Big AI News: Google Gemini 2.0 Date LEAKED! OpenAis New AI Project, New Robots, And More

So one of the first stories in AI wasn't in fact the new R1 model. There is another strange competitor that has seemingly come out of nowhere to challenge 01 preview. So someone tweeted yet another 7th grade math problem that only 01 mini can solve and they basically showed chat logs where O1 mini was able to be the only model that was able to get the answer correctly. But if we look at this model now, this model is one that's somehow manages to get the answer right. And it says check reasoning process. And the possible values that X can take to have exactly 30 zeros in the first 100 terms of the sequence are minus 3, 2, 6 and 7. And I'm not going to go into all of this reasoning details, but what I will say is that this is super interesting because the KPU is essentially Misa kpu. I'm not sure if that's how you say it, but this is a model that was actually announced to quite a long time ago. Now I'm going to show you guys exactly what I mean by that. So I think it was around five months ago or so five, six months ago, we basically got this bombshell screenshot from this lab that apparently managed to develop some kind of. I wouldn't say I don't even know if it's an algorithm, I don't know if it's a way to use AI identically. They haven't exactly revealed how they are using the model, but they seem to have some kind of framework that allows to prompt the model in such a way that it is able to get far, far better reasoning outputs than anything we've ever seen before. For example, you can see on the GSM 8K, you got like 97% on the math benchmark at the time. That was state of the art at the BBH. All of those ones, they're around to 95 to 100%. And I remember at the time of this being released, there was currently a lot of skepticism because, you know, you can't just come into the AI space, try drop a screenshot saying, look, my model beats yours on several benchmarks, and then just disappear into the wilderness. Nobody's going to believe that that thing even exists. But that's exactly what they did. But now they seem to have come up with something that actually works and is actually a real product. So a simpler question that they did ask this model was how can a person with no hands clean their ring finger? And then you can see that this model reasons with itself. And then it says this question can contains a logical impossibility. A person with no hands would not have ring fingers to clean. Since ring fingers are part of hands, it's similar to asking how a person with no feet tie their shoelaces. The premise is self contradicting because the absence of hands means the ring fingers do not exist. So I think this is the kind of model that trick questions simply don't get the model because I think that this is the response that we wouldn't really get from other models. I tried to use this question in Google Gemini and and you can see that the response I managed to get. I can also show you guys this one here where I actually decided to put this exact query into Claude from Anthropic. And you can see right here that the model didn't respond right. Now I personally do think that yes, this model does use some advanced reasoning if it's able to solve problems that only.01 has solved. But I do think one of the key things a lot of people are now going into that gray area with is missing the true intelligence of models due to, I wouldn't say flimsy benchmarks, but I would say due to vage benchmarks. Because for example, right here, what exactly is this question testing? Is it testing empathy? Is it testing reasoning? Because some people would argue that if a model says you can't wash your hands if you have no hands, yada yada yada, some people would say that's completely right, but other people would say that that is completely unempathetic and I'm going to show you guys why. So, so I don't know if you guys know about this model. This One is the R1 Lite model. This model is basically quite like OpenAI's 01 model and essentially we are able to see what the model is thinking before it manages to respond. So you can see right here that it literally states that how can a person with no hands clean their ring finger? And at first glance it seems pretty straightforward, but when you think about it, it's actually quite a complex problem. And it says what does it mean to have no hands? And, and does it mean that this person has no hands at all? Or perhaps they have some prosthetics or some sort of assistive devices? The answer might vary depending on the specifics, but for the sake of this discussion, I'll assume this person doesn't have hands and is looking for a way to clean their ring finger independently. So basically what I'm trying to say here is that if this question allows the model to basically assume that the person has some kind of prosthetic. I don't think that the model should be graded wrongly on that answer. I think we should have, you