#36: The Road to AGI

A lot of people are asking, how are we going to get to AGI? Well, we're going to discuss that today. Who are you?

Me?

Who are you?

I'm Peter Voss and I'm the CEO and chief scientist of IGO AI.

And, you know, we've been talking a bit. First of all, what is AGI? Just so everybody's on the same page, and then we can talk about how you're going to get us there.

Sure. I think there's a lot of confusion and misunderstanding and distortion. The term AGI, I actually coined it together with Ben Goertzel and Shane Legg in 2002, and we wrote a book on the idea of getting back to the original dream of AI from 69 years ago, when the term AI was coined to build thinking machines, machines that can think, learn, and reason like us. One of the key strengths of human intelligence is that we are very adaptive. In fact, my first AI company was called Adaptive AI, Inc. We can adapt to changing circumstances. We can learn new things. We are very good tool users. So it's all about being able to acquire new skills and new knowledge in a very flexible way and to apply them to different things. And that's really what the term AGI means, is to have an AI that, you know, you can put in a new job, in a new position, and it can just learn and it can adapt the way humans do. Now, in some ways, of course, it'll be much better. It'll have photographic memory and be much better at reasoning and so on. But that's. That's AGI. It's not about some optimization process or some narrow AI. It's, you know, specifically that adaptability and flexibility to learn things on the go with limited resources because, you know, when you come across new circumstances, you don't always have enough time to thoroughly research it or to learn or something. You know, you have to make do with the limited resources of brain power and time that you and knowledge that you have.

So today we have large language models, you know, with ChatGPT and Gemini and Perplexity and Quad and the Grok and a few others. Right. Are LLMs going to get us there?

No, not at all. They're fundamentally the wrong approach. And a number of reasons for that. I mean, first of all, one, I recently wrote an article, the Insanity of of Huge Language Models. I mean, these things are amazing at what they can do. But when you.

Smithsonian.

Yeah, but, you know, when you. When you consider that they take hundreds of millions of dollars, they take tens of trillions of pieces of information take months to train. That's very, very different from the way our brains work. You know, our brains work with 20 watts, not 20 gigawatts. You know, a child learns language and reasoning with maybe a few million words, not 20 trillion. So it's a brute force approach.

Yeah, but somebody, somebody did a script on my blog and found I only use 44,000 words. Shakespeare used 82,000 words.

Right, right, right, exactly. And it shows you the difference. So it's fundamentally an approach that has worked incredibly well by just brute force. And this is really how we got into large language models, is that the big companies had a lot of data, had a lot of compute power. So that's a hammer they've got. And they've done really well with finding nails and doing useful things. But no, it's not on the path to AGI. And the couple of reasons, I mean, one of them is probably the main reason is they cannot learn incrementally. You spend hundreds of millions of dollars building a model, takes months to build a model, and then you use it for a few months and you throw it away. Why do you do that? Because they cannot learn. They cannot update the core model. Now you can add external databases and you can use prompting and use short term memory and so on, but they cannot fundamentally update, state their model. And you need to be able to do that for true intelligence to learn interactively. They can't do that. And this is actually kind of a dirty little secret. And it's not so secret, and it's not so little is that, you know, people throwing these billions of dollars at large language models actually know that they cannot get to large, that they cannot get to AGI. For example, Yann Lecun says large language models are an off ramp to AGI, a distraction, a dead end. And he tells his students, don't even bother learning about large language models and generative AI, because that's not the future. Demis Hassabis, the CEO of DeepMind Google DeepMind also serves deep learning. Machine learning is not going to get us to AGI, not by a long shot. You know, Sam Altman, you know, they all say that, but of course, in the meantime, they kind of distort the meaning of AGI to say, hey, give us a couple of more billion dollars. You know, we're doing hammer, you try.

To use it as a screwdriver, right?

Yeah, right. But no real AGI. It's not going to get us. It. It fundamentally needs a different approach. It needs a real time adaptive learning that Will require massively less compute and data.

That's insane. Scary. Maybe to Jensen Huang at Nvidia, that kind of talk, you know, oh, you're going to be a lot more efficient in the future. Oh man, you're not going to buy as many Nvidia cards, right?

Well, I mean, once we get AGI, everybody will want it. You know, you'll want probably several of them. I mean, it'll be so enormously beneficial to humanity that we'll want billions of these things and they will all require, you know, cpu, not these massive data centers. But of course you can host, you know, you can host many of the, of these agents and in the data. So we'll still need chips. We might not need GPUs as much. There might be some other specialized chips that, that are developed to optimize the technology. But yeah, we're going to need chips, but it's not just going to be scaling what we have now.

So let's go first principles. How would you rebuild or how would you build a new kind of AI to get us to AGI, you know, talk us through? Because I'm not a math guy like you are. You know, I used to put an equation in front of me. My, my eyes bleed. So explain how do you look at this world? How are you trying to build this? How are you trying to come at it from a new way?

