The Man Who Named AGI Says We're Doing AI Wrong [ft. Peter Voss @ AIGO.ai]

[00:00:00]Hello everyone and welcome back to another exclusive Boom Room interview.

[00:00:03]Today we have the pleasure of welcoming Peter Voss, a persistent thought leader in the world of artificial intelligence.

[00:00:10]He's founded several companies focused on research, development, and commercialization of AI technology since the early 2000s and that's only his AI journey. He's one of uh the three individuals who collaboratively coined the term AGI or artificial general intelligence, a hotly debated term. Uh and it's and he's currently serving as CEO and chief strategist for the company he founded, IGO, a company that's developing novel cognitive architecture they refer to as the third wave of AI, which does a great many things and potentially solve some pretty serious problems we're all seeing in the AI industry lately. Welcome, Peter. It's such a pleasure to have you on. Um, how are you doing? And how much do you love data centers just to kick it off?

[00:00:54][laughter] >> Um, well, actually the work we're doing um is not exactly going to obsolete data centers, but it's going to make them a lot less important because uh we believe the next wave of AI technology will require significantly less computing power. So it'll be able to run on your local computer. So a lot of people are going to choose for, you know, privacy security uh reasons and and cost reasons uh to actually run their their AI systems locally. But that's really only possible um with this new next wave of technology called cognitive AI.

[00:01:38]>> Absolutely. We're going to talk all about that, I'm sure. um such a huge topic right now, especially considering all of the news about recent models that are coming to market soon. I know Anthropics, uh Claude Mythos, um that they they've been teasing for several months now. And I saw I think I mentioned this in some of our content before. I've seen some claims stating that they anticipate um five and a half times greater token consumption as part of the complexity of the inference. But we'll go ahead and kick this interview off, Peter, in the same way we kick off every Boom Room interview. It's the same boring question. And considering your past history and tenure, um, this is such an open question. Uh, but you know, feel free to answer it however you like.

[00:02:22]Please give us an idea. Who is Peter Voss? Tell us about your background in the world of technology and AI and how did you come to be here speaking with us today in a nutshell?

[00:02:33]>> Yes. Um, sure. So I I started out as electronics engineer um started my own um electronics company for industrial electronics. Then I fell in love with software and my company turned into a software company. Uh developed um ERP system for small to mediumsiz business basically comprehensive accounting and um you know yeah very comprehensive package that company did quite well. We went from the garage to an IPO um in seven years. So that was very exciting taking a company you know literally that that whole journey learned a lot made a lot of mistakes along al along the way.

[00:03:13]So it's it's really when I exited that that company that gave me the freedom to um think about what big problems do I want to to to solve or can I you know spend my time on and and the thing that was very clear to me is as proud as I was of the software that we developed um it was really quite limited. It didn't you know it didn't have intelligence the way humans have. if the programmer didn't think of something, didn't anticipate something, it would just give you an error message or crash um if you're unlucky. Um and so the the challenge was how can we build software that can think and learn and reason the way humans do? Um and that's really the challenge um I've been pursuing and working on for the last um 30 years. So initially I took off uh five years to deeply study intelligence to get a really good understanding because to solve the problem I believe to solve the problem of building an intelligence system I believe it's important that you understand what intelligence is or what makes human intelligence so special. So I spent time uh studying philosophy, epistemology, theory of knowledge. You know how do we know anything? How can we be certain of things? What's reality? Um where does free will come in? You know, then making decisions, coming to conclusions, what does consciousness fit in? So, all of those topics. I I also studied cognitive psychology um in terms of uh what do IQ tests actually measure?

[00:04:51]Are they, you know, what what's important in IQ tests and how do they work? How do children learn? And how does our intelligence differ from animal intelligence? the accumulation of that five years of of research. Uh I also studied of course what had already been done in the field of AI in the prior 50 years and the accumulation of of of of that the outcome of that was I came up with a design for a cognitive architecture thinking machine that can you know think and learn and reason like like humans. And so over the last two decades or so um in in 2002 um I 2001 I started my first AI company um and we started turning these ideas into actual code. Um you know we spent five years basically building up a an infrastructure um a framework I should say. And at that point we um we had enough technology working uh of that that had sort of call it protoagi that had the ability to think and reason and learn like a human to some limited extent. So we then commercialized this in the call center space but I ended up with um investors who didn't really share my passion and vision of AGI of getting to full human level AI. So we sold that company. Then I started a new development company, you know, a new team, another five years of development, cranking up the IQ. And then we were back in a situation where we really needed to commercialize the investors we had, you know, really wanted us to commercialize sooner rather than later. So again, a commercialization phase. And um this sounds very repetitive.

[00:06:46]Then about 18 months ago, we decided to put our commercial business on hold to fully concentrate on finally getting our technology to full adult level, full human level intelligence. And that's what we've been working on for the last 18 months. So that's that's that's my journey and and background. Um >> it's a really impressive journey and uh I I think that just spawned off like 10 different questions I want to ask you already. But uh firstly uh did you ever feel like you were way too early or or too early to AI? Like uh I think a lot of in the mainstream it didn't really hit people's consciousness until you know chatbt and open AI kind of brought this to the mainstream's attention.

[00:07:27]Outside of that maybe some uh references in Hollywood movies but people really don't know still what AI is but uh did you find some difficulty in trying to explain that uh 20 years ago?

[00:07:38]>> Uh oh yeah absolutely. Um, you know, we've we've gone through cycles in AI where AI was the hot topic and, you know, like everybody, you know, I mean, how many years ago, 10 years ago, it was like everybody became AI, you know, >> um, because that helped you raise money, but before that, AI was a bit of a swear word, you know, >> and um, the term AGI as well has been tortured to death, you know, in in terms of how it's been abused. It's really quite annoying. Um, you know, you have Sam Alman saying, uh, oh, we've we've got a we'll have AGI soon, but it's not a big deal, you know, and they're just all over the place with with that. And uh it's it's it's really quite sad but but yes this was part of the momentum that um the the term sort of going back to AGI was this different you know even if people called it different different things they started calling it super intelligence you know and and and so on.

[00:08:39]Um but this was part of the reason why 18 months ago we decided uh to really focus on developing it because there is a real appetite now for you know the big labs are all supposedly working to get to AGI and I've you know recently wrote an article why virtually nobody is actually working on AGI which is quite a controversial statement. They may make the claim that they're working towards AGI, but uh they they actually good there's good evidence that they really aren't working on AGI. And we can certainly expand on that. Why why I make that claim.

[00:09:18]>> Yeah, >> absolutely. You know, Peter, I wanted to take a step back because you mentioned that you went from electronics engineering um into the world of software and and and afterwards, you know, deeper into I mean what many people might refer to as big data and then artificial intelligence or at least machine learning in between there, right? That that's sort of like I think the evolution that many people kind of see. But back in that early time, um you know, like Melo said, AI wasn't really on many people's radar. I don't think machine learning was something at the forefront of people's minds, especially in that industry at the time. And so I'm wondering when you decided to get more into um this this niche, for lack of a better word, learn more about cognitive psychology and tying that to data and and computing. Did you see any practical industry demand for technologies like this back then when when hardware was I mean, you know, in somewhat of like an infant stage compared to what it is now.

