AI NEWS : Meta CANCELS New AI Model Release, Q-STAR BANNED, LLAMA 3 Release Date, and more

So, welcome back. Let's take a look at some of the most important AI stories, because there are a few that you don't want to miss. So one of the stories you actually don't want to miss is that Meta won't offer future multimodal models in the eu. This is pretty incredible. Meta will withhold its next multimodal AI model and future ones from customers in the EU because of what it says is a lack of clarity from regulators there. Now, this is pretty funny because if you've been on Twitter for quite some time, and when I say on Twitter or on X, what I actually mean is if you've been looking at the AI space and the kind of discourse and conversations that surround the AI space, one of the ones that is more popular is how EU tends to regulate everything, whilst America and the other nations tend to innovate in these other spaces. And it seems like this trend is continuing. So Meta are stating that they're going to withhold its next multimodal AI model and future ones from customers in the EU because of what it says is a lack of clarity from the regulators there. Axios has learned and it says the move sets up a showdown between Meta and EU regulators and it highlights the growing willingness among US tech giants to withhold products from European customers. Now, this is pretty crazy because you have to understand that the EU has been a little bit insane when it comes to regulating things. And of course here it also states, and this is something that I did cover a little bit, that Apple similarly said last month that it won't release its Apple intelligence features in Europe because of regulatory concerns. And I do wonder if some of the other companies like Google and OpenAI do tend to slow down their releases for certain features. For example, when Gemini 1.5 Pro was released, I do remember trying to access that in the playground and I do remember that some European regions couldn't access that software for quite some time. So I'm wondering if that is the case, but there wasn't any news on that. But the point is, is that it's not a great time to be a member of the eu because this is a serious piece of technology that is going to change the landscape. If you aren't familiar with what's going on with Meta, the model that's probably going to be released is probably llama 3. And I'm guessing that what we're seeing here is that it's going to be a multimodal model. Now this is going to be pretty crazy because they also say that it's not just that, it's the future ones too. So this is going to have some severe implications. And with AI regulation at this kind of level, I'm sure that this is going to possibly impact space. Now, of course there's going to be some companies that are still going to thrive for the likes of Mistral and other companies that operate in the eu, but it is going to be a little bit difficult to be someone who's trying to enjoy what the AI space does have to offer. Not to state that you can't access any of that within a VPN though. Now, if we dive further into the news, you can see that Meta plans to incorporate the new multimodal models, which are able to reason across video, audio, images and text in a wide range of products, including smartphones and its Meta Ray Ban Smart glasses. Honestly guys, this isn't a sponsored video, but if you haven't tried out the Meta Ray Ban Smart glasses, they're actually something that does work. And when I say something that does work, traditionally what we usually get from these products that are touted to be the next wave in terms of technology, AR and VR, meta smart glasses are something that I've used personally and have bought a pair. And I can say firsthand, well, they really do do exactly what they're supposed to. And I can imagine that if they do manage to get AI working in those glasses, I can truly imagine a future where those are an actually good product. And for those of you saying that, just think about for those of you who criticize the Rabbit R1 device, where many people were saying, why not just put this in an app? Why not just have this on your phone? Think about the glasses. You could have a vision AI system that's able to tell you exactly what you're seeing and give you updates and all sorts of different pieces of information that are relevant to your daily life. Now what's crazy is that it says Meta says its decision means that also EU companies will not be able to use the multimodal models even though they're being released under an open license. It could also prevent companies from outside the EU offering products and services in EU that make use of the new multimodal models. The company is also planning a larger text release only version of Islama 3 model