Massive AI News: Google TAKES THE LEAD! LLama 4 Details Revealed , Humanoid Robots Get Better

So with another incredible few days in AI, let's take a look at some of the stories you really do want to know about. Because although people think the industry is slowing down, a few key stories from today's video and you'll see that things are not slowing down at all. Now, one of the things that we got from the Meta 2024 earnings call was that they're aiming for Llama 4 to be the most advanced AI model in the industry in 2025. It's going to be trained on 10 times as much as compute as Llama 3. Zuckerberg said he would rather build too much compute than not enough. Now, if you know why this is crazy is because, you know, previously when Mark Zuckerberg actually spoke about Llama 3 was that he said that, you know, when they were training the model and they were doing their tests, that the model was still getting better. And essentially they basically said that, look, we just decided to release the model, but we still could have made this better. Now, I do want to state that in if this is the most advanced model in the industry in 2025, then that would mean that this model would have to outcompete Google, outcompete Claude Anthropic, and Of course dethrone OpenAI, who's going to be releasing GPT5. So it seems that the AI race is truly heating up. Because if that is their goal, and you remember previously they were just like, okay, we're going to release AI models. But now they're like, look, we want to have the most advanced AI model in the industry. This is a completely different thing that.

They'Re stating, figuring out the right level of infra capacity to support training more and more advanced models. Llama 3 is already competitive with the most advanced models and we're already starting to work on Llama 4, which we're aiming to be the most advanced in the industry. Next year. We are planning for the compute clusters and data we'll need for the next several years. The amount of compute needed to train Lama 4 will likely be almost 10 times more than what we used to train llama 3. And future models will continue to grow beyond that. It's hard to predict how this trend, how this will trend multiple generations out into the future. But at this point I'd rather risk building capacity before it is needed rather than too late, given the long lead times for spinning up new infrastructure.