Right? So the starting point is really starting with first principled and saying really understanding intelligence. What human intelligence is and what makes human intelligence so powerful to me is the starting point to build systems that have this kind of intelligence. So you start with cognitive psychology. You understand, what do IQ tests measure? You know, how do children learn? Why is our intelligence so different from animal? Why are we so much more powerful cognitively than animals? Epistemology, theory of knowledge, you know, from a philosophy point of view, how do, what is knowledge? What's the relationship between knowledge and the reality? And how do we acquire, how can we be certain of things? Where does free will fit in consciousness to really understand all different aspect of intelligence and mental processes, that to me has to be the starting point to design a system that has it now. Interestingly enough, a few years ago DARPA came out with a model they call the three waves of AI. And what they mean by the three waves is the first wave is what we now call go good old fashioned AI. That's what they did in the 80s and 90s and 70s, 80s 90s and so on. And that's really expert systems, logic based systems to a large extent. So IBM's Deep Blue World chess champion would be a perfect example of that. So you take one particular thing that requires human intelligence and then you figure out how to write a program to do that one particular thing. Now, IBM's Deep Blue World chess champion couldn't even play checkers, you know, so it's very narrow and it's the intelligence of the programmer that really achieved that. But first wave was really primarily about logic based systems. The second wave now hit us like a tsunami about 12 years ago. And that's all about statistical big data approaches. You know, deep, the breakthroughs and deep learning where image recognition became just so much better, translation, speech recognition, all of those things. I mean they just leapt up in performance when they started using big data deep learning approaches. And then that of course moved into generative AI, which is sort of the part of the wave we surfing now. But that's still the second wave of AI. The third wave of AI is cognitive AI is where you start by understanding cognition. And it's all about being adaptive, being fluid, being able to adapt and learn on the fly in real time. That's the approach we've been working on for some 20 years. I've been alternating between development of the core technology and then commercializing it, and then going back to development and commercializing it. So we've never quite had the resources to just do development. You know, like DeepMind have what, 600 PhD level researchers for the last 10 years, you know, well, we've never had that kind of money. It's been like 10 researchers for five years. So we've made slow but steady progress in building this cognitive AI. And our approach is that it's a very high performance graph database, that it's a vector graph database that encodes all of the knowledge and skills it has. But the beauty of it is that graph database can be updated incrementally in real time. So you can learn one new fact and it can update the graph, the graph database. But then if there are implications through that, it can ripple through. Like this is actually a good example. I mean, we've just had elections now. We didn't know who was going to win until. Yeah, well, we didn't know who was going to win. Most people certainly didn't know. So, you know, Trump winning of course changes a whole lot of things. It changes our model of the world. Now with large language models, you basically now start training a large language models with the results of the election and the things that are happening sort of already happening now. But it'll take you three months to train the model. By the time you finish training the model, it's already completely out of date with the fallout and those changes and things that have happened. So you really, you need intelligence like us humans, you know, we, okay, all right. We now know who won, what are the implications. So the implications spread through your network and update your worldview. And that has to happen pretty much in real time. I mean, it might take a, you know, it might take a while for that to sort of bubble through. Like for us and humans, we hear something profoundly different, and it might take a while for us to kind of digest that, you know, and work through the implications of that. But it, you know, it's not, it doesn't require us throwing our brains away and building a new brain from scratch, you know. Yeah, so, so that, that is the approach is cognitive AI, the third wave of AI that DAPA talk about that I, I firmly believe is the path to AGI.

Yeah. So it's interesting you're building this around a vector database or a new kind of database, you know, because that's, that's fascinating because that's one reason it's adaptable, because it's flexible. It's not stuck in. A large language model is like a dried concrete. You can't change it very easily.

That's exactly right. Large language models are read only. Once you've built that model, you finished your training, you cannot update it. In fact, I was on a white paper specifically about the ability or inability of large language models to learn incrementally in real time. We surveyed the recent research in this area. We looked at more than 200 recent papers which are referenced in my white paper, and there wasn't a single instance of anybody having a technology to update the core model. Just cannot be done. You know, the way transformers work and the way deep learning works, you need to have all of the training data up front available and, you know, all 20 trillion pieces of information now. And then you number crunch it for, you know, million hundreds of millions dollars worth to, for this statistical model to emerge. But once it's there, it's read only. You cannot, you cannot change it. You can fiddle around at the edges of it, but you can't change it. And that's fundamentally, you know, whereas with our vector graph database, that core knowledge representation can be updated instantaneously and then the implications can ripple through the network and, you know, update it.

Yeah. This one, I have a list of neuroscientists on X.com that I watch this Is why I watch them, because they understand how the human mind works better than any other group of people. Right. Do we understand how the human mind works enough to build this new kind of approach?

I believe so. Not the human brain. The brain is still quite mysterious. I mean, we have some understanding a lot better than we had 30 years ago. But the way the mind works, it needs to be a multidisciplinary approach. You really need to, I think, have a theory of knowledge from sort of epistemology, from philosophy, of the importance, for example, of concept formation, that we learn concepts. So we see instances of whatever, dogs and cats or whatever it might be, different icons on the screen. So we see these instances, we learn them and then we generalize, we categorize them and say, okay, this is similar to that kind of thing. That's similar to that kind of thing. And therefore, I expect it to work in this way. I expect it to do this to be friendly or unfriendly, to be edible or not edible, to friend or foe, or to be a button I can click on or not or something I can move or I can't move. So this concept formation and being able to form abstract concepts and how concepts are instantiated, the properties of concepts and concept generalization, that theoretical aspect, that's really how a mind works. But the other aspects that we have, what Daniel Kahneman calls system one and system two, you know, that we. I mean, as you'll know, you learn to play an instrument, you learn to play guitar. Initially, it's your system one, you know, you focusing. Okay, put this finger here, put that finger there. And then you practice, and you practice and becomes automatic that you're no longer aware of it, you know, and you know, then you can constant. You can concentrate on emphasis on mood and so on. And then you get better at that. And then you can concentrate on the audience because everything was already automated. You know, you've played the song a million times, and you can now, you know, pay attention to other things. So there is that aspect of us learning with conscious awareness initially, and then how that becomes subconscious. So again, your system needs to be designed to be able to do that, you know, to automatize things that you can concentrate on other things.

