Big AI News : OpenAI Teases o2, New Robots Advance, New AI Image Generator Crushes, Claudes New

So did you know this week that Unitree published this video, Learning in two Days? Human like natural walking, AI driven robots need to learn new actions as fast as possible. This is very important. So essentially what they've done here in this short video is they've trained their unitary G1 humanoid agent to be able to walk with a human like gait. Now you can see right here that this is impressive considering the kinds of abilities that the unitary G1 was able to do before. This is a stunning presentation of the acceleration that is possible when you do have a team of dedicated engineers that are able to constantly update the robot and see what is possible. Now with this, I think it's pretty incredible because I do remember that with this first area right here, this human like walking, this was something that was really tough for many companies to do. And of course, some people are saying that this isn't energy efficient, it's not the most flexible, yada, yada yada. But I do think what this does show us is just how quickly things are moving. Before, if we would have seen this human like gate, I'm sure that we most definitely would have been surprised because this is, you know, right now, only the second robot that I have personally seen across the Internet that is able to walk in such a manner that looks remarkably human. I remember the video I did recently on Engine AI, it was so viral that many people were speculating whether or not this thing was even CGI or not. But from this other company here, you can see that Unitree's, you know, humanoid robot platform is able to really manage to get this down. And I think this also shows us what is going to happen in the future, because when we see other robots do other things, we're going to see if those other robots from other, other companies are going to be able to apply those new capabilities. And of course, later on in the video, we do get the surprising nature of these robot dogs. And with the amount of reinforcement learning that these dogs do have in simulated virtual environments, when they do come back to the physical environment, they're able to do a variety of different tricks that seem wonderful and amazing. The robot is able to stand, you're able to kick the robot while it's on two small legs, and it's able to completely balance itself and immediately get back up. I think this is something that shows us that once these robots have, you know, a billion, a million trillion hours in the virtual world, I wonder how they will act in the physical world.

Now.

