#141 Peter Voss Reveals the Future of Artificial General Intelligence

0:00You want some grounding, you want grounding in senses and also you really do need for general intelligence, you do need some kind of a vision. It doesn't need to have the same, you know, sense acuity as humans necessarily, but you do need vision, I think, to have a sense of 2D, 3D world and also action movement. I think there'll be a time when it becomes really, really obvious to pretty much anybody who works with a system saying, wow, this is the real thing. It really can learn by itself, you know, and it still needs a bit more. How long will it take us to get there? Well, I usually answer that question.

0:33It's not a matter of time. It's a matter of money. Hi, I'm Craig Smith, and this is Eye on AI. This week, I talked to Peter Voss, a longtime researcher in the field of artificial general intelligence. He was one of the authors of a book titled Artificial General Intelligence, which helped popularize the term. And Peter has a company called iGo.ai, A-I-G-O.ai, which builds intelligent chatbots for industry. But Peter's real ambition is to build artificial general intelligence using cognitive architectures and graph knowledge bases, a strategy very different from the pre-trained transformer architecture is in use today and which many commentators have argued will never get us to AGI.

1:38I hope you find this conversation as fascinating as I did. This episode is sponsored by DataStax, the real-time AI company. With DataStax, any enterprise can mobilize real-time data and quickly build smart, high-growth applications at unlimited scale on any cloud. Companies building real-time generative AI apps can leverage the data stack's vector search capabilities to build LLMs, AI assistants, and more. Sign up now at the link in our show notes. I'm Peter Voss, I'm CEO and Chief Scientist of iGo.ai. and I got into AI, actually it's quite a long journey. I started off as an electronics engineer, started my own electronics company, then fell in love with software and my company turned into a software company.

2:38I developed a comprehensive ERP software package and that company was quite successful. We went from the garage to 400 people and did an IPO. So that was great fun and learned a lot, made a lot of mistakes and made some money and it when i exited that company it really occurred to me what project do i want to tackle now and it's you know it was obvious that software is not very smart by itself that it makes mistakes if a programmer doesn't think of something it'll just crash or give you an error message or something so that software really doesn't have common sense it can't reason so that's what i embarked on more than 20 years ago to figure out how to bring intelligence to software.

3:24So I took her five years to study all sorts of aspects, all different aspects of intelligence, starting with epistemology, theory of knowledge. How do we know anything? What is reality? How can we be sure of things? And, you know, things of that nature. And then also about cognition is what do IQ tests measure? How do children learn? How does our intelligence differ from animals? and things like that to really deeply understand what intelligence is all about. I obviously also studied work that had already been done in the field of AI. And then by 2001, I was ready to start my own AI company. And we spent several years basically doing R &D, taking the ideas that I'd formulated and turning them into actual code, into an actual framework, which we subsequently commercialized in a company called Smart Action to automate phone calls intelligently, basically with a brain.

4:27And my current company is basically the second generation of this technology. We took some more time to crank up the IQ of the system and current company, iGo AI, we call it a chatbot with a brain as opposed to all of the other chatbots out there that really don't have a brain. And that's where I am today. And really the mission I'm on is to develop AGI, to develop full human level general intelligence. Yeah. And we spoke last time, IGO is based on knowledge graphs. It's not a neural net or a deep learning model. Is that right? And just when you talk about a brain, a chatbot with a brain, the brain, you're really referring to the knowledge graph database.

5:23Is that right? Well, the knowledge graph is only one part of it. It's a very important part of it. But it's really the substrate of our brain. It's how short-term and long-term memory are encoded, you know, that all of the knowledge and skills are encoded. I look at it as much as a neural network, but in the sort of original traditional sense, as well as a knowledge graph. So it's a combination, you know, it's nodes and links that basically encode relationships between concepts. It's very, very different from what almost everybody else in the field of AI is working on. So what most people are working on is statistical AI or more recently generative AI, which is basically big data, big compute, you know, number crunching, building these large read-only models.

6:17And that's really where, you know, all the visible progress has been made and all the money has been made and all the oxygen has been sucked out of the air with that approach, basically a statistical big data approach. Whereas my approach, our approach has really from day one been cognitive AI. So the starting point is what does intelligence require? So the knowledge graph that is, as I say, the sort of substrate of our brain just encodes the knowledge and skills that it has. But the various algorithms and capabilities that the system has as deep understanding, deep parsing, context, reasoning, short-term memory, long-term memory, language generation, all the various skills that the system has.

