BIG AI NEWS: 10,000X Bigger Than GPT-4, AGI 2025, New Boston Dynamics Demo And More

So welcome back to the AI grid. In today's video, there was genuinely some insane news that I didn't think would happen for quite some time. But let's not waste any time and let's dive into all of the stories that you potentially did miss. So the first story that I am excited by is ideogram 2.0. Now, most people I know don't actually know what ideogram 2.0 is. It's their most advanced text to image model now available to all users for free. And I got to be honest, guys, this is not like a sponsored video. They're not, you know, paying me to say this or any of that stuff, but this software is just so good at what it does that I can't believe that more people aren't using this. Okay. And I'm going to show you some examples of why. This is probably, in my genuine opinion as someone who's, you know, done a lot of designs, you know, a while back and someone who, you know, you know, just does designs because, you know, sometimes it's fun to play around with a lot of stuff. You know, I do graphics for different things that I'm working on building sometimes. And this literally is basically text to Photoshop. If you've ever, you know, thought about what AI can actually do for people who don't actually know how to make graphics, and this is something that is, you know, it's different from Dall E, different from Mid Journey and today it went free for everyone, which is absolutely insane. And they introduced an iteration two of their model. And this is something that I think is going to be absolutely insane because it actually does solve a lot of the issues with generative AI. Like when people say that, okay, you know, I create an image and it's not got the right text or it's not got the right prompt adherence. This thing. And I'm going to be doing a more in depth video on this, this thing is genuinely good at following all your prompts and actually creating graphics that are actually usable on a really insane level. If you're a graphic designer, you know, someone working in the design field, the slash creates field, I genuinely think you should be paying attention to this because it's rather incredible. But if you're someone that's just building, this is definitely a useful tool. You can see here that it offers industry leading image generation capabilities. Apparently human evaluations consistently rate ideogram 2.0 significantly better than Flux Pro, which was only recently released. And you remember the comments about that. The reviews were completely raving about how good Flux Pro is. And of course it is really good. But you can see here that on alignment, on the overall preference, and on the text rendering in the arena, or whether they tested this model, it seems that ideogram 2.0 is just miles above what anyone is using. And I got to be honest with you guys, I think the people who made ideogram 2.0 potentially have some really good people on their team because they just know what users actually want. And as someone who's used all of these tools, this one just literally saves you so much time. Now I'm going to show you guys a few examples as to why this is so crazy. So I'm going to show you guys how insane this is. So this was a tweet that I saw from Tools Telegraph, and it said, testing Ideogram AI Design. Okay? The infographics and text were made with a single prompt and magic prompt on. Okay. And you can see right here, the first prompt. This is an infographic that was done. You can see a colorful infographic explaining the benefits of adopting animals. We with icons, charts, and easy to read text. And you can see that it literally manages to get pretty much all of this right, which is really nice. Then of course, a colorful infographic. And one of the things, if you're like, why are you showing me this infographic that AI made? Well, one of the things that people always complained about with AI was the fact that things like this. Let me just highlight this. Things like this, like more flexibility, global talent pool, better work, life, balance, these areas. Usually it would be gibberish. Like, usually what AI would do, be able to get the main text right, maybe like one or two words and then the rest of it would be gibberish. But you can see right here that this is like an infographic that you can completely go ahead and use, which is absolutely incredible and nice. And this is a model that's really, really fast. And not to mention it's free. So, I mean, there's no reason not to use this. And you can see right here the benefits of recycling. You can see that this is a nice infographic. And then there's another one here. How long would this take? I did some creative work before. I've worked with Illustrator before. And this stuff, I don't want to say it would take forever because it wouldn't take forever, but this is like an insane time saver. So, yeah, this is a pretty crazy tool. I know most people are going to gloss over this bit, but yeah, it's pretty Crazy. Now, in other AI image generation news, Midjourney actually are now releasing their web experience available to everyone. Previously this was only to people with a certain number of image generations, but now this is available to everyone. If you don't know what Mid Journey's website is, I'm going to show you guys what it is because it's pretty cool. But this is the web ui and I really like this because if you're trying to, you know, create something, you know, you basically, when you're creating something, you can, let's say, for example, we go to the settings right here. What you actually have is you have a lot more control for those people who can't be bothered to type stuff out. So, for example, you know how the Discord user interface is, you know, confusing and janky and just like you have to type in dash, dash ar, dash dash V, whatever it is, dash dash stylize. You can pretty much do anything you want here. You can see, you know, you can organize your images, you can explore people's different images, you can, you know, see what they've used. And it's just really, really like 10 times easier to use. Mid Journey, you can swap between versions 6.1, you can personalize it, set it to standard versus raw up. The weirdness, it's just, you know, so much more useful than just using Discord because Discord was so slow. And this one is just a lot more effective at getting it to do what you really need it to do. So there was one Wired article that actually spoke about anthropomorphization and emotional resilience. And this essentially explores problems that arise when users perceive AI in human terms, something that's going to be exacerbated by the human, like, voice mode. So basically, during Red teaming, they were basically saying that, you know, OpenAI researchers noticed instances of people were saying language like, this is our last day together. Which goes to show that humans have a tendency to. And the word just basically means that you're