The unitary robot wasn't the only robot showing us some new updates. We had the figure robot right here that was actually showing us some very interesting stability. You can all see that this robot was able to show us some dynamic stability. Basically, this is a really important feature for robots because essentially what robots are is they are currently delicate robots. And I mean that by like this just huge delicate machines, that thing is probably worth $250,000. It's an engineering feat. But if that thing can't take a shove, and if it can't take a push, then it becomes, you know, I don't want to say completely useless, but a lot less useful considering the fact that this kind of humanoid robot, it's supposed to operate in human areas and human environments. If you're, you know, a human, likely you've been in a crowded environment before, maybe someone shoved past you unintentionally, maybe you've been on a busy subway or a busy tube if you're in the uk. Either way, being able to handle such things where you're able to handle a push or a shove and do it, you know, really well in this form to be able to stabilize yourself almost immediately, that is something that ideally we want to see from humanoid robots as quickly as possible because this means these platforms are stable, they are rugged, and they're built for the real world. And the crazy thing about this as well is that like, remember how I said that this thing probably costs quarter of a million dollars? The ring that you guys need to understand is that when you see those demos where they have like a hook hooked up to the back of the robot, that is usually because, number one, the they don't want to drop the robot and, you know, lose a bunch of money in terms of, you know, the funding that they have because robots are really, really expensive. But of course, number two, robots usually just aren't very stable sometimes. In fact, most times they have problems with even getting these things to actually work when it comes to even walking. So they kind of need to hold it up. So this goes to show just how dynamically stable it is, the fact that you can walk past this robot, brush it, and it's able to stand up with no way of falling down. Did you also know that the Google deep learning researcher, one of the most respected ones, actually spoke about AGI? Now this is Francis Chollet. He actually has his own AGI test, which is the ARC AGI reasoning benchmark. And this is basically, you know, some would regard this as the gold standard for AGI because it doesn't involve, you know, simple LLM reasoning It's like a real test that, like, if you get above 80%, they're basically saying that, look, you have an AGI level system. So that test, you know, if that benchmark, if it's broken, that's going to be like an AGI day. Now he said something really fascinating because he also said that the first AGI system will be much worse than the previous AI systems at pretty much everything, except for tests that focus on measuring general intelligence. For example, in his ARC test, and he says, like a baby, it will be generally intelligent, be able to learn efficiently and adapt, but will lack any specialized competence. And he says, however, it will be able to use, he says, however, it will be able to learn using much less data and the capabilities it picks up will generalize much more strongly to the variability and unpredictability of the real world. And as a result, after a few years, it will beat previous AI systems at absolutely everything, including many problems that we thought were completely out of reach before. So this is where he's giving his insight into what he expects for the future on AGI. And, and this kind of information is important because a lot of people have this idea that we're going to train these models, we're going to use that data and, you know, everything's going to work out. But when we do get an AGI system, you have to think about how humans are humans. I mean, some people would argue that humans have a lot of training data, like Yann Lecun, but when we actually look at how many hours of training that it actually takes for humans to learn driving, humans can learn driving on 10 hours, 15 hours, but you know, a neural net, you know, like whatever Tesla is using currently, it takes them, you know, petabytes, terabytes, huge amounts of data for them to be able to get to where they are. And of course, currently they're still not completely everywhere. Full self driving just doesn't completely work. So the point is that, you know, when we actually have human level AI, it's going to be able to do, I'm not saying it's going to be able to do 10 to 20 hours of training, but imagine an AI system that, you know, only needs 100 hours of data, 100 hours of real world data, and it's able to go off and actually work really effectively. Like imagine what happens when you give that thing a million hours of data. It's going to be crazy. And what he also said was that the arrival of the first AGI will not be the, you know, most shattering earth event that people expect it will go entirely unnoticed by the general public. Now this is a question to you guys. What do you think about this? Because I personally agree to an extent, because when we look at how AI is advancing, most people don't know about Claude. Most people do not know about Perplexity. If you're watching this channel and you do know what those tools are, trust me, I can assure you that you are in a select few because you are essentially in a bubble of people that watch this kind of content and people that are actually focused on the AI developments. If you go into the real world, I've asked my friends tons of times, have you used Perplexity? Have you used Claude? Most of them have never heard of those tools and they aren't even bothered to try them. So it's something that probably will go unnoticed by the general public despite the significance of the event. But it is something that Francis Olay, this person, says that it's going to permeate throughout the. In maybe five to 10 years. So he says