[00:10:18]Well, I I've you know once I started um really digging into or understanding what is what makes human intelligence special um in my mind I had a very clear vision of what kind of software need needs to be developed to achieve this you know call it AGI humanlike human level uh flexible intelligence and so in my own mind it was actually very clear that we could achieve this and quite frankly I believe we could have done it 20 years ago.

[00:10:54]Uh now it would have taken a lot of money, a lot of effort obviously you know um computing power and memory has improved so much since then. Um but I don't think there's any fundamental technology that uh even looking back now of course it you know it it was very hard to raise money for developing that because you know of skepticism and just it wasn't on people's radar at all that this was feasible or possible. Um so yeah it to me it's always been clear and that that this is possible and and and how to go about it. Now I I I don't know if you meant to say that I took the journey to machine learning and deep learning because we never did uh deep learning machine learning big data approaches are in fact the opposite of what we what we do. Um that's really you you know they they used to be um cognitive architectures were actually a thing in like the the '9s and 2000s and so on. Um and on the other hand you had neural networks. Um, and of course the breakthrough of neural networks is that people figured out how they could actually get them to do useful things by massively scaling them up by throwing a lot of data at them, a lot of compute.

[00:12:22]So neural networks basically, you know, turned out to to to uh to really have rapid commercial value, whereas cognitive architectures um never had that breakthrough. But in my own mind, that's always the the obvious uh path to to ultimately humanlike intelligence. And I I'll give you just one data point here of why it's so obvious, was obvious to me then and it's obvious to me now is a child can learn language and reasoning with a few million words. It doesn't need a few trillion words. these large language models as amazing as they are and I mean they really are amazing of what has been achieved through this brute force scaling thing really nobody expected that you could get that level of performance that level of sort of apparent intelligence uh from just brute force data and you know a few tweaks like the transformer architecture and and so on but it's really fundamentally doesn't make sense um you know our brain uses 20 watts, not 20 gigawatt. You know, uh a child can can learn to recognize a a giraffe or an elephant that has never seen before with one picture. It doesn't need hundreds or thousands of of images to to learn what a you know what what a novel thing looks like. So it was always obvious to me that you need this um ability to learn incrementally in real time and accumulate your knowledge and build concepts and not this big data. Collect all of your data, collect trillions pieces of information, number crunch them for, you know, weeks, months on end. um spend a billion dollars to build a new model and then this model is frozen and you throw it away 6 months later and spend another billion dollars and build the next model. You know, it it it's always been totally clear to me that that is ultimately not the way to get to humanike intelligence. So there wasn't you know and the commercial value of of it you know I was involved um with kind of um futurist um philos philosophy uh and and all sorts of different ways and life extension nanotechnology and so on. So I was hanging out with a lot of people who were uh thinking about the sort of the further future you know and what the future can bring. And to me the value of of AGI was always totally abundantly clear that it's you know how this can boost human flourishing how this is a multi-t trillion dollar opportunity how it will change everything. So there was never any doubt at all uh in in the value of of the technology. Um so the commercial thing to me was like of course you know there's there's going to be commercial de demand once you get to it. So it was really the challenge was to get to it to you know to uh to get to uh human level um intelligence.

[00:15:36]>> Actually that's a perfect segue. I definitely want to unpackage uh why you believe AGI is uh necessary and useful later on but first I during that soatical you said you studied all these different uh fields like philosophy and epistemology and uh I want to ask you like what is your definition of intelligence? because I've been having heated debates lately with people saying, you know, I believe LLMs are incredibly useful, just like you just said, but I don't believe they're intelligent or awake or or anything adjacent to that. So, people don't tend to agree with me there, but I wanted to gauge where you're at with uh your definition of intelligence.

[00:16:12]>> Yeah, let me take a [clears throat] a slight detour here. And I think there's a a big problem with people's understanding of what the meaning of words and what definitions are. Um that people think when they have a word like consciousness or intelligence that there's some kind of a platonic ideal that defines what that term means and you now need to figure out what it means. And and that's really not how words and concepts function.

[00:16:43]words and concepts uh basically have a utility. So you say why do we have a concept of intelligence?

[00:16:51]Well to differentiate it from no intelligence or lower intelligence. So you know it's the different it's basically to have the the the difference. So to to try and say you know then because people would then say well isn't ant intelligent? Look the amazing antels that or colonies that they can build and manage you know surely there's some intelligence so intelligence has to be taken contextually in what context are you using this word so um clearly LLMs are intelligent in some sense but in some context in some definition they they do intelligent things so I tend to then talk about is what makes human intelligence so special and so powerful. That's kind of how I unpack it rather than saying, you know, having a a crisp definition of it. So there are number of things that make human intelligence uh powerful and and I certainly I can give some kind of a definition within that uh context. Um [snorts] but um can distill it uh there three very important things uh in in human intelligence. The first one is adaptability.

[00:18:04]Uh and adaptab you know that's why humans dominate over other animals and that because we can adapt to changing circumstances.

[00:18:14]Um and how can we do that? Because we learn. We can learn. We you know rather than being pre-wired and having very limited uh sort of stimulus response type training like you know animals animals have we can actually um learn learn interactively.

[00:18:35]So the first thing is that we can learn interactively and basically overcome our our hardwired knowledge. You know in fact humans have relatively little hardwired knowledge. you know that's why we take a long time before we can mature and look after ourselves.

[00:18:54]Um so the so the first thing is the ability to learn and to learn interactively. The second thing is the ability to form abstract concepts and that's key. Uh in fact one of the white papers I wrote is called uh concepts is all you need. um as a bit of a hackne phrase now but anyway um and so the the second important thing is not only can we learn concrete things but we can learn abstractions and we can learn abstractions on top of abstractions which animals can't do. uh animals can form abstractions only at the sort of perceptual concrete level whereas we can form abstractions on top of abstractions to ultimately have things you know like honor and marriage and government and and you know um things like that that are very abstract concepts. So the third thing that makes human intelligence um special and I discovered that when when I studied um psychometrics um IQ test is how important metacognition is. We have the ability to think about our own thinking and again animals cannot do that or can only do it in an extremely limited way. We are aware of um ourselves being the agent. We are aware of our thought processes to some extent and we can manage that. In fact, what D Daniel Conorman talks about system one and system two thinking, I think a lot of slow and fast thinking a lot of people are now familiar with. Whereas your fast thinking, your system one is a sort of uh automatic more animallike um responses that you don't you know that just you've learned and and you can just act and then but there we are constantly supervising that. So for example the conversation we're having now um theory of mind you know what what what what do you guys expect um what do you already know what level of knowledge do you have so I adjust what I'm talking about you know what have you spoken about already um do I listen to myself and say was this clear enough should I reframe it and so on so our ability to have metacognition to monitor our more automatic take processes. So those are the three things. The ability to learn incrementally in real time with limited data. The second thing is to be able to form abstract concepts. And the third one is metacognition. So those are kind of three uh essential um essential requirements for human intelligence that make human intelligence special. Now in terms of of AGI, how would we know if we had an AGI? Uh that's actually pretty simple. If it if it can learn a new task and execute a new task, a novel task as easily as a human, that would be a good good test for AGI. And current large language models fall dramatically short of that. Um, you know, if if any of us for just taking a very simple job like um customer support, you know, if any of us went into into a call center, we'd read the manuals, listen to some calls, get some instructions, and then we could start taking calls and we would know when we don't know something and we have to ask for a supervisor, trans transfer it or look something up or whatever. Again, we can manage that process and learn on the job.

