New AI Agent, GPT-5 Not That Good? 100 Billion Humanoid Robots, Mixutre Of AGENTS And More

With an absolutely incredible week in artificial intelligence, today's stories are going to be some of the ones that you really don't want to miss. So without wasting any more time, take a look at some of the most interesting stories that you really want to pay attention to. So one of the first AI stories, slash robotic stories that I found to be truly fascinating was this coming from Stanford University. Basically, Stanford University did a Collaboration with Google DeepMind before. If you remember, earlier this year, they did the mobile aloha thing. That was a very interesting project where they demonstrated the ability for robots to do autonomous tasks. Now, essentially what they're doing here is a different kind of thing, which is along the same kind of guidelines. And this is human plus. So they essentially have a humanoid robot, and they're trying to make them perform various tasks by imitating human actions. This involves creating a system where robots can actually observe human and mimic human movements and then use that data to perform the tasks autonomously. So this is an entirely new pipeline of training autonomous robots using an RGB camera that actually allows you to observe human body and hand movements in real time. And then that's how they collect the data. So essentially, the human motion data is collected using advanced pose estimation algorithms. And the robot then actually shadows these movements using a policy trained in a simulation environment. And right here you could actually see that this is the RGB camera, where this is it observing the actual human in action right there. So it's not the cameras on its head, but this camera right here that is, you know, enabling it to shadow it. And this actually, funnily enough, kind of gives me some real steel kind of vibes for any of you that have actually seen that movie before. So once you have this single RGB camera that is able to, you know, shadow humans, you can see the camera is once again located in this area here. And of course, the robot is then doing those actions. You can see in this scene that the camera is obviously located right here. It's actually pretty, pretty hard to see. And the base model that they've actually used on this thing, funnily enough, is actually a unitry robot. I do remember when they released this robot. I think this is the unitry H1. And this is, of course, an earlier model which goes to show the speed at which robotics advancements can allow for universities to be able to, I guess you could say, you know, use them and actually see how effective they are. Now, one of the things that they actually did talk about is they did actually talk about. I didn't mean to pause it there. But they did actually talk about how the fact that this robot doesn't actually have a lot of degrees of freedom. So essentially most. And what that means is that the unitary robot that they're using, it doesn't, you know, like, have the same range of motion that humans have. Like, humans have a lot like, a lot, a lot like a huge range of motion. Like, your shoulder joints are incredibly, you know, flexible and you can, you know, rotate them in, you know, various different ways. But this kind of robot, it wasn't really built for those kinds of things. So it's quite hard to do a lot of these tasks because the robot is very rigid in nature. But however, it's not stopping them from being able to train certain tasks and then of course, get policies that actually work effectively. So the real time teleoperation thing that they're doing here, this is where they're basically gathering data, because teleoperation in the other way is actually quite hard to do. So they're essentially inventing a new way for teleoperation data to be captured. And of course, you can see on the top of the, on the robot there, it actually does look like eyes of the robot. Like, funnily enough, it looks like a kind of droid. But these are just robots that they're using to, of course, capture the data. And essentially a low level policy is trained using reinforcement learning on a large dataset of human motion, which is actually 40 hours of diverse human activities. And I think when you go onto the actual website, their GitHub, you can see that the autonomous tasks that it's learned here are actually quite impressive now, of course, by the standard of the average person, because one thing that I do find interesting is that I do sometimes like to show people what's going on. And this is people that have literally no interest in robotics or AI, because I want to see, you know, how they look at, you know, robotics and how it's done. But you can see that, you know, some of these ones autonomously do actually look pretty interesting. I do have to say, folding clothes is definitely very, very hard. You can see right here, this one, I think it was rather impressive. You've got a high jump, which is completely autonomous. The robot is able to perform a jumping motion and then of course, able to stabilize itself after the resulting, you know, forces have acted on it, which is pretty, pretty hard considering that this robot is like super, super rigid. So the fact that it's able to do this with not the best, not the best kind of movement is rather Impressive. And all of these are fully autonomous. So this is the robot executing based on the policies is previously learned. So, so this is not teleoperated at all. This is fully 100%, you know, automated, I guess you could say. So right here you can see this is the warehouse autonomous. And then of course, we have, you know, unseen objects right here, which is rather, rather effective. So it looks like we are, you know, in a nice area. And this is, you know, the previous mobile aloha in a, you know, I guess you could say the teleoperation mode. So overall, it shows that there is a level of innovation going on at this lab that I think is really, really interesting. And, and I'm always intrigued to see what, you know, teams that do some really, really effective stuff come up with again and again and again because it shows, you know, a nice track record of the fact that they're continually, you know, innovating on new methods. So this is something really interesting. Now for the future, what I would love to see is that I would love to see the autonomous skills on a different hardware platform. For example, Unitree's newest robot. I would love to see, you know, some of the new policies trained on that because the robot is a lot more smooth. It's, you know, it's got more degrees of freedom. It's able to do a wide range of tasks. And I'm thinking that if we can get robots that are, you know, this effective, you guys can see right here, like, if it's this effective at, you know, jumping or whatnot on such a limited platform, not, not, not to, you know, hate on the platform or whatever, it is still pretty good. But I'm stating that when we get to true areas where we have, you know, true cheap robots that are flexible and we're able to use those for research, I think that's when, you know, things are going to be getting very, very incredible and very, very futuristic because it's going to a truly advanced level of research that we can, you know, explore with these robots. So autonomous humanoid shadowing, imitation from humans. I think human plus is, is fascinating. Now here's where we have a very, very fascinating thing. And this is Elon Musk actually speaking about the Tesla Optimus. So I chose this story to be next because I wanted to, you know, transition smoothly from the previous story on how we just spoke about humanoid robots doing some tasks autonomously. This is where Elon Musk actually speaks about the future of humanoid robotics. Now, I do want to say there's a quick caveat to this because of course, Elon Musk has been known to, I wouldn't say exaggerate his timelines, as some people would put it, but I would say he's pretty ambitious within, you know, what he wants to achieve in a short amount of time. So in this clip he basically talks about humanoid robots and being able to produce a 100 million per year and those doing pretty much every task that you can imagine on Earth. And Elon Musk has said some pretty ambitious things before, but I think that with the rate that technology is improving, I don't think 100 million is going to be pretty easy. But I do think that once certain levels of AI are achieved, like embodied AGI, I don't think that this is completely out of the spectrum. Within the next 10 years, I would say, or at least 20 years. So I think things are going to be drastic. But take a look because it's pretty.