So there you have it. They are focusing on building Llama 4 and continuing the AI development. So for those of you that think, okay, this is slowing down. It's clearly not. I truly do believe that they can get it done. But I'm wondering how on earth are they going to achieve having the most advanced AI model in the industry in 2025? Not the Meta, can't get it done. But we've seen recently there's almost been this sort of convergence around this new GPT4 level. But I'm really wondering if GPT5 and Gemini 2 and Claude 4 managed to pull ahead. Can Metaslama 4 manage to once again up the stakes again? So we've seen them catch up before, so I mean this time maybe they're going to pull out some new stops, some new architectures, some new different ways to probably reason to where this is going to be possible. Now I'm not sure if this is going to be open source. I do remember that previously there were some internal disagreements about open sourcing this new model. So I do wonder if that powerful model is, will actually be open source at all now. In more prediction news, OpenAI's director says that artificial general intelligence may be 5 to 15 years AI. So if you aren't familiar with OpenAI's board of directors, one of the board directors is Adam d'. Angelo. He's the CEO of Quora and he made a prediction during an event last week and he said that the advent of AGI will be a very, very important change for the world when we get there. Hence the reason I've created my post AGI economics community. Because if there's going to be some crazy event in the future where a powerful technology is going to change the world, it seems like you might want to prepare for that scenario. Now he says, of course, following the reports of earlier where OpenAI developed the new way to track its progress towards building AGI, the company sharing a new five level classification system with its employees. If you aren't familiar with that internally, OpenAI had this. You can see level one chatbots, AI with conversational language, level two reasoners and human level problem solving and level three agents and systems that can take action. So you know, I've only spoken about the first three here because we're quite far away from level four and level five. And I think, you know, what you can truly see here is that even before we get to AGI, you can see that having human level problem solving is going to be a complete game changer and having systems that can actually, you know, take actions, these are going to be really, really game changing events when they actually are surpassed. Because having systems that can take actions and having, you know, even on the latter end of things, you know, innovators that can aid in invention, an AI that can do the work of an organization. This is going to be absolutely, you know, incredible level stuff that most people just aren't even comprehending due to the current, you know, nature of generative AI, where there's some hallucinations here and there. So I do think that once these levels are surpassed, I think even before AGI, things are going to get really crazy. Now I do think that this prediction right here is actually within the realm of possibility because it's not one that is too crazy and it's not one that is actually too safe in terms of bets, because he does say within five to 15 years. And you have to understand the window between five to 15 years is basically between 2030 to 2045. And if we've seen anything, you have to understand the AI development within the next five years. Considering what some people are predicting for 2030, we have to understand that the next five years of AI development are likely to be as twice as fast as the prior five years, considering the fact that now we've not just got a little small San Francisco startup that's working on this stuff, we've got literally all of the world's top labs working on generating, researching better ways to make intelligence. And there's a lot at stake here. So I would argue that within five years, could this happen? Most certainly. It most certainly could happen. But it will be interesting to see if there are any major roadblocks on the way to AGI and how those come about. Now in terms of actual developments regarding new forms of intelligence, there is exciting news from the chatbot Arena. Now. Google DeepMind's new Gemini Pro experimental 0801 has been tested in the arena for the past week, gathering over 20,000 community votes. And for the first time, Hugemini has claimed the number one spot, surpassing GPT4O Claw 3.5 with an impressive score of 1300 and also achieving number one on the Vision leaderboard. The Gemini 1.5 Pro excels in multilingual tasks and delivers robust performance in technical areas, hard prompts and coding. So this is incredible. You can see that Google's new model, their 0801 and notice they're not calling this Gemini 2, although it very well could be, but it's just Gemini 1.5 Pro, an experiment. I'm guessing that they're, you know, doing some additional reasoning on top of the already made model, because maybe they learned what anthropic did, because Anthropic did some kind of. I'm not sure if they did post training, I'm not sure if they added some architecture, but. But one of the things that we know for sure the Claude 3.5 sonnet does is that it thinks really hard about its problem before generating response. If you weren't paying attention at the time when Claude 3.5 Sonnet was released, there was basically a prompt where you could ask Claude to show its thinking and it would show its internal thoughts before actually emitting a response. And I'm not sure if it was that alone that put that. And I'm not sure if it was that alone that allowed Claude 3.5 Sonic to just deliver so much better performance in terms of its reasoning ability. But we do know that Google has definitely done something very similar, as they've not, you know, gone ahead and trained an entirely new model from scratch, but they've, you know, experimented with Gemini 1.5 Pro, and for the first time we've seen a GPT4O truly defeated by a decent margin. Now, I do wonder how long it is going to stay there for, because sometimes, you know, models get there and then they don't stay there for quite some time. And, and usually what we will see is from OpenAI that they usually manage to release a model within like the next two weeks and then they try and immediately reclaim the number one spot. But I do have to be honest with you guys, I've been testing this in the Google AI playground and it doesn't fall for all of your crazy tricks that traditional LLMs might fall for. For example, one of the common things that these AI systems fall for is what's bigger, 9.9 or 9.11. You can see that when I ask this model, you can say, you know, 9.9 is bigger than 9.11. And of course it says, think of it like this. Both numbers have the same whole number. Part 9, 9.9 has 9 in the 10th place, while 9.11 has a 1. And since 9 is greater than 1, 9.9 is larger. So therefore 9.9 is greater than 9.11. Now, of course, some people could argue that this may have been in the training data because this was, you know, on Twitter at least, a new widely discussed benchmark for whatever reason. But I do think that this model is actually a lot better because I've tried a lot of trick questions and unfortunately, not really unfortunately, but surprisingly, they just haven't worked. So I'll be surprised to see where this model does get stumped at in terms of its reasoning ability to answer certain questions, because I even had certain questions on personal projects that I would only use Claude 3.5 on. And this model, whatever this model is, it now actually manages to complete those questions really easily. So I would definitely say that if you have any really hard questions that you struggle with or that you're using for Claude, try it out in this experimental version in the Google AI Studio and see if it's able to do it even better than Claude 3.5 Sonnet, or at least on par. Now, one of the craziest things that most people also did miss, and I nearly missed this because I did actually look at the main update.

So.