it took five to 150 years between the initial steam engine prototype and widespread economic disruption, but AGI will take five to 10 years for the same. So it's a much shorter timeline. But I do think that the future is going to be super interesting because every month this benchmark gets, you know, 2%, 3% more. So it will be interesting to see how new models do cope with that. Now, one of the most insane pieces of information that I just really wanted to talk about was the fact that Sam Altman actually says that O2 gets 105% on the GPQA. Now, of course this is, you know, it's not a leak because this is from Sam Altman himself. Of course this is, you know, speculation. He says, you know, I do appreciate that I can make myself laugh so hard. A nice way to go through life. But essentially, you know, Sam Altman has spoken about 02 pretty frequently recently and that kind of indicates to us that Zero1 is nearing its completion and O2 is probably nearing the end of its training date or training cycle or however, you know, they're working with the model. Which means, okay, surprisingly again with AI, that things are moving much faster than we did think. Because if Sam Altman is here stating that, you know, I heard O2 gets 105% on a pretty challenging benchmark and that means that things are moving really incredible. So I wouldn't be surprised if we do get O2 in 2025. And that was something that I'm not going to say is impossible, but this is something that I would not have expected to happen this early in the day. And for those of you who are thinking, okay, 105% on GPQA, I'm going to give you guys a bit more context with actually what that means. So PhD level science questions are basically the GPQA. And you can see on the right hand side exactly how much 01 improves over GPT 4.0. We can see that this improves significantly on all things. So look at GPT 4.0 versus if we look at 01 preview, then of course, if we look at 01, you can see that there's a stark improvement across the board. And of course you can see in the PhD level questions, which is GPQA, we can see GPT4O. Then of course 01 preview and of course 01. So it looks like AI is pretty much surpassing the expert humans at this level, which is pretty insane when you actually think about it. And if Sam Altman is saying that, look, the GPQA is just going to be all the way up here, which basically complete the benchmarks, that's going to be basically meaning that we're going to need new benchmarks after these inference time reasoning models are here. Which means that the kind of intelligence that we're going to get next year along with agents, is going to be, I think probably, I don't want to say the first mind blowing moment, but it might be the moment where people start to realize, okay, we're no longer the smartest things on the planet anymore and these things have a ridiculous amount of intelligence. Now, of course, it could just be Sam Altman hyping up his own company as of course, OpenAI always know a way to continue creating hype around their products. But I do believe that since we've entered a new paradigm and since that we've got such a major jump in the first iteration of the model, and the fact that they're saying, look, there's many other things that we can try, I still think that we have a long way to go in terms of what's to happen. Now what's crazy about all of this is that we actually did get a Ama with OpenAI, Sam Altman, Kevin Will Srinivas and Mark Chen. So this is something where they actually revealed a lot of information. Now one of the things they said okay, and I've done a video on this I should, should have released by now. But one of the things they actually said was they said okay. Well, they didn't say okay, Sam Altman said, okay, and if I can find this tweet. So the tweet is actually a comment, but basically he said something about saturating all the benchmarks. Now I'm going to show you guys this because it's insane. So you can see here he's basically saying that, look, a bold prediction for 20, 2025 is that they're going to saturate all the benchmarks, which basically means that all of those benchmarks that people see, you know, like this one's got 80, this one's got 70%, all of those benchmarks are going to be completely crushed by OpenAI. So this is something that I think is, you know, incredible when we start to think about how far we've come and the marginal improvements that we do get to see on these benchmarks. And I think if we're seeing the jump for what it is from 01 to GPT4O, I mean, I think it's going to be something that's incredible because most people haven't realized just how insane these models are. We don't have use for them on a day to day basis, which is why even when AGI gets here, most of us probably won't even realize. Of course, those of us in the AI space will pay attention. But the prediction for 2025, and remember, a lot happens in a year. Like, if you think about where we are now compared to where we were a year ago, in terms of reasoning, in terms of like, you know, inference, in terms of the kind of models that we had, in terms of the coding and like, you know, the entire, you know, company space that we had, things have moved very, very quickly. Now, if you want to see just how quickly OpenAI have been moving, you can see this graph here. It basically shows you how things have been accelerating. We can see that for the first time it was GPT1, GPT2, GPT3, which was, you know, pretty crazy. Like, you know, we got one a year, then of course we just went boom. We had chatgpt and then, you know, like a couple months after that we go GPT4. And now, guys, you can see that like every two, three months we're getting new products, we're getting new things, we're getting new updates and, you know, OpenAI are racing to stay ahead of the competition across all fronts and areas. So this is pretty crazy. So this is from a private OpenAI webinar, but I want you to take a look because you can see, okay, how much is going on. Like, like it's just it's crazy. It's genuinely crazy just how quickly things.