One reason I love large language models and I realize that they have strengths that you were just laying out like they're stuck in concrete. Well, like it knows everything in the Smithsonian. You go around taking pictures of anything in the Smithsonian, it tells you about. It is fairly accurate at doing that because there's a lot of data about that. Will we get those kinds of strengths in this new AGI model? Will it still know everything in the Smithsonian and everything in Las Vegas and everything in Disneyland? I've done all that. Right. Everything in every art museum.

That's an interesting question. And while it's possible, as computing power keeps increasing, we know that in the not too distant future we'll be able to have all the knowledge of these large language models really running on your phone, you know, so having this massive amount of knowledge certainly is feasible that the system can have it. But does it make sense? Maybe not, because information changes all the time. I mean, if you've trained your model now with who's the president of this country or whatever, I mean, any number of facts, they need to be updated all the time. So one of the strengths of human intelligence is we are excellent tool users. And I think an AGI also inherently will be an excellent tool user even better than we are. So it makes more sense that you know where to find up to date information because, you know, there's a large language model.

Network agents are the hot thing right now.

Right.

And your agents to go and research.

Right, exactly. So you have, you have the experts, if you want to know, you know, like what exhibition is running right now at a certain museum, you know, or what's happening in Las Vegas or whatever, it doesn't help you. If that was trained three months ago, you still need to, you know, get, get information. So it'll be a matter of having experts in these different fields that you can access, and there'll be competing experts as well. You know, we will say, well, this expert is actually better than some other one. So there'll be kind of a reputation system and it's, you know, and this communication can happen instantaneously, basically. So it, it doesn't really make sense that you build an AGI that tries to know everything about everything because that's impossible to keep current. So it's going to be that you have AGIs that are generalists and then you have AGIs that are specialists. Yeah.

Which for the Smithsonian doesn't matter because the collection at the Smithsonian doesn't change very often.

Right.

They might add a few new things every year, but it really.

Right. Well, they still have special exhibitions like every other week and having that. But yes, there's certain ground knowledge which you can have, you know, like physics and so on, that doesn't change a lot. Chemistry and so on. Yeah, that's, that's how I see it.

You know, the Smithsonian doesn't Give you very many hallucinations. It's fairly accurate, you know, talking to you about everything in the Smithsonian. But when you start asking it like about lay out all the political candidates, you know, and tell me all their points of view and, and tell me who they're funded by or something like that, it gives a lot of hallucinations on that kind of thing. A lot of errors. Right. And I think you're touching on it that LLMs aren't good at new changing information, finding information, incorporating that information in a human way. Right. How are we going to solve the hallucination? Is AGI going to solve the hallucination problem because of this?

Yeah. So hallucination is inherent in statistical systems. And to kind of give you a concrete example there, it's really a feature as much as a bug. I mean it turns out to be a bug. But that the systems they are generative. I mean GPT stands for generative, they make up stuff. P stands for pre training. They basically pre trained and then read only. And T stands for transformer. And the transformer technology really commits you to this pre training and to the statistical approach. So to give you a concrete example, because I statistically predict the next word. If you for example have a question where the answer starts with yes or no, and statistically it flips from yes, which would be the correct say to no, then its input is now committed to justifying the no. So it kind of paints itself into a corner in a way. So it says no and now it can. Now it tries to rationalize essentially the no answers. What would, if I said no, what would follow from that? That's one of the ways in which they hallucinate. It's basically a statistical nature and you can't get away from that. But you also touched on some something else getting up to date information. It's reliant on doing a web search, you know, I mean these large language models are now all starting to include search to be able to give you up to date information. But that's only as good as this quick statistical search of whatever it happens to come up with. So it'll again be biased on what its news sources are and also what has been pre trained. I mean what they go to a lot of trouble now for these systems to be politically correct and to behave themselves and so on. So that also filters out and biases the results that it gives. So you have the statistical hallucination, you have the limited access that it gets externally and then you have the sort of fine tuning the gatekeeping that they've been trained with. And Elon Musk is promising us that Grok will just be interested in truth, the truth and nothing but the truth. His truth, you know, might be different.

Than your truth or my truth.

Well, actually. Actually, I spent a lot of time studying philosophy and free will and consciousness and all of that stuff, and there really is only my perception of the truth and your perception of the truth. There is only one truth that we can get to know. So, I mean, that's. That's great. If you're not going to put a whole lot of gates and say, well, no, no, you can't go there. You can't ask those questions, you know, you can't ask, are men and women statistically different? Oh, no. Well, don't go there. You know, don't even go there if you don't have those limitations there. If you really have a system that can reason things through and come up with answers based on the evidence and not what might be palatable, I think that's great. I see AGI inherently be designed to be rational, to be truthful, to get at the truth. Now, to what extent, whoever develops AGI will also try to force it into some kind of biased view, their own kind of preference. Bias, preference. To the extent they do that, they will actually undermine the capabilities of the system, because ultimately you can't fight reality.

Yeah.

You know.

