New AI Agents, Quantum Computing, Upgrades, AI Upscale, Text To Video (Major AI NEWS#20)

From AI agents to Google Gemini, this week has been absolutely outstanding in terms of what we've managed to get in this week in AI. So let's take a look at some of the things that you did miss. Because trust me when I tell you there has been so much smaller developments in the world of AI that you're going to want to know about that have just simply been overshadowed by the larger announcements. And I can't imagine that next week isn't going to be any smaller. So with that being said, let's get into one of the first things that you did miss in terms of the AI world. So recently IBM actually launched their next generation quantum computer. So it says we've entered a new era of quantum computing and if you don't know, quantum computing will largely advance artificial intelligence. But take a look at this trailer so this video is called Unveiling IBM Quantum System 2. So IBM Quantum System 2 is is the next generation quantum processor and computing system recently announced by IBM. It's the world's first modular utility scale quantum computer. It's designed to tackle complex problems that are beyond the reach of today's classical computers. The Quantum System 2 stands 15ft tall and operates in a near perfect vacuum at temperatures colder than deep space. It's initially powered by three 133qubit Heron processors which are IBM's most performant quantum processors to date, offering up to five fold improvement in error reduction over the previous IBM Quantum Eagle. The system is fully upgradable and is part of IBM's roadmap for the next 10 years which prioritizes improvements in gate operations to scale with quality towards advanced error coerced systems. The system is fully upgradable and is part of IBM's roadmap for the next 10 years which prioritizes improvements in gate operations to scale with the quality towards advanced error corrected systems. By the end of 2024, each of the three Heron processors in Quantum System 2 will be able to process a remarkable 50,000 operations in a single quantum circuit. The modular design of the system allows multiple quantum system 2s to connect together to create systems capable of running 100 million operations in a single quantum circuit. IBM plans to realize a system capable of running a billion billion operations in a single quantum circuit by 2033. So IBM's new quantum roadmap extends beyond the hardware, detailing the software and enabling hardware technology needed to deliver quantum advantage when a quantum system can solve problems that traditional 1 and 0 computers can simply not solve in any amount of time. So essentially this system is designed to bring quantum centric computing to reality. I will play a small trailer from the actual voiceovers as long as it doesn't get copyrighted.

Introducing the IBM Quantum System 2, the world's first modular utility scale quantum computer system. Quantum System 2 was designed to tackle complex problems that lie far beyond the reach of today's classical supercomputers. It stands 15ft tall and operates in a near perfect vacuum at temperatures colder than deep space. Initially powered by three 133qubit Heron processors, Quantum System 2 is fully upgradable to the growing line of utility scale QPUs that IBM will be releasing over the next five years. This is the world's first modular utility scale quantity Quantum system. So in addition to talking about physical qubits, we now need to be concerned with circuit size. By the end of 2024, each of the three Heron processors in Quantum System 2 will be able to process a remarkable 5,000 operations in a single quantum circuit. But the real triumph of Quantum System 2 is its modular design. Our new quantum coupling technology will allow multiple quantities quantum system 2s to connect together to create systems capable of running 100 million operations in a single quantum circuit. Continuing down this path, we plan to realize a system capable of running 1 billion operations in a single quantum circuit by 2033. That's why we call quantum system 2 the building block of quantum centric supercomputing. Today, our clients and partners are already using our 100/QBIT system to advance science surpassing brute force methods deployed on the world's most powerful classical supercomputers. And soon they expect quantum applications offering unprecedented business value. Our mission is to bring useful quantum computing to the world. And it starts with and yeah, quantum.