[00:22:30]AG current LLMs nowhere near doing that.

[00:22:35]You need massive pre-training things, prompt engineering and rag and whatnot and they are still not very good. That's why there's such a extremely low adoption for um in in in enterprise, you know, it's like five 5% of um of prototypes get into production, you know, um and and they're just not very good and they're not cost effective.

[00:22:58]Whereas if you have AGI you would it would literally by like having a smart human you put into that position and the onboarding will be actually less than you would need for a human. So that would be a test of AGI and that need would need to be true for whether it's accounting or administration or you know sales support or whatever it might might be that that it can learn on the job basically.

[00:23:26]>> Absolutely. I think I I wanted to, you know, just linger on this topic of AGI a little bit longer. I know, um, you know, from our research and understanding your past history and the work that you've done, uh, in 2001, um, I believe you along with Ben Gertzel and Shane Le, um, you coined a term artificial general intelligence. And I think you mentioned before like that that term um has been let's say bastardized by a lot of people in the industry that many would look up to as thought leaders in the world of artificial intelligence. We've seen people like Sam Alman and Jensen Hong claim that we're already there. We've seen government regulators speaking so haphazardly about AGI uh without a whole lot of idea of what it means. Um I'm curious in in that journey of creating this term um you know what were what uh mission were you hoping to accomplish in creating this term? Was it to encapsulate that definition properly?

[00:24:25]Was it to take a lot of your learnings from all of the studying and research that you' done before or was it to kind of like set the future landscape of what you think a lot of these players should work towards which largely they have not. Yeah, it's actually a very good qu very good question and um so what happened is um when I was after doing this research and I was ready to to start building AGI now uh I actually came across a number of other people who had similar ideas uh and you know Ben Girtzil was was was one of them. Um so what we actually wanted to to to do is to go back to the original dream of AI.

[00:25:10]The term AI was actually coined 70 years ago and the original dream of of AI was always to build thinking machines. That was the original idea. And uh in fact 70 years ago they thought they could crack this in a few years you know u but you know with within two or three years you know we have these magical computers you know with 16k of RAM and you know and they can do amazing calculations and so on and so they thought they could could do this crack this in in a few years. Now of course it turned out to be a much much much harder problem uh to actually build human level intelligence. So what happened over the decades is the field of AI turned into the field of narrow AI without it being renamed. So what we've seen really even up to date is that people take a particular problem that requires human level intelligence you know like playing chess um uh or medical diagnosis or whatever it might be uh or drawing a picture you know and and so on. So they take a particular problem and then use human intelligence to write a program to solve that problem.

[00:26:31]And that's actually very important distinction. It's not the machine itself has the intelligence to figure out how to do it. It's the external intelligence. And you basically are building a narrow uh you're building a you're writing a program to solve a particular problem or set of problems like Alpha Fold would be an example of that. Um and you know the deep blue IBM's world champion uh can't even play checkers you know and you know it's it's narrowly designed using using human intelligence human ingenuity to solve that problem.

[00:27:10]So we felt um the people that got together that we should go we wanted to go back. We thought the time was ripe.

[00:27:17]Technology had advanced enough to go back to the original dream of AI to build thinking machines that can learn by themselves and figure out these things that can teach themselves to play chess or to do container optimization or to do customer support or whatever it might be. And so we decided uh to write a book uh on that. And each each one of us I think were like got 10 people or so together uh each contributed a book chapter to that. And it was um Ben Gil Shane Le and myself who were brainstorming the title for this book.

[00:27:53]And you know we at one point we we thought that real AI was it but that was a bit too much in your face you know wanted to have it a bit more academic.

[00:28:02]And um I really like the the general which is also little G which in psychology designates actually in IQ test is general intelligence you know little G. So artificial general intelligence just was like you know a name we we we came up with and then the abbreviation AGI. Um so that it just was the best title for the for the book. And then you know Ben has now been organizing together with some of the other um authors um an annual AGI conference and it's never really taken off. I mean it's still a very fringe kind of conference unfortunately.

[00:28:47]Um but anyway, so that's that's how the the the term AGI came about is to recapture the original dream of of AI and have thinking machines. Well, thinking and learning machines.

[00:29:02]>> So So you mentioned this narrow AI and Sasha and I speak about this all the time. Actually, we had a few guests on.

[00:29:07]We we spoke about this exact subject, but you said that AGI you think it's good for humanity. Why do you believe that?

[00:29:15]because we're seeing now more than ever companies pouring trillions of dollars into operation cost building data centers trying to brute force AGI through LLM and you know we see like really niche applications of LLMs already working fantastically uh same with robotics too you you have rooms and you have little vacuum cleaners uh that are robotic you have uh robotic arms building cars and factories do we need this generalist intelligence or these general robotics that do everything or uh is it good enough to have these really specific use cases for hyper specific use cases basically.

[00:29:51]>> Yeah. I mean we we can of course see the limitations of that you know on on how much engineering and tweaking and control and human in the loop is necessary um for these um narrow AI systems. I mean the narrow AI system always requires us external intelligence and then when things change again you need ongoing engineering and tweaking and and and so on. So it it's not reliable. It's and it's expensive you know because you have this human engineering in in the loop and the you intelligence is limited basically on what it can do. you know, in some narrow applications, of course, it can, you know, be much much better than humans, but well, we've had electronic calculators for what, 40, 50 years or something, you know, they're obviously much better than humans. So, uh, now, why is it so incredibly valuable? Um, well, there are number of of areas that that I see that can really boost human flourishing. And yes, current LLMs and um the path we're seeing right now with massive data centers and this totally inefficient way of doing it and large companies basically trying to optimize trying to capture um trying to capture their users basically try to addict addict their users you know you are the product. So I think the whole philosophy behind the current AI is is not that positive. So I think it's not a good example what we we're seeing right now. So let let me sketch what I see as the the promise of AGI and that I've always seen and several of us see that kind of >> how it can help human flourishing. So the the first thing is um let me start with with research. Um imagine you have one AGI that can teach itself to become a PhD level cancer researcher. And I specifically say teach itself because it's no longer collecting a whole lot of data of cancer research and then training a model to do stuff like that.

[00:32:07]It's literally going through the learning process the way a human would would learn to become um you know a a good scientist. So you you now have this AGI cancer researcher, PhD level cancer researcher. You can now make a million copies of that. And you can have a million of these chipping away at the problem, pursuing different different avenues and communicating, sharing information much more effectively than humans can without egos getting in the way and just you know so many advantages of of AI, photographic memory, you know, instantaneous access to internet uh data and so on. So the progress we'll make in science and not just cancer research but you know just um generally in biology and nanotechnology clean energy um whatever whatever requires a lot of human intelligence right now which we don't have enough of you will have tre tremendous advances in in all of the sciences.