know, questions that are really, really granular in terms of the preciseness and thus the expected output. So this is something that, I mean, I'm really excited for this because if the model is, you know, advanced at reasoning and is on O1's level, and if they do manage to get to even a fraction of these capabilities, I think this is going to completely change the game because they've got an agentic framework wrapped around AI and it's quite likely that that is going to boost model performance even more. Now in terms of model performance going even higher speed. Something that was rather crazy was of course this. So this is a LinkedIn post from Danbo Ren, the regional head of Google Cloud AI Customer Engineering, and he actually made a mistake that is now pretty viral on social media. So the crazy thing is is that he said Gemini experimental version 1121 has been gaining great momentum leading ranking on Chatbot arena and please stay tuned for Gemini 2.0 next release during the second week of December. Now, I do know that this information is quite likely true because this information was changed. So this original post that was posted on I think LinkedIn was actually edited to say stay tuned for Gemini soon. This was the original tweet. So this is something that is probably going to be real. And I'm not going to lie, we do know that Google Gemini is just around the corner because because they have been teasing it for quite some time and we've seen their recent models manage to surpass OpenAI on certain benchmarks. So I wouldn't be surprised if we do get Gemini 2.0 during the second week of December, meaning that would be between the 10th of December to the 13th of December. Now of course there is the possibility that they could potentially change this due to the fact that with AI releases there is a lot of competition for the spotlight. A lot of the times when we see AI models being released and there is OpenAI trying to push their models on the same exact date. So it's quite likely that if Google are going to be releasing Gemini 2, OpenAI might release something just to try and upstage Google as they have in the past. Literally the last time OpenAI upstaged Google was when Google released Search Grounding. It was something that I was supposed to cover, but due to all of the OpenAI news that day, I literally didn't even have time to so this, this is something that can happen again. Which means that during the second week of December, I think that it's quite likely we get Gemini 2 and it's also quite likely that we get a major OpenAI release because they usually try to upstage Google releases. Now, of course, Google might actually change the date to throw off OpenAI. So that is something to pay attention to because we know that these, you know, these companies are really, really fierce and, and they're really, really competing for a lot of stuff right now. So it's going to be super interesting to see how that releases. But if you haven't, I would recommend trying out Gemini Experimental because it is definitely one of the models that is a step ahead of other models. And for those of you that say it can't reason that well, I've actually changed my prompt structures when using Gemini and I've gotten better responses out of it. So I think a lot of the times when we think a model isn't that good, we just need to try different methods of prompting and then we can see that it usually manages to give us what we want. So what could OpenAI be doing to upstage Gemini? Well, apparently they are going to be launching a browser very soon. I don't know about you guys, but I think you guys might be paying no attention to the AI browser space. I've tried to talk to various individuals that are not in the AI space about the types of things that you can do with AI browsers and they just haven't been paying attention. So essentially what this article states here is that a of lot OpenAI is considering developing a web browser that it could combine with its chatbot and has separately discussed or struck deals to power search features for travel, food and retail websites, according to people who have seen prototypes or designs of the product. So essentially what we're looking at here is a situation where OpenAI are looking at once again to upstage Google by launching their own browser. Now this couldn't come at a worst time because I want to show you guys a piece of news that is pretty, pretty crazy. So one of the things that Google is unfortunately suffering from is apparently it's monopoly. So there was this recent case about Google's monopoly and apparently they might have to sell Google Chrome to end this search monopoly. So it will be interesting to see if Google has to sell chrome and then OpenAI steps in at that moment because that would be absolutely incredible. So things in the search space are really, really changing. So I think that this is of course something to pay attention to and there's also a decent amount of information that is with this.