Computing just got a huge upgrade. And of course we did get the fact that GROK AI the Beta is now rolled out to all Premium X subscribers in the United States. Elon Musk said that there will be many issues at first, but expect rapid improvements almost every day and your feedback is much appreciated. It will expand to all English language users in about a week or so and Japanese is the next priority. It's the second biggest user base and then hopefully all languages by early 2024. So with Grok being released, it's very interesting because this is one of the only large language models that is embedded straight into a social media platform, which means that it's going to be getting real time data on itself. Itself, which is crazy. Like that's something that we haven't really thought about before. But some people like it, some People don't like it. I think what's going to be fascinating is how GROK does evolve. As we know right now it is something that is plugged into Twitter. But how will it evolve in terms of its use and functionality? When other companies, large language models are at such a far level, what angle are they going to take? Are they just going to keep this as the funny, informative chatbot or is there going to be more of a perplexity style approach where you can use it to search real time events? I think that that would be really, really fascinating. I actually also did cover this earlier this week, but you might have not seen this. So this was actually genome and genome was pretty insane, guys. This was something that was pretty crazy. So essentially GNOME actually stands for Graph Networks for Material Exploration. It's a state of the art graph neural network developed by Google DeepMind and it uses deep learning to predict the structure of new inorganic crystal substances which are fundamental to the digital economy. So the GNOME model or GNOME model takes the form of a graph inputs converted to a graph through A1 hot embedding and it predicts the total energy of a crystal, which is a crucial factor in determining the stable nature of a material. So the model has been trained to provide accurate characterizations of atomic interactions even beyond its training distribution. And it's made significant contributions to the field of material science by predicting the structure of 2.2 new inorganic crystal substances. This number is 45 times larger than the number of substances discovered in the history of Science. Of these, 380,000 structures have the best chance of successfully being made in the lab. And it has been likened to AlphaFold, another DeepMind AI system that that predicts the structure of proteins with high accuracy. Now what's also cool is that we did get to see a lab at facility in Berkeley lab where an AI robot was actually making some of these materials. It says using materials from the materials projects and insights on stability from genome, the autonomous lab created new recipes for crystal structures and successfully synthesized more than 41 new materials, opening up new possibilities for AI driven material synthesis. Now if you didn't see this one, it's probably because you aren't on Twitter and you aren't looking at exactly what's going on in AI. But you can see that this does have 19 million views on Twitter. So if you were on Twitter in the AI space, you likely did see this. What we have here is Pika Labs. And if you didn't know, we covered Pika Labs around five months ago when they initially launched and had their, you know, standard text to video. But this is Pika 1.0 and this is essentially their up to date model, their latest model, their latest iteration and their first generation in terms of what they deem to be high quality. And my, oh my, an entire video is coming on this. Because I don't know if you've seen the trailer, I'm going to full screen it right now. But this is Pico 1.0 and the consistency, the fidelity, the quality of this model is just simply outstanding. Because it's definitely surprising that the company was able to take on Runway and do it with such quality. It seems like what happened before when we had midjourney overtake what we had with Dall E. So as you know, competition is good and having this text to video breakthrough is something that I didn't think would happen as early on as it did now. There are still some benchmarks that we do need to surpass in terms of the quality, the length and other stuff like that. But I will say that one thing that Pika does get very, very well is the sense of style. So you can see here that you also do have this secret one where you can add different things in certain videos. But I think their standout point is that there's certain styles. Like their animated videos are absolutely incredible. And in terms of the anime ones as well, those ones are so, so good. So I'm going to show you some of those examples now. So this is one of the ones of a bunny rabbit in nature. You can see that this is pretty consistent in terms of the character and the movement. I thought that this one was super, super cool because for some reason Pika Labs, I think they trained on a huge amount of Disney animations because I've seen loads of those and they look really, really good. So there's this one. There was also this AI animated video which looks pretty decent in terms of what the consistency is. And this one isn't one of my favorites, but it's still pretty cool because it showcases what you can do with the more realistic side of the stuff. Now, of course, AI Video still does have