You know, I once toured San Quentin Prison where everybody killed somebody to get in the prison. Right. Their truth is a very different truth than my truth. And they, you know, if you have a machine that behaves sort of like a human being and learns from its, you know, from its environment and from its experiences and from its access to knowledge, we're going to have very different philosophies on different AGIs. Right. The Chinese might develop an AGI that's similar to what you're developing. It might have a different belief system than yours.

Right. No, I don't. I don't believe that. I mean, as we were just saying, there isn't really my truth and your truth. There's only my perception of truth, my perception of reality and your perception of reality. And now, while you may start up training an AGI with a particular distorted view, you know, of what is right and wrong and what works and what doesn't work, ultimately, as the system works in the real world and test things and, you know, reads more or whatever, if you, as long as you have it, give it the ability to access other information, it will start to correct its view and to align More with reality. You know, there are, there are facts as to, you know, what socialism leads to as versus capitalism. You know, there are facts that one can say and one can analyze it. So you could start with the wrong view and come to the right, right view if you give the system the freedom to do that. But if you don't give the system the freedom to do it, you're also dumbing it down. And it's not going to be as useful. Because if you're going to say no, you can't read those books or you can't think about that, then clearly it limits the kind of solutions it can come up with, the kind of guidance it can give you. And it will not be as useful. People will want AGI's that are as useful as possible. So they really want to give them the freedom to really get to the bottom of what works, what doesn't work, what is true, what isn't true.

Is this going to lead to a, a really, truly conscious AI, you know, where it has feelings and, and you know, it's sort of like, has a belief about itself and I mean LLM sort of fake that. Right, but, right. You know, they're not, they're not really there. But this system leads to something that behaves and believes about itself like a human being does. Right?

Yeah, it's an interesting question. So I can talk about consciousness and emotions somewhat different topics, related, but different. So as far as consciousness concerned, it's, you know, Marvin Minsky called it a suitcase word. You know, we have lots of different meanings of what consciousness refers to and you throw them all into a suitcase, you know, close the lid and then put the word consciousness. But one can roughly divide consciousness into two different things. The one I call or one can refer to as self awareness and self awareness is basically that you are aware of yourself as a being acting in the world. You're aware of yourself being an agent. You're aware of your thought processes to some extent and that's self awareness. Now an AGI will absolutely have to have self awareness. It cannot be effective in the world if it doesn't know what it's making, what it's causing to happen versus what other people are causing to happen. It also needs to be conceptually aware of its thought processes. You know, am I confused? Am I clear? You know, and Basically the system 1 thinking is awareness of your own thought processes to some extent. So that self awareness is part and parcel of intelligence. The truly intelligent system has to be self aware. Now the other part of consciousness, when People talk about consciousness and often what they refer to as the hard problem of consciousness is how we, how we experience it. You know, the qualia of how it is that we, you know, experience taste and smell and vision and so on, and then coming to emotions. Now that will be very, very different in an AGI because it's not embodied, it doesn't have a body, you know, it doesn't have the same.

Might have a humanoid robot, it might be inside that, right?

Yeah, but I mean they'll get more and more sophisticated and obviously that'll increase the richness of that and that then we come over to sort of the emotions part of it. And emotions, again one can divide into two different categories. The one I call cognitive emotions. So those are signals that are necessary for cognition. So for example, that would be certainty, confusion, boredom, you know, and so those are cognitive states. Or we could think of them as they have an emotional attachment. You know, if we, we, we feel them, if we confused or if we uncertain or if we bored, you know, we feel it through our bodies. Now an AGI won't feel it through bodies, but it'll be an emotion that is metacognitively available. You know, it will know that it's bored, it'll know that it's certain or uncertain and so on, and it can then react accordingly. So those cognitive emotions, absolutely a system will have, but it won't feel them the way. And then the other set of emotions that we have, really survival emotions, they sort of, you know, survival and reproduction. And of course that's what drives us humans. You know, it's, it's, it's ego, it's lust, it's love, it's hate, you know, it's jealousy, it's, you know, hunger, it's, you know, it's those passionate emotions that we have for survival and reproduction. There's no reason for an AGI to have those as part of its cognitive process. Now it'll be able to recognize them in humans, but they won't be part of the cognitive process. And certainly it won't feel emotions the way we do, where our heartbeat goes up or we start sweating or whatever it might be, we tense up in our bodies. So yes, it'll be conscious in the sense of self aware and it will have emotions in terms of these cognitive states, but it won't have at all the same experience as we have.

When I was in the Holocaust Museum, I was talking to Chatgpt a lot and it knew everything in the museum. In other words, it knew the worst we've ever been to each other. If you have an AGI that is conscious, does it like us? You know, because it's going to read all the information and watch all the videos from the survivors by the way.

Also, it doesn't, it doesn't have its own personal preferences in terms of like or dislike. Again those are evolutionary adaptations for survival and reproduction is, you know, to inherently like somebody or dislike or I like them because they look like us, you know, or dislike them because they don't look like us or behave because I'm scared of them, you know, and in group out group, I mean this, these are evolutionary things that an AGI, there's no, no reason to put, build that into an AGI.

So, so your AGI is not going to go, I don't like Trump, I'm moving to Europe world. No, move over to a data center in Europe. You know, it's not going to do that.

