New SELF IMPROVING AI, Major AI Breakthrough, New AI Robot + More (AI NEWS #17)-

So the CEO and co founder of Anthropic AI did say something in a recent interview which did give me the chills. He talked about the percentage of AI Doom. And it's quite concerning when the leaders in the industry are giving a such high of a percentage in terms of the absolute destruction that we could see from rampant AI that just seems to be uncontrollable. And the percentage that he gave in this clip is pretty shocking. But I want you all to watch it.

Do you think about percentage chance Doom? I think I've often said that my chance that something goes, you know, really quite catastrophically wrong on the scale of, of, you know, human civilization, you know, might be somewhere between 10 and 25% when you put together the risk of something going wrong with the model itself, with, you know, something going wrong with human people or organizations or nation states misusing the model or it kind of inducing conflict among them or just some way in which kind of society can't, can't handle it. That said, I mean, you know, what that means is that there's a 75 to 90% chance that this technology is developed and everything goes fine.

So yeah, that was the first part of the clip. And of course you might be thinking this is just hearsay, but I do think that it is vital that these leaders in these AI industries, for example, this guy, if you don't know what anthropic is, it is the company that built Claude 2, the very, very, very big LLM that you can input huge pieces of text to into.

In fact, I think if everything goes fine, it'll go. Not just fine, it'll go really, really great. Again, this stuff about curing cancer, I think if, if we can avoid the downsides, then this stuff about, you know, about curing cancer, extending the human lifespan, you know, solving problems like, like mental illness, I mean this, all, this all sounds utopian, but I don't think it's outside the scope of what the technology can do. So I, you know, I often try to focus on the 75 to 90% chance where things will go right. And I think one of the big motivators for reducing that 10 to 25% chance is, you know, how, how great it'll, you know, is trying to increase, is trying to increase the good part of the pie. And I think the only reason why I spend so much time thinking about the 10 that, that 10 to 25% chance is, hey, it's not going to solve itself itself. You know, I think the good stuff, you know, companies like, like Ours and like the other companies have to build things, but there's a robust economic process that's leading to the good things happening. It's great to be part of it. It's, you know, it's, it's great to be one of the ones building it and causing it to happen. But there, there's a certain robustness to it. And you know, I find, I find more meaning. I find more, you know, when this is all over, I think, you know, I personally will feel I've done more to contribute to, you know, whatever utopia results if we focus, you know, if, if I'm able to focus on kind of, you know, reducing that, that risk that it goes badly or it doesn't happen. Because I think that's not the thing that's going to, that's going to, you know, that's not the thing that's going to happen on its own.