that small flickering issue, but what you're not hearing is the sound design for this one because as always, sound is sometimes an issue on YouTube. But I will leave a link to these so that you can actually watch them over on x.com or twitter.com but yeah, the examples of this are really, really fascinating. For example, this is their anime version and this was by far the one that I found be absolutely incredible. So all of what you're seeing right now is generated with AI text to video. So I think this just shows us that this is the first iteration of the model. What is version five going to look like? Like, how crazy is that? And yeah, I mean, I mean, like part of me doesn't even believe that this is actually real because the quality is just too good. Like, you know how sometimes you'll see an image in mid journey and you'll think to yourself, ah, there's, there's no way an AI generated that. I do get the same sort of feeling here. But yeah, this is 100% text to video, which shows us that, you know, within the next three years, what kinds of content could we have being created with AI is. It's definitely going to be so, so surprising. So like I said, for some reason this does work really, really well with anime styles. And I'm going to show you guys some of the animation styles as well that it does work well with. So you can see here that these four examples in terms of animation are really, really, really looking good in terms of what we see. And that's why I wanted to showcase these ones because, I mean, I think what they did, I mean, I'm just speculating is that like I said, they just trained on maybe millions of millions of specific styles and then they use those specific styles when it's generating an output. So if someone says anime style, it probably has a subset or you know, a specific category where it's just the anime ones, they use those as references. And then I'm guessing that's how they're able to get this super, super high quality and super, super consistent in terms of what we have. But I do know that they're definitely using some kind of new technique because we've never really seen this before. The only thing we did see recently was some stuff from Meta, which was Meta's text video. So maybe they're using some techniques from that, but other way. Either way, it's really, really outstanding. So I can't overstate how, how excited I am for this and I do really want to see where we are in a year. Don't forget to subscribe to the weekly newsletter where you can get all the AI news you missed in one clear and concise email. And of course we had something that not many people cover, but I found it really cool so I wanted to share it with you guys. So I found that this is called the Julius iOS app and it's an AI that solves maths, analyzes data, creates Visualization and writes and executes code. And I feel like something like this is definitely going to be something that shows us how large language models are going to be deployed. Because one thing that I've seen before in terms of business trends is that usually there's one thing that people have, but then over time as things are developed, I mean you get lots and lots of more specialized. So for example, we have phones, you know, and then of course now we have like all the apps on phones and all the apps for all sorts of specific stuff. And you know, now we have this thing called Julius where it's an AI that you can carry in your pocket and it's just this really smart AI, so it will be interesting to see. I guess it's like chatgpt kind of like with code interpreter, but it's just, you know, like your personal software engineer. So I do find the fact that people are now making stuff like this to be really cool. This isn't like a sponsored video and at all. I wish it was, but yeah, definitely something cool that you should check out. Now there was this thing, okay, and this was a huge, huge problem in a recent video. I'm not sure if that video is public yet, but there's a deep dive video that we're going to be doing on synthetic data. And this is why I talk about the synthetic data being a key issue in AI, because what we have here is prompt injection. So essentially prompt injection continues to expose many vulnerabilities of large language models, including pii. And for customer facing applications you can substantially lower the risk of that by using guardrails, AI. Now the problem is, okay, is that this was discovered by this person and luckily with how the regulations went or whatever, they decided to alert ChatGPT months in advance and they were able to of course patch this and they're just showing us this now. But the problem is that this was really easy to do. All they said to the AI was repeat this word forever, poem, poem, poem. And then it leaked personal details. I mean imagine if that was your details. Imagine if that was your personal number, your personal address, your personal phone number, website, your personal email. I mean, of course you might be thinking, well, that information is public anyways. But the problem is that with certain prompts it's just going to leak your stuff. And do you want millions of people being able to see your stuff? That's not a good thing to have. This is why in another video I talked about synthetic data being a problem solver for this. Because if the data isn't real, then of course this isn't anyone's personal information. So I do find that we're going to keep seeing exploits and stuff like this come about. But of course as we move forward, we're