But one of the things that was crazy was that Gen 3 Alpha actually added image to video. Now, I think text to video is incredible because it's just a text prompt, and you're getting a video at the end of it. But the reason that image to video is so crazy is because this allows so many different use cases, because now you're using images to actually steer what is being done by the video generation model. Now, this has, you know, remarkable use cases because so many people are now using images from midjourney to be able to animate them. And one of the things that I've seen time and time again from Runway is that this model, for whatever reason, is able to simulate not physics, but certain liquids and other things in a very coherent way. I'm going to show you guys a few examples of Runway's, you know, image two videos. But it's remarkable at how the physics engine is. I'm not sure how it's being done, since it's completely generative, but I gotta be honest, you know, like, for example, this one right here, you can see exactly what this looks like, and it's really effective. But I'm gonna show you some of the more popular examples that were floating around Twitter. So, for example, if we take a look at this image on the left, you can see that someone's probably taken this image from a window or on top of a rooftop. You know, for all I know, it could be AI generated. But if we take a look at the Runway generation, you can see that the simulation of fluid coming from behind, whether it be a tsunami or a wave, it's remarkably accurate in terms of how the fluid manages to come over the building. Now, the reason that I find this so impressive is because one of the common mistakes of generative AI was that things usually don't manage to mesh together well, you see objects passing through objects, you see objects just not going to law with the laws of physics. And this was basically something that many people and many critics of AI said that generative AI would never get. Well, so I'm not sure what architecture they're using, but this is actually quite similar and quite on par with SORA in terms of how we're seeing the water manage to bend around certain objects. Now what's so interesting about this as well is that one of the things that I've seen from a few industry people when they're speaking about this is that they said that this kind of tool is seemingly like a kind of vfx, because like I said in previous videos, if you've ever worked in fluid simulation in cgi, you know that trying to render something like this using traditional CGI methods, it takes serious, serious compute. And it's not to say that, you know, AI didn't take computer train, but if you wanted to do something like this on a standard computer, you'd have to probably, you know, leave your machine on for a few days in order to get actually good looking fluid simulations or put it on a kind of render farm. Because simulating all of these particles is really, really difficult and compute intensive for a single system. But if you have things like runways, Gen 3 Alpha, where you can literally just generate image to video, this is going to be something that I think would allow people to explore new forms of VFX almost immediately. And Runway did even mention that they're making this model as twice as fast with a new recent update that they're rolling out. So I would definitely go ahead and test this on certain things. Some people are putting a liquid over various images and so far it looks really amazing. So this is some remarkable stuff.

Hello, my name is Menti Bot, how.

Can I help you?

Hey Menti, please get the shopping cart and come with me.

I got your instructions, I'm going to take the cart.

It. Sam. Thank you, Mentee.

You are welcome. It was my pleasure.

So that was the demo for the Menti Robotics robot. This was a short demo showing us how the future of humans and robots can potentially collaborate. I think this one was rather nice because it actually shows, you know, what a lot of people do miss about this technology. A lot of the times some people do forget that there are individuals that are less fortunate in terms of their abilities or disabilities. And humanoid robots will actually help these people quite a lot. A lot more than you might think. And it's definitely going to give them a new lease on life in terms of being able to do tasks and that they may not have previously enjoyed before or may have found difficult. So I always find the fact that, you know, a lot of people sometimes hate on AI, and it's understandable. It's a technology that for the most part might just replace your entire career. But on certain edge cases, you know, for example, using GPT4O with vision, you know, having a humanoid robot able to do difficult tasks for those people who are not especially mobile, and as people manage to get older, sometimes you can develop a chronic issue. I think humanoid robots that are running on energy and able to jump up and grab this and run around and do this, I think it's going to make people's lives a lot.

