A.I Just Changed Everything. We are In Trouble.

So, ladies and gentlemen, we actually have a problem on our hands. This was an issue that I didn't talk about before because I didn't believe that the capabilities had surpassed the threshold needed for it to become a widespread issue. I do believe as someone who's, you know, been paying attention to the AI community and all of the things that go on, this is about to become a bigger issue than people do realize. And this is of course, the issue of super photorealistic AI image. Now people have spoken about this in the past for quite some time, but when they spoke about it, they didn't really understand how photos are realistic and what kind of issues actually drive problems. Now, one of the things that many people spoke about before was the fact that AI images would just flood the Internet and they were going to be super hyper realistic. Now, when you take a look at these images right here, whilst, yes, I would argue that these images are super realistic and show them to anyone three years ago or even two years ago, they would have said, wow, this is absolutely incredible. This is hyper realism. And now we will never know what is real from what is AI generated. Now I could agree with you to some extent, but I would also disagree with you and let me explain why I believe that midjourney wasn't realism, arguably photorealism, because there's two types of realism and, and it was only in the recent couple of days we entered the problematic kind of realism where you're truly not going to know what is AI generated or not. You see Mid Journey in its early days, and I say early days and that just essentially means, you know, six months ago, three months ago, that was the time when we had these kinds of images, okay, it was version five. And these kinds of images were the images that were realistic to the point where we were like, okay, they, this is pretty crazy that we now are not going to know if these professional images are AI generated. But the problem is, is that these images, okay, and I'm putting these on screen because these look commercially ready, like they just have that professional look to them. And the point is, is that, you know, the average person can't get their hands on these kinds of images. You're going to need a film camera, take a screenshot from a video. You're going to need like an actual photography camera to, to actually get these kinds of images. And this is why I didn't talk about this, because these kinds of images being on your feed, they just look like basic corporate images that you would see anyways, that there are literally millions of on like, Shutterstock or any of those basic websites. But this is the problem, okay, because recently there was an open source tool that was released that entered the threshold of Uncanny Valley. So when you take a look at these images, you're going to be like, what do you mean? These images aren't realistic? They're super realistic. Okay, Compare this realism that you're currently seeing on screen right now and compare it to this image right here. Those images are two different kinds of images. On this kind of image right here, this is what you call, like, you know, I guess you could say, you know, professional realism, where it looks as if, you know, a studio took it. But this kind of realism, okay, this realism where it looks as if an actual person took this image in the comfort of their own home. I'm not sure what kind of realism you want to call it. I'm not sure if there's a term that I'm not using, but this kind of realism is genuinely problematic because this kind of image is one that you truly cannot discern whether or not this image was AI generated or the image was created by an actual person who exists in reality. You can see here that there's also this image right here. And I'm going to show you guys some even more shocking examples and that just completely, you know, spin things on the head. But these images don't look entirely professional. You could argue that, yes, there is a little bit of softness on the cheeks and of course, on the skin, but that's only going to matter if you're someone like me who spent an unnormal amount of time looking at AI generated images. This is why I now think that we have entered the area where this genuinely becomes problematic. And a, as someone who enjoys AI development, I do wonder to the point that we've now released this kind of open source tool where we can get images that look like this. This is the point where you start to truly wonder what kind of things are going to be realistic versus what things are going to be AI generated. That kind of question being there is a genuine concern now and a genuine question now. If you're questioning, is it just digital photos from the early 2000s that you can make this. Okay, let's just write off that genre of photo. That's not the issue. The issue is that it can also do images that look like they're on the iPhone. Take a look at this. You can see that these are images that are made with Flux 1.0 Pro. And these are titled boring Snapchat photo circa 2015. I'm going to show you some of my examples in a moment because I almost didn't believe it when I saw it. I spoke about it in another video. But trust me when I say that these look completely hyper realistic. I mean if I was browsing social media and I saw this image, there is no indication that tells me this is AI generated at all. Previously, like I said before, if it was this image on social media you might just think, okay, it's just a corporate image or yeah, it's just a mid journey one, but this literally looks like someone took this on their iPhone 10 minutes ago. And there is no indication that this is an AI generated woman. Which brings a remarkable problem now into the space because we now have this issue of believing whether or not what we see online now other examples of this further exacerbate the kind of issue because this image right here just looks far too realistic. If we look at the color, the tone, and I was even talking know to someone on Twitter about this, stating that there are complete differences in realism. I would argue that this kind of realism right here is much more worse than this kind of realism right here. Because these images are images that look like they were set up. It looks like the lighting's been done, it looks like they've been edited in post. But this image, if someone was to send me this image, this is the kind of image that you can use to fool someone into doing a variety of different things. You could potentially get money from someone, you could potentially pretend you're in danger. I mean there's a billion different ways that