Fascinating when you think about the, the Optimus robot, which is really a humanoid robot that is intended to, you know, be able to do anything you want it to do, to be, you know, it's, you know, your companion. It can be at your house, it can sort of babysit your kids. It could teach them, be a teacher. It, you know, it can do factory stuff. Like, I think that the ultimate ratio of say, how many super useful humanoid helper droids do you want? Like, who doesn't want a C3PO, you know, you know, but a C3PO plus R2D2 plus, you know, plus plus it would be pretty awesome, I think. I think everyone in the world is going to want one, like literally everyone. And then there will be obviously robots in industry making stuff. And so, I mean, I think the ratio of humanoid robots to humans will probably be at least two to one, something like that. One to one for sure. So, which means like somewhere on the order of 10 billion humanoid robots, maybe maybe 20 or 30. And so then it's like, okay, well, let's say, you know, you kind of make, let's say the build rate is, I think the build rate will be probably something ultimately like a billion a year. Humanoid robots, like actually. And if Tesla just has a 10% share of that, and it might be a lot more than 10%, and there's, you know, who make like 100 million optimists units a year. I mean, for reference, the auto industry is roughly 100 million vehicles per year. So, you know, sort of similar ballpark, at least within an order of magnitude. I think we could make one for a cost of maybe at really high scale of about $10,000, it's smaller, it'll be less expensive than a car. So. And I think if you sell it for $20,000 or something, this is at large scale volume, Tesla would basically make about a trillion dollars now.