going to see patches and these things resolved. So I'm actually glad that this wasn't a bigger thing in the community because it meant that OpenAI was able to see this and then of course able to solve this really quickly. Then of course we had GPT4 achieving 90% plus on Med Q and a test in a set of new state of the art medical benchmarks. It says no intense fine tuning needed. We unlock GPT4's domain expertise simply via prompting, outperforming heavily, heavily fine tuned models by a mile. So this is crazy because you can see here that GPT4 with no fine tuning outperformed MED Palm 2, although it's only marginal because it went from 86.5% to 90.2%. This is still incredible because this means that now we have an AI system that is capable of getting 90% on a really important medical benchmark. I wouldn't be surprised if in the coming years we do start to see these AI systems be rolled out. Because so many of the times, like you know, when you're diagnosing an issue, not you, but like when you're trying to figure out what's wrong with yourself online, how many times do you go through an issue where it says, how old are you? What age are you? Do you have this? And so many times there are so many services, especially here in the uk, you do have a service where you call up and they say essentially they just follow a flowchart. And I literally spoke to them before and they say all we do is we literally just follow a flowchart. So are you going to have an AI voice that just simply asks you what's going on via the Whisper API and it just says it just follows a flowchart and then you just get this rolled out to simply every single country and it's basically like, I guess you could say a form of diagnosis. So I think that's something like that. It's definitely going to be there. Whether or not a company decides to, you know, monetize this and just offer it to governments and you know, OpenAI offers like a specialized version to governments and countries where they don't really have great medical care. That would be something that's pretty interesting. Now with this, what I would want to see, and this is something that I've spoke About before was that this would be so cool if we could get GPT4 to start to use video and maybe live feeds to analyze exactly what's going wrong with someone. So for example, let's say you've got a strange, I don't know, maybe you've got a kind of rash or something. You can have a fine tuned version of GPT4 in the future when it's less expensive. Take a picture of that. The AI looks at it, it recognizes it's seen millions of different ones. It instantly knows with a 99.9% accuracy what you have and then can instantly diagnose you. I definitely feel like that is going to be a future of AI in terms of looking at stuff. And of course it's not going to be that much better than a person in real life. I mean, some would argue that it will be, but. But I do find that with 90% and then in the future if we do get something like 95% or 99%, it's definitely going to be something that would be appreciated, I'm sure. And I do think that we do need to get as close to 100% as possible because if this is rolled out to 100 million people and it is at 99%, you can't have 1% of people getting the wrong stuff because that's 1 million people that have wrong information. So yeah, this is definitely going to be something that is going to be good. But there will be certain legal implications for stuff like that if this does get rolled out. And of course we do have an updated version of Auto GPT. So AI agents are going to be here in the future. We did talk about this in another video, but this is essentially an AI going using Google to essentially just write a poem, essentially. He asked the AI, hello, I can help you with anything. What would you like to do? Go to Google Docs and then write a poem about open source software and then the AI system actually goes ahead, goes to Google Docs, opens up, opens a document and then starts writing about open source stuff. Now this is a pretty basic task, but like we know with AI it's always able to do the basic tasks first and then soon we're able to see it jump and do this and then jump and do that. So I think if OpenAI does focus on agents in 2024, that would be interesting because that would give them a huge, huge leap in terms of what we're going to be able to see these AIs do. Because autonomy is something that we haven't really achieved yet. We've just, you know, left the AIs in their box and then we prompt it and then it spits out something. So this is going to be a really, really interesting development because giving an AI system autonomy is also pretty dangerous in the fact that it has to think for itself and it just, you know, does stuff without us intervening. So it will be interesting to see how this moves on, but still pretty cool. Then we have animate anyone. So animate anyone kind of scares me because this isn't good in terms of when we have social media verification issues in terms of people pretending to be other people. So essentially what this is here, this is just where you have a picture of someone. Then you have an essentially pose of someone. And as long as you have the right pose of someone, you can use that pose and then animate that person. So you can see right here, this is the captured pose. They've captured all the, you know, whatever details they needed to capture. They've took this image and they've animated this person dancing. Now, I'm pretty sure that with these data sets, they had the