people are going to use these in adverse ways for society. And I'm struggling to think about the actual benefits of these kinds of images. Yes, you can use open source tools for research and of course there's always the argument of the fact that yes, open source tools foster a great ecosystem, but I am struggling to find a good use for, for photorealistic images that allow people to create images that are ones like these. Now of course you could argue that this is just a byproduct of the training data. Maybe they had a few images of the training data and with careful prompt engineering you get images that look just like this. And I do have to be honest. These images, if I'm being completely honest, are somewhat cherry picked from the data, but it doesn't really matter. The fact is, is that if you told me three years ago this image is completely AI generated and, and this image right here is also AI generated, I would have said get out of here. That is completely impossible. You have no idea what you're talking about. AI isn't going to evolve that fast. But here we are now and we are in this place where these kinds of images are ones that would even fool me. And I look at AI generated images all the time. So I mean, it's going to become definitely harder to kind of realize what is AI generated versus what isn't. We've also got this image that was floating around as well and it has that truly stunning, hyper realistic look to where you really don't tell that this is AI generated at all. And like I said before, midjourney had that really, really smooth and professional look, whereas this one literally looks like it was taken on an iPhone. And I wouldn't have questioned this image at all, which, you know, makes me even wonder, am I going to have to now look at every single badge of on someone's shirt to see if the text makes sense? And I mean, one thing we do know, and this is the biggest thing, is that this is the worst it will ever be. This is the worst images will ever be. And I mean, if we take a look at the rate of development, it was only two years ago that we had literally avocados on a chair. And now we've got AI image generation to the point where it looks like it was just taken two seconds ago. Now there's a lot of things being discussed here and this video isn't all doom and gloom because as you know, the technology evolves, so do the detection. There was a detection method, so someone made with Claude, they said hack to tell if an image is AI generated, bump up the saturation and look at the microphone patches in the teeth. I wrote the code completely in Claude and it's a public artifact that you can try. Now I think that this is, you know, remarkable in terms of the fact that people can easily create softwares and apps like Claude. And I think this thing is quite good. But this also doesn't work every single time. And the fact is we're going to have to make sure that we're developing programs to sufficiently track whether or not images are AI generated as quickly as we're developing the tools to generate those AI generated images. You can see right here the difference between the saturation and an image that wasn't AI generated versus one that wasn't. But unfortunately, later on in the thread it did show an example of, of the fact that this doesn't work across all images. Now this brings even more credence, even more credence to the Dead Internet theory. And the dead Internet theory is the conspiracy that basically says that the online Internet is now consisting of mainly bot activity and automatically generated content manipulated by algorithmic curation to intentionally manipulate the population and minimize human activity. I've got to be honest, I think bots are going to be a huge problem in the future and that in 10, 20 years from now, now the Dead Internet theory is going to become a larger issue. I do think that there are going to be different ways, maybe through the blockchain ways for verification. Of course, if you already know, Sam Altman has, you know, spoken about Worldcoin. But before we dive into ways to actually fix and remedy the issue, take a look at this. Because this was why I made this video, because spoke about the images first. Sure. Images are one thing. Okay, let's say we decide to remove images that are things we believe. Boom. Just like that, images are gone from human consciousness. We now no longer trust images of things we see. What about video? Well, video is here too. If you haven't seen this demo, I'm sure many of you have seen it before. But this is not Elon Musk. And I've got to be honest, if I was to see this person livestreaming me before I knew that the software existed, I, I would have genuinely believed that this is Elon Musk. And I'm someone who's relatively skeptical to all sorts of technologies and videos and things that I see online because I always believe that you shouldn't just believe immediately what you see online. You should always fact check things because there are so much of things that aren't true. And one of the things that happens online is that we just read the title of things and we take that at face value. But this isn't Elon Musk. This is someone using an open source software and that they're using with their webcam. I mean, let's say, for example, that this person was using this adversariously. Let's say that they wanted to defraud some investors. Let's say that someone wanted to sign a deal for their company and they were like, look, we've got Elon Musk on the webcam and we're able to, you know, talk about this product or whatever, yada, yada yada. You can imagine investors throwing millions of dollars. And this is just one example. Imagine someone calling you, saying invest in my company, yada yada yada. And the thing is, is that this was just from a single image. Now, I don't know about you guys, I Don't have any images of my face online. But for those of you that do, you have to start to be concerned about, okay, I have HD images of my face online. Someone can take this, take my likeness and get a webcam and completely photorealistically imitate me without my consent. And the reason I made this video is because this entire demo is the point to where I was like, okay, we've just hit that threshold where things are about to start to get crazy. So this video is a warning to those of you in the AI space because I know some of you guys like the tools, you like the tech, you like the news. But you have to remember that this tool is impacting everything that is pretty much digital. And in doing so, there are certain use cases that whilst, yes, I mean, I'm not sure what this would be good for, there are certain use cases that you have to be aware about so