A trillion dollars of profit a year is pretty insane. But yeah, I, I get, I guess we're gonna have to see because a lot of the things in the future that are based on, I wouldn't say baseless claims like some angry commenters do say, I've seen that on various different forums and, you know, tweets and stuff like that. But yeah, I mean, we have to understand that the future is going to get pretty crazy in the future because if we actually, you know, look into the future and say, okay, let's say we have like AGI, like a real AGI and it's actually embodied and it's actually able to do stuff, and how valuable would that be? Would it be more valuable than the car? And honestly, I think at that level, like at where society would be, I do think it would be more valuable than a car because I think that, you know, having your own, you know, person basically for $20,000, that can do pretty much anything that's going to be, you know, smarter than you. It could, you know, clean your house whenever you wanted. Not like there's not going to be some specialist made version of it. But I do think that, you know, having that around, you know, is going to be kind of interesting because I do think there's probably going to be specialized versions, whether they're like, you know, girlfriend robots or, you know, husband robots. But I definitely feel like with the way society is going to be, they're definitely going to be like, employed. It's probably mainly going to be like B2B where Tesla, you know, has them working in factories and they're just deployed for maybe like government projects to build more data centers or, you know, to build more companies for new projects, whether it be, you know, Dyson spheres. I don't want to get into, you know, some crazy stuff. But the point is that if we actually imagine a future where a Tesla humanoid robot is as good as the average human, or like, let's say 20% better with their hands. I mean, you have to think about how quickly society is going to evolve because we can just basically, okay, just hear me out for a second. You can basically print employees that would work 24, 7 and do whatever you want that would need any breaks and wouldn't get sick. So you have to Think about that on a really, really high level. And those employees can even make more employees. So pretty, pretty crazy stuff when you start to think about how the scale of this could be over the next decade or even 20 years. So now this is something that actually caught me off guard. Some former meta researchers actually released a new agent and this was actually rather impressive. I'm going to show you guys a demo and then I'm going to give my opinions because I do think that this is going to be the year of agents. But this surprised me because I think, okay, and maybe, just maybe I think that what's going on right now is that, you know, other companies are starting to catch up to OpenAI and starting to eat into their market share. Not that this is that big in terms it didn't get that much of a reception on social media. But the point I'm trying to make is that this is something that I thought that I would see OpenAI demo first, but I guess they're still doing iterative deployment. But anyways. Zeta Labs Jace.

Hi. Hi. I'm Fredrik, co founder of zettalabs. I. I'm here to introduce Jace, your AI coworker. With just one instruction, Jace can handle any task for you. For example, Jace can plan a full fledged trip. It will ask you for all the necessary details and make all reservations for you. This is possible thanks to our autonomous WebAgent 1 model AWAOne. With AWA1, Jace can control and perform actions in a browser just like a human would. Let's take a look at the first few steps in the Live view. Jace is starting on Google, going to the Airbnb website, entering all the details and clicking search. Jace isn't just an assistant, it can take on entire roles. Imagine a recruiter who never sleeps. Jace finds candidates, schedules, interviews and manages the hiring process effortlessly. While using Jace, we quickly realized we could take it a step further. So we asked Jace to create a company on its own. With a simple prompt, it created a plan, registered the business, found the first client and made its first revenue. Jace still has its limitations. It can struggle with complicated tasks and the current browsing speed is somewhat slow. We are working very hard to make it faster and more reliable. We can't wait to see how Jace will help you in your life.

