HUGE AI NEWS : Ex OpenAI Researcher REVEALS What AGI Will Be Like, New Humanoids And More

So, coming in with the story number one, we have a new Chinese humanoid robot. They've honestly been developing these at quite a rapid pace and now we're starting to see the different ways in which they are managing to make these robots even more flexible. We all know that there is actually an initiative for 2025 for China to be the leader in terms of robotics. And today we got Limx Dynamics, which is a Shenzhen based startup foundation founded in 2022 and it's actually, you know, got success with its new hominoid that they are essentially aiming to rival Boston Dynamics. So if you aren't familiar with this current pose that it is doing, this was something that Boston Dynamics robot, the Atlas robot, the very famous one, actually did in a demo video and that was something that this is trying to replicate. So it's pretty interesting because they can quickly, quickly replicate many of the things that are going on in the US and, and considering the rapid pace technology, I wouldn't be surprised if many of the robots that we do see. And as I was saying, it's going to be really interesting to see how the rapid advancements in humanoid robot technology manages to spread overseas. We all know that humanoid robots are going to get a huge boost in 2025 due to the physics simulations actually being incredibly more efficient in terms of their speed. And we're going to likely see humanoid robots perform in ways that we've never seen before. It's going to be really interesting to see China also manage to mass produce these robots and actually start to have them working in factories. And considering the fact that in China they manage to have really, really efficient manufacturing, it's going to be quite surprising if they do manage to pull off some really, really effective robots at an industrial scale. That is going to be something I'm looking out for because we've already seen some instances of that. So that's going to be interesting. So point number two is rather interesting because in the past couple of days in the AI community, what I've noticed is that there is a very, very clear thing going on where individuals are starting to talk about artificial superintelligence more and more in a realistic term, it's no longer seen as something that is a fad or something that is this exaggerated term that is somewhat clickbait, but actually as a real possibility for future AI systems. And Logan Kilpatrick, the Google lead product lead for Google AI has actually said that straight shot to ASI is looking more and more probable by the month. This is what Ilya saw now he does further expand on this, but this is a significant tweet because this is something that we hadn't really seen before across the wider community. The only person that was really talking about artificial superintelligence at the time was of course Ilya Sutskeva. And even remarks or claims of AGI were being currently disregarded as as people who were far too hyped up on AI hype and AI news and just seriously in the AI bubble. But right now you can see that other labs and other individuals are talking about superintelligence. Of course this is referring to Eli Satskova's mission trying to refer to artificial superintelligence. And if this is true, it kind of does make sense considering the fact that that is quite likely what Elias is pursuing. And considering the fact that Google released their recent model, which is actually also something like Test Time Compute. Now them starting to say that, look, ASI is on the table. It kind of seems that we're all headed in general direction now with artificial superintelligence. He actually goes on to explain a bit more. He talks about, you know, how he founded the company to do a straight shot to artificial superintelligence, no products. And then of course he before saw this as unlikely to work since if you get the firewheel spinning on models, you can actually build a real moat. But the success of scaling Test Time Compute, which of course Ilya Sutskever saw early signs of, is a good indication that this direct path of just continuing to scale up might actually work. He says we're still going to get AGI, but unlike the consensus from four years ago that it would be the inflection point moment in history, it's likely going to look like just another product release with many other iterations and smaller other options within a short period of time. Now, if you're wondering what's actually going to happen in 2025, I did do a video covering Sam Altman's predictions. But if you do want to know a little bit more, we can take a look at some of the common themes. Of course he says artificial general intelligence. This is something that Sam Altman alluded to in a podcast. Well, he didn't really allude to.

He.