soon and and that will be made available for customers and companies in the eu, Meta said. And the main issue that is plaguing meta is EU's GDPR law, the EU's existing data protection law. Now I do have to be honest sometimes this thing is quite annoying when I'm trying to access American websites that do have interesting content and then I just see a data protection law pop up saying, sorry, this content isn't available in your location. But interestingly enough it says the United Kingdom has has a nearly identical law to gdpr. But Meta says it isn't seeing the same level of regulatory uncertainty and plans to launch its new model for UK users. So it seems that the EU has a choice to make. Are they going to allow certain exceptions for certain companies or certain products, or is llama 4 going to be largely restricted in the EU? It's going to be kind of interesting considering the pace of AI development is one that continues to advance rapidly and the space is evolving very quickly with new regulations being drawn up fairly quickly. But I will pay attention to this now in more Meta news They're actually planning to Release the largest LLAMA 3 model on July 23rd Meta Platforms Plan to release the largest version of its open source Llama 3 model on July 23. According to a Meta employee, this version, with 405 billion parameters or the settings that determine how the model responds to questions, will also be multimodal, meaning that it will be able to understand and generate images and text to information previously reported. Now if you hadn't been paying attention, Meta's Llama 3 release was absolutely outstanding because they managed to raise the bar on what we thought possible for open source. If you didn't watch the video I made, there's a rundown of that video, the link somewhere in the description, and trust me when I say it, you're going to be really surprised from what Meta AI has been cooking. One of the craziest things from the LLAMA 3 information drop was the fact that LLAMA 3 was still learning when they were training it, but of course they decided to cut the training and then of course wrap up the model, of course doing some safety testing and other things. Now I think this is going to be truly fascinating because this is going to be one of the times where we might have a GPT4 level model that is truly open sourced. And for the first time it looks like we're going to get some truly competitive open source models on the scene that are comparable to Yi's models and of course Mistral's mixture of experts. It's going to be kind of fascinating to see also where this model does rank on the chatbot arena, because if we take a look at Meta's actual self reported benchmarks on their webpage, we can see that this is something that is truly incredible. Taking a look at Meta's llama 3 and comparing even the 70 billion parameter model to state of the art systems that we do have right now, like Gemini Pro 1.5 and Claude 3 Sonnet, you can see that it does clearly hold its own. So the 405 billion parameter model at as some people may speculate, might just be an open source version of GPT4. If this is an open source version of GPT4O manages to catch up to Claude 3.5 sonnet, this would actually shock the industry because it's going to be a time where the dynamics have shifted, open source has caught up and now some of the industry leaders are going to have to pivot make better models. So this paper wasn't covered as much. I mean the AI space right now doesn't seem to have as much hype around it, so. But of course there are still people publishing papers and there is still a lot of information that is out there that shows us that we are still moving towards a landscape where these models are going to continue to get smarter and new paradigms are being unlocked. Introducing a mixture of a million experts. You can see here that this is fascinating. It says the recent discovery of fine grained mixture of expert scaling law show that higher granularity leads to better performance. However, existing mixtures of experts architectures models are limited to a number of experts, small number of experts due to computational and optimization challenges. This paper introduces the Peer parameter Efficient Expert Retrieval, a novel layer design that utilizes the product key technique sparse retrieval from a vast pool of tiny experts. Over a million experts on language modeling tasks demonstrate that peer layers outperform dense FFWs and coarse grained MOEs in terms of performance compute trade offs. Long story short, I know that this sounds pretty complicated, but if you read what this breakdown on Twitter talks about and it actually includes an expert from the paper, simply by adding new experts and regularizing them properly, MOE