Googles AI CEO Just Revealed AGI Details...

I think there's still two or three big innovations needed from here to we get to AGI. And that's why I'm more of an on a 10 year timescale than others. Some of my colleagues and peers and other, and some of our competitors have much shorter timelines than that. But, But I think 10 years is, is about right.

So Demis Hassabis recently gave an interview where he actually gave us insights to how the AGI architecture is being built. And not only that, he also gave us the timeline to which he thinks AGI will happen. So I think this is one of the most insightful interviews because we get direct statements on the AGI architecture and of course this may for some individuals reset their timelines and reset their expectations and filter through the AI hype in terms of what is going to be there. Now Demis Hassabis is Of course the CEO of Google DeepMind's research lab, which is of course the, the company that essentially made Google Gemini, which is an outstanding model and they've made tons of breakthroughs in AI. So this is clearly someone who knows a substantial amount of AI and has impacted the space more than you can imagine. So I'm going to take a look at some of these statements, break them down and try and understand where we are headed in terms of everything related to artificial general intelligence.

LLM closer to AGI, I mean, it feels to me closer to interacting with a human, which feels to me what AGI is. But is it actually, I think that the multimodal and these days LLMs is not even the right word because they're not just large language models, they're multimodal. So for example, our lighthouse model Gemini is multimodal from the beginning. So they can cope with any, you know, any input. So you know, vision, audio, video, code, all of these things as well as text. So I think my view is that that's going to be a key component of an AGI system, but probably not enough on its own. I think there's still two or three big innovations needed from here to we get to AGI. And that's why I'm more of an on a 10 year timescale than others. Some of my colleagues and peers and some of our competitors have much shorter timelines than that. But I think 10 years is about right.

So one of the things I actually wanted to talk about was the fact that Demis Hassabis has a 10 year timeline. Now later on I'm going to talk about how the fact that of course there's going to be different levels to AGI. But I think one of the most interesting things is that different companies have internal, different models of what's happening. Recently I made an hour long video dissecting Dario Amodei's views on artificial general intelligence and powerful AI. And of course, he is the CEO of Anthropic, which is the, you know, company that produces the chatbot Claude and that company. The CEO actually thinks we could get AGI and as early as 2026. Although he explicitly does state that he dislikes the term AGI and he refers to what he calls powerful AI. But nevertheless, I think it's rather fascinating that he has such a shorter timeframe than some of these other AI CEOs. Now, some people might argue that these other companies need to say that this AI is just around the corner so that they can get increased funding. And you know what, that does make sense when you look at what other individuals are saying, quite like Sam Altman. Sam Altman has a more even extreme view, which is that, you know, superintelligence is going to be here in a few thousand days, which is even not on the realm of AGI, but artificial super intelligence. So I think it's pretty interesting, although some people would argue that maybe these statements actually reflect where these companies are in terms of their real research. And the thing with this is that we don't actually know where these companies are with their product cycles. Every piece of research is now closed. But this is just something that just gets even more interesting. So of course I don't think this is a huge dampener on the AI space in terms of AI hype. But AGI being 10 years away, I still think what people might fail to understand is that as we move forward in terms of the timelines, it's going to be pretty, it's going to be like a gray area because what we will have is we will have an increased level, level of capabilities year on year. So when AGI is directly achieved, it will be quite hard to say when this does occur. Of course there are different graphs and different classifications for different AI systems, but it does remain a mystery as many different individuals have different definitions for AGI. But I think having an actual artificial general intelligence that doesn't hallucinate, it doesn't make some of the fundamental mistakes of generative AI and LLMs in today's architecture. I think that is something that of course might need a few more architecture breakthroughs that Demis Hassabis talks about. Now, what are those breakthroughs going to look like? Of course, we don't know just yet, but I do think that potentially some of the companies that are working on this stuff may have already made those breakthroughs and are probably scaling through those. Because one thing that we have to remember is that AI is now closed off. Whereas previously we used to have an open research ecosystem where a lot of the research would propagate through the community and be widely shared. But with companies like OpenAI and Anthropic being closed, a lot of that research doesn't really get shared around as much. So it will be kind of interesting to see how those breakthroughs occur. Of course, recently OpenAI had their recent breakthroughs with Test Time, Compute, and of course Google had some additional research to support that kind of information. So it will be interesting to see how these models evolve in 10 years time, because I can really, really imagine these models being truly more advanced than they already are and working through some of today's limitations. And this is where Demis Hasabis actually talks about how a lot of the products that are being built right now for consumers are essentially what is needed for AGI. When you build a product that has audio, video, understanding, image understanding, and you're building humanoid robots and that kind of stuff, it becomes easier to sort of build AGI at the same time as providing value to your custom.

You need for products are actually about 90% similar to the type of research you would do anyway towards AGI. So those things have converged a lot. Whereas if five years ago or 10 years ago, if you wanted to build AI into a product, you'd have to, because the general systems and the learning systems weren't quite good enough yet. You'd have to go back to, you know, logic networks, expert systems. So, you know, the kind of Alexa era of assistants, for example, all built still on that old type of technology, which is why they were brittle and they don't generalize and they're not that useful in the end. Whereas the new generation of assistants built on these learning systems are going to be way, way more capable. So it's actually very exciting. And I actually see things like Gemini and also our own vision of a multimodal assistant in the future called Astra, at the moment as on the critical path towards an AGI system, because they actually push research towards that. Should we watch a video? Sure. We've got a video of Astra working.

