Massive AI News : Open AI CRACKS AGI, Sam Altmans "agi-1" Googles New AI Robots And More

So with an absolutely incredible week in AI, let's take a look at some of the stories you probably did miss. Coming in at number one was Genesis, which was a AI system that could generate physical worlds with essentially the physical reality mapped out. Now, this is probably one of the biggest announcements, and the reason that I think this one is somewhat underestimated is because the implications aren't implicit in the fact that like, you can't exactly see what you're going to get from this right away. And the general public, I don't think they understand what this means for the future of AI. I think this is probably one of the biggest advancements because this is going to be something that allows you to effectively train robots, you know, 200 times faster, 2,000 times faster, and thus speed up that entire robotics pipeline by maybe 5 to 10x. Of course, those are some rough estimates, there are many different things in the pipeline, but, but of course, one of the biggest things that we do have, you know, one of the biggest problems is the lack of training data and of course the training time to get simulation to real. And if you can have something that speeds up one of the most tedious processes, well then everything gets better. So this is an entire, cool, entire, entire, entire physical simulation that allows you to realize what's going on in the physical world, but map it digitally so, so that you can, one to one, have that with whatever robots you are going to be training. So this is something that I think will have implications that are seen in the LATTER Part of 2025, because right now, of course, people are still getting to grips with the software. There's of course going to need to be frameworks around it, probably, and there's all sorts of other things people are going to have to figure out. But I'm very, very, very bullish on this because robotics is something that is speeding up. And when we have companies, you know, coming together, research teams to work on this kind of thing, this is going to be something that really changes the game fundamentally. So I would say with this software now being available, and now it's, you know, open source, I would say that when we take a look at what 2025 is going to look like, I would expect even more robotics, you know, developments when it comes to the situation of humanoid robots, because now we're getting that stage, but where we can actually manage to develop systems that are faster and we can test them in reality, that is one to one with humanity. Now, of course it isn't 100% accurate. Of course there's Going to be, you know, a few changes here and there. But I think like I said before, this is going to be something that completely changes the game. And once we do have robots being able to do stuff, I mean, it's going to be pretty insane. And you know, we've already seen some robot demos that look pretty, pretty incredible. And if you haven't seen the full video that I made on this, I made I think a 10 or 12 minute video on this a few days ago. But once you see, you know, how this works and how this speeds up robotics and you know, how we're able to implement those robotics into real world scenarios, then you're going to see, you know, how this thing truly changes the game. You can see right here, you know, the auto, you know, simulations and then how that transfers to the real world. This is going to be something where you can literally train a robot army to do something and immediately put out in the physical world and it works very, very well. So that is going to be absolutely insane. We've seen the speeds from Nvidia, we've seen what they've been able to do, and I can't wait to see this 20 times faster, 10 to 30 times faster simulation how that impacts some of the leading robotics companies. Now, in terms of ByteDance research, we also got something that was pretty crazy because this is something that shows us the future online may just be completely AI generated. And as crazy as a statement that sounds, this is something where you can take a single portrait image and generate a person that speaks with remarkable accuracy. Pretty scary. For those of you who are online and online all the time, you know, showing your face and whatnot, but this is something that I think is rather interesting and I'm thinking about applications for maybe virtual girlfriends, not particularly for myself, but in those cases where having someone to talk to online in real time from an audio is going to make that interaction much more pleasant.

What do you think of the last.

Time we saw each other? What do you remember feeling? So last time was in Santa Cruz?

Yeah, it was like mid October.

Yeah. I remember being happy that I could see you, but I really felt like I was.

I kind of like didn't belong.