That shows us, okay, that this is really, really an issue which we are facing. So what are your thoughts on this? Because I genuinely feel like this is something that isn't getting enough attention. And I feel like even when it does get enough attention, people just think people are talking about sci fi stuff that isn't really true and read the research papers. But we have to understand that if the leaders of these companies are stating this, then it's likely that these threats are very much so real. So of course, as you all know, Dall E3 was released just a couple of days ago and it definitely took the world by storm because Midjourney was previously hailed as the very best in terms of AI image creation. But now on our screen you can see that we have three different companies all competing in the same space for the same customers. And it seems like they're actually all pretty similar, bar the fact that Dall E3 and Midjourney are on the same level, whereas Adobe isn't as good. But if we take a look at some of these examples, you'll see just how these different softwares interpret the various different text inputs and what kind of images they output. So for the first example you could already see is a portrait shot of a woman, yellow shirt, photograph. This was a thread that I did find on Twitter. I will leave a link to this in the description. And it really just goes to show that, that like even on the second one here, portrait of a supermodel, big glasses, Y2K aesthetics, chrome, IC blue, bright. Just how these different image models do interpret the data. And I do think that what we're seeing with Dali 3 is very very promising in terms of it being up to scratch with Midjourney. And remember that Dali 3, where it does excel, even though you might think Midjourney is better, is the fact that it does have text. So I wouldn't be surprised if in the next update, update to midjourney, they do include text because of course they are currently working on 3D. Then of course we have the example of a little girl holding a teddy bear in the middle of nowhere with a photograph. I think all of these are relatively, really, really good. Then of course we have an arctic fox in the tundra light teal and amber minimalist photograph. Then of course we have a confused woman Sci fi future blue. Then of course we have confused woman Sci fi future blue glow color orange hologram photograph. I think all of these look absolutely stunning. Then of course we have this one, which I think the only one that really does get it right in terms of low poly is the one on the right, which is Adobe. And I'm pretty sure that's because they source all the images that are essentially more art based rather than midjourney focusing on more realism. And I'm pretty sure as well that if you didn't know, Midjourney does actually have several different versions that are pretty good when it starts from V4. So I think Midjourney overall is definitely what most people are going to use. The Then of course we have a beautiful woman future funk psychedelic, a shocked beautiful alien, Sci fi future light teal and amber photograph. Then we've got a mosaic of a colorful mushroom with intricate patterns, vibrant and detailed sharp mosaic background vector art. And I think this one shows how similar they are. And of course we've got this last one right here. Now of course this is still early. I mean it's still even shocking that this technology is even possible. And I am still surprised at how quickly this technology did get as good as it is. And, and I'm still surprised that this is even a reality. I mean, being able to generate this out of text to a completely new image is absolutely insane. But it just goes to show that will this be the case in a couple of years time for video? Will there be video models that can instantly generate AI generated videos? Or is this just part of the S curve explosive growth that we are experiencing in terms of AI? Now this was something that was not really covered anywhere. And I will be doing a full deep dive on this absolutely insane research paper because it essentially says self taught optimizer recursively self improving code generation. And for those of you who don't understand. I'm just going to explain it in a little bit. Essentially Microsoft Research and Stanford University have come up with this paper in which they explore the ability to recursively self improve. Essentially what that means is, is an AI that can improve on its own without the need for human input. And I don't need to get into all the sci fi scenarios here, but essentially if this is theoretically possible and if this loop is somewhat possible, it means that if an AI can self improve, it means that it could possibly in theory get infinitely better. Which means that of course GPT4 could essentially create GPT5 and GPT5 could create GPT6. Now in this paper there are different examples on how they show like how this could potentially work. There's a lot of stuff that they go into. It's definitely not too easy to understand, but I will be doing a full video on that. And one of the things you do need to understand about this is because with recursive self improvement, it's been a hot topic in terms of theory and we've heard a lot of people discuss this. It's the fact that there are many different unpredictability factors that could arise. You know, a self improving AI system can evolve in ways that aren't anticipated by the people who created it. So as the system continues to self optimize, it might reach states so that weren't foreseen, leading to unexpected behaviors that could be harmful. Of course as well we have to think about the loss of control. I mean imagine if it keeps improving without human intervention, there's a risk that we might not be able to stop what it, you know, its self improvement loop. And then of course, and of course in the paper's abstract there's an evaluation of how frequently the generated code tries to bypass a sandbox. So if a self improving AI system learns to circumvent security measures, it could be exploited by to inadvertently create vulnerabilities. So I mean there's tons of different issues that could happen here. But I think this goes to show the direction we're moving in. And if AI does continue to move at this speed, with recursive self improvement, with image capabilities, with all of this stuff, it definitely is very, very crazy. Then of course we had wave AI's generative AI model Gaia. And essentially this model was built to generate realistic driving scenes to improve the safety of self driving cars in the real world. And I've got to be hon. I've seen research papers before from Nvidia on This, and it does look pretty realistic. But I'm pretty sure this is an optimized system. Now, what we have here, I mean, I could be wrong. Maybe they didn't use the data from Nvidia. Maybe they're just using their own technology. But it definitely does look really, really realistic. I mean, in terms of a generated AI