He actually stated it that we're going to get that in 2025. That of course will be fascinating. He also said AI agents a much better 4.0 upgrade, which is definitely long awaited for the people who sincerely wanted GPT5. Much better memory. This is of course something that I spoke about before. Much better memory is Certainly coming. A longer context window is definitely something that he wants and a lot of people do need. A longer context window allows you to essentially input more in, which is of course really effective for doing a variety of different tasks. I have to be honest with you guys, using Google's 2 million context length window is something that is remarkable for a variety of different tasks. A deep research feature. And this is really interesting because this actually takes aim at Google's. I wouldn't say it's one of their core offerings, but genuinely, it's the most underrated AI tool right now. I made a video on it, I made some super prompts from it, and I think it's probably going to be one of the biggest tools of 2025, because it's basically AI agents doing research for you, but people don't know that yet. It's just called deep research. So I think something like that is going to be really, really crazy, especially when OpenAI release it, of course, a better Sora that is, of course. So with a better Sora, that's going to be really interesting because one thing that people don't understand is that when SORA was the leading model, OpenAI didn't really have any pressure to release anything better. But now that they're not the leading model, we're actually going to have to see if SORA is actually better than VO2, because which is going to be a really, really hard feat. So when SORA gets released, it's quite likely that it will be better than Google's VO2, which, when you think about what that means for the video creation industry, that is going to be something that's absolutely incredible. So that is going to be something super interesting that I can't wait to see. It also says, interestingly, many great updates we have coming were not mentioned at all or very little. And this is really fascinating because Sam Altman actually says that many great updates that they have coming weren't mentioned at all or very little. So it's quite clear that OpenAI's roadmap in 2025 is going to be one that is one is full of surprises. So that means that, you know, just keep your eye out for anything that they might release, because it's quite likely that we're going to get caught off guard again with a tool or product or a service that we probably do need, but haven't yet requested. And this is something that I think one of the key things that makes OpenAI so successful is that they actually don't just make AI stuff, they they actually make them into really good products, which is why of course, when you use ChatGPT, the user interface, for example, even when you're using Sora, it's actually remarkably more effective than any other AI tool out there. And that's the reason that I think this company will still be successful. Now, unfortunately for OpenAI, if we're actually talking about the company being successful, we can actually take a look at what the godfather of AI has to say about OpenAI. The godfather of AI is Geoffrey Hinton actually throws support behind Elon Musk's lawsuit against OpenAI. And they're actually supporting Elon Musk's lawsuit seeking a preliminary injunction that would stop OpenAI from becoming a fully for profit company. If you aren't caught up with the news, unfortunately for OpenAI, they are trying to become a for profit company in order to further the mission. And of course Elon Musk has filed an injunction to stop that because essentially OpenAI were a charity and he poured millions of dollars into said charity. And now that the said charity has taken those donations and has turned that into a for profit company, it's essentially something that they just aren't happy about, which is of course understandable. Now it might just be the fact that Elon Musk is a little bit jealous that of course OpenAI are a bit ahead and he's trying to slow them down. But it's very interesting to see that the godfather of AI is now supporting this. We can see that he says OpenAI was founded and as explicitly a safety focused nonprofit and made a variety of safety related promises in its charter and it said it received numerous tax and other benefits from its non profit status. Allowing it to tear all of that up when it becomes inconvenient sends a very bad message to other actors in the ecosystem. And this is very true. If we have a company that starts out as a charity and then over time manages to just say, you know what, now let's become a for profit. That is a very bad precedent for companies because other individuals will try and use this, this structure to essentially ensure that their company is going to be successful as well. So this is going to have all sorts of implications. If it does go through that OpenAI is allowed to do a for profit conversion. That's going to have to be really technical because they don't want to open the floodgates for, you know, various individuals starting their own non profits and then turning them into for profits. I'm not someone who's, you know, completely educated on the law But I do know that when you do allow certain things, it essentially sets a precedent, meaning that other people are going to try to do the same thing as well. And of course, landmark cases do get referred to all of the time. So this is going to be something that is really, really important. I truly am not sure of the decision of this, but it will be very interesting to see. I mean, both of these companies are probably going to have the best lawyers on either side. So it's going to be super interesting to see where OpenAI manages to fall. Now here we have Sebastian Bubek, who is someone that actually worked on the BI series of models. These were the smaller series of models from Microsoft that basically showed how important training data was. And in this interview clip he actually talks about the recent study in a video which I covered, where we actually do get to see the fact that O1 surpasses doctors, well, physicians in terms of diagnoses in certain domains, which is truly fascinating. And it's a remarkable jump, an improvement from GPT4. And I think it sets a really interesting statement for the future because if these diagnosis tools are that much better at humans, the question I have to ask myself is at what point does it become reckless to not use the AI tool?

Just recently there was a study that came out on medical diagnosis comparing human doctors to GPT4. Human doctors on that benchmark were getting 75%. Okay, so three out of four were correct. GPT4 gets nine out of ten on medical diagnosis. Like this is something that we care about, I care about in the real world. And GPT4 is already beyond human. Now in that same study, they are already foreshadowing what happens with O1 and saying that there are now cases where human Doctors are at 50% and O1 is at 90%. I mean, this is just, you know, completely incredible.