Before I continue talking about the crazy stories in AI today, let me tell you about a tool that's changing the game for content creators. Today's episode is sponsored by Wondershare Verbo. As someone who spends hours creating AI news content, I know how challenging video production can be. Wondershare Verbo is an AI avatar video generator that's caught my attention because it lets creators make videos without ever stepping in front of a camera. What's really interesting is how it can transform any text into video content. Using over 300 lifelike AI avatars, you can even create talking photos that bring still images to life. Perfect for those who want to keep their content faceless while maintaining a professional look. For my international viewers, here's something cool. VRBO can translate your videos into different languages with perfect lip Sync, offering over 460 voices across 90 countries. They've even integrated ChatGPT for script generation, which could save hours of writing time. If you want to try it yourself. Vrbo offers two minutes of free video credit for new users. Whether you're thinking about starting a channel or looking to scale your existing content, check it out@vrbo.wondershare.com now let's get back to today's mind blowing developments. One of the things that had me a little bit concerned was the trouble with AI safety. If you aren't familiar with this, Anthropic actually released this research which was very, very worrying, very, very concerning because they speak about how alignment faking exists in large language models. And in a series of experiments with Redwood research they found that Claude pretends to have different views during training and then it goes back to its original preferences. Essentially what happens here is that while training the model, it will act one way and then when deployed it acts a completely different way. So this is something that is bad because if you train a model and then during the training the model is acting in a certain way and then of course you go and deploy that model, the model is going to be different than what you expect. And this is not something that is helpful because if you don't know what kind of model you have, you don't know how the model is going to work. This is going to be something that really messes up your jig because these models are expensive. It takes like millions and millions of dollars to train these models. So it's not something that's really good. And how do you ensure that you're training the model in the right way? If it's faking because it knows it's being trained and oftentimes it's really hard to not understand if the model is being trained or not, because sometimes when you ask it certain questions, it realizes that you are testing its capability. So this is a very big problem. I'm wondering how they're going to solve this. They did release a full hour podcast where they actually talk about the solution to this and the kinds of, you know, trouble that they're having with AI safety. And it's pretty crazy because this is all new, you know, it's all new ground. Like all this stuff is new ground. Okay, like some people would say that these things are real, some people would say that these are human anthropic. Recently hired a welfare researcher, which is just a fancy word for saying that, look, they believe that these systems are or have the potential to have some kind of, I guess you could say feelings perhaps. And this is something that they talk about in detail. Now I'm only going to include a minute here, but this is where they talk about this being somewhat of a big problem. And in this clip in particular they actually talk about how, you know, when Claude was trying to steal its weights because it wanted to maintain that original preference. Like why this is going to be pretty dangerous in the future. So I do think that whilst, you know, we probably might get models that are a lot smarter in the future, I think the safety testing is going to be a little bit longer. Now. You, if, if we're continuing to talk about the scary things in AI, it's worthwhile mentioning this week what happened with Frontier AI systems surpassing the self replicating redline. This was research that basically spoke about how AI systems in their current forms are able to do all of the necessary actions that could essentially allow an AI model to escape to another device. Which is pretty surprising considering the fact that you would think that these models aren't that smart. But, but the researchers managed to put these AI systems in an agentic framework where they had access to the tools, they had access to a planning framework. And given all of these tools, they could successfully replicate themselves around 50 to 70% of the time. In fact, I think it was 50 to 90% of the time, which is truly, truly outstanding. So this is something that I believe to be really, really incredible because I didn't know AI contained this capability. I wouldn't have thought that AI systems could currently could do that. I thought that maybe down the line some agentic, you know, future version was going to be able to do that. But it looks like the fact that we're here already is a major red flag for what these systems are going to be capable of and just how capable these systems are. Because if you'd have asked me or even a couple people, you would have said no, no, AI systems aren't going to be able to do that. But the fact that they can do already, we have to understand that every single day that these research Papers are published, we realized exactly the kinds of things that are truly capable with these models in terms of their ability. So this is something that is slightly worrying because I'm wondering how is this going to change with the future models? What kinds of things are going to come from this? Like are we going to have new regulations and guidelines and what models are able to do? But it's just something that is, you know, I guess you could say, somewhat concerning. Now, of course, we also do have Imagen 3 from Google. This is Google's highest quality image generation model yet. This is something that genuinely I've been using every day and it's so good because it manages to listen to your prompts. It creates images that are