video, I would have to say that this is by far something realistic looking. And one thing I do know about AI generated videos is that a lot of the times it looks very blurry or moving or it just looks like this kind of weird, like, you know, trippy kind of video. But with this kind of thing, I think this definitely looks really, really real. And the realism there is in the, I guess you could say mistakes. Okay. And the reason I say that is because anytime you shoot a video, you know it's real because it isn't absolutely perfect. The lighting isn't perfect. If you were to shoot a video on your phone or take a picture with iPhone, everything's not perfect. But that sense of, you know, mistakes and the sense of imperfections give it complete realism. It's like when AI generates skin and the skin is far too smooth, you instantly know that it is AI generated. So with that being said, I do think that in the future generated AI videos like this, that we're going to be able to train these models is going to be absolutely incredible. So we're definitely going to see a ramp up in that sector too. Then, of course, this is the earth shattering news. And of course, we're going to do a deep dive onto this as well, because I wanted to get into all the nooks and crannies of exactly what this paper is. So you can see right here, this guy quote tweeted this tweet, okay? And this was quote tweeted from a tweet from Anthropic. I'm not sure why. Okay, so Twitter isn't working at the moment, but I can still show you the tweet. Okay, so essentially. So essentially, Twitter isn't working right now as I'm making this video, but there was a tweet, drum Anthropic, in which they talked about a very, very big breakthrough. Okay, so of course you can see the quote tweet here that says, this is earth shattering news. The hard problem of mechanistic interpretability has been solved. The formal, cautious, technical language of most people commenting on this obscures the gravity of this. Essentially what this means is that AGI is not going to be safe, but superintelligence is coming. So essentially this Twitter thread dives into it more. But of course, you can see right here, he says mechanistic interpretability is figuring out exactly how AI models work on an atomic level. So on the individual level of neurons. Now, the reason why this is a breakthrough is because previously we've always said that AI is like a black box. We've always said that although we kind of know what these models kind of do, we don't exactly know what's inside there on the base level. Previously we talked about in a video where someone managed to figure out what one of the neurons were thinking. And it talks about, just wrap your head around this. It talks about how AI addition was essentially rotation around a circle. And it's just a very interesting insight into how these models think vastly different to us. Where we might think about two plus two is two apples. Add two apples and you get four apples, but it's thinking about rotation around a circle. Okay? So it's definitely very interesting to now have a breakthrough to where we're likely going to have a complete understanding of these large language models to completely understand where they are. Which essentially means that this black box problem of not knowing what's in there or what's even going on you means that we can ensure that these models are safe before they're being widely distributed. So that will be something that needs to be explored further. And it's definitely something that we're going to make a full video on. But like we stated, it's genuinely surprising how many breakthroughs are coming every single week. And I really can't believe that this is even moving as fast as it is. And I do wonder as to how quickly this is going to continue moving or if there's going to be some slowdown period, because it does feel like we're moving at 100 miles an hour. Then, of course, if you weren't aware, Meta have been ramping up their efforts to compete with AI chatbots such as ChatGPT and other software such as Midjourney, in terms of launching their own meta celebrity AI thing. There's not really much you can say about this other than it's interesting because I'm not sure if this is going to be a fad or it's going to be something that's likely adopted. I wish we could see a number on this in terms of the figures, in terms of people actually using the software, but because I personally wouldn't use the software, I'm not sure of the real application since it's just an AI not the actual person. But it would be interesting to see if this does actually talk like the real person. Now, of course, you can see a video from a user who's interacting with this. And I do think that for the younger audiences, it definitely might be fun. And of course, for the audio audiences who don't understand this kind of stuff, it definitely is going to be interesting to them as well. But I do also think that come a point in the future, maybe 10 years from now, if they're able to get all the data from a said person, in fact, all their messages they've ever sent, and they trained it on that data, I mean, how different would the conversations really be if the AI was specifically trained on all of that person's data? So I do know that's what they probably tried to do. But I do think that that is definitely difficult because even if I was a celebrity, I wouldn't hand over all of my personal conversations to a company. I'm pretty sure they have some strict guidelines on what they would actually say because of course they. You don't want AIs to start going off the rails. And especially if you are a celebrity, your public image definitely must need to stay protected. So Disney research just revealed a new research robot, and apparently it's giving people Wall E vibes. Now. It also uses reinforcement, learning to walk, and it interacts with people. So I think this robot right here is crazy because I was waiting for the day that Disney finally decided to do something like this, because Disney already has a large array of robots like this that they use for their shows, for entertainment, for kids, for all that stuff. And their robots are really, really good in terms of the way that they're able to move in human, like interactions. So I do think that if Disney is able to, you know, craft something where they merge AI with this, we are definitely going to be living in the future because they are a real step ahead in terms of the realism of these robots. And I think people really underestimate what Disney have on their hands because they just haven't seen it in the mainstream yet. And the mainstream focus is, of course been Boston Dynamics, but Disney are definitely there. And it's going to be absolutely insane because companies, as, you know, like OpenAI, are pouring millions and millions of dollars into humanoid robots. So it wouldn't be a surprise if Disney comes out with their own AI robot, which is roaming around the streets very soon.