And I'm not going to spend too long on this because we've already covered this. But I mean, just take a look at the paper. It's about like superhuman performance of a large language model on the reasoning tasks of a physician. They tested them in four categories. I mean like, you know, when you look at, you know, 01 preview and guys, this is just 01 preview, okay? This is not like the second iteration of the model. Just imagine the third iteration of O1, the fourth iteration with test time compute. I mean, how accurate is that model going to be? Especially if it's fine tuned on, you know, medical diagnoses and stuff like that. So I mean, it's going to be really interesting Especially when you can have a AI that's able to generate synthetic data and then train on that synthetic data. So even if there is a super rare case, the AI is able to, you know, pull it from its training data, that wouldn't exist in the real world. That really would have. Dr. Stump. So I think this is genuinely a silent revolution that's going on because when we even check right here, you can see 01 preview in terms of the score out of 100. I mean physicians and GPT4, you can see the kind of results they get. An O1 preview, okay, preview is literally all the way up there. I mean it's, it's truly remarkable on, on landmark diagnostic cases as well. We can see O1 preview performs nearly at 100. It's just something that is, yeah, pretty incredible stuff. I would say that this is a good sign though for anyone, you know, with kind of health issues, because I think right now we're just getting the diagnoses, but eventually on the tail end of this, we are going to get some probably significant medical breakthroughs by the end of this decade or by the end of next decade, certainly. So it's going to be really interesting at a time for prosperity, hopefully. Now, interestingly, there was this model called DeepSeek V3. And this is something that I do want to say essentially shook the AI community. This was because this is a model that basically managed to perform on par with a lot of the current LLMs. But that wasn't the only thing that this model managed to do. This model currently managed to essentially get the same benchmarks at a fraction of the cost. I've heard that the cost of this model is around 20 times cheaper than GPT4. And from my sources, many people are currently switching for this to their API usage. Now, Claude 3.5 Sonnet still remains number one in terms of the coding capabilities, but from what I'm hearing is that many people are now switching to this model in terms of their GPT4 applications. Now this is something that is pretty good because if you've ever wanted to build an AI applications, you and maybe the, I guess you could say API costs were something that were a factor. Now Deepseek v3 is coming in as a very cheap alternative that allows you to rapidly prototype what you want without facing heavy inference costs. So this is going to be a true shake up to the market. And there was also some information regarding this as well that was a little bit controversial. So essentially if you asked this model what it is, it said, I am a Large language model developed by ChatGPT. That is what the model would output. Now, of course, that doesn't really make sense. If the model is Deep Seq v3, it should say, I'm Deep Seq v3 developed by, you know, the company DeepSeek. So there was some information going on out there that basically said that they trained on OpenAI's model's output, which is of course against OpenAI's terms of service. The only thing is, I'm not sure how OpenAI would even recognize that a company is able to do that. I'm sure they've got advanced ways to do that, but considering the state of AI right now, I'm not sure that's something that OpenAI would even want to pursue at the moment. But when we do take a look at the benchmarks, you can see here that it does seem to basically outperform GPT4.0 on a variety of different benchmarks. And if we're being honest about what GPT4O is currently, it isn't the best model in terms of an LLM that you could use. You know, Google's newer model is better. I actually use Google's model every day. Now, of course, for coding, you've got Claude 3.5 Sonnet. That model is, you know, better in coding aspect. So it's something that just manages to, you know, put OpenAI's leadership, of course, in a state of contention because we have a model that is cheaper and basically on the same level. So people are always trying to find ways to save money. And if you can, you know, find a way to save money by using this, of course that is something that people might do. Now, whether or not people are going to go to the hassle of actually going to a deep CP3 account because you know what OpenAI are essentially doing. You know, they've got advanced voice mode, they've got a variety of different other things that still make them a very good product to use. I still think that the market has allowed shakeups that are really fascinating. Now, in addition, we also do have this thing called Reproduce, which is where there's step by step blueprints that reveal how to recreate OpenAI's reasoning abilities from scratch. So there's a paper that actually presents a roadmap to reproduce the O1 model using reinforcement learning, focusing on policy initialization, reward design, search and components. So this is going to be interesting. I'm actually going to be covering this in a video probably tomorrow. But this is fascinating because if the secret of this is out, it basically means that the secret of AGI is out. And if that is the case, and that means we are about to get AGI globally, and if we're about to get AGI globally, then that means that there is about to be the acceleration that we've never seen before, because none of these countries want to lose the AI risk. Now, if we're talking about, you know, large language models and different architectures, and of course the O1 architecture, there is this fascinating paper from Meta called LCM Large Concept Models. And with large concept