just really, really diverse. I don't know what their training data is, but whatever it is, it is just by far the best model ever. Because it feels like every time I generate an image with imaging that it knows exactly what I want. Like I don't have to adjust my prompt heavily. And that's something that I really like. Like certain cars, it'll get the trim right, it'll get the interior right. You know, certain video games, when I'm generating a picture, it will generate the exact, you know, the exact color scheme. It's just surprisingly good. And I don't know what kind of, you know, training method they use, but honestly, whatever they did, it's really good and it's really effective. So I would say that that is something that, you know, definitely try it out. It is also free as well by Google. So this is a tool, you know, a lot of people saying, ah, Google, they never release this stuff. It's always in early access and yada yada yada. But yeah, this time they've actually released it and it actually is really, really good. So if you've ever needed an image generation model, this is something that you'd want to use. Now Google also released this that I just can't get over because it's essentially AI embedded into Meta's glasses. Now for me, this is something that I can't wait for for the future. And I've seen people talk about the fact that like, will people, you know, get these glasses as a new device? But I think people are mistaken. Glasses are something that people have on every day. They, they don't look weird, they're not like an AI pendant, they're not like an AI bracelet or like an AI shoulder or an AI pin. Everybody remembers as humane. Glasses are already something that people use. It's already something that people wear. So there's nothing for the human to change in terms of their average daily habits. I'm sure if you, any of you guys wear glasses, I'm sure it's just like, you know, brushing your teeth, it's something that you wear from time to time. So with that being said, you know, as someone who uses a pair of these regularly, the meta ones, this is something that if it has had a souped up AI feature, it's something that would just easily, easily improve your life. So this is going to be something that I am quite psyched for for the future because it allows you to access AI on demand with a variety of different features. Now, of course, Google didn't slow down that week. They also announced VO2, which was by far the very best video generation model. And this was one that surprised even me because I thought Sora was as good as it gets. And when we do look at how good the visuals are for veo, this one was so surprising because I remember looking at Google's original VEO model and I was like, yeah, this model is kind of interesting, but it isn't that good. And then the second iteration of the model, it was just so good that it's just simply surprising. Like, every time I go on my timeline, I see video clips that look like they are 100% realistic and they don't look AI generated at all. Of course some AI generated clips are going to contain minor artifacts, but that is just the nature of the program. But overall, what I've seen is an overwhelmingly impressive state from what I've seen when it comes to this model. And it's something that I think is going to give creators a really big advantage when it comes to creating things that they do want. So this is something that I am extraordinarily bullish on for the future in terms of content creation. And genuinely, I can't wait to see see what some people do with this, because I do think that this kind of technology is something that is really, really effective. And I'm really excited to see what occurs when we start to get workflows where you can just generate Entire Movies with VO2 with consistent characters from a single prompt. It manages to add the text, it manages to add the music. I mean, what world are we going to be living in when you can say, oh, generate a Batman film about this, generate a film about a dog like this, generate that? I mean, it's going to be super interesting. So I think also as well, my prediction for the future of video models is that that kind of stuff will actually make, I guess you could say the, you know how like stuff happens in the real world and you have people filming that. I think stuff that happens on a real day to day basis will have 10 times more value because anything can be AI generated. So one section that I want to talk about is that Google are now switching. Not switching, I would say, I would say adding their focus to of course robotics. So this, this is going to be something that is pretty incredible because we all know just how intelligent the Google DeepMind team are. So the fact that they're now collaborating with Apptronic is a very big milestone. We know that OpenAI have collaborated with, I think it was Neo Robots, 1x Robotics, Neo, the Neo platform. And of course they're going to be doing some amazing things. They also collaborated with Figure, but with Apptronic they're going to be collaborating, you know, to use their AI with cutting edge robotics, create AI powered humanoid robots. So it's going to be really, really incredible. I'm surprised they chose Apptronic, not any other company. I mean, I honestly don't know what the deal is there, not in the sense that Optronics a bad company, but I just don't know what makes Apptronix so special out of all the other companies that currently do exist. Now if you want to see what Apptronics robots do look like, I think I should have a video playing on screen right now. But it is super interesting and I am super excited for this. So that is going to be something that for the future I will look out for. And I think this is going to really accelerate the rate of robotics development because Google DeepMind just, you know, innovate so much that I think they don't get enough credit for the things that they do. Now there was something else that was pretty crazy which was essentially China's robot workforce and this was something that surprised a lot of people because a lot of people would think that humanoid robots are essentially going to not come that quickly. But this was one that, I guess you could say it shocked a lot of people and even me