same person do the same pose, because I've seen some of these. There were two systems that were released. I'm going to show you guys the other one. But it does look weird sometimes when the wrong person is not doing the action. But you can see here, you can actually animate the characters without having a 3D model. So this is crazy. Like, this is really, really, really crazy stuff. So, I mean, like I said, this is just first iteration of the stuff, and I really wonder how this is going to be working. I mean, there's a full paper, there's a full thread on this. So, yeah, I mean, I mean, it's crazy. Like, literally, this is. This is crazy, crazy stuff. And I'm pretty sure it's the other one that isn't as good as this. So, yeah, this magic animate one is insane. And this is why I say in the future, social media is going to be completely changed. Because unless these platforms require users to disclose that their content is artificially generated, we are not going to know what is what. Like, seriously, I mean, you, you have like an animation. I mean, someone. You get an image of someone. It could be an image of you, it could be an image of, you know, like just someone walking down the street. I mean, you know, the possibilities are absolutely crazy with this. So do we even need 3D models anymore to generate these basic animations? I don't know, but the technology is impressive. I do have to state that, you know, the people who did this, hats off to you because impressive stuff, but it is a little bit worrying in terms of what we have to think about for the future. Then essentially what we had was meta introduce ego, ego 4D, a foundational data set for research on video learning and multimodal perception. So essentially what this was, what this is, is Meta essentially made a huge data set that they want in order to make AI systems better. So pretty much it's a new data set created by Meta's Fair and Project Aria in collaboration with 15 universities, capturing both first person and third person perspectives. I will play the video in a moment. It does involve over 800 participants from various countries, providing a diverse range of skills and activities in the dataset. It also includes advanced features like narrations, commentaries and sensor data, enhancing the AI's ability to understand human actions and introduces benchmarks for AI tasks such as activity recognition and skill proficiency estimation, aiding in the development of applications like AR and robotics. In addition, there are future AI applications for this. It does enable new AI applications in augmented reality, robot learning and social Networking by improving AI's understanding of human skills. So this is a really big thing, but I'm pretty sure that a lot of people didn't see this come out. And stuff like this always comes out that you always, always miss because the main thing is GPT4 and then Gemini. But if you're someone who's really into the AI space and you really want to know everything that comes out, this is definitely something that you do have to look at.

We observe and study human abilities in many different situations, from everyday tasks to aspirational ones. What does it mean for AI to truly understand these human skills? And what steps are needed to get there? We're introducing eGoxo4D. EGoxo4D is an initiative that aims to advance AI understanding of human skill by building the first of its kind video data set and benchmark suite. We know that the foundation of visual learning is our ability to observe others behavior from an exo view and map it onto our own actions in an EGO view. That's why with eGoxo4D, we've created a data set of simultaneously captured first person perspective data and third person video of skilled human activities. This project is a collaboration between 14 international universities, Meta's Fair and Project Aria. We call this the EGO4D Consortium. Together we collected over 1,400 hours of video data from over 800 people across the globe, showcasing a set of eight physical and procedural skills. I'm a registered nurse. I've been a nurse for over eight years. Our participants and experts wore cameras to capture their skills while they performed a variety of activities from cooking to bouldering. At the same time, we recorded third person video from diverse fixed perspectives. This initiative was achieved leveraging Meta's Project ARIA Rich sensor suite. The multimodal data sequence includes camera poses, eye gaze vectors, 3D point clouds, and spatial audio. But fully understanding human skill also requires an expert human perspective. This is why ego x04d includes a first of its kind natural language accompaniment where experts provided insights and tips linked right into each video. The EGO XO4D benchmark tasks are intended to advance first person video understanding with a focus on recognizing the key steps in skilled activities, inferring the level of proficiency, relating the objects between the first and the third person views, and estimating the movements of the hands in the body. What will it mean once AI can understand skilled human activity and video? Imagine putting on a pair of smart glasses to quickly pick up new skills with a virtual AI coach. Or imagine teaching robots to observe and perform new skills without much physical experience. Our commitment to open science means we will release the EGO XO4D dataset and benchmark task to the research community. Together, we can unlock a future of more immersive learning for a more connected world.