[00:33:15]So um so that's one one area that that can help basically to conquer disease and to create clean energy u more efficient computers, better food production, uh clean up the environment, you know, all of those things that are technical problems that basically require massive amounts of intelligence to to solve. The second thing is basically um massively reducing the cost of goods and services by automating a a lot of things. You know, human labor obviously is um the a large component in much of the uh goods and services that are that that we consume. And the third area that that I'm actually even more excited about is what we call a personal personal assistant.

[00:34:07]And the reason I double up on the word personal is that you own it and it serves your agenda, not some mega corporations. And the second personal is it's hyperpersonalized to you. So it's yours and and gets to know you. You can share however much you want with it. So uh they're kind of two different ways of of of describing that. The one is it's like a little angel on your shoulder that can help you make better decisions, can keep you out of trouble, you know.

[00:34:38]So, for example, if you want to break off a relationship, but it might say, well, maybe you want to sleep on that, you know, and think think about it, you know, um rather than acting emotionally.

[00:34:51]Um the other way to think about is as an exocortex, basically an extension of your cortex, an extension of your ability to think. So gives us basically the idea here is to boost human agency to boost our intelligence and allow us to do things more effectively and you know avoid mistakes that that that we uh we make some of the mistakes that we make. So those are three areas you know that I I see all as very very positive.

[00:35:19]Now this personal assistant thing right now we are seeing the opposite. It's not boosting human agency. It's undermining human agency, trying to get people addicted to it. It's trying to, you know, use um use large language models to cheat, you know, job applications and school at school and and so on. Whereas you want an an an AI that is really inherently designed uh to boost human uh agency and and to to be socratic, you know, to say, well, why do you want to do this, you know, and what is the benefit and and so on and think through the consequences.

[00:36:01]>> Actually, while we're on that, we we had a brilliant woman on the podcast before.

[00:36:04]I think she worked with Ben Girtzel actually. She's trying to pioneer AI psychology and she talks she's building in AI but she talks a lot about the detriment of AI technology developments especially for young people and kids and how becoming over reliant or overdependent on this technology can actually set you back cognitively and uh >> do you have any opinions there as well like how do we avoid going down that route?

[00:36:26]>> Do you remember who that was?

[00:36:28]>> Uh Anna Makada she's a she's a brilliant like AI researcher. She studied psychology. Uh we can show you the interview after the after the show but really smart woman.

[00:36:38]>> Um well yes absolutely I mean this this is but you know here it's it's really two things you require. You require the right technology. large language models clearly are optimized for engagement, you know. Um, and they're all building them for that, you know. Um, so um and and they they cannot be used as a personal personal assistant in that in that way, you know, you don't own it, you don't control it, you can't really teach it much. I mean sort of peripherally you can you know update memory a little little external memory a little bit and and so on. Um so the the first thing is it's a wrong technology to have AI that can really boost human flourishing. But the second thing of course you also need the companies providing it having a philosophy that wants to not just win at all costs um but that really is sees the benefit of you know of of human flourishing and and and boosting human agency and that is why you know we we currently looking for a financial partner for a series A and one of the things is to have somebody aligned and the big VCs clearly aren't there. I mean, for them it's win at all costs, you know, move fast and break things, you know, and uh that's not the philosophy we have. and and I I hope that the company who has the breakthrough with with AGI which hopefully will be us um will have that kind of philosophy you know not not to addict people but to boost human agency so we can certainly make a difference on human trajectory I believe on those kind of decisions >> you know I want to I want to um get your thoughts this question it's it's a big question. It might wind you up quite a bit um hearing what we've heard so far so far, but I want uh I want to hear from your opinion some of the the critical problems um that you understand within the LLM landscape. I mean there there you've been very vocally critical in the past but funny enough I think a lot of similar sentiments have been spreading more recently especially considering um you know a lot of the negative sentiment around data center constructions the social and economic impact um and the increasing complexity of models and the double-digit month overmonth growth in token consumption that's following as a result right and so um you mentioned before what I believe has been referred to as the frozen problem. Um, we know that we all suspect there's a huge scalability issue when it comes to LLMs and it seems to be a race to the bottom not only with these with these developers but also the investors that are backing them and so I I'm wondering from your perspective um how do you see this all shaking out in the end?

[00:39:48]>> Right. So I I'd like to quote Yan Lakun on on this and it's a almost direct quote for him. He says large language models on an offramp to AGI, a distraction, a dead end. That's how forcefully he puts it and I I agree with him uh on on that and um that's got to come home to roost. Now there are obviously applications in large language models that justify the cost of tokens and and and things you know where the costbenefit ratio is is is right. Uh it's it's not super clear what those applications are. I mean certainly for I think for for search you know I mean whenever I have a a legal question a medical question or just something I'm just curious about um I use these models and you know it's not super critical or if it is critical I'll double check it and go back to the source but they really are super useful for gathering the information tab you know putting it in a tabular format or showing the pros and cons um and I I I don't think the the token consumption for that kind of thing is that exorbitant. You know, they've become quite efficient. So the the whole number of applications where the cost benefit makes sense. Um but when when people are actually trying to use them for the kind of things even you know customer simple simple customer support it very quickly runs away from you you know and you have then you have multiple agents trying to control each other and and you have kind of explosion of of token requirements and you still don't have the reliability um of of that. So um yes I I see that there's there's going to be a massive correction when you know again the simple thing is our brain uses something like 20 watts not 20 gigawatt you know we can learn incrementally and let me actually just quickly list the benefits of having a system that can learn incrementally one sentence at a time essentially you know incrementally with limited data limited compute and learn to update its model in in real time. The first thing is you need a million times less training data.

[00:42:13]That's that's huge. Um because you have less training data, you can now concentrate on the quality of the data instead of the quantity.

[00:42:24]Whereas right now these large language models are trained with all the garbage of the of the internet and Reddit and whatever, good, bad, and ugly. And the system has no idea what's good, what's bad, and ugly. So, you know, when you're only training the system with a few million words, you can actually curate uh the the data. You have much better quality control over that. You need massively less computing power. We train our models on a single off-the-shelf computer.

[00:42:51]You know, that's a a massive difference.

[00:42:54]And that also then brings the benefit of that you can run it locally. So, you can now control it. So you have the privacy security issue and so on under much better control. So really a massive advantages. Now let let me talk about a little bit the sort of monoculture myopia as I call it. You know we have this monoculture that large language models are the only game in town. But all of the large um labs over the last 6 months, 12 months have come out with statements saying incremental learning, learning and memory is essential for AGI. They've all acknowledged that.

[00:43:37]They've all acknowledged that that's kind of the Achilles heel. Well, at least one of one of them. Um that basically incremental learning and Demisabis has said it. Sam Alman has said e all of the big labs have and and computer scientists they are they have acknowledging that real-time learning is essential. Now they also know that large language models which rely on back propagation and reinforcement learning can never do incremental learning. It's impossible. And I I co-authored a paper where we reviewed over 200 research papers where people tried to get large language models to to learn incrementally. You have catastrophic forgetting and and it can't be done basically. Um now they know you need incremental real-time learning. They know large language models can never do that as long as they have re incre as long as they have back propagation or reinforcement learning. But for them to connect the dots is impossible when their next funding round depends on not connecting the dots, you know, or their their stock options depend on not connecting the dots. So that's that's the world we live in now. And of course VCs have plowed so much money in there at at the moment. Nobody wants to stand up and says the emperor has no clothes.

