Big AI News : Googles Gemini -2 , Claude 3.5 Opus, GPT 4.5 Soon, Apples Stunning Research

In today's video, there was actually so much news this week that I couldn't even cover most of it. So I'm not going to waste any of your time. I would say that this is largely one of the most dramatic weeks for AI, and you guys are about to see why. There have been a billion different stories that are just insane. So let's take a look at the first story. Okay, so I know this, I guess, technically isn't AI news, but it actually falls on the branch of AI drama. So we actually have Geoffrey Hinton, okay, still stating that he is proud that one of his students fired Sam Altman, because Sam is much less concerned with AI safety than profits. Now, I think this is absolutely insane because firstly, Geoffrey Hinton saying, this is just incredible because this guy is literally the godfather of AI. For most of you guys who don't know why, the significance of this, this is basically like saying, I don't know. I don't know how I can show you guys the significance of this statement, but this is like the goal. Godfather of AI. Guys, just think of how much weight there is behind this man's name in terms of AI. And he's saying that he's glad that his students. And when he's saying his students, he's actually referring to Elia Sutskova. For those of you who aren't familiar with the drama that happened at OpenAI, of course, because Sam Altman is much less concerned with AI safety than profit. So he's basically saying that, look, okay, I'm glad you fired Sam Altman, because Sam Altman was doing some crazy, crazy stuff. And this is insane, because one of the things that we know about, you know, the AI drama was that the majority of people within the industry, they didn't actually talk about this stuff. So it's actually crazy that we have literally the godfather of AI. And there was some big news as well that happened with him this week, which is even better. But I'm going to show you guys what he said, because it's crazy. It's crazy, crazy, crazy stuff. Because it's been a crazy week.

I'd also like to acknowledge my students. I was particularly fortunate to have many very clever students, much cleverer than me, who actually made things work. They've gone on to do great things. I'm particularly proud of the fact that one of my students fired Sam Altman. And I think I better leave it there and leave it for questions. So OpenAI was set up with a big emphasis on safety. Its primary objective was to develop Artificial general intelligence and ensure that it was safe. One of my former students in Yasutskova was the chief scientist and over time it turned out that Sam Altman was much less concerned with safety than with profits. And I think that's unfortunate.

Now what was even crazier this week for Geoffrey E. Hinton, and not just only him, was that he won the Nobel Prize for Science. Now I saw that there were a bit of, you know, ruffles in the community about like why do they win this for science, yada yada yada. But I think that this is, as some have mentioned, part of the wider trend where we will start to see artificial intelligence probably start to eat other categories in other fields. This is because, as you know, AI enabled science is cross platform, meaning that the breakthroughs that we have in AI can be applied to multiple fields. Fields like AlphaFold and predicting the structure of proteins is something that can be greatly applied to science. And it's one of those things that in the future going to be more than likely that we're going to start to see these things continue to happen as AI manages to enable certain breakthroughs in a science. And this is where we got Demis Hassabis also receiving the Nobel Prize in Chemistry. And he actually speaks more about this.

We got a call from Sweden and you know, most scientists understand what that means and actually for me it came through via my wife's computer because I couldn't get hold of our numbers. They didn't have either round numbers actually and I had to give them your number. So it was quite, it was pretty funny and pretty hectic as well. But it was obviously super surreal and actually my mind just went totally blank. I mean Nobel I think originally set what the prizes that he set back 100 plus years ago. So obviously there was no computer science. But I think it's been pretty amazing to see the effect of AI on the other sciences and as a tool, you know, and I think we'll probably start seeing more of that but, and you know, it's just great to see that the impact that finally after a lot of us have been working in this field for 20 plus years now and, or a lot more in case of Geoff Hinton and John Hopfield. And finally I think this is recognition of the actual real world impact these types of technologies and algorithms are having.