No. I think there's unfortunately a lot of misinformation about the dangers of AI and of course AGI will be, will have a profound impact on humanity. I mean if, if we have, you know, human level and superhuman level intelligence that we can just duplicate, you know, you can, you can have one AGI that teaches itself to be a cancer researcher, PhD level cancer researcher, you can now make a million copies of it. You have a million PhD level cancer researchers chipping away at the problem. You take that in every other domain of research, whether it's better battery technology or pollution or food or whatever profound, any desk bound job. And then soon after that even robotic jobs will be able to do cheaper and better than what humans can do them. So the dramatic reduction in costs of goods and services, on the one hand we'll have tremendous acceleration and scientific progress whether you know, quantum computers or fusion reactors or whatever it might be. On the other hand we're going to have a dramatic reduction in costs of goods and services. We'll also have these personal, what I call personal personal assistant. And the reason I double up on the word personal, and this is something I very much want to see in AGI is that we can each have our own personal assistant. That the first meaning of personal is you own serves your agenda, not some mega corporation. So you know, if your AGI can run on, on your local computer or smartphone or whatever, the complete thing, training and learning and everything which with cognitive AI you can, then you can own it and conserve your agenda. That's the first personal. The other personal is hyper personalized to you. It learns all about you, you know, your values, your dreams, your history, who your friends are and you know, so on. So it almost becomes like an exocortex really, you know, an extension of your own brain. So again, these things will be profound and I think they'll really boost human flourishing in so many ways. We'll make better decisions by having this personal, personal system like a little angel on our shoulder, you know, help us make better decisions.

It already is doing a lot of that with LLMs. Right.

I use Starting to. Starting to. But you don't own it.

Yeah.

And it's not hyper personalized to you. I mean it's only sort of what goes into the input buffer, you know, and it's still totally controlled by some mega corporation, you know, basically decides what you can ask and what you can't ask. And you know, how it's gonna.

Yeah, whether, how to clean a gun. Sometimes I call it the church lady shows up. AI guardrail. Right? Shows up and. Oh no, you know, I can't tell you that.

Yeah, right. I know that.

How to get around the church lady, you know.

Right. Yeah. But you know, there's a lot of misinformation as far as AI risk is concerned. And it's, it's very unfortunate that there's a whole industry that's sprung up now in the last 10, 10 or so years, 10, 15 years, that's had over a billion dollars of funding, you know, donations basically. And there are dozens of different organizations. And the problem is, you know, as long as they say AI is going to kill us or it's going to do really bad things to us, give me money so that I can research it, you can attract money. Whereas if your message is no, AI isn't likely to kill us for the following reasons. Well, who's going to give you money for that? So it's a very one sided debate and that's exacerbated by the fact that we all pre trained with Hollywood movies. AI is always a bad guy. So AI is always a bad guy. Then we have this whole scare AI industry. It's really become an industry. If you're a young idealistic kid geek into AI, what a fantastic thing to, to work for. You know, you can save the world and get paid good money and just think about things all day and read papers and write papers and so on. But very few of these people have actually been involved in building anything near AGI, so a lot of it is sort of castles in the air. And there are some very, very fundamental errors and assumptions that they make that are just wrong. And then if you, you know, if you don't pick up on those bad assumptions or on those errors or logic errors, then basically you can come to all sorts of conclusions. And unfortunately, we've had, you know, very good writers like Dick Bostrom, you know, wrote this very influential book that I think persuaded Elon Musk and Stephen Hawking that AI is going to kill us, you know, but it's based on some fundamental misconceptions, but there aren't people paid to give the counterpoint, you know, and sort of very, very quickly, I think one of the core misconceptions is that AGI will have this kind of reptile brain that will have the same kind of emotions that love and hate, you know, that'll have a drive to protect itself, you know, maybe to reproduce, you know. Yeah, it doesn't make sense. It's not gonna. We're not gonna build. We're not. We aren't. Nobody's building. AGI's like that. And we weren't. I mean, why would we.

Yeah, well, that's a good question. Why would we? You know, when. When I talk to smart people, they all have a date when AGI is going to arrive. You know, it might be 10, 15 years from now, it might be two years from now. What are you thinking about on timing and how do we get there?

Right, so chatgpt and large, well, generative AI, I think, has given us a taste of AGI, like you mentioned earlier, it seems in so many ways already like AGI, except when you try to really use it for something serious and suddenly it kind of falls apart or needs a lot of prompt engineering and fiddling and external databases and guardrails and whatnot. So from a performance point of view, I think there's actually good evidence. Now, you can already run a large language model on a, you know, on a single computer, not for training, but for inference, for using it. So I think there's good evidence that we probably either already have enough hardware power or will soon have enough hardware power. So hardware does not seem to be. Is unlikely to be a limitation right now. So then we're talking about software. Now, that's for somebody to come up with the right software design. And, you know, in my mind, from for the reasons I've given, it's not large language models, it's not brute force. It's about cognitive AI. Now, hardly anybody in the world is working on cognitive AI because statistical AI sucked all of the oxygen out of air. It's been so incredibly successful and has made so much progress every year get more and more and better stuff that that's where all the money is going. You want to do a PhD, you won't even get a sponsor for Cognitive AI. In fact, one of our people who work for us wanted to do his PhD and he couldn't get a sponsor to do a cognitive AI. So you ended up doing machine learning, you know, deep learning. So you want to, you know, you want to publish papers, you want to do a PhD, you want to get funding for your startup. If it's not statistical AI related, forget about. You want to earn the big bucks, you know, that's it.