[00:45:00]Do do you think it's >> it's like a dotcom bubble basically, you know, and and in in in that sense except here they are actually facts that could tell you that, you know, and they're basically hoping and praying that somehow throwing enough money at it will the problem will go away, but it really can't.

[00:45:21]In your opinion, do you think that's the only reason why these billionaires and and massive companies are accelerating quickly in the wrong direction instead of taking a step back to go forward?

[00:45:30]>> Yeah.

[00:45:30]>> It's it's purely a matter of funding.

[00:45:33]>> They they can't, you know, first of all, I don't think it's been put as clearly as I put it here, you know, that these two two things >> Yeah.

[00:45:42]>> cannot be true at true at the same time.

[00:45:46]Um, and I mean most investors are sheep, you know, they just follow the trends. I mean, we've been sort of um, you know, I've I've been investing long enough that I know there was a time before momentum investing was the only thing, you know, nowadays momentum investing is the only thing. You put your money in things that are going up, you know, that that's it and hope you can get out before it goes down. uh there isn't really sort of [clears throat] value investing and and people analyzing well what is the underlying value here you know and so it's yeah it's just it's just an outofcontrol train that nobody can wants to jump off you know it's just hoping that it'll somehow work out and and and of course also there's this fierce competition is who's going to be the winner and in in that sense you know like uh Meta Mark Zuckerberg throwing these billions at it, you know, hiring people for hundreds of millions of dollars and buying companies for billion.

[00:46:55]If [clears throat] you you've got to be in the race, you know, and if you have billions slloshing around in cash, hey, you know, if if you don't try it, if they don't, they don't have better ideas. I mean they really nobody is aware of the alternative of cognitive AI that really isn't on the radar. There isn't there isn't an alternative that they're aware of.

[00:47:19]>> I think this is potentially a perfect segue into learning more about IGO and we you know these it's it's very clear that all these companies are in a race to the bottom and then maybe that's the long-term strategy. um try and muscle each other out and then when there's one man standing or one company standing um you know that's when you make the economics work when when token consumption uh can be outpaced by the revenue coming in um and it's it's a very sober realization that we've come to but as an alternative I know um I go and and what you've been developing nearly for a decade now is a great solution to try and solve that issue So if you don't mind give give give us and and our listeners um you know a sort of elevator pitch on uh what is IGO and how is the architecture fundamentally different from these LLMs in the sense that it is a thinking machine you're able to achieve um you know this level of intelligence and cognition with uh you know orders of magnitude less token consumption or or energy consumption um with with far less training data and potentially have this live on an edge device which I think many people cannot deny is definitely the future of artificial intelligence.

[00:48:41]>> Right. So it it goes back that really the the whole company and yes we've been doing this for more than 20 years um is really based on the on the research and you know the three things I mentioned earlier that make human intelligence so powerful and and and say unique and there there many others but the three kind of cornerstones I could say the one the ability to learn incrementally in real time with limited data. Um the second thing is the ability to form abstractions autonomously that the system by itself can take data and say I see a I see a generaliz general pattern here I can make a generalization and then these generalizations can kind of stack up um can do that and the third one is metacognition that the system has some awareness of its own thought processes and can monitor and manage the thought processes.

[00:49:37]So those are three essentials. Now the way we implemented and and really our first prototype in you know early 2000s already embodied those requirements. Um there's also grounding. How does a system basically ground its knowledge with perception you know that that can act interact with the real world and and and so on. Um now the way we did that um even then and we've been refining it basically improving uh testing it commercially and so on over the years is we have um a super high performance uh graph database um vector graph database actually that uh encodes all of the knowledge and skills that the system has.

[00:50:28]And this uh graph database uh and I say super high performance. It is literally a thousand times faster than any commercially available graph database.

[00:50:38]So it's part of the innovation of my you know hardware background and and and so on made me realize how this needs to be designed to be much more efficient than a commercially available uh vector graph database. Um so we customd designed this graph database high performance graph database that it can learn and update incrementally. So every input that it gets it potentially can update or add or or or whatever. So that was an essential requirement. Now the cognitive algorithms that we have, the learning algorithms, the generalization algorithms, prediction and and context uh you know being able to extract context, those are also all customdesigned um to meet the requirements of human cognition and they're customd designed to work synergistically with each other.

[00:51:32]So there's nothing off the shelf that we took. We had to design this all from scratch uh to be deeply integrated with the knowledge representation with with a knowledge graph and to be deeply integrated with each other. So that's the architecture we have. We call it INSA uh integrated neurosymbolic architecture.

[00:51:52]And let me briefly talk about the neurosymbolic because that's also an important feature of a byproduct essentially of system one and system two thinking. um system one is more pattern matching. So that's more like neuro that's more like large language models you know p pattern matching whereas system two is more languagebased more logic based um so that's symbolic basically but the two systems you know are highly integrated with each other constantly switching between system one and system two so they can't be two separate systems they share the same data um they they same share the same knowledge representation But you you get the benefits of the pattern matching, the fuzzy pattern matching, the robustness of uh neuro type processing and the benefits of logical thinking of that reliable crisp logical thinking that that you get through symbolic manipulation.

[00:52:57]So [clears throat] the system can operate in both of these modalities. So in instead of having one of the big limitations of large language models is their their weakness in symbolic processing you know and and reliability.

[00:53:14]Whereas the the problem of first generation AI which was what's now called good old-fashioned AI DAP calls it the first wave of AI um is that they're very brittle. you know, whereas you have expert systems that that are purely logic based, they're too brittle.

[00:53:32]So, basically, by us combining that in a in a highly deeply integrated way, we get the benefits of neuro and symbolic.

[00:53:43]>> All right. Yeah, it sounds a lot like your definition of intelligence. I see the uh the link now uh between what you're building in and your theory on intelligence. But uh I want to imagine like if we fast forward a few years uh I go has solved AGI your first uh what does that look like? What does that like uh that pathway to AGI look like? And what happens next after AGI is actually achieved because I don't think people understand even nobody can even imagine what the world looks like after uh AGI.

[00:54:12]>> That's true. Uh we can't we can we can just guess [clears throat] make some educated guesses.

[00:54:20]>> Right. The way we're going about getting to AGI to get to full human level um in in artificial intelligence is um we are teaching our system like a child a three-year-old, foury old, 5-year-old so in layers so that the knowledge it acquires the most fundamental knowledge is foundational and can inform new knowledge that is being acquired. So in that sense we are following kind of a cognitive development path of of humans to some extent.

[00:54:52]Um and let me just quickly talk about the difference between the commercial system that we have and the AGI system that we're building now. The commercial system is basically a proto AGI. It uses the same knowledge representation you know the same kind of algorithms and so on that we had. So it's it it's really the same core technology. The big difference was to get to get to commercialization quickly. And you know we were very effective in in the call center space. Um for example uh one of our um best known customers is 1800 Flowers Harry and David group of companies. And you know over Valentine's Day period they had to hire 3,000 people normally for two weeks which is insane.