So that's definitely by far one of my favorite things to happen because, because seeing people that are getting the recognition they deserve for the kind of work that they've been putting in for 20 plus years is just like a feel good story, you know. Now I did cover this in a previous video, but since this is a video where I'm talking about all AI news, I'm going to cover this quickly for the next two minutes again. But this is basically Daario's Amodei's thing where he talks about machines of loving grace, how AI could transform the world for the better. So in this, basically what he's actually talking about is how AI can transform the world in terms of the good possibilities. I know most people like to think about AI safety, what could happen, yada yada yada, but he's actually thinking about the good areas where AI is going to have a dramatic effect on multiple areas that I genuinely didn't even think about. And it was really thought provoking. I did an hour long video breaking it down. So I will leave a link to that if you want to see the entire thing. But I think one of the biggest big takeaways for a lot of people that really did surprise them was the fact that if you look at where he goes to powerful AI, this is where he actually talks about his definition of powerful AI, and he says that like I think powerful AI, which is, you know, he, he dislikes the term AGI and I think I'm starting to agree because when we actually talk about the term AGI system that can really do anything, I think it's becoming more true that it's quite harder to get AGI, but it's easier to get powerful AI. And I know it's kind of hard to explain, but just, let's just focus on what's going on here. But he basically says that this type of AI, okay, AGI, powerful AI, however, we want to slice the cake. He says that this could come as early as 2026, though there are ways that it could take much longer. So if this does come at 2026, which right now we're in 2024, okay, there's about two and a half months left until the end of this year, which essentially means that we only really have pretty much like a year for this thing to happen. That's like, you know, let's say a year and six months, because I said it could come as early as 2026. Let's think about like early 2026. That's only literally a year and two. Two months. Okay, a year and three months. Like a year and three months is nothing in the terms of technological development. So going from where we are now, where we have reasoning models, we don't even have agents yet, reliable scalable agents yet. We could have powerful AI in a year and three months. That is something that I think it sets the time frame because this isn't an article that is quite hypey in the terms of like, oh, we're going to have powerful AI by this year and it's going to be able to do all these sorts of things. This is Daria Amodei, the company that's not really focused on AI hype. They've made Claude 3.5 sonnet, which is, according to a new recent study that I was looking at, it actually is still one of the best models out there, despite. I know there's a lot of leaderboards that say this ELO is going to be yada, yada, yada, but yeah, I think it's going to be really surprising because having these models come out relatively soon is going to be quite a shock to many people. And this is why I say, you know, most people don't realize exactly what's going on in the future. Now, of course, he says, okay, by powerful AI, I have in mind an AI model that's similar today's to today's LLMs in form, though it might be based on a different architecture. And that is something I'm going to get into a moment. Okay. But it might involve several interacting models and might be trained differently with the following properties. So this is where he's basically saying this is how AGI will actually be. So when you have the concept of agipowerful AI, this is what you're going to be able to see. So he says, in terms of pure intelligence, it's smarter than a Nobel Prize winner across the most relevant fields, fields like biology, programming, math. Okay, that means it can prove unsolved mathematical theorems, write extremely good novels, I think we can get that easily here. And writing difficult code bases from scratch, that is something that I do believe. I don't know if it's going to be able to prove mathematical theorems, but the crazy thing about this is that Terence Stao did actually say that something like this is going to be able to happen anyways, okay? And he's literally one of the most gifted mathematicians to ever exist.

Apparently we prove one theorem at a time, or maybe two or three if you are efficient. But with AIs, you could imagine in the future, instead of trying to prove one, solve one problem or something, you take a class of 1000 similar problems and you say, okay, I'm going to tell your AI, try to solve these 1000 problems with this technique. And it will report back, oh, I could solve 35% of these problems with this technique. What about this technique? Then I can solve this percentage of problems, or if I combine them, I can do this. You could start exploring the space of problems rather than just each problem separately. And this is something that you just, you either cannot do right now or you do over a process of decades, with dozens and dozens of papers slowly figuring out what you can and can't do with various techniques, but with these tools you could really do. I mean, you could really sort of start doing mathematics on a scale which is really not really unprecedented. So, I mean, the future is going to be really exciting, I think. You know, I mean, we'll still also be proving theorems the old fashioned way. In fact, we'll have to, because we can't, we won't be able to guide these AIs unless we also know how to do these things ourselves. But we'll be able to do lots of things that we can't do right now.