models, this is essentially a kind of model that basically tries to make the model think how humans think. So, you know, how, you know, large language models think in, you know, it's tokenized and just the tokenizer stuff. Basically, you know, this is probably a better explanation. This is an AI researcher and he said that I'm more and more confident that tokenization will be gone because of course humans don't think in tokens. Tokens are hard coded abstractions in LLMs that lead to weird behavior. You know, LLMs can solve PhD math questions but cannot answer. Is 9.9 bigger than 9.11? So meta is shifting, you know, LLMs to LCMs, which are, you know, changing from next token prediction to next concept prediction. And the concept is treated as a sentence, you know, representing an idea, representing an abstract idea or action. And I think this is something that's, you know, really interesting to see if it gets embedded into other architectures because we all know that the tokenization feature, it will not feature, you know, it's a part of the architecture that is something that has caused problems before. And we've seen Meta previously tackle these kinds of things where they're looking at things different to everyone else. So I think if Meta can of course continue to push that, it's going to open up a really nice area in terms of AI research because we're going to get areas that most people currently aren't focused on. And this is something that's really important because if you want to get truly smart systems, having them thinking in different ways or ways that are really similar to us is going to really advance that. Now we also got this clip that has been doing the rounds on Twitter and this is where we get the ex OpenAI employee, Daniel Cocotaldjo, actually explaining what AGI looks like when it's going to be released. So this is a super fascinating interview because I don't think there have been many individuals from Frontier Labs actually releasing the kind of information we get in this short one minute clip. And this is something that is fascinating to me because one of the things as well is, you know, it kind of coincides with what Sam Altman said previously where he basically said that, you know, we're gonna get AGI and it's just gonna be like another day for the average person. So take a look at what he says here.

What you would see eventually is some sort of very exciting announcements and some sort of very exciting products being rolled out and things like that into the world. But by the time that's happened, you know, but by the time you actually see the plumber, the robot plumber walking around, AGI research will have been automated for like a whole year or more. And the things living on the data centers will be super intelligent, like literally across the board, qualitatively, quantitatively better than the best humans at everything. And they won't just have been designing robot plumbers, like there's a million of them. They'll be working on lots of things at the same time. They'll be the robot plumbers, but there'll also be new weapons, new drones, there'll also be all sorts of new software. There'll also be all sorts of new basic science that will have been figured out, new types of physics, new, you know, chemistry, new biology. It'll be this giant explosion of amazing new technologies that will be sort of like the, the scientific and industrial revolutions. Like if you think about what happened in 500 years, from like 1500 to 2000, it's just this insane explosion of intellectual progress in science and then also like physical progress in terms of like machinery and tech, you know, tools and stuff like that. I'm imagining a similar sort of explosion happening over the course of maybe two years, all under the control of one AI project, probably.

So for those of you who are wondering how AGI gets here, that is going to be something that is super intriguing to me because of course I would love to know what life is like when AJ gets here. And of course we do a lot of speculation on this channel, so it's going to be super interesting to see if his predictions actually do come true, if in the sense that we're going to start to see a variety of different inventions and products getting released. And then if we're actually able to tell if, of course, AI research has been automated. Now, interestingly enough, gearing up into 2025, we can see that the Google CEO Sundar Pichai has actually told employees to gear up for a big 2025 as the stakes are high. Now this is super interesting because we've seen that in the last few weeks, in the last quarter of 2024, Google really showed everyone that they are still the kings of this AI stuff. And I think this was surprising to many people in the AI industry because what we saw from Google before was rushing to get products out. We saw sloppy releases, we saw delayed releases, and we even saw releases that just didn't happen altogether. At Google's IO conference, we saw coming later this year more times than we had actual product releases. But with the recent Google vo, with the recent new model, with the recent thinking model, I mean, it's been absolutely an onslaught of different products that we've had. It's been nothing but gems that we've literally had with Deep Research. I mean, the list just goes on and on and on. Google have just, you know, raised the bar to, I think, definitely the gold standard now in terms of across the board AI, in terms of usability. So it's going to be super interesting to see where they stand. Now. If you actually read further into this article, we can see that, you know, it talks about how 2025 is going to be a critical year and it's important that they internalize the urgency of this movement and they need to move faster as a company, as these stakes are high and that these are disruptive moments in 2025, they need to be relentlessly focused on unlocking the benefits of this technology and solve real problems. Now this is good for us because of course it means that we're going to be getting some killer