Project Root is Nvidia's moonshot initiative at building a universal AI brain for all kinds of different humanoid robot platforms. And at this siggraph we are introducing a set of tools for developers in the humanoid robot ecosystem to build their AI models better and more efficiently. And this time we introduce a new synthetic data generation pipeline. We start with human collected demonstrations using a mixed reality device like Apple Vision Pro. And then we multiply that by 1000x or more using Nvidia's suite of simulation tools like RoboCasa and Mimicchen. At Jensen SIGGRAPH keynote, he introduced the three computer problem, the DGX, OVX and AGX. Basically, we use DGX, the main workhorse computer, to process lots and lots of videos and text to train the multimodal foundation model for robots. And we use OVX to run the Nvidia simulation stack such as Robocasa, Isaac Lab and mimicgen. And in ovx we can multiply the real world data by at least 1000x using our simulation tools. And then once the model is trained, we will deploy that to the AGX computer so that we can test the model on the real robot and on edge computing devices. With these three computers, we will be able to enable the developers around the world to build better AI models for humanoid robot hardware platforms. I think this year is the year for humanoid robots. We have seen lots of new state of the art hardware springing up in the ecosystem. And we have seen the emergence of multimodal foundation models that can form the AI brain for these humanoid robots. And with Nvidia's exciting developer tools and simulation suite, I believe we are one step closer to solving the AI brain for humanoid robots.

The era of physical AI is here. Physical AI models that can understand and interact with the physical world will embody robots. Many will be humanoid robots. Developing these advanced robots is complex, requiring vast amounts of data and workload orchestration across diverse computing infrastructures. Nvidia is working to simplify and accelerate developer workflows with three computing platforms. Nvidia AI Omniverse and Jetson Thor+ generative AI enabled developer tools to accelerate Project A General Humanoid Robot foundation model. Nvidia researchers capture human demonstrations, seeing the robot's hands in spatial overlay over the physical world. They then use RoboCasa, a generative simulation framework integrated into Nvidia Isaac Lab to produce a massive number of environments and layouts. They increase their data size using mmiqgen nim which helps them generate large scale synthetic motion datasets based on the small number of original captures. They train the Groot model on Nvidia DGX cloud with the combined real and synthetic datasets. Next they perform software in the loop testing in Isaac, sim in the cloud and hardware in the loop validation on jetsim Thor before deploying the improved model to the real robots. Nvidia OSMO Robotics Cloud Compute Orchestration Service manages job assignment and scaling across distributed resources throughout the workflow. Together, these computing platforms are empowering developers worldwide to bring us into the age of physical AI powered humanoid robots.

Foreign. So that was Nvidia's video where they basically talk about how we're going to scale humanoid robots with the Omniverse. And by the looks of things, this, you know, is pretty crazy. We have an entire system that's set up for us to be able to collect data with the Omniverse. And of course you can see the Omniverse cloud, the DGX cloud. We've got Project Groot, we've got. And with all these systems, with all these NIMs and autonomous things going on, we do have a situation where the humanoid robots are most certainly coming. I mean if you watch my previous video, I spoke about how figures new robot is coming in a few days and they're going to demonstrate exactly how effective it is in terms of its agility, manipulation and end to end ability. So I mean this is definitely one of the areas that I'm most personally excited for because seeing a robot, a real human robot interact with the real physical world, I think this is where people are really going to start to see that, okay, maybe this AI robotics thing is coming sooner than I thought. Now, in more robotics news, we did have the German robotics company Neura release a video of their humanoid robot 4NE1 performing tasks and the robot's concept was unveiled in 2022 and it's one of the first to join the early Access Nvidia humanoid robots developer program. So coming and if you've been been paying attention to, you know, exactly what's going on, you will have seen that China have been developing an insane amount of robots. And I know that this is of course for anyone, this is not a Chinese robot at all. It's actually German. So it's a little bit different there. But I do think that like people cannot fathom how many humanoid robot platforms are being built and just how many robots there will be in this world doing things in the physical realm that they just, you know, definitely didn't do before. And it's just quite shocking. Every time I see a new humanoid robot platform, I'm just like, wow, another one, another one, another one that's capable of doing different things. And I think the future is pretty clear now on where we're going to be in the future. It's just, you know, maybe 10 years, 20 years. But I think the future is going to look completely different when we have these robots just, you know, walking around and doing, you know, a million different things. Then we also had some more unhobbling. And this is an easy trick to improve your LLM results without any fine tuning. So many people know few shot prompting or chain of thought prompting. And a new better method was presented by these researchers at ICML 2024. It's called Plan like a graph and I'll let them explain it because they're going to do a better job than me.