putting human like characteristics onto something that isn't human. And you're perceiving it wrong because you kind of want an emotional connection with it. It's just the way how humans work. And it's just humans have a tendency to do that. A lot of times we might see an object, we might see an animal, and then we kind of anthropomorphize the thing and think, okay, this animal is smiling or it's happy. But oftentimes we do get that wrong. So they notice this. And the reason I'm thinking why on earth AI news? Well, if you don't know, GPT4 voice mode should be rolling out soon and with that I think it's going to impact society a lot more than people think. Like I think a lot of people are downplaying how much this is going to impact society. Like I personally believe that this is probably going to be, I don't want to say it's going to be bigger than GPT5 because GPT5 if it like holds with regards to expectations of what that software should be capable of, I mean it's going to be incredible. But I'm talking about in the sense that the average person using an AI system GPT4 voice mode is going to be far more useful than a system that's capable of human level reasoning because the majority of tasks that people use are broad, simple based tasks that you don't need incredible reliability on. But in order to get to that mass market you do need something that is really, really effective and that is exactly what GPT4O voice mode is. There are just so many different applications and I think it's going to be really crazy and I think you know how like when social media came about and you know, people are saying that there's this loneliness and now it destroyed society and how there's all of these, you know, extra issues. I think there's going to a lot of issues that are going to, I don't want to say tear the fabric of society apart but it's going to make society a little bit weird and I'm not sure if we're kind of prepared for that. But anyways, it's coming. So you know, this is, you know how living with technology. Basically what I'm talking about is like, you know, you're going to have people that are going to get really addicted to this stuff. Like people are already on their screens, doom scrolling addicted to social media and that kind of thing. So I mean when you have an AI robot that you can literally talk to, sounds human, can make funny voices, knows everything, responds instantly, is always there, you know, can play any role, loves you a lot. Like there's even that ChatGpt Dan thing and that was just the base text to speech level. Like I made a tutorial on this because I thought, you know, a few people just wanted to know but the video got 40,000 views so this was like, you know, just a quick tutorial on how to actually use a prompt and it was blowing up on TikTok because people would like want to use chatgpt because there were girls that were feeling lonely and just isolated. And, you know, there was this prompt that basically, you know, turns the AI into like a boyfriend that they all want. So it's like, I mean, it's insane. Like, I wonder how the dynamics are really going to shift. And if you're thinking this is just something that, you know, I'm just talking about, it's just a prediction, yada, yada, yada, guys, you know, some people are already addicted to character AI and this is something that has literally been, you know, very prevalent considering the fact that character AI has month over month gained users and has been, you know, crazy with the younger demographic. Like, you know, people here help with addiction to character AI. This app is causing addictions to people. Why is everyone. I'm addicted. Are you addicted? This can be scarily addictive. Like, it's pretty crazy that this is going to happen because this is like a new medium that people haven't consumed content through, and it's not really content, but this is like a whole new medium that, you know, hasn't really been tested on society just yet. And, you know, with character AI, they're largely probably going to get an integration with GPT4O voice mode. And I can't imagine what, what that's going to do. So, so Epoch AI, and this is arguably one of the biggest stories this week. 21 hours ago, they examined whether the constraints on power on chip manufacturing and on training data or data center latencies might hinder AI growth. Now, the reason that this is rather important to discuss is because many people are wondering, is AI going to taper off in the later years? Are we just on this curve where AI growth looks something like this rather than something like this? And basically what their analysis suggests is that AI scaling can continue its current trend through to 2030. Now, that is of course, something that most people did think. Of course, there are many different bottlenecks and stuff like that. But here's the most insane statement. So it says here, okay, that training state of the art AI models requires a massive amount of computation, which is growing by four times every year. And if this trend continues, we will eventually see training runs that are 10,000 times larger than GPT4 by the end of this decade. Guys, that is a remarkable feat. 10,000 times larger than GPT4 by the end of the decade. Now, the reason this is so crazy is because you remember there was that graphic that people kept showing around and, and the graphic was basically of, you know, a model that was like 10 times larger than GPT 3.5. And of course it then showed like, GPT 4 being like this tiny ball and then GPT 5 being like this giant one. But the largest trainable model by 2030 is. It's absolutely insane, guys. Like, this is insane. Like it's themes, okay? And this is not just like some random company that's done some research. They've done like quite a lot, okay? They've done all the research needed. So this isn't some prediction that isn't going to come true. It's within a high degree of certainty. They're basically saying that we can very likely train models that would exceed GPT4 in scale to the same degree that GPT4 eclipses GPT2. So you can see right here that by 2030, the largest trainable model is going to be 10,000 times larger than GPT4, which, I mean, it's kind of hard to fathom how big that is going to be and what the implications of that are. But essentially the article breaks down into synthetic data. The compute the amount of tokens on the Internet, just like all of the unexplored areas and ways that they're going to be able to train these models. Now, I think by 2030, I'm not even sure if we're going to be training models that big because maybe we figure out different architectures, maybe things are, you know, a lot more efficient by then. But I think it's something that is remarkably incredible, the fact that, you know, having a model that's 10,000 times larger than GPT4 was something that we didn't even think about. And I mean, if we look at the scale up, that would mean that potentially looking around, considering GPT2, that