models can adapt to continuous data streams. Freezing old experts and updating only the new ones prevents catastrophic forgetting and maintains plasticity by design. In lifelong learning settings, the data stream can be indefinitely long or never ending, necessitating an expanding pool of experts, or as some would put it, an AI that learns over time. So basically this paper reduces inference cost and memory usage scales to millions of experts and just happens to overcome the catastrophic forgetting and enable lifelong learning from the model. So this could be a truly game changing piece of research. Now something that I did also want to show you was a piece of footage that essentially allows you to understand the pace of AI development. On the top here, what you can see is a AI generated pizza commercial for Domino's or whichever company that it might be. Now, it is assisted because at the time, the text that you are seeing on screen is not AI generated. This is, of course, in After Effects or whatever, video editing software. But you can see the discrepancy between the quality in just a year. This is in the video area as well, which is why I do believe that many people are struggling to see how in just three years or even five years, there might be a rapidly evolving landscape in terms of how much AI can develop. Now, I do know that this is video, and you might be thinking, okay, video has its limits. We are just on the middle of the S curve, so jumps like this are to be expected. And this is going to be something that, you know is of course typical of what you might see during rapid growth. I would agree to some extent with that. But of course I would also say that if we look back, you can remember the current sentiment around AI video. Many people were speculating that AI video would take several years and several longer years than traditional text because text was way easier. But we can see now that video has largely been conquered and the consciousness surrounding what's possible with AI was once again reset. Our brains were allowed to adjust to the timeline. That this isn't something that is incredible, although it fundamentally truly is. So with that being said, I do think that within the next two years, there's likely to be a very large discrepancy between what we think is smart now and what the future models are truly capable of. If you remember when GPT4 was released, so many people were astounded by the capabilities. But now people literally go into GPT4 and they say, wow, this model is so dumb. I can't believe I'm using this every single day. It's wild how humans manage to adapt to new technologies. And in the realm of AI, it's an even shorter timeframe than traditional technology. Now, in more AI regulation news, we have Mark Anderson and Ben Horowitz saying that when they met the White House officials to discuss AI, they said they could classify any area of math that they think is leading in a bad direction and make it a state secret and that it will end. Now, I want you to watch the clip first, because I did some research on this and this is real, and it's pretty crazy about where things might end up in terms of the AI space. One thing that most people aren't Thinking about is that AI truly has true capabilities and true power in terms of the unfiltered possibilities that this technology could present us with. And in doing so, that does open up the nation to some very catastrophic risks. And that's what they're talking about.

We argued in our meeting with the White House on AI policy was, you know, look, there are going to be, there are going to be issues that come from AI, but let's. They should be regulated. The regulation should happen at the application level, not at the technology level. Yeah. And the reason for that is, is, is, is basically because, like to regulate math. Right. At the, at the level of the algorithms. Like to regulate. And he argued with me when I said that. Well, so, yes, so, so, so, so Ben basically said, look, it doesn't make sense because to regulate AI at the technology level, you're regulating math. And of course we're not going to do that. Like that doesn't make any sense. And you'll recall that what they said was, no, actually we can classify math. We can classify math. And literally this was. This is verbat. This is verbatim. This is, this is. We, we did. We, we cl. We classified a whole entire areas of physics in the nuclear era and, and made them state secrets, like of the, of the, like theoretical physics. Yeah, science, physics. We totally classified them and made them state secrets and that research vanished. And we are absolutely capable of doing that again for AI. We will classify any area of math that we think is leading in a bad direction and it will end.

