Sam Altman Says "Ai Progress Will Be IMMENSE From here, Llam 3.1 BEATEN Already! Google news..

From Sam Altman stating that from today, AI progress is going to be immense. And to Google releasing a new incredible announcement regarding their Frontier models and even another model that surpasses Armor 3.1, an open source model. There's tons of news that you probably did miss during this week in AI. So one of the first pieces of news that we actually got this week that's rather fascinating is SearchGPT. If you didn't know, there has been a real demand for products that can actually search the web much better than traditional search systems. If you've ever used Google, you'll know that it currently just isn't that good anymore. And now essentially SearchGPT is basically searching with an AI system that browses the web and finds the easy relevant sources that are just what you need. So you can see right here, you can search for basically whatever you need and you're able to find everything instantly. Now, I do think that this is largely the future of search. And Google did try and implement this on, although to their credit it wasn't ready for release, but they still tried to release it in competition with some other people. But essentially what's crazy about this is they are actually testing SearchGPT with a small group of users and publishers to get feedback. While this prototype is temporary, we plan to integrate the best of these features directly into ChatGPT in the future. And of course, if you're interested in testing the prototype, you can actually sign up for the waitlist. I can see here basically how this works, and I'm going to show you all a tool that you can use right now, which is actually as effective as SearchGPT. For those of you that want to use something that's quite like this, because there are several online tools that you can now use and one of them is actually quite popular. But I do think that we're likely to see more and more tools like this be popular in the AI space. Because one thing that AIs are good at, you know, current large language models is that they are good at summarizing large pieces of information. And what is search? Search is lots and lots of different pieces of information. And, and if you summarize it, you summarize the links, you do get something that is rather effective. So you can see right here someone searching for music festivals in North Carolina or whatever. And you can see this is the kind of search that goes on here. Now some people actually have gotten access to this. So this is going to be something that, you know, you do want to check your email if you did sign up because some people actually did get access today. So if you did sign up, check your email because you might have access today. But I'm going to show you guys a quick tool that you can actually use that is actually as effective as this. In fact, I do think it is a little bit more effective than this. For example, I can ask it completely anything. So I can say, what is the current state of the art LLM model? And I can just question mark and I don't even need to spell anything right. You can see it's researching, it's acquiring some sources and you can see the top large language models of 2024. And it's giving me all of these results, which is really cool. Most of you probably would have heard about Perplexity by now, but you can see right here that it gives you links where you can copy and paste them if you want to credit the sources of the. Of course, if you want to be like, okay, where did that source come from? Click that link there. And then of course what you can do is you can just copy all of this text. So you can copy this, paste this into an article, you can share this, you can rewrite it, you can say, okay, rewrite it with this model, that model, I mean, and you can continue going on with other questions. This is literally what I use every day. And I do think that once GPT search gets to this level, most people will probably opt for this other than ChatGPT. Whilst yes, there are some things that are bad with this, but I do think that if you're actively doing research, like on a topic, or if you're doing research for something that's a little bit obscure and you want to ask a direct question, this is something that's really good because sometimes you ask a direct question to Google and the search results just, just, they just don't help. Now Google actually did release a bunch of news this week. One of the things that they actually did release was 1.5 flash in Gemini. Now the reason I'm actually talking about this is because 1.5 flash in Gemini is actually free. And this is cool because this is a really quick model that has a four times longer context window and on the app this thing is blazingly fast. So if you, if you've ever wanted to use Gemini and you didn't want to sign up for their Pro subscription, I know most people don't have many AI subscriptions like myself, but if you did want to sign up for that subscription, you can now go ahead and get Gemini Flash for free. So you should be able to use that in your apps, available in 230 different countries today. Then we had Google presenting their amazing stunning Alpha Proof and Alpha Geometry 2, which literally did the impossible for certain people. So essentially they're presenting the first AI to solve international mathematical Olympiad problems at a silver medalist level. Now, I'm actually going to do a longer video on this because the implications of this are actually staggering. As in, I was doing some research and I found out that, you know, this connects to that and that connects that. And this is like, this is a huge deal, like, not to state that, you know, most people, you know, like this got 1.4 million views, a few decent retweets and stuff like that. But the point is, because it's not an actual product