could be like GPT6 or GPT7 type capabilities, which, you know, looking into the future, I don't even know what that currently looks like. Is that AGI? Is that asi? I think it's probably AGI ASI at that point. So, I mean, it's going to be pretty crazy. Maybe this is how big ASI is, maybe that's how big AGI is, but I will be most certainly paying attention to exactly how this entire area manages to heat up. And I will be doing a further deep dive onto this because there's a lot of data that is just truly fascinating about this entire thing. Now, I do apologize for this video being, you know, nearly awful quality, terrible quality, but this is, you know, the only available video that exists. And thanks to the humanoid hub for bringing this to my attention. But what we have here is literally, no, this is not like an AI generated video. For those of you that might think this is like some, you know, low poly AI generated stuff. This is the Boston Dynamics new spot humanoid robot doing some impressive, you know, capabilities right here. And what they're showing off is they're showing off how this robot can move and just how human like, its movements are. Now, the video that we're seeing is part of a longer research video in which they're going over how the robot moves, how effective it is, and just a bunch of other stuff regarding robotics. So what this is showing us is that when they do do a demo like they have done in the past, it's probably going to break the Internet once again. Because if you remember, the previous demos that we did see for these robots, I mean, they were nothing like we've ever seen. And, you know, even going back to look at them, their agility, the flexibility, the, you know, fluidity of these robots was just something that we couldn't see. And this is, you know, video. The video is lagging. So your video, like your phone, your laptop, whatever you're watching this on, isn't, you know, messing up. This is a, you know, low resolution video. But this is the full video for those of you who are interested, because we can see the robot here doing a variety of different tasks in teleoperation. You can see right here. So I'm guessing that they're getting this robot ready for certain tasks. And I mean, it's going to be kind of interesting because, you know, this is something that I think is probably the leader in terms of the fluidity of the robot. And like I said before, in terms of what this robot could, you know, feasibly do, if we're actually looking at robots that are truly, truly effective in humanoid robotics. So I'm super excited for this because I don't think people have paid attention enough to this. And you can see right here that, you know, the video once again is lagging a bit, which is, you know, quite unfortunate. But this video, it's around, you know, an hour long, doesn't really have that much views, but it's really fascinating to just see how they're continuing to develop these robots and, you know, push robotics further than most people have even realized. Now, in more robotics news, there was of course, the Unitree G1, which is leaping mass production. Now, this isn't as good as the Boston Dynamics robot, but I'm not gonna lie, because this robot is actually a lot cheaper. And because this robot is, you know, going to be mass produced. You could argue that the unitary robots are somewhat more impressive than the Boston Dynamics ones. Now, I know it's an insane take, but when we actually take a look at how quickly they've managed to develop this robot and how quickly they've managed to get this robot to do a bunch of different tasks, such as walking up to, you know, the stairs one leg at a time, like most robots, what they'll actually do is they'll literally, like, take a step here and then two steps, then two steps, but it's like 1, 2, 3. Like, it's. It's insane, like, how crazy that is. And this is, you know, all end to end with deep reinforcement learning. And remember, robotics is extraordinarily hard. Like, this is not software. You know, you can't just, you know, fix code and just retry again. This is something that works with the real physical limitations of the world. And having a humanoid robot that is that effective is. It's remarkably incredible as to, you know, how far we're going to move. And seeing a robot like that be that stable, I got to be honest, guys, if some people got pushed like that, they're going straight over. This isn't like a meme or something, but these robots are remarkably flexible. And I can't wait to see further iterations of the product because I feel like while yes, we might be at, like, that start point where, you know, the race is off, like, a lot of people have realized that, okay, this is a real thing. Like, the ChatGPT movement definitely, you know, solidified what we already know. And the fact that AI is, you know, combined with robotics is going to be completely huge. But with all these companies now trying to be the leader in their space, I mean, it's just going to result in a lot more technology that's a lot more effective and a lot more competitive in terms of how well they're able to build the products and deliver on those promises. So I think this robot here, this platform as well also being rather cheap, is going to do remarkable wonders for research and development areas, because a lot of those robots that you see doing amazing demos are just so expensive that it becomes, you know, not really feasible to use them in the research areas where students and researchers are trying to use them. There was also Apptronic, which was showcasing a new demo in which it was basically showing how well it was able to adapt to a new task. So you can see right here that essentially this is the Apptronic robot. This is another humanoid robot. That is being developed by a company aiming to automate tedious warehouse tasks. And essentially the CEO and founder is talking about, you know, how Apptronics Apollo robot performs pick and place tasks of deformable items autonomously. And this is made with just, you know, ten hours of training. So we're making breakthroughs on all fronts here, like in this entire industry. It's pretty crazy to see. And Apptronic has had the humanoid platform for quite some time, and they've been developing something that is actually quite underrated in terms of how well it manages to, you know, adapt tasks. And I've seen another demo, which I covered in a video a few months ago, of this robot actually being extraordinary fluid and it being able to interact with humans in a really nice way. So it will be intriguing to see further developments from this company because they've already shown some things that were really promising. Now, something I should have covered early on in the week was that Google's Project Astra is coming to their Gemini AI assistant. This is something that is currently live in Android phones, and I'm going to show you guys how effective this is after you guys see this quick demo.