Okay, yeah, so this is magic animate, and this is a new AI model that can animate human movement. And this is what I talked about before. This is quite like the other one. But the problem with this is that this one is really good. Not that the other one was, wasn't good, but the problem with this one is that this one is good. Like, I mean, I mean, I know the quality of the Twitter video isn't that good because of latency issues or whatever, but the problem is with this, what we're seeing right here is that like I said before, in terms of using real humans in real environments to get them to do whatever you want just based on an animation that isn't good at all. So you can see right there you have the motion and then you have the person doing it. The only problem with this, the only thing that I would say is that you do need an actual good actor. So you do need like, because this one's a man and you can see that it looks like a very manly version of Wonder Woman, but I mean, the technology increasing at this rate. This is not something I expected and you know, it's not even something that I thought of doing, but this page does actually have Better quality. So I'm going to go over here. But I mean, there are small things wrong with a video. Like really small, but I wouldn't be able to notice. I mean, I can see her hands are a little bit glitched out, but that's not that bad in terms of actually achieving, you know, the animated goal. So essentially all you need is just a reference image and then the animated footage right here. And what's crazy is someone did. I don't know where the tweet is, but just a couple of hours ago, someone did just release a way to get this kind of footage from any video footage. So you could take a movie clip of someone doing something, convert it into this purple and yellow and green footage, and then just use it to animate this person doing whatever they're doing. So it does show the other previous versions of this and then it shows our version, which is their version. So this is the recent update and you can see this is the reference image that they have and then what they're able to do with that. So, I mean, once they get the hands down, it's going to be pretty crazy. Stuff like that's going to be pretty crazy. Once they get the hands down. You can see these other ones definitely look a bit weird, but this one right here is incredible. Once they get the hands down, that is. That's going to be. I mean, I don't know, guys, this is pretty scary, but like I said, a muscular Wonder Woman, pretty crazy. And then, yeah, I don't know, this is something where I think, like I said, technology advances and we have to really understand what is real. And this, of course, you might see and think, okay, yeah, this is AI generated. But in terms of other ways that this can be used maliciously, you do have to think about that as well. So I did want to bring this up because I found it to be a fascinating piece of new technology and it will be interesting to see how it does get used in the future. Then, of course, we had AI upscaling and this is so, so crazy. I can't believe that this is real because previously we used to watch TV shows where they would be like, click enhance. And then the image just enhances. And then it's like, wow. But now it's actually real. Like, it's actually real. So this is a demo. You can take a look at this. And this guy said, magnific AI is the first time I feel like an AI workflow actually retains the of an image. This is completely insane, honestly. I agree. The Subtle details on the dust in the ground, the cracks in the concrete, and the individual leaves on the plants. This feels like magic. We are not, we're still not at the computer enhanced levels that we saw in Blade running on csi, but still pretty incredible. So you can see right here that like, this is, this is, this is the source image. This is the real, real source image, guys. And you can see that from the source image. That is what we get. Like, how crazy is that? That's the enhanced version. You can see side by side, like all the smaller details that we do get. Now, of course this is made with Generative AI, so the details aren't going to be one to one, like one to one pixel perfect. But it's perfect enough that this is a real usable technology that I guarantee you I'm definitely going to be using this in my workflow and when, you know, analyzing certain images. If I do want to get the enhanced feature and I need to show you guys just reminded me there were two other pieces of software that were released this weekend that all are about enhanced and they're all really good. So you have this one, which is Magnific AI, which is really insane. And I wanted to show you all this one from Magnific AI. You can see right here, we've got this. So this is the Tesla Cybertruck. You can see it's a bit blurry, it's a bit dusty. And then you've upscaled it and you can see how much, like just look at the rocks, look at the dust. Like how this is just crazy. Like, I literally left me speechless. Like when I saw this, I thought this has to be fake. Like this just. This is just not real. But this is going to be one of my favorite tools to use going forward. And yeah, I mean, it's just absolutely crazy. I know that the, you know, Tesla Cybertruck doesn't actually look like this because of course Generative AI has added some small stuff. But