He also says here, so these are going to be the modalities, okay? He says that it has, you know, all the interfaces available to a human working virtually, including text, which is obvious, which we have now, audio, video, mouse and keyboard control, and Internet access. If powerful AI has Internet access, we might want to be a bit careful about that. But let's not make this an AI safety conversation. So it says it can engage in any actions, communications or remote operations enabled by this interface, including taking actions on the Internet, giving directions to humans, ordering materials, directing experiments, watching videos, making videos and so on. And it does all of these tasks, again, remember these tasks that we just talked about, taking all of these actions on a computer, Computer exceeding that of the most capable humans in the world. So in a year and three months we're going to have AI that is largely more capable than pretty much anything that a human can do on a computer, which is going to be a dramatic way that I think the majority of people change computers. Certainly there's going to be this huge shift for how individuals use the computer, maybe still as a tool, but a lot more effectiveness is going to be coming out of that. And here he says it does not just passively answer questions. Instead, it can be given tasks that take hours to days or weeks to complete, then go off and does those tasks autonomously. So this is the agentic nature of the AI that he's talking about, which means that of course, by 2026, if 2025 is the year of agents, 2026 is probably when they get really, really good. And he also does say that it does not have a physical embodiment other than living on a computer screen, but can control existing physical tools like robots or laboratory equipment through a computer. And in theory, it could even design robots for itself to use. I think that is insane for 2026, for how quickly things are moving. And that's why, guys, I'm literally so focused on AI, because I don't think most people realize that a few months in AI, sometimes it seems like there's not a lot of news. But like I said before, sometimes there are just areas where a lot of companies are releasing the stuff that they've still been working on and these guys are competing with each other on the fasting, moving technology. So this stuff is going to move and has moved quicker than most people have thought. Which is why I'm constantly getting blown away by the timelines and having my expectations reset for what I think about what AI can do. I remember it was only earlier this year that we saw Sora and we were like, oh my gosh, I can't even believe AI can really do this stuff. Now of course, we do have the statement that powerful AI might be here in 2026. Some people are quite skeptical about this. And there was a recent paper that I did a video on that actually blew up quite a bit and it actually talks about the opposite of this and that we might be a little bit far from AGI and a little farther away than we did think. So if you watch the video, you're going to know exactly what I'm talking about. But Apple's AI research team released this paper. Now this paper has the AI community divided this. And I'm going to summarize it all for you guys because I've already done a 25 minute deep dive. But basically they have this paper, okay, and this paper is where they say, can large language models truly reason or are they just sophisticated pattern matches? The reason they're asking this question is because they want to find out if these large language models actually understand how to solve problems in a step by step manner that allows them to genuinely be smart and genuinely have reasoning capabilities. So they ran a study and in that study they, what they did was they had the very famous benchmark, the GSM 8K, okay, which most models perform very well at. Okay? So it's just 8,000 grade school math questions which are really easy to solve with varying levels of difficulty and all they did, okay, which is insane. When I first Read this. I was like, I can't believe this is true. So all they did was they changed around the value. So if they said Sophie watches her nephew, it might have been Mark watches his brother. And, and if he stacks 31 cookies, it might have been 24 oranges. All they did was change those values, change those values around. And the resulting experiment was that like, okay, if these models are truly having reasoning capabilities and they're not memorizing the test with the test results, surely they should be able to solve these questions. Okay? And so from that, the results were incredible. Okay, they showed that there was like a 10 to 20% drop across a various level of different models, basically stating that, look, if these models were truly understanding the reasoning capabilities, we should be able to switch out the values and the names and these models should be able to still get the same results. And what's crazy, they're like, okay, maybe the numbers that we switched may have made the difficulty a bit too hard. What if we just change the names? And if they just change the names in these tests, the model's accuracy dropped by 10%, which is not good. Why would a model on a math question, why would the accuracy drop more than 10% if qualitative data was changed? Does that make any sense to you guys? It's not like they change the values on the numbers. A 10% drop just means, honestly, I don't know what else this could mean, but that these models must be memorizing the test to some extent or there might be some data contamination which leads them to believe that, look, these things might not be as smart as we think they are. So this isn't some random research