Google DeepMind research prototype that represents our vision of the future of AI assistants. It explores just how helpful an assistant can be when it understands the context you're in and responds naturally. One of the first places you'll see Project Aster come to life will be right in Gemini Live, where you'll be able to share your camera during a conversation with Gemini. So you can show Gemini exactly, exactly where you're stuck in your calculus homer. Ask for help with the next step of your furniture assembly, or brainstorm ideas for getting the neighbors together when inspiration strikes. Or in this case, just drives up. And we're bringing your most useful apps into Gemini Live too, so it can help you take action during your conversations and messages, and pull information from apps like Google Calendar. So right from Gemini Live, you can text the details about this business to your neighbor and cross check your calendar, all without opening another app. Now, we can't wait to bring these capabilities to live for everyone, so stay tuned. And our vision for the assistant doesn't stop there. We're evolving Gemini to be even more agentive, to tackle complex problems with advanced reasoning, planning and memory. So you'll be able to think multiple steps ahead. And Gemini will get things done on your behalf under your supervision. That's the promise of a true AI assistant.

So you could see that one of the things they spoke about was reasoning, planning and memory and this is basically the golden three things that AI needs to solve in order to actually become useful in the real world. Currently, AI doesn't really have any of these in terms of planning, memory and thinking multiple steps ahead. If you want AI to do that, you're going to have to kind of build an agentic framework to actually get that to work. But there is of course Google's Project Astra, Google's Gemini Live, which is basically where you can talk to Gemini and have a natural language conversation with it. So I'm going to show you guys. I'm not sure if you just heard that, but this is essentially Google Gemini Live. So I'm going to.