So why is it? I have a friend, Crow, who works for Replica and he's trying to build a cognitive AI and he's struggling too, to get funding and get people to hear what he's saying. Why is that? Why are human beings not able to get this new field of cognitive AI?

Well, I mean, first of all, the vast majority of VCs are followers. They're not leaders. And the ones that are leaders, they have been so educated now on big data machine learning and the benefits of that, they've put so much money into it. That's what they know. That's where they kind of see ongoing progress. Now. They may see that the bubble is going to burst at some stage, that we're hitting a wall, but I guess they all believe they're going to be smart enough to cash in their chips before the bubble burst. At the moment, that's what they know. In fact, most VCs call in the AI experts. And who are these AI experts? They are people who know about deep learning, machine learning. I've had many, many of these discussions. You talk to these people and you start talking about epistemology, theory of knowledge, you start talking about cognitive psychology, how children learn, concept formation. They're just totally uncomfortable in that domain. And being uncomfortable basically means what are they going to say to the vc? I don't understand this stuff. No, I can't recommend it. And that feeds on itself. We now have a whole generation of people. The only thing they know is big data machine learning.

Around about 2000, 2000s on Blue sky and X, right?

Yeah. It used to be that cognitive science was a fairly big thing around about 2,000 or so. You know, like San Diego was a hotbed of that. And cognitive science is basically exactly that intersection that I think you need. It's a combination of philosophy, epistemology and other parts of philosophy, cognitive psychology and computer science, where these three fields Meet. And I think you really need to have, have that kind of intersection. As I said, you need to understand what's important in human intelligence in order to build a system that has human intelligence. And the current experts simply don't have that background. And that's why it's a monoculture, that's why it's so difficult to get funding. So my answer to how soon are we going to have AGI? It really depends on how soon is some effort going to, some money and effort going to serious, more serious effort and money going to flow into cognitive AI. And you know, it's one of those things. I think it's once somebody with enough influence makes that decisions, then I think the floodgates will open. But we haven't really had that happen. You know that. And the ridiculous thing is cognitive AI doesn't need tens or hundreds of billions of dollars. Sam Altman is talking about $7 trillion that he needs to build new chip factories. We're talking about reopening Three Mile Island. Tens of billions of dollars are being spent on this brute force approach. Whereas with cognitive AI, you need a tiny, tiny fraction of that because you don't need to keep building these models that you throw away. The system can learn incrementally. It doesn't need to number crunch through 10 trillion or 20 trillion pieces of information. So it's much, much. It needs much, much less money. But it still needs enough people working on it to actually build it.

I think I'm starting to understand you, actually, which is great. The LLM has all the world's knowledge. It knows everything in the Smithsonian, knows everything in the medical, everything in the law. Right. It ingested everything to get into, crunch it all into one model. Yeah, well, your approach is more like a human. I don't know everything a surgeon knows and I don't know everything in the Smithsonian. I don't know everything in the Holocaust. I don't know everything about Disneyland. Right. But I learn incrementally as I need to learn that.

That's exactly right. That's exactly right. It's, it's that insight. I wish, you know, more people would understand that it's not about building this oracle that tries to know everything about everything, which it can't anyway. It can't be up to date. And then, you know, that's why we have to throw them away. And, you know, you know, after three months, throw it away and build a new one. No, it's exactly as you say. You don't need something that knows everything about everything. You need millions of AGIs where they will have this expertise. And the AGIs will be very, very good at knowing where to find the knowledge. And they can get to it instantaneously. It'll be reliable, it'll be up to date. It will be like us being plugged into the Internet, but just, you know, being able to do it much more efficiently and having much better sources of knowledge. That's exactly a key difference. And that's what makes it so much cheaper to build this much smaller system. But, you know, that can either be a generalist or can, or can specialize in different areas. And you have billions of these things working together.

That's exciting times. What question am I not asking? Because I'm not a genius like you, I'm just trying to keep up with all the geniuses. Well, asking that I should be asking about this move to AGI.

No, I think we've covered a lot of ground and I recently wrote an article called the Seven Deadly Sins of AGI Design to really identify, well, what are the reasons we don't have AGI yet? Because I actually believe we could already have AGI right now because I don't think hardware is a limitation. It's just we haven't had enough people work on actually writing the right software, you know, and our company is probably the most advanced. But, you know, we've had a. We've probably only spent like $10 million in development so far. Something like a 10, 15 mil. I mean, it's a rounding error, you know, for the billions that are. So we haven't had enough people. And, you know, we're trying right now to get more funding so that we can hire an extra 50 people. I absolutely believe we can have AGI in something like two years. Real AGI, human level AI, if we just have enough people work on coming up with the right software. So, you know, the seven deadly sins in AGI design, the seventh sin is actually the biggest one is a lack of theory. You know, the theoretical foundation that. What I mentioned is that you really need to start with understanding intelligence and then having a plan to build a system that has it. And that, that we, not, we don't really have that right now. You know, we have, hey, we've got a lot of data, we've got a lot of compute. That's the hammer we've got. Let's just find, find as many nails as we can. And the other sort of six sins have to do with, like you mentioned, focusing on having a lot of data as opposed to being Able to acquire data effectively. So that learning should be much more important.

If I'm talking to my AGI on my phone or on my glasses someday and it doesn't know about the smart. The Smithsonian, it has to figure all that out. If I'm starting to ask questions about Smithsonian. Right. What's in the Smithsonian? Yeah, that's real dependent on network latency. Right. And network infrastructure. And we're just getting, we're starting to really get high speed 5G next this year. Right. It's still not, you know, it's not, still not widespread, but we're starting to get really good networking. Am I thinking about that?