Are evolving in that the pace is actually accelerating. You know, if you think about OpenAI going back to 2018, we launched GPT1. It was almost every year we would come out with a new frontier model. And this gave people and companies a lot of time at this point. You know, it's just kind of early developers a chance to react and think how these models would perform in the products they're building. You know, fast forward. We launched it in November 2022, so almost two years ago. But when you actually kind of fast forward really into that 2023 into 2024, you know, things to start accelerating faster. Japgb Enterprises hit 1 recently at the end of 2023. But then we started launching the GPTs and we wanted GTB store at 2024. We came out with Teams. Another business product, 4O, which is our multimodal modal GPT4, came out in May. 01 just came out, our first model capable of reasoning a couple of months ago. We just launched Canvas a cube a couple of weeks ago. Searches came out today. I included a lot of major products and frontier models, but there's hundreds of other product updates that happened.

So, yeah, you guys can see here that the pace isn't slowing down. In fact, I would argue that it's speeding up. And I really do want to see what next year has in store for us because we have AI agents. They're going to be absolutely incredible. Of course, like I said before, Sam Altman has been talking a lot about O2. We can see right here that he says, you know, when someone said, okay, and this is the craziest part, this is the next story, okay, they actually released accidentally 01, okay. And I got to use it. I'm going to show you guys how crazy it is. So someone said, okay, unleash the full 01. Currently. That's referring to the fact that we don't yet have the 01 that we need. Currently we do have 01 Preview, which is like a, you know, a distilled version of the Strawberry model. And of course, okay, he's stating that, look, okay, it's not going to be that much longer until we get 01. Okay, but how long until the calls are going to be unleash? 02? So we've seen two references on O2 just today, like I said before, could be AI hype, but I don't think it is completely, given the fact that a large majority of this paradigm is inference time, where you're not training the model for as long, but you're just spending that inference on actually when it asks the question, it spends that time reasoning throughout the solution. So it seems that maybe O2 is already here. And if he's already saying that, look, Our goal in 2025 is to saturate the benchmarks, it's quite clear that they have a path that they, you know, are feeling really confident in. And it seems like, you know, O1 is just the beginning of that path. Now, like I said, now another company managed to release something that I personally believe is really effective. So basically CLAUDE released the PDF capabilities in, you know, their app. And this is game changing because I was always a little bit skeptical of Claude using PDFs, because if you've ever used PDFs with Claude, one of the major problems that it actually did was it didn't actually look at the images. What it did was it actually. So what it actually does. Okay. And what it actually would do was it would extract all the text from the PDF and it wouldn't go systematically and actually read that PDF. And as someone that would want to work with long context documents, this was something that's just remarkably frustrating. So it wasn't able to actually analyze the visual elements in association with the actual text, which meant that the reasoning was quite unfortunate. And this was a small thing because when you could input your PDF on the right hand side, you would see that it would say extracted text, yada, yada yada, and then you would be able to see the text it extracted. And oftentimes I'd be like, this text is formatted. Weirdly. This doesn't look really right. I can't believe this is doing this. But now if you want upgraded PDF features, if you've been using PDFs and you've been getting some subpar responses, Claude now has this feature where it examines everything. So I'm thinking that, you know, in fact, I don't want to say I know how to do any, because I don't. Okay, maybe they're just putting it through the image model, but either way, this is a good feature and I'm definitely gonna. Now, for those of you that watched my previous video, okay, remember that how we spoke about there was this Recraft V3 model that was crushing benchmarks and we had no idea with as to, you know, what on earth this model is. Now I'm surprised that a company from London based AI graphic design startup managed to just create a model that completely demolishes everything. Okay. And it demolishes flux 1.1 pro, ideogram v2 midjourney, and even stability AI. So I don't know how they managed to get their hands on this model. I don't know how much training they've done, but whatever technique they've used, whatever they have done, it is working because they've managed to do something that is crazy. They've taken on these big companies and they have won the arena. Weight is incredible. And when I show you the images that this is able to do, you're going to see that this is absolutely insane. So shout out to Fofrai for these images. You can see on the left we have 1.1 pro, which is flux, which is a really good model. Not to say the Flux is bad, but just take a look at what happens when you use Recraft. Now, the prompt here is a photo of a wet cat, but if we go to the right, you can see that that looks a lot more like a realistic photo of a wet cat. So this one here, we've got a cat. Looks amazing, looks hd, it's in the rain, but it's not exactly a wet cat. That right there is exactly what a wet cat would look like, especially after, you know, cats are mainly running away from you and not trying to have their buffs. And you can see here that the quality for these models is absolutely incredible. Incredible. So you can see that this kind of quality looks incredible. It looks insane. And this is something that looks really nice. Now, there's also some more examples. Simple language model tasks. This page contains the number of. And the model is able to say that the word Strawberry has three Rs, which is absolutely insane. Then, of course, someone says that the next, you know, image prompt is that this page contains the result of two plus five. And you can see that it says two plus five equals seven. That just looks really super realistic there. Like, if someone sent me this image, I don't know how I would tell that that's completely AI. Only, in fact, the pen, I don't even know. Maybe I've never seen that pen before, but I'm not like a pen expert, so I can't really pinpoint that. Then it says write two adjectives in English. And of course, you can see it has two right there. And then what is the name of the U.S. president? And you can see Joe Biden. So this is something that other models just simply cannot do. And I think their success boils down to the fact that, like, maybe considering. And this is just something that's Just came to my head now. But considering the fact that the company is from a graphic design based startup, considering the fact that they understand the design elements, I think that gives them the edge when looking at how to select certain images. And of course, you know, the kinds of design choices and realism that they do make. I think a lot of companies, you know, they, they have a lot of, from people that are in certain, you know, industries, like they're just focusing on getting people that can really train these models really well. But it's good to have a designer here because, you know, when you've got a graphic design firm, these are going to be the people that are, you know, the end users of some of these products. So it makes them really integral to, you know, how the thing is developed. And we can see here, like, this is pretty crazy. I'm going to show you guys one more example and this one should really show you guys where the model is excelling. So the prompt is Squid in a kitchen. We can see that there's a squid. It's on the kitchen. It looks super realistic, as super realistic as it does. And then of course, on the right hand side, you can see we've got Squid in the kitchen by Flux 1.1 Pro. So this is a model I can't wait to test on new prompts to see how realistic it is, where it excels, where it fails, if it passes my own personal benchmarks, and if I can use this for certain tasks. Now for some more creative news, we did also have Runway get advanced camera control being now available for Gen 3 Alpha Turbo. So this is something that I think is going to change the game because this allows an incredible level of user creativity. Now you can put the camera to the right, the camera, you know, up, down, up, you know, left, right, up, down, flip it all around and you can get multiple different angles of something that is yours. And why this is incredible is because a lot of the times with Runway, we can't really control well. We used to not be able to control where that camera went. So having this feature is going to, you know, open up next level, you know, for creators that really want to work with this kind of technology. So this is something that I think, you know, changes the game. You can see that it results in a lot more dynamic scenery. And these shots that you're able to do are just a lot better in terms of what we're able to get because the model is able to be like, okay, I want to move left. Okay, I want to move right. It's not just completely random. For example, this, you know, this, this image right here of the sushi. You can see exactly how this works and exactly, you know, how crazy it is in terms of the kinds of realism that we will get. So, I mean, the kinds of pipelines that we're going to get when it comes to AI video. I mean, you know, a lot of people think it's just text to video, but Runway, I think what they're doing, and I think this is the smartest thing they are doing, is that they're developing the kinds of tools that allow people to really direct their own short films, their own movies. And I think that is, you know, really important because it's not just, hey, this is, you know, AI text to video. It's, hey, you can move the camera left, you can move the camera right. You can have full creative control and freedom over the kinds of shots that you do want to have. And this is where someone is zooming out completely to reveal a, you know, you know, an abandoned city or, you know, a city that's just been through ruins and chaos. So, I mean, this kind of thing, I mean, it's really, really game changing. Like we're about to. I wouldn't be surprised if by next year we get our like full, you know, fledged first AI movie or we get like a full, you know, YouTube video, like YouTube channel that was just like full of AI, you know, shorts or whatever, of people just using AI to create really cool and interesting content. So I will really love to see exactly what that looks like in the future. Now, Mustafa Suleiman actually said that memory and vision are coming to AI where it can remember everything about what you learn and see. And I'm going to include this clip because after I show you guys another clip, you're going to see that. Whoa. Okay. And from the comments that I've seen, a lot of people are requesting this feature, but I'm going to show you guys that thing in a moment.