products. And like I said before, a lot of the things that they've released have been really tremendous in terms of my productivity and just downright the things I've been able to do. You know, I've spoken about the fact that Deep Research is an amazing feature. Also the fact that, you know, the agent browser, when I've seen the demos of that, it looks really, really amazing. And I can't imagine the amount of time I'm going to save with that. So it's going to be super interesting because of course, you know, if you thought that 2024 was crazy, they are saying that 2025 is the critical year. So if you're someone who's a Google fanboy, or Maybe even an OpenAI fanboy, definitely keep your eyes on Google because they're going to have a lot of products that are coming out. And so here you can see that they actually said they know that they have some work to do in 2025 to close the gap and establish a leadership position there as well. But of course they know that you know that they want to be number one. You can see right here that they said that I expect some back and forth in 2025, but I think will be state of the art. So Sundar Pichai is basically stating here that in 2025 he expects them to actually be the number one AI company in terms of overall market share, in terms of just how good everything is. And honestly, by the end of 2024 we could argue that Google is number one because when we looked at the benchmarks, when we looked at the product releases, I mean the stuff they did was really, really impressive. So, you know, right here they're stating that, you know, he acknowledged that Google has to had to play catch up, but he says in history you don't always need to be the first, but you have to execute well and be really best in class as a product. And in 2025 that's all it's all about. So for you consumers out there, you know, don't sleep on your Google devices just yet. I'm sure there's going to be some major updates coming as well. And of course with this, we do have this individual, Jun Yangling, who is, you know, building Qin models at Alibaba, actually responded to this tweet where someone said they want a Sonnet Level 70B model, which is essentially just a smaller large language model like ChatGPT. Someone said that they wanted a sonnet level version of that and he said wait on me. So I mean it's going to be really interesting to see if we manage to cram more intelligence into These smaller models. A 70B Claude 3.5 sonnet level model. I'm not sure if they can pull that off because I think some things are just achieved through scale. I and I think even if you try to be as efficient as you can, you might not be able to just get that kind of intelligence out of the model from that size, if you guys know what I mean. But then again, we've seen, you know, a variety of different things become more efficient, become faster. So I wouldn't be surprised if this is something that occurs. And I'm never going to bet against the trend of AI because I've done that in the past and it doesn't work well and AI is an exponential trend. So by the end of 2025 we could be in a remarkably different position. Now. In addition, what we also got this week was something rather interesting. We finally got the definition of AGI and this one was really good because we now know that OpenAI have defined AGI as a system capable of generating $100 billion in profits. Now, I don't think this is like, you know, the specific definition of AGI. I think this is more like a corporate cookie cutter term just to please the individuals at Microsoft. Because OpenAI's deal with Microsoft is, you know, it's kind of on shaky grounds based on all the things going on. But, you know, they're negotiating the agreement and I think Microsoft, of course, wants, of course, to retain control over some of this AI stuff. But it's going to be super interesting to see where things head with them, because, of course, I think that, like, a lot of people are starting to say that, you know, you've compared Microsoft and you've compared Google, but Microsoft doesn't have their own models. So, like, once Microsoft's contracts end with OpenAI, what are they really going to do as a company? And that is, of course, you know, something that brings us to the question, but it's not that crazy. It's not that big of a deal. I'm sure they'll work out some agreement. And one of the last stories we have here is actually super interesting. So. So we actually have this video of a humanoid Robot released by Matrix 1, and this one's rather fascinating. And maybe other companies can take a leaf out of this book, because I. I'm more intrigued about this company than I am any other company. So this is a company that I don't even know the name of. I'm guessing it's called Matrix. And they've released this robot that I've only seen one video on. So the captions you're seeing on screen, I didn't make those. I just found this video on Twitter. I couldn't find the original video. I Googled this robot. Nothing has come up so far. But the technology in this robot does look futuristic. It does look really advanced. I mean, I haven't really seen any kind of humanoid that looks like this. I think the last robot I saw that looks like this mildly is the, you know, Honda Asimo. And that was over perhaps 20 years ago. But this robot right here, it literally looks like it's from the future. And the fingers look really flexible. In fact, if I had to say it, it kind of resembles maybe an astronaut suit like the SpaceX one. But there isn't much to go off here other than this looks really effective. I mean, if this is able to do some remarkable tasks, it's going to definitely feel like the future. I do, of course, wonder how much this kind of product is going to cost. But of course, you know, you can keep up with the progress with that account on Twitter that manages to cover all of that stuff at Cyber Robo. I will leave a link to that stuff as well. And hopefully you guys have had an amazing week. And hopefully you guys have had an amazing New Year's as we go on into 2025.