So one of the moves that OpenAI.

Has recently taken was that they've actually hired two people that were instrumental to Google's success. And I think this is rather interesting considering the fact that there are some reports that they aren't even close to, to a browser. So I mean, it's going to be super interesting to see exactly what goes on here. I mean, when taking a look at this, I am really surprised that OpenAI is trying to really upstage them on like just every single metric that there is. And I mean it's, it's so surprising because when we look at Google's market cap, it's like something, I don't even know if it's a trillion dollars. I think last time I checked it was a trillion dollars. But, but a trillion dollar company getting, you know, taken on by this small, fast moving company that's just a few billion dollars is absolutely incredible. And I do wonder if Google are going to take some steps to really change this. And we can see here, recently this was the person that, you know, worked on the VP of engineering for Google and worked at the browser company, you know, software engineer. And now this person is working at OpenAI. So this is going to be really interesting to see what kind of things that they do. Now the article does state that OpenAI isn't remotely close to launching a browser, but we do know that OpenAI tends to work with secrecy, so I wouldn't be surprised if they do. Now, of course, in the browser space there was also this piece of information which is basically where you can now search like a pro with Perplexity Pro Search. So I'm not sure if you guys know what Perplexity Browser is. A lot of people that I personally know do know what this is, but essentially what we do have here is a way for you guys to shop more effectively. So a lot of the times the reason that people actually use Perplexity is because this is the kind of model that allows you to find things really quickly and it means that you can save a lot of time. Now with this, what you can do is you can use the AI Vision model to snap a picture of something and then you can search for this model and then you can essentially go ahead and buy that specific model. So you can say, I need this item, but I need it under a hundred dollars. I need it in this way, I need it in this color, I need it. And you can simply use this model to buy and Check out really quickly. So I am wondering if this is going to be the future of E commerce where you literally just say, okay, I'm looking for a shirt. I want it to be this texture that, you know, material. I want it to be, you know, this thick or this thin. And then the AI gives you a bunch of different products from a bunch of different websites and then you literally just click and it's like, oh, I know exactly what shirt you're like. And it basically shows you one you buy and, and it gets straight delivered to your door. I mean, I wouldn't be surprised if there are some generative apps out there that are being built like that. In fact, I think I did see one prototype of a company that was building something like that. Like literally, you literally just ask for what you want. Like, oh, I need a new washing machine. Oh, can it be $400? Because a lot of the times, and this is something that I've personally experienced myself, is that you spend so much time searching for the perfect product online, wouldn't it be so much easier to just say, I need a product that is xyz. This AI searches across the Internet and you get a response of the top 10 things that are specific to you. So I will love to see if this product does get implemented. So then of course we had this chart that basically shows the next phases and the next growth areas of AI. So we've got three phases of AI. You can see we've got this phase one here. Of course someone else is going to explain it in a minute. But of course you can see here we've got phase one. So you can see right here we've got phase one, which is research attention is all you need, all of those papers. Phase two, which is what we're just ending. So you can see GPT3, GPT4, Claude3. You can see that this was a really, really nice area. This is another S curve. And you can see this is actually quite like the chart that I put in another video. This is basically where we're on this new S curve. And you can see that O1 is basically at the bottom of this S curve. So this is basically going to be where we are in phase three. So you can see, yeah, this is basically similar to what I showed the other day. Obviously this is a lot more detailed, but, but you can see that there's going to be a lot more innovations in scaling coming along the way. And basically this was from the scale AI CEO Alexander Wang. And he's basically saying the scaling phase has entered and we're moving towards the innovating phase where other breakthroughs are going to lead to superintelligence in six years or less.

I like to sort of break down and contextualize the modern era of AI into sort of three major phases. So the first phase was research, which really was roughly 2012 until 2020 18. So basically going from the first deep neural network, Alexnet, which was really trained to do basic image recognition. This was the era when all AI could do was tell you if there were cats in YouTube videos. And that was this incredibly powerful thing all the way up until the first GPT model. So the transformer and the first GPT model trained by alec Radford at OpenAI, that kicked off the next phase, scaling, scaling. So the first six years, 2012 to 2018 was sort of this research phase, then 2018 to 2024, really up until today is the scaling phase. You know, the amount of resources going into these models grew more than 10,000 fold over this time period, last year or this year, I should say something like $200 billion is aggregate going into training these models. And this has resulted in incredible advancements in performance. You know, we've gone from GPT1, which was nothing special, to O1, which is sort of PhD level in math and computer science. And then O1, I think, kicks off this new phase of we sort of think about as an innovation era, which is basically now until superintelligence. We'll see if that's six years or maybe even a bit less than that. The hallmark, I think here is that we're spending $200 billion on the models. We probably can't spend a lot more than that, probably can't spend 200 trillion on the models. So there's only limited amounts of scaling left in terms of orders of magnitude. And so we need corresponding innovations to sort of come alongside. Definitely advanced reasoning and test time compute is one of those. And we think there's probably a few more handful of others that will get us to superintelligence.

Now, interestingly enough, the Cohere CEO Aidan Gomez actually says that scaling models has entered the flat part of the curve. But these models are already so smart and that it's actually taking them a little bit of time to assess their actual outputs. And this is now going to be applied to different research domains. So it's interesting to see what other CEOs are saying about the same paradigm that we're currently entering. Where are we in scaling loss?