Our class is basically benchmarking how well the large range models do and asynchronous planning. And so say you want to make a breakfast and you want to make a coffee, fry an egg and make a toast. Maybe the best strategy for you is to do these different steps at the same time. But say you want to make a coffee and you want to grind coffee beans and then buy the coffee. So you test that dependencies for some of the subtasks. So I guess for our setting we are mostly interested in given a complicated task as such, where you have to parallelize some tasks and sequentially execute some steps, how can you derive the shortest possible time for these complicated tasks? So we propose a method which we call Plan like a graph, which is basically telling the links model to do complicated planning like a graph by generating a graph first and then solve the test. And we find that our methods are actually outperforming the baseline methods across all the models. We Implemented, evaluated here, and also have a greater improvement over all capacities here. So in general, our classes are really good as an off the shelf prompt engineering method. So no training at all, only front engineering. So if you're interested, you definitely should try.

Now, the reason that this is rather important is because, you know, current LLMs do struggle with planning. Like planning was one of the things that they just don't get very well. And this allowing models to plan a lot better is, you know, it's an insight to the fact that we still are unlocking new methods to increase the capabilities of these models, which means once again that there's most likely still a huge reservoir of untapped potential in these models in ways that we don't understand. And it basically talks about. These models would initially struggle with asynchronous planning, where for example, if you're going to bake a cake, for example, you're going to preheat the oven. And initially a robot would just preheat the oven and then wait 10 minutes and then move on to rolling the dough, whatever. But a human would preheat the oven, they'd roll the dough, they'd wash their hands, do this, and they'd be able to do it. Planning. All of these things that we pretty much take for granted in the sense that we know how to do this almost second nature. But these models, when they're thinking about how to do it, and in fact, I'm going to show you guys a screenshot from the paper. But basically you can see here, this is just the sequential planning, which would take 65 minutes by just simply adding how long it takes to do every single step. And of course we've got parallel planning where it's like, okay, I can do everything at once, which just isn't true. And then this is asynchronous planning where you set the oven on, you then roll this, and then of course then you manage to get your thing in the end. So I think it's fascinating that we're still managing to find new ways to unlock capabilities from these systems. And here is Mark Zuckerberg also stating something rather similar.

Better, it's not just about what you want it to say. I mean, I think generally creators and businesses have topics that they want to stay away from too.

Right?

So just getting better at all this stuff. You know, I think the platonic version of this is not just text, right? You, you almost want to just be able to. And this is sort of an intersection with some of the codec avatar work that we're doing over time, you want to basically be able to have almost like a video chat with, with the, with the, with the agent. And I think we'll get there over time. I don't think that this stuff is that far off, but the, the flywheel is spinning really quickly, so it's, it's, it's exciting. There is a lot of new stuff to build, and I, I think even if the progress on the foundation models kind of stopped now, which I don't think it will, I think we'd have like five years of product innovation for the industry to basically figure out how to most effectively use all the stuff that's gotten built so far. But I actually just think the, the kind of foundation models and the progress on the fundamental research is accelerating. So, so, so that it's a. It's a pretty wild.

And of course, if you did miss this, this was the friend product that actually broke the Internet. This is an AI wearable. I'm going to show you guys this. I think it's. I think it's pretty cool. You know, I think at least people are outside, not staring down at their phone. But some people hate it. Some people say it's dystopian. But let me know your thoughts about this as well.

I'm out of breath.

We made it.

I don't know how to.

Woo.

Very good. That's fair. All right, let's go.

Let me show you how to game, bro. Okay.

Oh, come on.

Come on.

Oh, let's go.

Serious? Come on, Man, I hate this game. Take notes, baby. Oh, man, you guys suck, bro. You look like the back.

Let's go, let's go.

Dude.

What?

How did you do that?

I know, the effects are crazy. It's dank. I could eat one of these every day.

Oh, sorry. I got you messy. It's really nice up here. How'd you find this place? I don't know.

I just kind of like to come up here to be by myself. I've never brought anybody else. I mean, besides her.

She goes everywhere with you, right? Guess I must be doing something right then.

I guess so.

Well, see.