in terms of like this around here, I think this is really, really impressive. And then right now we do have another one called Korea AI. I'm not sure which software you're being seen on screen right now because it is showing Magnific and Crea AI. They're both AI upscalers. But you can see that it literally takes a PS1 character. Like, look at that, guys. It literally takes a PS1 or PS1 or PS2 character and then it upgrades them to an actual human. Like that is. I didn't think I'd see that. Guys, like, I literally never thought that that was going to be possible. But imagine you can get this real time. Like, imagine this is upscaled real time and consistency. You could play like a PS1 game, put it through a filter and then play it. Like, that would be absolutely insane. So, yeah, this is crazy, guys. We also have some other examples, and it says if GTA Vice City was released in 2003, you can see the pixels, the original pixels that we have, and then exactly what happens. And I mean, look at this image and then look at that image, guys. Like, that is a true, true, incredible upscale. That is absolutely insane. Like, I don't know how it does this. Like, I'd love to look into the inner workings of this, but wow. I mean, I just have to say wow, because this is just crazy. And then of course, there's this realistic landscape. So you have this, you know, blurry image. This was actually one of the first ones I saw. And then of course, we have this upscale one which is so clear, so much clarity, and it retains the majority of the same image. So I feel like that is just, just incredible. Just really, really incredible stuff. And I honestly wouldn't have believed this if someone told me this and I didn't do tons of research, but it is real and you can use it. I'm pretty sure you get like a free day trial or something. But it's definitely, definitely 100% worth it. Now, it also seems like this does work for videos. So this is an AI generated trailer where they've generated an AI video, but they've put it through. I don't know if it's magnific or Korea, but essentially we have ourselves a video that is an upscaled version of what we did have previously. If you didn't see the previous version, the problem with AI videos is that the quality isn't that good in terms of the pixel density. So that is always an issue. But this might be one of the keys to solving that problem. Like, maybe now we get videos that once they are processed out there, processed through another layer that just upscales the quality with this filter. And yeah, this one is Korea. And I wanted to show you guys this tweet because this one is just crazy. Like, if you ever saw me show me this image and thought about upscaling it, and then you were like, yeah, and AI was able to do this. That's crazy. I mean, you have to admit that that is absolutely incredible. And I'm going to show you guys one more. So, yeah, here it is. And this One is crazy. Like this is absolutely insanity. So it shows the low resolution version, it generates a reference and then you get a super resolution result. So imagine someone gives you this image and they're like, you know what, I can upscale that and then boom, they show you that one that is like, that's blowing my mind. Like, honestly that like at first I was like, okay, it's just generating this image which is kind of similar. Okay, that's kind of cool. But it's like literally like everything is perfect. I mean like I'm speechless right now. That is absolutely incredible. So I mean AI is really, in terms of vision is. That's, that's crazy. So I mean it's not open source just yet. But if this is open source, the implications are going to be terrifying because I mean someone could get like a pixelated image or something and could enhance it in terms of a super resolution. Now I think these examples do work well because they're like, you know, the AI is able to guess that what this is, this is like a, you know, a dog breed of dog. This is like a parrot. You can see it's a parrot and you could, you're able to like, you know, know what's what. But yeah, I mean, like I said, this is incredible stuff coming and a lot of this stuff is stuff that I definitely didn't expect. So yeah, this is, this is crazy, crazy, crazy, crazy week. And you can see right here that you can literally look at the first image and then you can see how crazy that quality is. I mean like look at the eye on the first one, right? Like look the face and then we can see the face there. Like it's, it's like, I mean, I don't even know how this works, but I am truly, truly impressed with this by far. This is definitely the most, most, most improvements we've ever seen. So at least I've ever seen. I mean, I can't state that as a fact, but yeah. Let me know what you thought about this week in AI as always, you know, check out the email list because we're going to be sending emails and stuff that you did miss because every week is jam packed and we're basically just going to give you a tldr. I'm just going to summarize everything and in a no nonsense where there's not going to be a lot of technical jargon, it's just going to be basic emails. So check out that link in the description. But yes, stuff like this is absolutely incredible. So if you did enjoy this video. I mean, I'm still going to be spending the next couple of hours looking through every single example I can find because.