company. Later on in the thread here you can say, okay, that it says, overall, okay, this is pretty crazy. Overall, we found no evidence of formal reasoning in large language models like open source models like Llama, Phi Gemma, Mistral, leading closed source models even in GPT4,01 and 01 series. And they say that this behavior is better explained by sophisticated pattern matching so fragile that in fact changing names can alter results by 10%. Okay, and this is why, okay, I think it's going to maybe mark a new paradigm for people who are trying to solve these issues is because they're saying that, look, we can increase the data, we can increase the parameters and compute or even use better training data, but this is only going to give us better pattern matches, which means that maybe we might actually start to get some new architectures being worked on. Because this is something that actually is quite surprising. And I think they're probably is going to be another follow up paper where these ideas are largely either cemented or largely debunked because it will have severe implications for the industry if just changing the names on tests results in a 10% drop overall for large language models. Like if someone asked you a question and just switched out the names, I don't think that you wouldn't be able to answer those things before now. In some AI news and speculation, we do have the very famous Jimmy Apples, someone who has always been accurate with their leaks, talking about the new models. Okay, this is someone that's been quite accurate across many different AI leaks. And Jimmy Apples is talking about how potentially we should have a 4.5 model in October, which is probably either today or Thursday. So across this week and across the next couple of weeks, potentially look out for a OpenAI release which is going to be a 4.5 level model. Now, the reason that this model might actually be released this week is because some individuals have been speculating that Anthropic are going to be releasing their model Opus 3.5. Those of you who don't know what this model is, this model will be largely their Successor to the 3.5 sonnet series. And there are pretty high expectations for this model. After 3.5 sonnet blew out of the water, many people's expectations, I mean, most people weren't expecting 3.5 Sonnet to be doing that much, but it seemed to work really, really well. So if Anthropic can once again scale whatever they did with Claude 3.5 sonnet and put that into 3.5 opus, which is their most intelligent model, that is going to be really, really something. And when I say most intelligent model, I just mean that it's the largest in the family. And whatever methods they're using to train a smaller version of the model, they should use that to train the bigger version. So this is where you've got, you can see right here that Jimmy Apples is tweeting about Anthropic proudly presents. And this was tweeted on October 10th. So it could be today, which is October 15th, or it could be this coming Thursday. And I think that this is something that is largely going to happen because I've heard a few whispers in the AI community. And the thing is as well is that if Anthropic do do release something like this, it might blow ChatGPT GPT4O out of the water. And I'm not talking about advanced mode and all the great things that ChatGPT is doing. But Anthropic seem to be largely just focused on the raw intelligence of the model. Which is why Claude 3.5 Sonnet, the coding abilities have been so good that people have been able to build ridiculous programs and enable this entire ecosystem of no code developers that, that are doing incredible things. So I think if Anthropic does release Opus 3.5, OpenAI are quite likely to drop their GPT 4.5 in response to that, considering the fact that OpenAI is always in a battle with these other labs for market share and of course recognition. Now we also got a screenshot that is a picture of the official Logan kay, the Google DeepMind then in San Francisco teasing us with the next Gemini. You can see that there is a presentation that says where are we headed? We can clearly see that Gemini 2.0 is on the board. Okay, so this gives us the future of Google. I mean largely we do know exactly where the future is headed. But of course this is a concrete view. You can see that We've got Gemini 2.0 Next Generation Frontier models which are to be decided of course, tune everything, larger models, multi turn vision audio embeddings, et cetera. And of course we've got the new models which are, you know, like I said before, Gemini 2.0 Veo, which is Google's stunning video model, which is, you know, nowhere to be seen at the moment. Imagen 3, which is, I'm not going to lie guys, okay, no glaze to Google, seriously. And then probably the best image model. And I say that because I've been using it on a variety of different tasks and it, it just solves them. Like I don't know what Google did, but the text is good, the 3D rendering in certain, you know, graphics is good. Like it can do games really well. I don't know why it could do games really well, but it can literally, you know, make a screenshot of any game and it looks just realistic like beyond anything I've seen. And it can make like images that you would take on your phone, but it's just, it's just insane. Like I've been using it, Imagen 3 for a lot of social media posts, for a lot of my friends who want images, and it's been absolutely crazy. So I don't know what they cooked up with Imagen, but it is absolutely insane. And of course they talk about Notebook 11 now in more Google news, Imagen 3 is actually available to all Gemini users and it's the highest quality image generation model yet. And Like I said before, I