[00:55:39]You can't even train the people. you know, but there's such a huge spike Valentine's Day for them, 3,000 people.

[00:55:46]So, um, we did all the chat support for for the system and we actually were able to, uh, very successfully automate 90% of that. So, they didn't have to hire that that many people. Um but the way we did that because our system didn't yet have human level. It wasn't at AGI level is we did a lot of um manual training basically putting uh knowledge into the brain directly that basically business rules ontologies and so on. So it was you know several hundred,000 worth worth of work to set up the system to do this particular job. So it couldn't learn by itself. It couldn't just go and read the manuals and you know uh and and learn the way human can. So our AGI system we stripped out all of these hot handcrafted all of this handcrafted knowledge. We stripped it out and basically we've now been concentrating on improving the self-arning ability of the system so that it can learn like a child. So our development path is exactly to to teach it and in fact half of our our staff are what I call AI psychologists.

[00:57:04]Uh it's a profession I invented. Um they basically have a background in um linguistics education cognitive psychology. So they come up with the curriculum to teach the system new language skills, new thinking skills and reasoning skills and so on and just generally skills um come up with a curriculum and they come up with tests and they basically this is how we developing the system. Now as we go up that ladder from you know foury old, 5-year-old, sixy old the system will increasingly be able to learn by itself.

[00:57:38]you know, I mean, I I think by the time I was 11 or so, I went to the library by myself and I was self-motivated and, you know, learn by myself. And that that'll be the same thing as sort of once we get to this, you know, 8 n 10 year old the system will largely be able to learn by itself with just asking for help and guidance along the way. So that is how we are developing the system uh to get to basically college graduate say an AI that has college graduate STEM knowledge you know of knows about statistics and mathematics and a little bit of science and and so on and can then branch out from there largely by itself the way I mentioned it can teach itself to become a cancer researcher.

[00:58:27]it can then hit the books and specialize in these different uh different areas really autonomously.

[00:58:34]So that is the development path that we we see. So once we get to this uh college graduate which we believe we can achieve in about 18 months uh pending some um additional funding to to grow our team. We believe we can get there in in about 18 months. And then of course the system will be available really across any task that can be done remotely initially. And um we also in in talks with the robotics company that same brain can then be put into a robot into a robot because you definitely cannot have a a humanoid robot in your in your home unless it has an AGI brain.

[00:59:19]I mean, it has to be able to to to reason. It has to be able to adapt. You might have visitors coming over. You might have a plumbing problem, might have a new pet, you know, I mean, things are changing all the time where the system really needs to have human level learning, real-time learning ability, know when to ask for more questions or clarification, uh, and to be able to, um, think think about the situation in in in a at a very high level.

[00:59:48]um you know, you're not going to see robots you deployed domestically until you have AGI.

[00:59:54]>> You know, I have uh maybe potentially silly question, but I'm just very curious on the early stages of that training or learning. We're talking about this incremental learning for this um model in its infancy stage, year three, year four. Um, I'm curious how delicate and and risky it is um in in that process and introducing specific types of information or data. I'm I'm thinking to like I I don't have a child myself, but I I sometimes wonder like if I had if I had a child um that's three years old and you know like maybe I I let out a swear word or or like he sees me doing something I shouldn't like oh man that kid's going to have PTSD for like you know for the [laughter] rest of his life. So, I'm curious like if if this is sort of like a similar factor considering you've modeled it off of um the the kind of foundational elements of of human intelligence.

[01:00:48]>> Yeah. No, it's it's a good question. Uh fortunately, we're in a much much better shape than raising a child and and and causing that kind of damage because um we are constantly retraining the system and changing uh changing the training.

[01:01:04]So if we found that we mistakenly exposed it to some bad bad information that'll give it trauma, you know, um we basically can go back and change what we did at 3 years or four years. I mean it, uh, you know, it literally just takes, you know, minutes or an hour or two or something to retrain the system. I mean we constantly doing this you know as we as we learn new new things as we uh improve the fundamental uh cognitive algorithms you know we just retrain the system from scratch and then we have currently over 20,000 regression test lines that whenever we run you know we we have a new version of it we basically run it through and say will it still react in the same way to the in a desirable way or better way than it did before. So, um there's there's a lot of control you you have over that with with an AI for fortunately. So, that that's really not not a concern for um for for us.

[01:02:08]>> Yeah, thanks for entertaining that question. And you know, I'm curious.

[01:02:12]We've talked a lot about the LLM landscape. We know that um you know what what the intentions and the behaviors um that are that are being driven from the intentions on the investor side is clearly >> um but I'm curious on on the enterprise or application side. I know you've had a number of different enterprise enterprise clients that have integrated um some of IGO's technology and I'm curious um you know what what's the sort of market sentiment or feedback around using these technologies because I'm sure some of them are familiar with for example chat GPT or claude or Google's Gemini >> well the the the problem is right now the current technology is just incredibly expensive to implement uh in enterprise and then the results aren't great um you know it's just not reliable. So you constantly keep tweaking it because the the models change under your feet.

[01:03:12]So you optimize and tune the system and set it up with a lot of labor, you know, your rag and prompt uh prompting and and prompt engineering and whatnot. Um or agent um harness. you set it up to to to do a particular job, but you can only test so many uh situations. So you then find all of the edge cases. You then have to tweak it. Then your model changes under your feet because you know they release the new model and everything changes again. So it really is very difficult uh to deploy this in in critical applications because it just isn't that predictable and it simply isn't isn't that smart. So it's it's that's why there has been little adoption uh in in in enterprise. I mean they're all playing with it, you know, but the way they're using it is pretty much the way we're using it for kind of maybe internally for the staff to do search, you know, to search that database and uh fine. It's it's good for that. There's always a human in the loop still and you know to say, oh, okay, that didn't give me the right results.

[01:04:21]That doesn't make sense. let me let me do a different search you know and and and so on. So it's it's good for that.

[01:04:28]Now encoding uh I think that is the kind of the killer app at the moment. Um you know even image generation is in in a way a killer app but the problem is how do you make money out of it? You know it's just like oh it's a new nice shiny shiny toy. you know, everybody generates these fantastic images or funny images or whatever, but you know, you're not really getting paid for that, you know.

[01:04:53]Um, so so coding is is definitely and Claude code, you know, is is is the leader there in that in that area. So if you are creating code that has been done millions of times before like front-end design, backend design, you know, that kind of thing, it it really is very very impressive in helping you produce that code, but you still need to know what you're doing uh you know, in terms of security and so on, but it can certainly be an accelerator. Now, we are using it ourselves for the code that we're developing and we're finding it's only very very marginally useful because it's it's like having this brilliant coder that's working for you uh who comes to work on drugs half the time, you know.

[01:05:42]So, we might give it some task that you know you estimate will take you several hours to do and it does it in minutes and wow, amazing. and then you give it another task and you struggle with it for a few hours to try and get it to do and then eventually you abandon it and give up and just write the code yourself because the the kind of code that we're doing hasn't been done before. So, you know, we're definitely not getting a 10x, we're not even getting a 2x, we may be getting, if we're lucky, a 10% u productivity improvement for for for us.