Like, how much capability improvement do you.

Expect over the next few years we're.

Pretty far along, I would say, like we're starting to enter into a sort of flat part of the curve and we're certainly past the point where if you just interact with a model, you can know how smart it is. Like the vibe checks, they're losing utility. And so instead what you need to do is you need to get experts to measure within very specific domains like physics, math, chemistry, biology. You need to get experts to actually assess the quality of these models because the average person can't tell the difference at this stage between generations. Yes, there's still much more to go do, but those gains are going to be felt in very specialized areas and have impacts on more researchy, more researchy domains. I think for enterprises and the general sorts of tasks that they want to automate or tools that they want to build, the technology is already good enough or close enough that a little bit of customization will get them there.

Now there was also this company that.

Actually went under the radar. So they, this is Studio, a platform.

For developers to run AI automations at scale and run A H, which is apparently the most advanced AI agent to date, which is going to be navigating the web and interfaces through pixel level interpolation and semantic understanding. And you can now turn your instructions into human like action with precision.

So for those of you guys who.

Are wondering why we haven't managed to get certain agents yet, other companies are also working on certain integrations with certain applications. And you can see that this is a company that is working slowly behind the scenes that is able to perform really, really well at a bunch of different tasks. I've noticed that a lot of these companies don't just come out with their stuff. They usually have this in, I guess you could say like alpha or beta. And of course this is one of the companies that has it right now. And I think they're alpha or beta stage. So this is going to be something that I think, think I'm excited for a lot because while yes, AI agents are really good and yes, they work really well, a lot of the times we end up using agentic workflows and there are really specific use cases that are just limited by an AI agent not being able to access a webpage, not being able to access your computer, to be able to do certain things. And so it is quite frustrating. Now of course, like I said before.

Reliability is going to be one of.

The main things for these agents. I mean, if you have an AI agent that is ordering you food, what.

If it orders the wrong thing.

What if it spends too much money? What if it posts something on social media that is completely false and it's not supposed to? These models sometimes do hallucinate. So I really can't wait to see what this is. I've personally signed up for the beta. I ideally want to do a video on this as soon as possible. But if you're actually wondering about how well this works, we can see that if we look at the success rate for this kind of agent, we can see that Runner H is 66.9% and clawed computer use is at 52%. So currently it seems like they are state of the art. So remember how OpenAI said that they were going to release an AI agent in around, I think, January of next year? So it seems like this company has put them in a bit of a predicament because they have another wall to climb in terms of usability. So OpenAI is going to have to improve their AI agent in order for.

Them to want to release theirs, because they're going to have to make sure.

That theirs is also the best.

So it's going to be super interesting.

To see see exactly how that works. Another thing that happened this week was.

The fact that GPT4O managed to get.

This creative writing update. This was posted on Twitter by Kyle.

Shannon and I wasn't truly aware. So basically the reason this was so surprising was the fact that, like, because I don't really use ChatGPT for creative writing all the time, I'm usually using something like Claude. You kind of get caught off guard when it has these kind of updates and that are, I guess you could say, qualitative. It's quite hard because it's not like a math benchmark where it got the question wrong before and now it got the question right. This is the kind of benchmark where it just is a little bit more creative. And it's kind of that benchmark that, you know, you could only sort of talk about, but you couldn't really demonstrate. But this example right here is one that is really good because it shows you exactly how creative the model is now. And I think the reason that OpenAI did this is because one of the things that a lot of people have spoken about is the fact that Claude is just so much more creative than ChatGPT. And it was one of the things that even I was speaking to some developers and some people and they were just basically saying that, look, this model is just not creative at all. This model cannot just talk the way that Claude does. So I don't know how they managed to do this or how they managed to train this, but you can see that when you're using writing styles. Now, I would definitely try and benchmark Claude Claude and Chat GPT to see if you're able to get your models back into OpenAI or if Claude is still superior. So it will be interesting to see what the writing community is saying, but so far it seems pretty interesting. So this is definitely something that you might want to test out. So now, there was also, interestingly, another robot called Proxy. It's a collaborate robot built for the environment and it's built for the real world. So they made this little ad. I'm going to do probably a full video on this, but. But I wanted to showcase this because I wanted to segue into something directly after this.