that people can go ahead and immediately use, this is something that people are quite underestimating. But I have to tell you guys that this is truly, truly staggering. The fact that this got a silver medal and was only one point away from a gold medal is insane. Because one of the key predictions that most people had on Metaculus, which is a forecasting website, was that they were basically wondering, are we even going to get a bronze medal by 2025? And AI managed to get silver, basically near gold in 2024. So this is something that is rather insightful because it gives us an updated perception on our AI timelines. Now it's a new frontier in mathematical reasoning. And Google has been doing some incredible research, which I will dive into another video. I'll probably release that one in the next couple of days. But just know that things are about to get crazy. And this is where Sam Altman also shares the same view. AI progress will be immense from here, and AI will be a critical national security issue. I wrote an op ed for the Washington Post about why the United States needs to maintain its lead in developing AI rather than leave a vacuum for authoritarian government. So this is the article here. Who will control the future of AI? A democratic vision for artificial intelligence must prevail over an authoritarian one. Now, this entire article is by Sam Altman, and he says that, you know, this is the urgent question of our time. The rapid progress being made on artificial intelligence means that we face a strategic choice about what kind of world we're all going to live in. Will it be one in which the United States and allied nations advance a global AI that spreads the technology's benefits and open access to it, or authoritarian one in which nations or movements that don't share our values use AI? To cement and expand their power. There is no third option, and it's time to decide which path to take. The United States currently has a lead in AI development, but continued leadership is far from guaranteed. Authoritarian governments are willing to spend enormous amounts of money to catch up and ultimately overtake us. Russian dictator Vladimir Putin has darkly warned that the country that wins the AI race will become ruler of the world. And the People's Republic of China has said that it aims to become the global leader in AI by 2030. Clearly, what we're stating here, seeing here, is that these other global leaders truly do understand the true value that AI has. And I'm not saying value in terms of it providing to distribute knowledge and all those great things that me and you can benefit from. I'm talking about the fact that having access to a system that is far more capable than anything we've ever had before is something that any of these leaders of other nations are going to be speeding towards. There is no way that they are going to be slowing down. And we have to find a way to ensure that we actually do maintain this lead or else things could get dark pretty quick. If they manage to take the lead on AI, they will force US companies and those other nations to share user data, leveraging the technology to develop new ways of spying on their own citizens, or creating next generation cyber weapons to use against other countries. Whilst that might seem like some kind of strange future, I don't think that this stuff is far off. I mean, imagine a superintelligent system that's able to design weapon systems and new technologies that are able to cripple nations and even destabilize economies and countries with a few lines of code. These are things that are genuinely plausible. I mean, it was only the other day that we literally saw CrowdStrike cripple certain parts of the economy as things granted to a halt just because of a simple software update. Imagine if some cyber attacks were able to do something on a lot larger scale from a foreign nation. These are things that we truly do have to think about. And one of the things that already does happen is that a lot of these authoritarian governments already do spy on their own citizens. And trust me, you don't want that kind of technology proliferating through the hands of those in power. Now it says that the first chapter of AI is already written. Certain systems such as ChatGPT and Copilot are functioning as limited assistants. But of course, more advances will soon follow and will usher in a decisive period in the history of human society. Now There are a lot of things that are going on here and even discusses saying that I've spoken in the past about creating something akin to the International Atomic Energy Agency for AI. But that is just one potential model. One option could knit together the network of AI safety institutes being built in countries such as Japan and Britain and create an investment fund that countries committed to abiding by democratic AI protocols could draw from to expand their domestic computer capacities. And there are just many different things here that he discusses about the future of AI. It is going to be really hard to kind of figure out how this actually works, because I think that this is a battle for power. As many have said, you know, whoever controls these superintelligent systems will control the world. And it was even recently I was making a video where someone who predicted accurately the past few years of AI development basically said that once you get to AGI and once you get to ASI shortly thereafter at in a year or two, it is very likely that you're going to have godlike powers over those who don't. I mean, all sorts of crazy implications there. Now, we also got Audio Flamingo from Nvidia, and this is a model that understands audio beyond transcriptions. Here's Nvidia's Audio Flamingo model. And of course, it says, turn up the volume to just listen to Audio Flamingo narrate those scenes in a place.