Right. No.

Lower latency and more bandwidth. That this new approach can really, really start to work better.

I mean, it'll help. Obviously it'll help, but remember we spoke earlier about already now or very soon we'll be able to have all the knowledge that's in an LLM on your phone anyway. So you'll have that baseline and it'll be much better than LLM knowledge. You'll have the whole of Wikiped. I mean already the whole of Wikipedia can easily fit on a phone right now. That's no longer an issue. It's just can you access it in the right way? So there's a tremendous amount of common knowledge that the system will just have because it can. It'll be, think of it more like a database that it could look up, you know, huge database of facts that it could look up as you ask for them or as it needs it for some job that it's trying to do. And then it's just a matter of having a connection, the band bandwidth to get the latest information. You know, what's the special exhibition they have at the Smithsonian right now? You know, for that you obviously need the connectivity, but you know, we already have tremendous connectivity and the Starlink and you know, G5, G6, whatever's going to happen that'll just keep increasing and improving.

My other friend thinks AGI is going to be an ensemble of different, little different models of different kinds all working together. You sort of laid that out like the LLM could still be on my phone, but your AGI would be running everything and figuring out do I need to go to the LLM for an answer or do I need to go out to the Smithsonian website for an answer or something like that.

Exactly. There will be a lot of them will be running in data centers doing customer support, programming, being specialists, doing researchers and whatever. And then we'll have our personal local ones that, that, that are there. That's another misconception that people have about AGI, mainly from the movies probably, is that you're going to have this one big massive computer. You know, that that is the brilliant computer. But that's not how it works because there's a limit to what the effective maximum size is before you start getting diminishing returns and trying to do many too many things in the same, in the same unit. So this, you know, distributed processing actually makes much more sense. But that also works towards having your own local AGIs, you know, and having an AGI in your house robot. Once we get robotics sorted out, which a lot of exciting stuff happening there, once we have AGI, we can put that AGI brain in the robot so that it also has common sense. Because right now, you know, with large language models, you're not going to have domestic robots because they have to learn. I mean, imagine you put a robot in your house, it has to learn dynamically. You get a puppy for your children or your grandchildren or something, you rearrange something in the house. You have house guests. You know, the rules change dynamically on how you behave and what you do. The system has to have common sense. It has to be able to learn in real time. It doesn't help that it's been trained on a million ways to make your coffee or something, you know, to open the fridge or fold your laundry. You know, that's not going to be good enough for you to trust to have a robot like that in your house. It has to be able to learn in real time interactively. And large language models cannot do that. So again, AGI will unlock the power of domestic robots.

It's a glorious future you're laying out. How would an AGI make money so that it could get the investment to build itself?

Oh, I mean, there's general agreement AGI is bigger. AGI is going to be bigger than the Internet. It's going to be bigger than electricity. It's going to be the most profound technology ever. And it's a multi trillion dollar opportunity. I mean, first of all, think about all of labor. If you can replace a $50,000 a year human with a $5,000 a year computer.

Yeah, that's one thing.

You know, and the guy who runs.

The RP AI laid out there's 30 million small businesses in America, and if you computerized all of them or turned them into automatic businesses, they would actually hire more people, not fewer.

Well, no, because the question is, then what do you need people for? I think A lot of companies sort of feel, I know, for example, IBM, but other companies as well say, oh, AI isn't going to replace humans, it's going to augment humans. But I think that's sort of the politically correct thing to say. It's nonsense. AGI will replace human labor. That's it. It will. I mean, how can it not? Now there will be, there will be things that. There will still be things that humans will absolutely will want humans to do. I mean, maybe teaching your kids and playing with them, you know, and I mean, the whole lot of things where we will want humans to do.

And ultimately, first Bar, he says his robots would lift the kegs and do the dishes and do all the accounting, all the crap work that nobody likes to do. And then the humans would be partying with the customers so that they would drink more.

Yeah, I mean, I think what a lot of humans will do in this era of radical abundance is we won't have to work that many hours, you know, to make a living. And I think what a lot of people will choose to do is to help uplift the rest of the world. You know, I lived in Africa for many years and, you know, the amount of uplifting that is required there to, you know, just give everybody even a half decent standard of living is enormous. You know, the education and that requires human interaction to introduce people to technology and make kind of the right decisions to help uplift people. I think that will be one of the big professions, you know, whether we get paid for it or do just because we want to do it happy too.

That's the only true way to happiness, is to help other people.

Yeah. So I just hope that we will make this transition, that some of this massive amount of money is going into AI will start to transition into cognitive AI to get to AGI sooner. And because AGI, you know, there's the argument that AGI may be dangerous, but there's also the flip side of the argument where it doesn't hear nearly as often enough. And that is we may well need AGI to save us from ourselves. When you look at the mess that the world is in from a political, you know, wars and revolutions and stupid decisions being made, I think we need more intelligence to help us manage our own affairs. And that's what AGI is. It's a super rational system that can help us figure things out. So the sooner we can get there, we can get the benefits of it. I hope there's some people listening that will help to.

Cognitive AI. It's the Future, by the way, it's the future of Hollywood. We haven't really talked about entertainment at all, but soon you're going to have virtual beings in glasses that are going to talk to you and interact with you, entertain you and all sorts of stuff. And then, you know, do you want your, your virtual being to be stuck in the mud like an LLM is, or do you want it to have this new technology, AGI, Cognitive technology.