don't know what Google cooked up with this model, but for me personally, I use this model in a lot of different scenarios and I think I don't know what they cooked up with this. I think it's just because Google largely has better data. With YouTube and stuff like that and Google Photos, I think they were just able to cook up a better model. So I would definitely give this a try. If you have access to Gemini, if you're someone that uses AI images on a day to day basis, try it out and see if it works for you. Because there's all of these models, they're all trained with different data sets and you get to see where it excels and there's certain specific use cases that you can use it for that other models just don't excel. If you did miss it. Earlier this week we did have the Tesla event which was truly fascinating. We got to see the Robo Taxi and and of course we got to see our friend Optimus being deployed in various different areas. Now I got to say a lot of people were quite confused about Tesla's Optimus robots. Like for example right here, these robots are clearly fully autonomous in their walking. As someone who pays attention to AI, I know that this stage is of course fully autonomous, but there were stages where the robot was teleoperated and this actually led to a little bit of a backlash with individuals claiming that they were fooled. So slash duped by Elon Musk. As with prior events, he may have tried to wow investors with flashy demos, but I don't think so. I think they just wanted to collect more data and if it was fully autonomous, I would say that they would say so. My thing is now is that if you see a robotics demo, or any kind of demo, always presume that it is teleoperated unless said otherwise. For example, in these videos, if it was fully autonomous, we would see it on the right hand side of the screen. So saying fully autonomous at one time speed or whatever it is. So of course the general public don't really understand that and that leads people to believe that it's something that fooled them. I would kind of disagree because it didn't really seem like that from the marketing. They just didn't say that it was completely teleoperated because there was no need. But anyways, this just gives us a glimpse into the future because I think that this is what the future might look like. Interestingly enough, now we also got Dreamina V2, a powerful all in one video generator. Basically, ByteDance, the parent company of TikTok, is an AI currently in beta testing for version 2.0. This offers a variety of different AI video features. And like I said before, AI video is getting absolutely insane. I genuinely believe that, like within the next two years, we're going to be able to prompt probably an entire film and we're going to have a lot more character consistency and control in terms of what we see from these AI things. So, I mean, I think within two years we're probably going to be able to get our own shows from a single prompt. Although I do think that probably the largest thing that's probably going to stop this from happening is of course the compute costs if these companies can't figure out how to make this really cheaply, because we even saw Meta's moviegen be released in outstanding quality, but it was really expensive to do. And that was also something that we saw that kind of hindered OpenAI's saw a release. So it will be interesting to see if we can manage to figure out how to do this more efficiently, because then we're going to start to get some really creative films. And I think what most people don't understand is how many teams go into an actual TV production show like you've got. I don't even want to talk about the executive producer, the producer, the writer, the script editor. There's just a billion different things that go into doing this and having people create their own TV shows, I mean, it's going to be interesting, okay. There's going to be a lot of content to consume and in the future, so it will be interesting to see how this entire space develops. Now, we also had the Generative video also get some other areas. So for example, for those of you who like AI video, Adobe also announced their Firefly video model, which was pretty surprising because this model is really good. And the most surprising thing about this model is that they actually were able to train this model with their own, not their own data, but commercial use data that they actually have the rights to, and they were able to get the quality to look outstandingly good. So I can't believe that this is here, because this now means that the creator space, in terms of individual creators that actually want to create their own AI films, is going to completely expand because not only did they have this text to Video Beta in Adobe Firefly, this is natively baked into Premiere Pro. In the new version in the future, it's going to be natively baked in, which means a lot more people are going to be creating things like AI videos and AI images, and just generally expanding what we already know. So we're going to see a lot more creative pieces of content. And for all the pieces of content that we do see, we see the content that looks really, really effective in terms of the character consistency, the shot consistency, and just how these scenes are. So this is a, of course, video, not video company, actually a company that works with these tools first. So seeing this is quite surprising. And I think they have the, you know, best ability in order to actually get this right in terms of the user experience, because I think other companies, they do very well at making the underlying technology. But companies like Adobe are the end product where users really use the technology. Preparation.