Every day we wake up and we move. You see them every day, although you may not notice them. The early morning cruise, the weekend shifts, the late night teams. Technology has transformed the world since I was first inspired by robots as a kid. But surprisingly simple manual tasks still take so much of our time and energy. We're designing these robots to help us, not replace us. In every sector, Proxy moves and swerves naturally, in sync with the people working around it. And as AI and hardware evolve together, so will the ways it can handle items in the world. I'm here to help. It allows us to focus on those things that make us uniquely human, communicate, collaborate, and solve problems. A future that's even more alive with the help of robots.

Now, the reason I wanted to show you guys this robot quickly was because this robot isn't like your traditional humanoid robot. This is the kind of robot that I guess is taking on a different form. Now, one of the biggest debates that's been happening in the AI robotics community was the fact that there are humanoid robots, but they're apparently useless. Okay, now let me just. Let me just. Let me just get to exactly what I'm trying to say here. Like, a lot of people basically say that, like, okay, we're creating humanoid robots. They look cool, they look amazing, but they're just really inefficient in terms of their form factor. Like, creating a robot that is a human form just doesn't make sense. Like, why not give the robot actual wheels? And this is where the second clip comes in, which is what Nvidia CEO was basically stating about humanoid robots today. And he was basically stating that, look, humanoid robots are the only type of robots that can scale to extremely high volumes because they can be deployed to the world as is. And it's basically saying that, look, if we make humanoid robots, then we don't actually have to change the world to fit their environment.

What is missing, of course, for robotics is an AI that understands the physical world. ChatGPT, or large language models today understands the cognitive knowledge and intelligence, but it doesn't understand physical intelligence. It doesn't understand necessarily that when I set the cup down that it's not going to go through that table. And so we need to teach an AI how to understand physical intelligence. Well, let me demonstrate in fact, that we're making good progress. Well, one of the demonstration that all of you have probably seen is you could generate using generative AI text to video. And I can surely generate a video that starts out with a picture of myself. And you prompt, Jensen, pick up the coffee cup and take a sip. Well, if I can prompt the AI to go and pick up the coffee cup, why can't I then generate the tokens to cause a robotic arm to go pick it up? And so the gap, the leap from where we are with generative AI and general robotics is very, very close. And so I, I'm very excited about this area. There are three types of robots that can be manufactured in high volume and pretty much only three. All of the other types of robots that we've seen historically, you know, robotics has been around for a long time, but it's very hard to scale into high volume. Scaling in the high volume is important because you need the technology flywheel, the high volume allows you to, to generate high R and D, which allows you to make great technology breakthroughs, which makes better products, which causes the volume to be even higher. And that flywheel, that R and D flywheel is vital to any industry. There are only three robots you could really do this for, but two of them are going to be the highest volume. And the reason for that is these three robots can all be deployed into the world as it is today. We call it brownfield deployment. Well, the three are cars because we created the world for cars over the course of the last 250 years. And the second drones, because, you know, the sky is fairly unlimited. But the highest volume one, of course, is humano robots. And it's because we created the world for ourselves. And with these three type of robots, we can, we can pretty much scale robotics into extremely high volume. And that's one of the advantages that, that a manufacturing ecosystem like this one, you know, really has.

And the actual founder slash CEO of figure Brett Adcock says here yes, and many people are still missing the point here.

Humanoid robots are a brown filled solution.