The moment. Because it's frustrating that it forgets that you told it that you were going out to dinner last night and that, you know, you had an undergrad at xyz.

I mean, so the bad assistant.

Yeah, it's kind of. That's boring, right? And so you don't bother really giving it your all. Just think about the moment when, like, it really does remember everything that you've said and it's able to retrieve from that history of your conversations accurately and.

The timeline for, for that.

I think that we're 18 months away from really Getting that to a workable level, if you look out far enough, yes. It's quite likely that your AI is going to learn to click on buttons in your browser, enter text in text entry boxes. It's certainly going to be able to read the page and understand the images and all of the text on the page. That's going to be a very different, you know, format. I think one of the things that's going to come out quite soon, I would say, is this idea of Copilot's vision.

Vision.

It's going to really see what you see like, and so you'll be able to have a companion that really browses with looking at the images you see.

Physically or see what you see is.

How you use, I think, at first digitally.

Right.

But then very shortly after that, there will be a visual awareness, whether it's through earbuds or glasses.

Yeah, earbuds with video in them.

That's.

Yeah, I think that's going to be a new format.

Is these.

These visual.

Now, if you aren't familiar with who Mustafa Suhman is, he's currently the head of Inflection AI. Well, in fact, that's not true. He used to be the head of Inflection AI. He's currently the head of Microsoft AI. Okay. Microsoft are doing some pretty cool things with their product called Co Pilot. So Co Pilot is basically going to be what Advanced Voice Mode should have been. So what we're going to have here is an AI that can see your browser in real time. So as I'm making this video or as I'm making other videos, I can just message Co Pilot. It's going to be able to see exactly what I'm doing. And it's going. I'm going to be like, hey, what do you think about this video? What do you think about, you know, this script? Do you think this is good? Do you think that's. Well, there's going to be like an entire AI system that lives in your computer that you can talk to while you're browsing. I mean, I think it's going to change, like, quite a lot of things because, I mean, you know, it's. It's actually pretty useful when you're working just to have, like an ear that can also see what you're doing. Because one of the things that I've realized is that now that Advanced Voice Mode is on desktop, it's really cool to like, you know, talk back and forth between ideas, you know, navigate, you know, some of the ideas that you're doing. But you know, once you have vision and you're able to show it certain things, it's going to be able to be like, okay, this is really cool. This is really nice. I like this, I like that. And I think that just completely changes the game because I won't have to completely copy and paste in everything in. It's going to see what's on my screen by it's going to be by my side. So it's pretty crazy. So I mean, I can't wait for this. Genuinely, I can't wait for this. And if you guys enjoyed this video as well, don't forget to subscribe to my newsletter. I'm going to be trying to because since there's so many stories I've been uploading so much, I think it's best if every single week I summarize them. It's going to be completely free. No subscription required. Cancel at any time, zero payment. It's just AI news that you most likely did miss. And there's still stories that I didn't include in today's video simply because I just don't have. Like this video would be an hour long if I covered every story within the last three days. So if you guys want that, let me know in the comment section below and I'll see you guys in the next video.