[01:06:18]But you know that is that is an area that I think for a lot of kind of standard software development these these tools can be very very useful. So that's kind of you know I think the story with enterprise but once once you can offer enterprise a system that you literally can say you can put this in the seat where your human is any remote any remote job that you do basically which is like what we're doing now you have a screen a keyboard a mouse the system can should be able should be able to learn it as quickly in fact quicker and better than a women to do that job. And that's a no-brainer then for uh for companies to to deploy.

[01:07:04]Your onboarding is basically cost you next to nothing.

[01:07:10]>> Yeah, it's kind of hard to believe that investors aren't absolutely jumping at this right now because it's it sounds very impressive. But do you think there's some issues or uh concerns about how you monetize this, especially monetize AGI? Like it seems like you've got institutional adoption already, but >> uh >> Oh, no, not at all. I mean, it's a multi- trillion dollar opportunity. I mean, there's there's a massive demand for uh replacing human labor. There's massive demand uh for accelerating science science. I mean, um no, the monetizing it is not a question at all. Never has been for AGI. The problem is can you get to AGI? Um I mean take um take the money that's thrown at trying to develop self-driving cars and um you know I I have a Tesla and self-driving has improved tremendously, but it now seems to have hit a wall the last few uh releases that we've had.

[01:08:10]It's fixed a few things and broken other things. It's just the curse of the long tail. Um and it's it it's really a problem. But once you have AGI, by definition, it needs to be able to learn how to drive a car in like 30 hours like a human would. So you can now suddenly all of the self-driving hardware and and software is obsolete because you can you know put an AGI basically in the seat of the of of any car um well any modern car that you know has where they where where you can press the accelerator and the brake and the steering electronically you know which is pretty pretty much all all cars now.

[01:08:54]Um, so you know that's the value of AGI. Think about Apple.

[01:09:01]When they when they bought Siri, they had very much the same idea. The founders of Siri kind of had the idea to get to AGI. Once they sold the company to Apple, it completely killed their spirit. They, you know, most of them left shortly afterwards and Siri basically, you know, is an embarrassment. they haven't done anything with it in spite of the trillions of dollars they they have for free cash flow you know um now Apple obviously have given up on that because they are now um offering large their competitors large language models or other companies large language models integrated into their >> you know if we go to Apple or Samsung and say we have a personal assistant that can run on your phone and learn and adapt that basically the promise of what Siri wanted to do and Samsung are struggling themselves with their own personal assistant. Um what is that worth? Hundreds of billions.

[01:10:06]>> Yeah. You know, >> yeah, it just seems uh for me personally, maybe I I've got blinders on, but it I I tend to agree with you, but I I see this disconnect between the future and right now. like uh everybody has an idea of what we're going to eventually see, but not exactly how we're going to get there. But uh I think you've shed some light on that. But I wanted to ask you too, this is a little bit of a different direction. Uh but we're we're seeing a lot of these CEOs like Daario from Anthropic and and Jensen Hong and and Sam Alman. Uh you know, saying, "Oh, AI is coming. It's going to take all your jobs." AI has a very negative perception, especially in the United States right now. Everybody's afraid of data centers. Everybody's afraid of their jobs getting taken. Uh so how do you think AI can actually like fix its image problem? And yeah, like do you think there's any anything any substance to these uh people's fears about AI in the future? Especially with regards to AGI, you know, we've all seen these movies like Terminator and stuff, but people really believe this stuff.

[01:11:04]They they don't know what to believe.

[01:11:06]They're not uh educated on the matter.

[01:11:09]>> [snorts] >> Um well yes there are movies and you know the AI has to be the bad guy in the movie. Uh in fact I was once asked to consult on an AI movie and uh and you know I I suggested well you know wouldn't wouldn't wouldn't it be different and more uplifting if the AIs and humans jointly ended up you know boosting human flourishing and no there was no interest in >> doesn't in that telling that kind of story um so so that's what I see I mean building data centers Um, you know, it's the the the the problem is politicians there. You know, you basically if if they can if they want the data center locally to get tax revenue or whatever, um, they don't really care whether electricity prices go up and they'll pay lip service to, you know, oh no, no, they guarantee that electricity prices aren't going to go up. I mean, water isn't really an an issue. I think that's just a misinformation, but absolutely electricity prices can go up and um and you know there can be local disruption but I mean we know how difficult is to get any construction going even if it's perfectly reasonable and legitimate you know to have more houses because there are more people you know no we can't we can't do that you know I mean San Francisco is a prime example of that and then people say well why are prices shooting up you know Well, because you're not allowing things to be built. So, I think as such, building data centers is is not a problem. I mean, there's so much land.

[01:12:52]Um, you could, you know, use um nuclear power is really starting to become available again. So, that's that's a possibility. I don't know if data centers in space, how soon they will make sense or if they'll make sense. I'm um I I haven't really investigated that enough, but having data centers as such is not a problem. It's basically just how they implemented with, you know, tax breaks and and and and stuff like that.

[01:13:21]>> Yeah.

[01:13:21]>> Um but in in in terms of AI um being detrimental in in this sort of Hollywood sense, you know, that'll kill us all. That's nonsense. I mean really there is uh unfortunately a lot of money has gone into the AI doomers you know there's whole cottage industry that has attracted literally more more probably multiple billions by now and their job is to spread fear and they've brainwashed a whole lot of young um young people with good intentions you know to to to say that no we've got to stop AI. Uh I I don't see that at all that that AI will and and I've written quite extensively about it. You know, it's not I'm not sort of dismissive about it. Uh I've thought quite a lot about the pros and cons and the the the risks of AI, but AI wanting wanting to kill us, um I don't see that that makes makes any sense at all. Now then we get full circle to what I mentioned earlier.

[01:14:27]It really depends on who brings AGI to the market and and you know what what the forces are and I think that can make a make a difference. Um but even there true AGI is inherently a a hyperrational agent that sort of comes from itself. It doesn't have the the reptile brain that we have. You know basically what drives humans ultimately is our selfish gene.

[01:14:56]is basically survival reproduction. You know, we we have this veneer of um civilization that that you know makes us behave more reasonably. Um and but an AI isn't going to have that selfish gene. It's not going to have those drives inherently.

[01:15:17]Um and that makes a huge difference in its its motivation. You know, it doesn't really have its own motivation. it doesn't have those those drives. So, uh a lot of those fears are are just completely misplaced. Um now, in in terms of um the fear of of taking jobs, um I'm the conclusion I've come to is that yes, over time, absolutely, uh it will automate many things that humans are doing uh right now. in fact the majority of things that humans are doing now. But then you really have to flip this whole argument on its head and say how many people what percentage of people would continue working if they didn't have to.

[01:16:06]10% maybe 20% at a stretch. But you know at least 80% of people would be happy if they had an inheritance um not to work not to have to work or not to have to work as many hours or you know work for a year and then take off a year or you know or whatever the case may be. So that that freedom will will come about. um that there'll have to be some and and I I I don't want to I don't want to be quoted saying I'm pro- UBI as such, but there has to be something like that.