So I mean, the last claim on the video is pretty crazy if it was true because they said this was an agent that was able to create an llc. You can see here it, you know, you can see like it's being able to plan it out And I think the reason I think this is going to get a lot crazier is because a lot of the information that we have about AI agents and what they're going to be able to do. The problem right now, which is why, you know, a lot of startups, like I think it's adept AI that you haven't really seen many, you know, announcements from on real, on really good stuff, is because agents struggle with planning. Like they just struggle with multi step reasoning, multi step planning. So I think if we do have a model like GPT5 that is able to solve that, then this thing right here is going to get really, really crazy. But you can see it's able to do some very interesting things. You can say, here are some of the best registered agent services you can see. Registered agent, this, this, whatever. And then it says appoint your registered agent for llc. And then you can see these are all of the things that it was able to do. So I mean, it's going to be kind of interesting to see if this is something that actually works. If you do go onto their web page, you can actually join the wait list. But I don't expect this to be open anytime soon. Like, I've seen a lot of AI agent things at the moment, but I do know that unless you kind of, I wouldn't say invent a new architecture, but like you're truly, truly training a new model or you kind of do something truly innovative, you know, things aren't going to get crazy just yet. But I do think that by next year, this time we will have probably really, really, really, really capable agents that are able to do quite a lot. Because from what we've seen, certain frameworks even surrounding the LLM technologies that we have today do provide like a noticeable jump just from the base level of GPT4. So with whatever next model does get released, which whatever model is at the frontier standard, I think it's going to be interesting to see where these agents are placed in the hierarchy. Now, something that was, I want to say a surprise, but in hindsight it's not a surprise because we kind of knew this was going to happen. So OpenAI actually added a former National Security Agency to its board. So you can see that this guy is pretty, pretty pivotal in things. And the reason I say that this was something that is a surprise, but also isn't a surprise because recently in a interview on the Dwarkresh Patel podcast, Leopold aschenbrenner, the former OpenAI employee that did an entire essay on the future of AGI and artificial superintelligence and how we're probably going to get to AGI by 2027 actually said that labs are going to be nationalized, which essentially just means that there's going to be quite a lot of government oversight. And right here we can see that this is probably one of the first steps. We have someone who basically was very, very connected in the US government now working at OpenAI. So this is a pivotal piece of information because it goes to show that some of the predictions from Leopold Aschenbrenner's interview and of course his document on the future of artificial intelligence are already coming true. Not only a week after they were released. But this is always something that, you know, a few people, including myself, have echoed in the community. Because if you just think about like the future of where these systems are going to get to, these systems will shift the dynamic of power to whoever who holds them. And that, you know, sets a remarkable standard for what is going to be necessary in order to actually take control of these systems and in order to ensure that they don't pose any threat to national security. And of course the public. There are a lot of people who are discussing these claims on Twitter. You can see Kim.com says OpenAI just hired the guy who was in charge of a mass surveillance at the nsa. He outsourced the illegal mass spying against Americans to British spy agencies to circumvent US law. He gave them unlimited spying to US networks. Tells you all you need to know about OpenAI. This is referring to the huge drama that happened previously with the NSA when basically we found out that the National Security Agency was basically just spying on Americans on their conversations on the phone. And there was this huge entire thing. Don't really wanted to get into all of that, but it, its big point.

But overall this doesn't come of any surprise.

I mean, like I said before, I do also believe that, you know, how we currently get these leaks or whatever. I do believe that by like 2025, I don't believe that we're going to be getting any leaks about future models. I know that it might seem like it, but I just think that by the time these labs are nationalized and you know, the true scale of these things are realized, this is going to be some kind of Manhattan Project where the, the majority of the information is going to be under serious, serious lock and key, meaning that unless your security clearance is above a certain level, you are not going to have access to what the next level of systems are being, you know, are being done. And there's probably even some Project Manhattan going on right now within OpenAI that we have no idea about. And if you're not sure what Project Manhattan was, basically it was a secret, basically town facility where they developed the first nuclear bomb. And of course, they had to do this in secret because they needed to get to it, because before other countries did, and with the way timelines are going right now, I mean, it would make sense for there to be some kind of, not national security effort, but some kind of effort in order to create AGI as quick as possible. Maybe in some field somewhere, maybe not at Los Alamos, but if you watched Opener, you would know about this, but maybe somewhere in America. And I truly do believe that, because it does make sense. Some people would even argue that whoever gets to ASI first is going to control the world. And that's something that's been ushered by people who do work at OpenAI. So this kind of trend is going to continue. Now, here's where we had something that was a little bit confusing because this was a clip from Mirma, not Mirma Mira Maratti. And essentially what she said in this clip confused a lot of people because she said that inside the labs, the models that they have are not that far ahead from the ones that are, you know, out in the public. But listen to what she says, because I'm going to play it once, then I'm going to replay it to you because I think there might be a mistake on the majority of people's part who are interpreting this clip.

Inside the labs, we have these capable models, and they're not that far ahead from what the public has access to for free. And that's a completely different trajectory for bringing technology into the world that what we've seen historically. And it's a great opportunity because it brings people along, it gives them intuitive sense for the capabilities and risks and allows people to prepare for the advent of this, of bringing advanced AI into the world. And obviously, the opportunities are huge. Now, it's normal that we talk a lot about the risks because they're so important.