Beyond imagination, where the horizon kisses the heavens. One man dares to journey where few have ventured. Armed with nothing but his wit and an unyielding spirit, he seeks the answers to mysteries that lie beyond the stars.

I'm not sure how much that audio you did here, it was actually quite low. But essentially what we have here is a new AI tool that can actually truly understand audio just beyond the transcriptions of that audio, which is different because it allows us to understand audio on a different level. So Audio Flamingo essentially narratives the scenes and basically tells you exactly what's going on. Now, it's not just something that describes the audio to you. You can essentially say, okay, what type of scene could this voice be used in? So it's able to describe exactly how the voice is, how it works, and it's got that understanding of what kind of voices should be used in what kind of scenes. And of course, what's really nice about this first part, something that I really like here, is that it says, what do you mostly hear in this audio? And it says you mostly hear the sound of a man speaking intermittently through the audio. Now, something that I thought was really cool. Okay, was that this was actually able to briefly describe the background noise and it basically says, look, the background noise is continuous throughout the audio and that it could be the ambient sound of an outdoor environment. So sometimes you might not hear something in an audio clip and you might be able to ask audio Flamingo, hey, what was it? I heard that at this sound. What sound is this? What should this be used for? Should this sound be used here? So I think this kind of model is actually rather fascinating and it's kind of thing I actually didn't really see coming. And it's actually really cool from Nvidia. So I wanted to include that in today's video. And then of course we have Elon Musk saying that Xai's new supercomputer in Memphis was installed in just 19 days and will be used to train Grok 3, which is expected by December and will be the most powerful AI in the world.

So we really need, we have a lot of catching up to do to companies relative to companies that have been around for five or 10 or 20 years. We're catching up fast. I think with the blast of improvement of XAI is faster than any other company out there. We just completed the. We just, we were just able to install and bring online a massive new training center that we like as mentioned, we're building in Memphis and it's from getting hardware installation to it. Beginning training was only 19 days and that's the fastest by far that anyone been able to do that. So we're moving quickly, but we're still catching up. Grok2 actually finished training now. Grok2 was training with roughly 15,000 GPUs and there H100. So Grok2 finished training about a month ago. We're doing what's called fine tuning, fixing bugs and whatnot. So we'll release Grok 2 which will be should be on par with or close to a GPT4 and that hopefully we release that next month. Then what we're doing in the Memphis data center is we're actually training Grok 3. So that'll probably finish training in about three or four months. And then there'll be some fine tuning and bug fixing and whatnot. And we're hoping to release Grok 3 by December. And Grok 3 should be the most powerful AI in the world at that point.