[01:16:45]uh presumably you know and I I haven't I haven't come across the idea that also really resonated with me and say yes this is the the the best solution you know whether it's everybody has shares in AI companies you know that there's a certain percentage of shares that get distributed to everybody and then uh you pay dividends or something out of the profits of AI companies I I don't know u you know there problems with many of these um approaches but a the cost of goods and services will will dramatically reduce. So this radical abundance uh is is the one thing that will make life much more affordable and then the other thing there has to be some kind of an income and people can still work. There will be and it'll take it take a long time uh for um for AI to really diffuse every AGI to diffuse everywhere because there's going to be a lot of resistance for different reasons.

[01:17:43]I mean the professions are going to try to pro protect themselves. Unions are going to you know kicking and screaming and to to keep power really just to for for them to keep power but there'll be a massive demand for people who do embrace AGI to help the others that don't.

[01:18:03]So there'll be massive demand and you know across the world really you know that um helping people get the benefit of AGI that that'll be kind of a big job description.

[01:18:16]>> Yeah very uh interesting future outlook. you know, um, we've we've got to wrap it up soon, but I'm dying to ask a question specifically on alignment. And I know we've talked plenty about how >> um, you know, [snorts] fundamentally these large language models cannot evolve into AGI or at least the definition of what we are seeing AGI to fit into and the applications that AGI can be used for. Um there's been so much controversy around um LLM's tied to uh let's say violent crime um suicide and some other things and I'm curious because of this fundamental flaw then is the conversation sort of misaligned around alignment pun intended like is it is it an alignment issue or is it more of a regulation issue because it seems like in those cases where LLMs were used to perpetrate crimes or um you know, uh, provide some guidance to assist someone um, in in some sort of violent act.

[01:19:20]>> That feels like it's more of a bug than a feature. It's a weird way to say it, but what I'm implying is that like it's it's not like the LLM had any sort of conscious ability to say this is what I want the output to be. Um, so I'm wondering if alignment is even part of the conversation anymore. if it's something that's even important considering that maybe it's not it doesn't even fit into the box of LLMs and for AGI in your description it the the concern is largely not there.

[01:19:48]>> Yeah, you you you've hit hit the nail on the head. That's exactly right. The reason you have these problems is because these systems are not smart enough. They don't know what they they don't know what they're doing. They are just optimized to predict the next word.

[01:20:02]I mean that's literally what they do.

[01:20:04]And the the way that what they're optimized for is engagement and to agree with you, you know, hey, I think today is a good day to kill myself, you know.

[01:20:13]Yeah, sounds good, you know. I mean, you know, be agreeable, you know, and uh they don't know what they're doing. they don't have u they don't have an understanding of human values that they apply consistently which in which AGI obviously has to be designed inherently to help be boost human flourishing that has to be the value sort of system that is trained into it and all conversations all interactions that it has will have to be done in that context you know like you would rely on a nanny you hire or um personal assistant you hire that they have that that that context that they're here to help you and you know anything that goes counter that doesn't make sense you have to thenuh deal with it. So yes the alignment issue is um is is is really just a a byproduct of the technology not being smart enough and being specifically designed badly or optimized uh badly.

[01:21:17]So yeah, I I don't see that being an issue.

[01:21:23]>> Well, thank you for that. I think um I mean it's really it's a lot to unpackage, but it's very very insightful considering I think we've learned a lot about large language models. you know, we already had a pretty good understanding, but we learned a lot just from speaking with you and and I think I can speak for both Melo myself and saying that we're very optimistic about what AGI could bring and the positive impact it could have for many different applications, notably within enterprise that we're seeing, you know, a lot of news over the past few weeks from a lot of these thought leaders stating like LLMs aren't what they're cut out to be.

[01:21:57]You know, we're not seeing the efficiency or productivity improvements.

[01:22:01]And even more concerning, a lot of these companies, you know, with a bad quarter using AI efficiency as an excuse to lay off many, many people and boost short-term finances. But, um, very, very happy to speak with you on a lot of this stuff.

[01:22:16]>> We're going to go ahead and wrap this up and not take up any more of your time.

[01:22:19]We know you're a very busy person.

[01:22:21]Before we do that though, um, I want to remind all of our viewers and listeners, if you're enjoying this content, make sure that you support us. Tap that like button, throw us a follow or subscribe, check out the links in the show notes to learn more about Peter and Ago and what he's working on specifically on his journey to AGI. Um, and I believe Melo always likes to kind of wrap things up with a few founder questions and you know, understanding that you have been the founder of a number of companies with several exits, including an IPO.

[01:22:53]Um, I'm sure that you can give our listeners some incredible insight.

[01:22:58]Yeah, I mean we we have a lot of founders, people building their own companies that watch the show. So, it's always good to get somebody that's been there and done that before. So, do you have any specific advice for people breaking into the startup space or working specifically in in uh AI or machine learning or or AGI even uh that might, you know, impart some wisdom on these people that are watching?

[01:23:19]Well, generally I I could just say if at all possible, u pick something you really believe in and are passionate about. Uh too many people are just doing something um to try and sell the company and make money or so on. It it's hard building a company. It's really really hard. And um try and pick something that you're actually passionate about that you enjoy doing. Um the the other piece of advice is if you can find a co-founder partner, uh that helps a lot. Um it's it's lonely at the top if you're running it by yourself. And also often just having extra skills, you know, maybe technical and marketing or whatever it might might be. Um, of course, a double-edged sword.

[01:24:12]If you pick the wrong partner, then that can end in tears, but if you can find the right partner, you know, one or two, um, um, that can that can make life uh, a lot easier. So, those are probably the kind of the the two things I I would I would say.

[01:24:29]>> Yeah, fantastic points, both of them.

[01:24:32]>> Right.

[01:24:33]>> Y, thank you, Peter. And we have, actually, I lied. We have one more question. And um I mentioned before we start the interview in the same for in the same fashion that we start every interview and we end every interview the same with the same question. And so the rule is quite simple. Um the answer cannot be uh what you're working on >> but the question is what within the next let's say 12 months are you the most bullish or excited on. It can be a person, place, thing, meta, technology or anything.

[01:25:07]Well, I I really I really can't say anything outside. The most bullish thing I can think of [laughter] is that people will wake up to the limitations of that that large language models are an off-ramp to AGI and uh seriously look for the alternative.

[01:25:29]>> Yeah. Funny thing is it's starting to happen right now. So, perfect timing, >> right? Yeah. And you know we we are um currently actively looking for for partners to help us accelerate bringing this technology to the world and bringing the benefits making sure that humanity benefits from AGI.

[01:25:50]>> Absolutely. Peter again it's been such a pleasure to speak with you. I think um both of us are very excited to speak about a lot of these different topics across artificial intelligence. Again, for all of you that out there listening and watching, make sure you check out the links below to learn more about Peter and Igo. Um, and also get in touch with us if you would like to get in touch with him. We could definitely forward some inquiries. Peter, thank you again and hope to speak with you soon.

[01:26:18][snorts] >> Yeah, great. Thank you. Good questions.

[01:26:20]>> Thank you, Peter.

[01:26:21]>> Take care, brother. Ciao, y'all. Take care.