a malicious actor to lie about its provenance, so (by 2026) I expect that most people will rarely be confused about things like that.

@DanielParker interesting argument. I remember an old ssc post about the prevalence of lying and how that applies to reddit. I recall the conclusion was that even though pathological +high IQ + basically evil is a rare combination, people like that can produce a lot of very viral content. So his conclusion is that surprisingly, we should believe that the majority of viral, disagreeable, super distressing stuff you find online to be produced by people like that, nevertheless.

Not saying it has any impact on this debate. I'm just saying there may be a situation where 97% of content shared still is legit, but because of amplification, when you consider it by reach and view rate, the more fake stuff could be much more common.

@DanielParker I would bet against that

@Lorxus Agreed. It's joever.

https://twitter.com/nickfloats/status/1689105620652199936?s=20

High speed cam footage in question shows:

• complex and surprising airfoil control and flexing

• distinctive high-contrast lighting needed to hit 6k FPS (especially on any specular reflections)

• plenty of unaesthetic uncontrolled takeoffs/abdomen strikes/proleg flopping

All of these are consistent with a high effort real-world measurement of weird ground truth. Neither de novo extrapolation from basic principles, nor interpolation from existing data would have these characteristics.

Unless it's a "Kubrick staging the moon landing" scenario, but why try for realism that hard?

In short - why would your first thought on what's clearly raw experimental data be "this is fake"?

@alexkropivny consider that this very video can be included in the training dataset, that complex physics simulations and articulated figure animations already exist, and that photorealistic CGI in general is already effectively believable

@cloudprism A reincarnation of Stanley Kubrick monopolizing HPC clusters to do FEA on moth legs and ray trace accurate floodlight reflections on metamaterials is feasible. Why would it be likely?

@alexkropivny I may be wrong but I think the point in question is about whether it would be plausibly unbelievable to the average person, not whether the ultra-high-resolution version would be detectibly fake to an expert of its subject

Anybody have evidence of good AI-created video? I just don’t see how this is viewed as so inevitable (88% really?) It could be pretty difficult IMO. Coherent video needs a lot more intelligence than images

@DylanSlagh yes but scaling laws are a thing including on video and video generation has been getting better even if it's still not that great(see the Runway stuff) and there's tons of video data out there to train on, so it's mostly a question of waiting untill people have enough compute to just train really big video models, and for that 3 years is a long time.

Basically video is pretty difficult, and generating images is pretty difficult, as is predicting the next token.

We have been advancing pretty fast on those anyway because we can in fact just throw increasingly more orders of magnitude of compute at it, and have lots of peoplke working on algoritmic improvements and there's lots of orders of magnitude of room for improvement there.

@VictorLevoso

@firstuserhere @VictorLevoso

Will this resolve yes if it can't give say 15 examples

@BenQ read the description. Market is about authors subjective opinion whether this tweet was a good take. (Ex: AI-generated video really is that good). The "15" bit isn't the important part.

Funny how this is aligned perfectly with the 2024 version of this market rn. Should be higher

Why is there a spike? I cannot find anything even remotely good enough for the spike to be justified.

@OliviaSaber There’s been a silent back and forth between people who think that this should be around 65% and people who think it’s more like 90%, so I think the current value of this market tells you more about who has more cash at any given time than anything underlying. It’ll eventually converge one way or another, but the time is still a while out.

@Gabrielle That sounds just right, well written. I will put up a big YES limit order at 72% so if someone with cash wants to take it, feel free

@OliviaSaber Buttockscocktoasten always puts all his money into yes rationalussy. Pretty inelastic demand for yes shares.

@JonathanRay Who is Buttockscocktoasten and why does this person matter here? What does Rationalussy mean? I - feel so confused

@OliviaSaber I think that comment was meant for a different market

@Gabrielle Can you please explain how the position works? I read that you probably think its a 65% currently, but also expect it to increase or decrease over time. If you already put this at 65%, then I don't think you'll predict it to decrease over time. If you only predict it to increase over time, then why do you hold 24,000 shares of NO?

@OliviaSaber The reason I bought NO is because I actually think this market is currently supposed to be even lower at around 50%, and that AI - enthusiasts are people with a lot of mana, and have ability to bet it up high. Over time, I expect this market to come to 65% but go back down after that once it is clear that amount of data is still 2 OOM away; the efficiency is not quite there yet, and probably will barely be available in time for resolution

@OliviaSaber I think that the current "correct" probability is somewhere around 65%, so when the market is higher than that number I buy NO shares. The market has been consistently in the 70%-85% range, so I've bought a lot of NO shares. "A lot" is relative though; my portfolio is Ṁ130,000 and I've only put in Ṁ6,866, at an average of 72%, so it's not like I'm really betting hard. (If my math is correct, the Kelly criterion suggests that I should bet Ṁ22000, which would triple my investment.)

If the market were lower than 65% though, I would buy YES shares, but no one else is betting it that low. I guess there are too many AI enthusiasts and not enough AI pessimists watching this market, or maybe I'm just wrong.

When I say that it'll eventually converge, I mean that by the time the market closes it should be clear whether or not the tweet did hold up, so the market will have gone to either near 0% or near 100%. I'm currently hoping that it goes to 0%, since I have a bunch of NO shares, but it's possible that a bunch of AI pessimists will start betting in the market, and I'll sell them all of my NO shares and buy YES shares instead.

My guess of 65% is a rough guess, but it's based on seeing that we're still pretty far away from the goal and videos are difficult and require lots of training data, but also seeing that people are using a lot of fancy tricks to achieve video without training new models. My odds will go up if I see lots of impressive AI videos, and go down if I continue to just see the lackluster attempts at videos that don't get anywhere near the goal.

@Gabrielle At least it will converge eventually. Unlike the AI xRisk markets 😂😂