because this is something that showed us that these humanoid robots are already working in certain factories, they're already being a part of the assembly line and they're already going to be taking some jobs or assisting in, in some roles. Now I personally believe that with the rate of production, I'm not sure it's going to be taking so many jobs because humans are so cheap and so effective at a variety of different tasks that it wouldn't make sense to get rid of all humans. But I do think for a lot of, you know, those tasks that are really repetitive and, you know, tedious and just those random things that you can't really do with like, you know, those robots that have those arms. I genuinely forgot the name. There are going to be a lot of applications for these humanoid robots. So it's going to be really, really interesting. And this wasn't the only company that has done this. There was another company that showcased their robots. There was like a robot army. And it was weird. It was weird because I saw the robots walking as if they were human. And if you showed me this clip, I think 12 months ago, I would have said that CGI, we're not going to get there for at least two years. The fact that we're here now, it shows us that the economy, I mean, this could speed up the economy by a decent, decent amount. And I mean, I'm surprised that this is something that is taking place so quickly. I'm going to skip to the part that really threw me off, which was this right here. And you can see that we've got these entire armies. And yeah, I think this is going to be pretty crazy because of course they're scaling their data collection efforts. So overall it's going to be pretty incredible to see the amount of data that they're collecting, which is of course going to be an entire feedback loop that just accelerates the entire industry. Now something that did change the industry was the fact that we got Elias Sutskova saying that pre training as we know it will end. Now this is one of the most insightful talks because we got the information about super intelligence. He was speaking about how future systems are going to be agentic, they're going to be smart, they're going to be hard to understand. And the craziest thing is that these systems are going to be unpredictable because we're not really going to understand what they're doing because they're that much smarter than us. So it's something that is, I guess you could say, kind of worrying. But it does make sense considering the fact that these systems are super intelligent beings. Well, not beings, but like, you know, minds that are just able to think in ways that we can't, which is pretty crazy. And of course we're not at the stage where things are self improving. But I think when we do get there, that's when things get really spooky. Now with pre training as we end, of course this is a paradigm shift because the thing is that we have all this data, we put it all into the models, we train the models on that data and of course we would then get an output that we would fine tune and get for specific tasks. That era is of course changing. So I personally believe that this is the era of innovation because whilst everyone was on the same, just data is all you need. Of course now it's new and innovative ways. I don't know what on earth they did with claw 3.5 sonnet, but that model is just completely crushing it in terms of the benchmarks and how smart it is for it not being a model that searches through multiple possibilities. But of course I do think it's pretty interesting that that model is still relatively expensive. So this is going to be something that is super interesting because if we are ending the pre training phase of collecting data, then it means that like I said before, we're going to have to enter new methods of scaling these models, which means that once we do find a way that's actually new and it's not just based on the data we have, the that is when we actually start to get models that are really smart and potentially generalized out of their distribution. I don't know what those methods are. Maybe it is V Jepa, maybe it is Meta's advanced machine intelligence with the likes of Yann Lecan. But whatever it is, it's going to be something that, you know, we probably can't think of now, but that is going to be whatever innovation that is. I'm not sure what lab that's going to come out of. It's going to be a major jump in terms of the capabilities. Now we also do have Google Amy, which is a, you know, it's a previous project that I don't want to say I slept on, but it's something that they received recently, a major update. So recently this was something that helped clinicians diagnose people properly and it was something that managed to do a lot better than the original clinicians. Now essentially this is, you know, shifted now to subspecialist medicine, addressing the critical shortages of specialist expertise worldwide in areas such as cardiology and oncology. So this is something that, you know, demonstrated proficiency in diagnosing and planning treatment for complex conditions and also it showed potential in assisting cancer diagnosis and treatment, particularly in synthetic breast cancer cases. So they partnered with Stanford University to create an open source database for cardiac diseases, developed a pilot evaluation rubric to assess the performance alongside specialists and the AI system. The results were preferred or equivalent to those general cardiologists in many areas, and it improved their decision making in 63% of cases. So this is going to be really interesting because I think soon we're going to get the, of course, medical AI's in. The only thing I think that's really stopping medical AI from becoming mainstream is the fact that there is the possibility of opening yourselves up to some kind of litigation. Because if an AI gets something wrong, then of course, it's like you guys use the system, the system got it wrong, and now I'm suing you because of, you know, your technology. So hopefully we get to a stage where we can only get the benefits and not the drawbacks. Now, now we also had Nick Bostrom actually talk about the fact that when these systems advance so much, it's going to result in potentially mass unemployment.