Sounds like you're excited about Gemini Live. What do you want to do with it?

Well, I want to let my subscribers know what you're capable of. What is Gemini Live capable of?

Gemini Live lets you chat with me in a natural way, just like you're talking to a friend. You can ask follow up questions, change topics, and go back and forth as much as you'd like. Since I'm always learning, this is a great way to help me get better at having conversations.

What do you mean by you're always learning? Is your memory updating or are you just like, what's going on, buddy?

So I'm always learning new things, which helps me get better at understanding what you're saying and giving you helpful and informative answers. It's a little different from just updating my memory. Think of it more like I'm constantly expanding my knowledge and skills.

Okay, so can I interrupt you?

Absolutely, you can interrupt me anytime. Just say whatever's on your mind.

Okay, so I want to make this video the best video ever about AI that happened this week.

Sounds like a cool.

Okay, just make it more concise, bro.

Right, so you want a video about the coolest AI stuff from this week. How about we focus on the top three or four biggest breakthroughs? We could find some clips or graphics to go along with it.

Okay, so I didn't mean to bore you guys there too much, but you guys can get the gist of what Google is trying to do. That is Gemini Live. If you don't have it, it's currently available only on Android devices. So if you're on iPhone, you might not see it. So that is why. But I do think that Google making an effort to roll out this software as early as possible is admirable and I do think that we also need to pay attention to this. Now, I'm not sure if I covered this before, but if I did cover this before, it doesn't matter because it's still rather important. But this is not a statement from a Google employee, but kind of like a hint towards the future. So of course a Strawberry man, as many people are calling him, he says just something for me, something that can go off, think over multiple steps, download some reports, compare, analyze, and then get back to me in a few hours at similar level that you expect with a high agency PhD. And of course, Logan Kilpatrick says 2025. And funny enough, the reason I said that that's rather interesting is because that is a time when Greg Brockman comes back from his sabbatical. Jimmy Apples has long had the date 2025 in his bio for quite some time. I'm not sure if that's when we're going to get AGI. Anything's possible. But Logan Kilpatrick is the lead product for Google AI Studio and he actually did work at OpenAI. So I'm wondering maybe are there going to be some, you know, big training runs finished around that time and do they already know exactly how those systems are going to be? Definitely. Kind of sets the tone for the future.