Now some people might think that this is a bad thing for certain areas to be classified. I don't think this is a bad thing at all. If we even remember what some of the leaks of Q Star were even claiming to be, those things could fundamentally just ruin society. Now essentially, let me give you guys a quick rundown of what that is. There was a claim that the Q, which is the infamous paper and the reason that this letter was such a wake up call is because this letter, if the claims were true, which is 50, 50 at this point, because there was a letter written to the board, we just don't know if this was the letter. But the problem was is that if this letter is true, and let's say that an AI model did manage to figure out, okay, a text only attack that could essentially manage to decrypt the entire world's encryption system, this would fundamentally break society. If we look at how encryption systems work, this would mean, you know, banks, websites, passwords, all of these things could be unlocked by A certain AI system. Now, of course, if this is true, and if OpenAI or whatever advanced lab was managing to develop this, then of course that research would surely be classified. And I do think that that is a good thing. And before you guys think that, okay, why would that be a good thing? Just imagine for a split second someone gets their hand on that kind of information that in the slightest is isn't a good person, and they just purely want bad things. You have to understand that kind of technology escaping a lab is going to have severe ramifications. And I can't even fathom what that could even look like. Now, of course, there are other areas that could be impacted, but the point here is that the nationalization of labs is not going to be something that is never going to happen and is always going to be on the cards. And if you think this hasn't happened before, as they spoke about in the interview, during the Cold War, certain areas of physics research, especially those related to nuclear technology, were classified. And of course, like I said before, these were classified for several reasons, of course, mainly due to safety. And I can imagine the similar scenario happening with AI. And as Leopold Aschenbrenner said, a former OpenAI employee working on superalignment aligning superhuman AI systems said that, you know, as the race to AGI intensifies, the national security state will get involved. No startup can handle superintelligence. And of course, he said, here I find it an insane proposition that the US Government will let a random San Francisco startup develop superintelligence. Imagine if we had developed atomic bombs just by letting Uber improvise, basically stating that, look, if we get superhuman systems, there is no way that the US government isn't going to have some kind of control over that. So just be prepared for that when it comes. Now, this is a piece of AI news that kind of went under the radar, but I managed to catch it because I was scouring the Internet for this, okay? And this is Samsung's new AI sketch to draw. Now, this doesn't seem that crazy, but. But to me personally, I think this is largely one of the largest pieces of misinformation that you could probably actually have. Now, I know you might be thinking, what on earth are you talking about? This is just a sketch to draw tool. This is a very basic tool. But I think that this is the kind of tool that people will use to generate actual misinformation. And I'm going to show you guys why, okay? So essentially, if you don't know what this tool is Basically, you, you can take an existing picture and then you can see right here that the person has sketched a bee onto this image. Now, once they've sketched the bee onto this image, AI does its thing and then it now implants an actual, you know, creature that looks photorealistic and embeds it into your image. Now, the reason I've said that this is a lot crazier than what most people think is because midjourney and other AI generated images, they generate AI in the entire image. Sometimes it looks realistic, some images look weird, and they don't contain snapshots of the real life. But if we can now input AI generated things into a real image, then we're going to get some pretty realistic stuff. And I think this is going to be even crazier when it comes to video. So if you can composite AI generated stuff in video, I think that stuff is going to be a little bit crazier than most people think. And let me show you guys why. Okay, so this person took a picture of this park and they managed to quickly, within seconds, draw a boat here. And they were able to literally get that boat into that frame and it looks pretty realistic. Like if someone sent me that picture, I'd believe that the boat was there. If they said, oh my God, I can't believe this boat was there. Now, of course, you might be thinking, okay, this is just going to be used for sending pictures to loved ones and of course, families and friends. And. And usually when you do have applications like this where you are able to generate AI generated images, especially on a smartphone, there's usually some kind of digital ID so that when that image is published to any major social media platform, it's usually immediately tagged. The problem is, is that I do think people are going to develop software like this so that you can actually change what's going on in a real AI generated image or even in a real actual image app. The point I'm trying to say is that someone could actually take an image of something real that's happening and then doctor the image immediately and quickly and state that this is the real image of what actually happened. Now, I don't think that's going to happen with the case with the Samsung phones, because obviously it's on device. But I do think that once we do start seeing some other tools developed that are more so video and in other realistic ways, it's going to be similar to midjourney developing super realistic Snapchat photos. In the last piece of news, some people are waiting for the AI bubble to burst, Although I don't think so. I do think that, yes, there are hype cycles and there are areas where AI might be overvalued, but a bubble is far too early to call. If you've seen what happened in the late 1990s before with the Internet bubble, people were literally just putting.com to their website and they were gaining billion dollar valuations. I think the AI market might be overhyped in just certain areas, but we can't deny the fact that this is truly transformative technology that's going to radically change the future. And you have to remember people have been talking about AI for decades. People haven't been just recently jumping on the bandwagon. Whilst, yes, there has been more investment and a lot more eyes on the scene, the narrative of what this technology is going to do hasn't changed for the past 20 years.