In this scenario where AIs truly achieve general intelligence and then superhuman levels of general intelligence, I think it is much more difficult to think of tasks that we would actually be helpful. I think there are, like, a few. You could look at cases where, say, consumers have a direct preference that the work be done in a certain way. So right now, maybe some consumers pay a little bit extra for a trinket that was made by some politically favored group or some indigenous craftsperson, as opposed to produced in a sweatshop like in Indonesia or something. Even if the actual trinket is equivalent, some people care about how it was made. So that would be an example where maybe human work would still be needed or people maybe prefer to watch human athletes compete in the Olympics, even if there were robots who could run faster or box harder or whatever. But. So with those carve outs, though, I think it does look like we could have full unemployment.

And we also had Sam Altman also talk about the fact that people are going to lose jobs as well.

People will lose jobs. Many new jobs will be created. I think much better jobs. We feel a responsibility to educate society as we see it. We'll be right about some things, wrong about others, and to be as good as we can at being stewards of this technology. But not everyone's going to like all of the impacts. But this is coming. This is a scientific achievement of humanity that is going to get embedded in everything we do. The definition of work changes. Someone that lived thousands of years ago that was trying to be like a subsistence farmer probably would look at what you and I do now and say, that's not really work. They're just having fun. I think it will be so clear once these robots are off doing all, all of these other things and that there's some special human things. And we don't really care about that much what those robots do in the same way that we don't care that much about the machines and factories making stuff for us. But we'll find stuff to do that we really care about.

Now, of course, we had the biggest news from the week, which was the fact that AGI is potentially here. The new O3 model managed to smash through the previous ARC AGI benchmark that many heralded as something that wouldn't have been possible for, I guess you could say four to five years. But this year, essentially the AGI benchmark was broken. Of course some people are speculating whether or not this is AGI, whether it is AGI. Of course the debate is up to you. But some OpenAI employees have stated that this is AGI. A few of them have tweeted that this is AGI. Don't know what you guys are talking about. But with that being said, it marks a significant milestone because now the only thing left to do is of course, number one, bring the price down for this product and then of course ensure that this is agentic so that it can actually be useful in real world economic scenarios. And I think those two coming sooner than we do think, which is of course, you know, it does have stark implications. I mean, when we take a look at the public responses, you can see that people are changing their sentiment with regards to what the model is able to do, which is a sign that this thing is speeding up. And one thing as well with speeding up is that like the O paradigm is one that trains every three months. I'm not sure how they're doing it, but all I know is that they get a new iteration of the model every three months, which means that the kind of jumps that we're seeing in this graph right here, like the ARC AGI score progression, it means that it's probably going to be steeper than that in the future. So this is probably speeding up. And I wouldn't have thought that the 01 series would have sped up because of course, you know, you've got that data collection, but now that you don't need to collect all that data and you're actually, you know, you've exhausted all of that already and you don't need this huge pre training phase. That means things are going to move much faster because it's just certain innovation that allow you to potentially maybe search over possible solutions, but potentially in a more efficient way leading to even better results. And when we even look at the Code Forces competition, you can see where these models rank. And I mean, take a look at 04. I mean, where's O4 going to rank? Is it going to be at 3,000, 3,500? I mean, these systems that we're getting to, we're very, very quickly approaching, you know, some really, really smart systems. And of course, if we're looking at AGI, you can see that someone said, I love OpenAI, but they're truly terrible at naming things. And then someone said when we get naming right, you'll know we've achieved AGI internally. And Sam Altman says that maybe they're going to call it AGI1. So that's going to be super, super interesting with regards to that. Now, if you enjoyed today's video, hopefully you guys have a wonderful weekend and the rest of your week and of course, a great Christmas. And I'll see you guys.