you can be like, okay, maybe I won't trust the webcam of, you know, the certain person doing this, or maybe I won't trust this video of this person doing this. Because I now know that open source tools exist to where we can get images that are completely photorealistic. And of course videos, and not only videos, but live videos to where live streaming immediately onto someone's face with lighting and angles and, you know, wrinkles and all of that good stuff is completely photorealistic. Now like I said before, this is not closed source. It's not like Microsoft's VASA one where they had that technology and they were like, look, we are not open sourcing this technology. We are not releasing this because we know that this is going to be, you know, bad for society. This is open source. And it's open source with the, you know, you know, with the Cliff notes basically saying that users of this software are expected to use this software for responsibly while abiding the local law. And you know, we can all know that the average person is probably not going to abide by the local law. Sure, you might use this to troll your friend that Elon Musk's on the live stream. But I'm guessing that there is the large potential that this is going to be having some sort of misuse. Of course it said if a face of a real person is being used, users are suggested to get. It doesn't even say users are expected. It says users are suggested to get to get consent from the concerned person and clearly mentioned that it is a deep fake when posting content online. And it says developers of the software will not be responsible for actions of end users. They're basically saying, like, look, we made this, but whatever happens, happens. It's now in your hand. And like I said before, these kinds of technologies, you know, they're only going to get better. This is Microsoft's Vaasa one where they're just taking a driving and I guess you could argue that this one doesn't look as realistic, but I've seen instances where people manage to lower down the resolution and you really can't tell. But Microsoft said, you know, hey, we're not releasing this. We know that this is going to have some adverse effects. But here's the thing, okay? You know, individuals, institutions, research organizations, they catch up, they open source this stuff and eventually we're going to be leading to a society where it's very hard to discover what is true from what is not. Now, of course, there was also this image floating around to which someone added video. And I would say that, yes, this looks a bit AI generated, but I've put here, remember, it's only going to get better, which is the big issue. Okay, it's only going to get better. Now, there was this video of someone, and apologies because I honestly should have the source here, but talking about how they managed to have watermarks, the stuff.

That I've been working on has been, you know, looking at how can we use some of it has been looking at how can we use cryptography to mitigate some of the near term harms from AI. So one of the main things that I did at OpenAI is I came up with a scheme for watermarking the outputs of large language models. This means inserting a hidden statistical signal into the choice of words that the model generates, by which you could later prove to high confidence that yes, this came from this AI model. So this could be useful most obviously for catching all the millions of students who are now using ChatGPT to do their homework for them. That's very common, even if it's not the end of the world. But more seriously, all the people who want to use AI for election misinformation, for spam, propaganda, fraud, deepfakes, you know, all this malicious behavior was possible before AI, but AI could just massively decrease the cost of it, right? And if you could detect what came from AI and what didn't, that would simultaneously address all these different categories of misuse.

Now, I personally wanted to test this out. So I use Mid Journey and I used Fluked and I was honestly shocked by some of the images that I was able to Generate using a few easy prompts. And you can see right here that I've got this Snapchat photo of this person. If someone sent me this or I saw this on Twitter, I think that this is completely realistic. There are no instances that show me that this is AI generated whatsoever. And the same goes for this image. The tone, the color correction, just everything. I guess maybe if I decided to analyze how smooth the, the shirt was, I could say that yes, it probably kind of does have that AI generated smoothness, but this is on the kind of details that most humans are just not going to look at. And I mean, are you really going to spend your entire life analyzing someone's skin to see whether or not the image is AI generated? I personally don't think anyone has enough time for that. So I mean, it kind of brings us to the issue where currently we're moving into this scary period where the, the Internet is going to be largely somewhat fake. And also there's going to be this issue of people not understanding what is real from what is not real. And it brings in this other problem which I think Sam Altman has previously spoken about. But if we think about the future to where we know that there's going to be billions of agents, there's going to be millions of humanoid robots, there's going to be millions of AI generated images, humans are going to be a drop in the ocean of all that stuff. We have to think about how on earth are we going to verify when a human makes a piece of content versus an AI controlling a social media account. And I think whilst yes, worldcoin has its issues and whilst this entire thing there's red flags and people talk about privacy and security and all those kinds of issues on society, I think we are going to need some kind of solution that allows humans to verify themselves online in a way that is scalable and that doesn't remove everyone's privacy. Because one of the things people are scared about is everyone being know on this kind of digital system and no way of remaining anonymous and protecting your privacy. But you also do have the issue of the Internet being this, you know, black box of individuals where you don't know whether or not you're interacting with an LLM with a photorealistic person that looks like one of these two. And considering you can now get consistent characters and you can now train AI systems on how people talk, you're never going to really know who it is that you're talking with. So I mean, we're really entering this gray area and I thought that, you know, I should make this video because this is something that you most certainly should be aware of.