Don't remember, Google actually has a product called Google Astra. This is basically where it's an AI agent that can have some of the early things that you would expect from an AGI level system has the ability to see. It also has the ability to memorize. And although this product isn't released yet, I do think that once this product is out there, once we do have more things that allow us to scale these kinds of of products, I think once it hits the general public, it's going to be a really interesting reaction.

Okay, let's do some tests. Tell me when you see something that makes sound. I see a speaker which makes sound. What is that part of the speaker called? That is the tweeter. It produces high frequency sounds. Give me a creative alliteration about these creative crayons. Color cheerfully. They certainly craft colorful creations. What does that part of the code do? This code defines encryption and decryption functions. It seems to use AESCBC encryption to encode and decode data based on a key and an initialization vector. Iv? That's right. What neighborhood do you think I'm in? This appears to be the King's Cross area of London. It is known for its railway station and transportation connections. Do you remember where you saw my glasses? Yes, I do. Your glasses were on the desk near a red apple. What can I add here to make this system faster? Adding a cache between the server and database could improve speed. What does this remind you of? Schrodinger's cat. All right, give me a band name for this duo. Golden Stripes. Nice. Thanks. Gemini.

So that's just the beginning of, you know, basically a universal assistance that can help you in your everyday life.

We're not here to make this system.

Maybe we can stop the video here. So, you know, and there'll be different form factors too. So you can see it on your phone, you can see it on the glass. Like, you know, and I can't tell you how amazing that would be. You know, if we went back five years and you told me we would be at this point where you just, you know, you just point at something with a camera and it fully understands the spatial context around you. It's pretty incredible, right? Like it's sort of got concepts and it understands what objects are and, you know, even recognize the neighborhood we were in just from sort of a random view out of the window. Things like memory for, you know, where you left something that could be extremely useful as well as an assistant. You know, personalization. All of these things are coming in this, what I would call the next generation of assistant. I call it kind of universal assistant because I imagine you taking it around everywhere with you, you know, on different devices. It's the same assistant, whether it's playing a game with you or it's helping with your work on your desktop or you know, traveling around with you with you on a mobile device.

So I do think that this is truly going to be incredible use for the future in terms of how we interact with software and computers because AI agents and AI assistants are completely getting rebranded. When you can interact with a computer and it knows all of your history, all of your memory, your past, previous conversations, it's going to be a much more natural and human like experience that's going to result in a much more enjoyable and fluid experience. I truly believe that this is going to be the next stage in AI that allows people to realize the true power of these systems and it's going to onboard a lot more people that are not as technically inclined. Of course, this is where Demis Hassabis actually talks about the kinds of breakthroughs needed to actually achieve AGI. And currently these current systems just don't have those capabilities. Things like true reasoning and true planning and true memory are things that humans can do pretty easily. But these systems really do struggle with.

Well, we definitely need these systems and all of you, I'm sure have used the various, you know, state of the art chat boss today. They're very passive, these systems. They're Q and A systems, okay, so they're pretty useful for answering a question and maybe doing a bit of research, summarizing some text, something like that. What we want next is more agent based systems that are able to achieve certain goals or tasks that you give it. That's certainly what a useful assistant digital assistant would need to do. You know, plan a holiday, plan your trip around a city, tell you, you know, book you tickets for something. And so they need to be able to act in the world and carry actions and do planning. So we need planning, reasoning actions, we need better memory, we need personalization so it kind of understands your preferences and remembers what you've told it and what you like. So all of those technologies are needed now the way the shorthand I give for that is some of our games programs like AlphaGo that beat the world champion at Go, has planning and reasoning in it, or albeit in this narrow domain of a game. So we have to bring those technologies and apply them now to multimodal models like Gemini that are basically models of the world as you've just seen it understands the world around it. But how do you do planning in the messy real world as opposed to kind of clean setting like a game? So that's The I think the next big breakthrough that's needed.

So this is where we actually get some information on the current debate that is being had in the scientific community with regards to how we get to truly intelligent models. Do you cram everything into a model and the model can completely do everything or do you have this neuro symbolic approach where an AI can act as sort of like a brain and that it has all of this smaller specialized AIs that can go ahead and do different tasks. I think that's kind of what is already happening with the early approaches which was kind of the breakthrough that made GPT4 so good. Of course, if you aren't familiar, GPT4 had a mixtures of experts approach which I think was 16 experts that were essentially smaller models that were just experts on things like math, writing, coding. And anytime a query was presented to the model it was routed to one of these smaller experts. So I think this is going to potentially happen on a larger scale. But, but it will be interesting to see how that occurs combined with tool use. It's going to be pretty incredible to.

Apply, you know, AlphaGo levels of chess playing and source and protein. So, so. Yes, exactly. So there's two ways that could happen. It's very interesting debate at the moment we're having internally and just in the research community is so you can imagine one of the big things that you want your general agent system to do is to do tool use, to use tools. And those tools could be hardware, so like robotics or things in the physical world. But they could also be of course other pieces of software. So maybe like a calculator, something like that. But they could also be other AI systems. So you could have a, you could imagine a general AI system like the brain then calling something like AlphaFold or AlphaGo to play go or to fold a protein or, or because it's all digital, you could imagine folding that, that capability into the general brain, into Gemini. But then that has trade offs because then are you overloading it with specialized information, say too many chess games and then that makes it worse at language. So you have to, it's an open research question whether you want to separate it into a tool, even an AI tool that a general AI can use in that specialized situation or do you want to upstream that into the main, the main system? And for some things you want to upstream into the main system like coding and mathematics because it turns out if you put it in the main system it actually makes it better at everything. So they're sort of like, so you know Lots of you are studying theories of learning and child development and things like that to actually sort of think through what sorts of things may actually be a general purpose and better off in the main system than in the periphery. Tools.