So basically, this clip here has a lot of people, you know, I guess you could say rattled, because they're thinking, wait a minute, I thought OpenAI was so far ahead of everyone else that, you know, they were on GPT5 and GPT6 according to a lot of the leaks and according to, you know, just thinking that if OpenAI had GPT4 in 2022, in 2024, they must have a truly advanced model. Now, maybe she misspoke Maybe this isn't true, but I think either one thing has happened here. Either she spoke the truth about what.

She'S seen, or potentially just maybe there could be some overhype going on. Okay. And I don't know. Okay. I think on one hand I think the most likely thing that's going on here is that. And I know this didn't mean to drop that, but I know that this doesn't seem likely, but I do think that it's just potentially just poor communication from OpenAI. One of the things that we actually got to see from OpenAI's recent fiascos in the past 12 months is that a lot of the times people don't even know what's going on at this company. And it's crazy to say that, but there are a lot of things that go on between CEOs, CEO like even when Sam Altman was fired, a lot of people didn't even see that coming. Of course there might be a separate board issue, but there were so many different isolated incidents that I think even if this statement is true, I still do think that they are probably working on some super advanced models that maybe she wasn't thinking about at that moment in time. Not to say that she's completely wrong, but I just think the fact that if we saw not only yesterday, the day before that, the months leading up to this, that people such as Microsoft CTO was literally stating that GPT5 from what he's seen has been the level of a PhD researcher, it could be able to do your PhD. And then Sam Altman saying that GPT4 is dumb, and then many other, you know, OpenAI employees very, very much so, hinting at the fact that GPT5 is going to be a remarkable jump, and Sam Altman stating that the jump is going to be quite similar. I mean, stating that there's not that much to advance right now. I mean, there's quite a disconnect there. But the only thing that we can truly say is that when GPT5 or whatever next level system is released, then we're truly going to know what these systems are capable of or if we've hit a plateau. Because right now that statement is definitely confusing and I guess we'll have to see if there is a plateau or what OpenAI is truly doing. So it will be interesting to see if this statement does hold up. But I'm not paying too much attention to it because this hasn't been the overall theme from just the many sources that I've seen have operated on OpenAI. So will be interesting to see exactly how this progresses. So then we had something here that I think was truly, truly fascinating because this, once again, another reference to Leopold.

Aschenbrenner shows that, you know, the kind of scaffolding and all these kinds of.

Other techniques that you can apply on.

Top of large language models can effectively.

Be used in order to push them.

Further ahead in terms of their capabilities.

So right here you can see that say we introduce mixture of agents and.

Approach to harness the collective strengths of.

Multiple LLMs to improve state of the art quality. And we provide a reference implementation together, mixture of the arts, which leverages several open source LLM agents to achieve a.

Score of 65.1% on Alpaca Eval 2.0.

Surpassing the prior leader, GPT4. The long story short here guys, is that basically they had and they use these open source models to surpass GPT4O.

On this pretty difficult benchmark.

So that is pretty crazy guys, they used open source models. They basically just used an entirely new framework of, I guess you could say, interacting with these models and in doing so they managed to push the bar ahead. And guys, they weren't even using models.

Like Gemini, Claude and GPT4.

These were open source models which aren't anywhere near the level of GPT4.

So how did this work?

So basically what they did was they used a mixture of agents. You can kind of think of it as like this.

So basically we're going to get into.

That in a minute.

But you can see right here, you.

Can see that the comparisons between the single model versus the other kinds of models is pretty intense.

So what you can see right here guys, is that we have the single.

Model right here and this is in yellow. You can see the yellow, okay, is.

A lot less than the blue and the blue is with the responses from the other models. So basically overall they were able to.

Improve the responses of these models by.

Combining their effectiveness together to get further and further.

And you guys can see that there.

Is a huge, huge increase in these kind of gaps where these models actually work. So this kind of thing is very, very effective. But how on earth did this work? So basically what they did, they decided.

To kind of like organize these robo.