I mean, Elon Musk is truly actually stacking up his compute stack. You can see here that he actually says nice work by XAI team and the X team and Nvidia at supporting companies getting the Memphis supercluster training started at 4:20 in the morning. And of course, he says with 100,000 liquid cooled H1 hundreds on a single RDMA fabric, it's the most powerful AI training cluster in the world. So this is going to be pretty crazy because a lot of the people that are at these frontier labs have realized that, yes, compute is going to be the new oil, and being able to rapidly train models and successive iterations of those models is going to be key in maintaining a competitive edge in AI. So I do wonder if this is going to stay the same, because of course we do have Microsoft's computers coming online very soon. They are slowly building up their compute. I've got a video on that. There are some really key details that many people did miss, but they are also building up an immense AI training cluster. So I'm wondering at that time where these supercomputers are completely built, how often are we going to get model upgrades? And I think that is going to truly accelerate the timeline because previously GPT4 took a really long time to train. But once we have these supercomputers online and we've got increased levels of efficiency in these chips, we're going to be going from months to train a model to just two months or even one month, which would mean that there are just going to be a lot more models being trained and that's going to give us a quicker cycle in terms of product releases, iterations of generations of models. So it's going to be really interesting to see how that all plays out. Now, this is something that most people did miss, and this is unfortunate because I did test this model and it's good. Like really good. Okay, so it said today we're announcing Mistral Large 2, the new generation of our flagship model. And compared to its predecessor, Mistral Large 2 is significantly more capable in code generation, mathematics and reasoning. And it also provides much stronger multilingual support and advanced function calling capabilities. So Mistral Large 2 is insane. Like this is insane, because if you aren't aware, this is an open source model that was released the day after llama 3.1 70 billion parameters was released. And of course llama 3.1, 405 billion parameters. You can see right here that Mistral Large actually has less parameters than llama, the newer Llama, but you can see that it actually performs a lot better on those tasks, which is insane because it means that you get a more effective model on better tasks, which is just effective overall. So this is Just absolutely incredible. Now I've used this on some personal benchmarks for certain tasks that I do and some models that usually struggle to get them right. I'm surprised that these smaller models are actually starting to ramp up in terms of reasoning. And it's fascinating because sometimes I would look at these 70 billion 8 billion parameter models and I'd be like, ah, these models are only good for asking them how they're doing. Sometimes I ask them these super difficult questions with multi step reasoning and sometimes they manage to get them right. I'm like, okay, this is going to be pretty crazy because now I have a model that is so much cheaper than anything out there that this means that I can use this a lot more effectively. I can put it into certain pipelines and workflows. So Mistrolarge 2 is probably right now one of the most underrated models that I would say is out there. And you can see right now on the Humaneval, it does really well. It's arguably just right next to GPT4O. You can see right there on some of these other benchmarks that it also is rather effective as well. So you can see right there on the average, it's Basically just under GPT4O on Python, C, Java, TypeScript, PHP and all these things. It's actually really effective at coding as well. So this is going to be super interesting to test and see how it does, see how effective it is. But by far, guys, I would definitely try and test this out on whatever AI hosting platform you're on. It might be po, it might be Perplexity, it might be Grok, but Mistral's large two to open source is going to be insane. And not only is this great, but I know that there are going to be some fine tuned releases of this. There are going to be some people that go out and fine tune this. So I'm excited for what the ecosystem does with this. And then here we have Mark Zuckerberg on the billions of AI agents.

I think we're going to live in a world where there are going to be hundreds of millions of billions of different AI agents, eventually probably more AI agents than there are people in the world. A lot of what we're focused on is giving every creator and every small business the ability to create AI agents for themselves, making it so that every person on our platforms can create their own AI agents that they want to interact with.

And if you think about it, these.

Are just huge spaces, right? So there are hundreds of millions of small businesses in the world and one of the things I think is really important is basically making it so with a relatively small amount of work, a business can basically a few taps stand up, an AI agent for themselves that can do customer support, sales, communicate with all their people, all their customers. I kind of think that every business in the future, just like they have an email address and a website and a social media presence today, I think every business is going to have an AI agent that their customers can talk to in the future.

And that future of AI agents being there in the future, I don't think it's that far away. And I think it's going to be as normal as just having a social media account. And this is why I think the future might just be, you know, billions and billions of AI agents just all interacting with, with each other, you know, based on every single person who's on social media or every single business and they're just interacting and exchanging information. I think it's going to be a super effective economy and it's going to be really interesting to see how it works now. Also, if you didn't know cling, which is the text to image or image to video model, or text to video model, this model here is actually now globally available. So you can actually just create an account with cling and you.