You need no changes to our world for a humanoid to do useful work. And many will ask, is human the ideal form? This is asking the wrong question. Humans are an odd form, but that's not the point. The point is that we built an entire world to conform to our form factor. And humanoids can scale to tens of billions of units because of this. And it's just hard engineering from here. And of course, and of course I actually added the CEO there because, because they actually recently had an Update. So figure 2 is now an autonomous fleet with 400% faster and 7 times higher success rate. And they're committed to delivering millions of useful robots to companies and homes worldwide. And their autonomous fleet operations is really, really reliable. Now you can see that they're running 1,000 placements a day in the BMW factory and they are running all day and all night. And this is only improving because of course, as they are doing this, they are currently able to collect more data and they're going to be able to improve their models overall. So it's really interesting to see how robotics is developing, especially since when we look at this company, like figure I remember literally looking at the press release for the time that they were just raising funding for this company and now they have a robot not only like 18 months later that is in the factory, they're on their second iteration, they've got the third one under works and this one is able to do a lot of reliable stuff in BMW factory. So if this is the kind of progress we're seeing in 18 months, it kind of makes me think, okay, what kind of progress are we going to see with these humanoid robots in the next 10 years that's going to really change the way the world is. Because over the last six to 12 months, I think humanoid robots has taken a really huge step and definitely in the right direction. And of course, Demas Hassabis, in a recent interview actually spoke about how we are on the brink of a new golden era of discovery pushed by science and exponential advancements in AI.

I think we are on the cusp of that. I really do feel like we're on the brink of a new golden era of discovery. Like, like the whole of today's, you know, symposium is called. And I think what we need is a lot more interdisciplinary science. So using AI, bringing in the right way, asking the right questions with domain experts. And I think this, it's, it's almost limitless what its applications could Be and of course AI itself itself is a scientific discipline, is improving all the time. So there's, there's applying today's technologies directly to the other fields and then there's also continuing to improve AI itself. And that, that as well is, is a sort of exponential improvement. So, you know, there's a lot of, I think, progress to be made in just the next few years.

Then of course we had Nvidia Edifier, which is a multimodal architecture for developing visual generative AI models for image 3, 3D, 360 HDRi and physically based rendering, which is PBR materials. And this uses the Nvidia AI foundry and service providers can train and customize Edify models to build commercially viable services built on Nvidia's NIM. So this is something that allows you to create 3D models in a really effective way simply by using natural language. So for those of you that are thinking about getting into certain fields, but you think, ah, maybe it's a little bit too difficult. For me, this is the kind of thing that is definitely bringing down the skill gap in terms of what your creativity allows you to do. Didn't really want to leave the video on a sort of a, you know, bad note here, but this is the last story of the video. But I want to talk to you guys about this because this is the chart that is recently going viral again. And this was something that I've already covered in a video. I had an entire series on my channel dedicated to post Agile Economics. I'm no longer focused on that. But since this graph has once again become a talking point in the community, I just want to explain what this actually means. So this is basically a graph from the IMF. They did a report in 2023 and they basically spoke about how AGI is going to change the future of wages. And basically what they're stating is that we have two scenarios here. We have baseline AGI, which is in 20 years, and then we've got aggressive AGI, which is in five years from now, which is basically 2029, which is the latest date that most people working at these AI labs are predicting. So I think what we can see here is that, okay, in an aggressive AGI scenario and in the base AGI scenario, the productivity of the average human goes up, which makes sense because if you have AGI, it's able to do more. The productivity goes up. And then of course we have this second graph, which is of course the wages. And this is where things start to get a little bit concerning because you can see that aggressive AGI, right here is a situation where, yes, your wages go up first, but then at the peak, just before AGI is really, really realized, you can see that your wages start to plummet down to where they were originally. And this thing also happens with baseline AGI. So basically what this actually means is that when your wages go up, your wages go up, because, of course, your wages traditionally go up. You can see the blue line is just what happens with no AI. Your wages are traditionally going to go up with inflation. I mean, let's not get into politics here, but that's what's supposed to happen. And then what we do have is with aggressive AGI, your wages go up normally, and then they plummet down. Because, you know, these companies are going to basically be using AGI to do basically everything. And then what we have is a price war for goods and services that basically brings the wages down for the average person to, you know, really interesting rates. Because I don't know how the world is going to work at this point. Like, I really don't know how the world is going to work at this point if AGI is going to be doing a lot of the tasks. So it will be interesting. I mean, one saving grace for those of you who don't really like AI is the fact that hallucinations is still gonna be a problem. So it seems like if you're a specialist in your field, you're gonna still be largely valued. But it's gonna be really interesting to see how wages and how this entire economy comes, because I've seen tons and tons of statements, and it's not hype, it's not fake. It is a real thing that these policymakers are going to have to think about. So there's a graph there that I wanted to share with you all and just explain to you exactly what it means.