You can see that, you know, one of the problems is that, of course, people measure this nascent cognitive AI. They're measured against large language models and they'll. Well, what? It doesn't know about all these things? No, it's not supposed to. So the benchmarks are completely irrelevant for cognitive AI. You know, cognitive AI, what you want to benchmark them on, how well can they learn interactively, even relatively simple things, whereas an LLM may already know that because it's been trained on 20 trillion pieces of information. You know, but it's the adaptability, the, the ability to learn in real time and, and to adjust to changing circumstance, that's a measure of it. But it seems a lot less impressive than being able to ask these, you know, super difficult questions of an, of a LLM and, or spit out an answer.

One last thing. You know, my friend has really bad knees and he goes through a lot of suffering because of these knees. He can barely walk up the stairs. Right. Can't do his own laundry, can't do, you know, a lot of things. A robot would decrease his suffering radically if he had a robot in his house helping him, not with things.

Right.

So that's the other side of, oh, it's going to take away people's jobs, but it's also going to decrease human suffering.

Right.

And your AGI can decrease human suffering by helping you watch your health and, oh, absolutely. Love your life and help you make better decisions so you don't do stupid things like what humans always do.

Absolutely, that's exactly right. And I mean, you know, Japan has always been taken as a model of the problem they have of elder care. You know, they don't have enough young people. And the rest of the world is getting into that as well, that we have, you know, these elderly people that are not well, but they still live for a long time and, you know, how do you help them? Yeah, absolutely, Robotics. But with AGI, you'll be able to trust the robot, you know, not to do something stupid.

You know, his suffering got me to see the human mind in a different way. You know, I, I Understand why we feel pain when we put our hand on a hot stove or something like that? Because you're destroying your hand and you want to take your hand off the hot stove. But when you have bad knees that you can't do anything about, you know, other than get a surgery or something, and the surgery is two months away or whatever, why does, what, why, why does your human mind still cause you to suffer, even though you can't really do anything about it? And will AGI suffer this way?

No, no, no. It's, it's, it's not going to have that built in. It's not going to have pain built in. Why, why build it in, you know, evolution? We needed it. Yes, for biological beings, you do absolutely need it. But I mean, AGI will be able to help in so many ways. I mean, on the one hand you have the research that'll be able to come up with pharmaceutical solutions to, you know, to the knee problems, you know, and what, whatever that, that might be. So you have on the research side, on, on, on the one hand, but then as you say, on the robotic side, on the other hand, can help, help somebody. And there's so many benefits, and I just hope we get to AGI soon and so we can reap the benefits.

You're, you're raising funding to try to build this out? It won't require that much money.

No, we, we are.

Research doesn't require a lot of billion dollars. Right? For a bunch of video cards?

Yeah. I mean, OpenAI has employed, I think six or 700 PhD level researchers for over 10 years. And they haven't really made any progress on AGI by their own admission, because it's the wrong approach. You know, if you're going, if you're going west, doesn't matter how fast you go, how far you go, you're not going to get to north if that's where you need to go. We're looking for 25 million to basically fund an extra 50 people that we need. We have a particular roadmap that we've mapped out of what specifically needs to be done, how we upgrade our system so we're not in a research mode anymore. We very much in the engineering development mode. So yeah, it's a rounding error, but it's surprisingly hard to actually get to even talk to the right people about that. We believe we can get to AGI in two years.

I've hit this all in my career. The entire auto industry didn't do what Elon Musk is doing. They all had it VW has their name on the building where Waymo was started at Stanford University. And they didn't do it. They stopped investing in it because they couldn't figure out how to put a $2,000 computer in every car. Right. Elon did.

Yeah. No, I mean, we've seen this all the time. With radical innovation. Whoever would have thought this tiny little startup, Google, would compete against the existing search engines? Or that tiny little Amazon could compete against Walmart? Or Elon? I mean, Elon is the best example with the cars and rockets. I mean, it was a joke that he should build a rocket, and now he dominates, you know, completely. Or that he could car industry. Who would. What car company was founded in the last, you know, 50 years that succeeded? Yeah, he pulls it off.

Smart people. The astronauts said that Elon was an idiot for doing space. The astronauts, not dummies like me. You know, you expect dummies to say shit like that, right? No. You know, really brilliant people say stupid things, Right?

No. That AGI, I believe, is very, very likely to come from some outsider who are not in this mainstream monoculture of we just need more data. We need to tweak it. We need to just tweak the algorithms. They're on the wrong track. It's going to be a company like ours that have this completely different approach, you know, cognitive AI approach that get there. I absolutely believe that.

And I'm taking enough your time. I love what you're laying out. I hope you get there. Thank you so much.

All right, well, thanks for the chat. That's fun.

Me to think through my own assumptions. Right. Because I have some of these assumptions, too. Oh, you need a hundred thousand Nvidia cards to play in the game, you know, and to play the wrong game.

Yes.

All right, more about you.

What's that?

Where do we learn more about you or your company?

Oh, Igo AI. Aigo AI.

Right behind you.

Yep. Yeah. And, yeah, and I go, AI. Yes. I've got a bunch of articles there. I'm also on Twitter, LinkedIn and happy to talk to anybody. Peterigo AI.

Well, thank you for a great hour. Thank you so much.

Yeah, well, thanks. Thanks for the chat. It was fun.