7:08They basically work either part of the skill set that is in the knowledge graph or they are additional algorithms that work on the knowledge graph. So the brain itself really is, it's a cognitive architecture that has all of the components required for cognition. Yeah. On cognitive architectures, maybe you can talk a little bit about that, what they are and what some of the popular architectures are and what architecture you use. Yes. So cognitive architectures have actually been around for quite a long time, for 30 or more years. And a lot of them are based, well, they're different camps. Some of them are based on really logic programming, on logic inference and having logic as their base.

8:05Others really started off with biology and saying sort of neural networks and, you know, how does the brain work and can we simulate that in a cognitive architecture? But the common denominator is basically the cognitive architecture tries to embody all of the requirements for cognition. so it's typically you know input there's input processing and then there's output and that that loop is is closed and then there might be some higher level reasoning uh involved now you know when when i mentioned cognitive architectures people often say well we've tried that for 30 years and it hasn't worked you know but of course go back 10 or 15 years and people would have said exactly the same thing about neural networks you know hey we've tried neural networks for 30 years and they haven't worked well sometimes things don't work until they do and there are actually several reasons I think why cognitive architectures haven't come to the forefront one of one of them already mentioned is that the success of statistical big data approaches has you know it has been so successful that it sucked all of the oxygen out of the air and basically if you want to get something funded it's got to be you know statistical AI if if you want to write get a PhD it's that if you want to earn big bucks that's really what the fields you have to work in so I think cognitive architectures have suffered from from that but also they have been I believe two sort of technical aspects that we've overcome that I think have hampered them.

9:49And one of them is the knowledge representation itself of, you know, what we refer to the knowledge graph or knowledge representation, really has to be of super high performance. And we benchmarked our knowledge graph against graph databases, state of the art graph databases, and we are literally three orders of magnitude faster, a thousand times faster. So something that'll take one second of our system to compute and respond to, you know, involving a knowledge graph would take a thousand seconds. And of course, that doesn't make it viable at all, you know, for an interactive system. So, and related to that is that the various components that I mentioned, like deep parsing, understanding, short-term memory, context, reasoning, and all of those need to be deeply integrated with each other and with the knowledge graph so we developed all of these components ourselves from the ground up whereas traditional many of the traditional cognitive architectures have a very modular approach but you know which is typically a good engineering approach so you say hey we need a parser the stanford parser that's good parser we'll use the stanford parser we need a knowledge graph we'll use some knowledge graph and we need The reasoning engine will use somebody's reasoning engine.

11:13But the overall system really suffers from that because cognition really requires these systems to work together. When you hear a sentence, it might require short-term memory to say, you know, say, Bob's not coming along. Okay, did we talk about Bob or something related to Bob recently? No. Long-term memory, do we know what Bobs do we know? Reasoning, what kind of Bob would make sense for the context we're in? Well, it might be the family dog that's not coming along to the picnic next weekend, you know. So you really need to have this deep integration of all of the different components for cognitive architecture, I think, to work effectively.

11:53We've actually spoken to, you know, R &D teams at, you know, various large companies at Microsoft and Oracle. And so where they've tried to integrate graph databases into their language systems, and they just couldn't get them to work because of this performance and integration issue. That's interesting. Yeah, so that's IGO. you start with a model that you've pre-trained on a certain amount of data that gives it an understanding of basic relationships and language. Is that right? And then the company or person using it, as they use it, they add to the knowledge base. or to the knowledge graph through through their discussion or their questions and answers yes absolutely so i can probably best illustrate that you know with one of our customers because they are really logically three layers physically they they all all the same but logically they separate into three layers um you know one of our big customers is 1-800-Flowers a group of companies Harry and David and Popcorn Factory and so on.

13:21And they wanted a hyper-personalized concierge type personal assistant, you know, for each of their 20 million customers potentially. So we have the core knowledge base of our system that basically knows about people and places and how to start a conversation and, you know, how to reason and basically to handle language, sort of a core competency um you know you could equate that taking somebody sort of just starting college or something of course it doesn't have the broader knowledge base but you know converse conversational knowledge and then um for each particular company we then add the next knowledge layer the ontology or knowledge layer required for that company you know what are the products for that company what are their business rules and the integration to apis to their backend system is also handled at that layer where you can dynamically get, you know, the customer's order status or their history or whatever, which may be processed on another system.

14:26So the knowledge graph can basically integrate dynamically for company specific things. And then the third layer is per individual customer of what you learn in interacting with a customer. They might tell you, you know, I want to buy a present for my sister's anniversary. And then the system will remember that for that customer and potentially be able to utilize it, you know, if there's a follow-up question or next year or whatever the case may be. So it's basically the core knowledge and skills that are available to all users of our system. Then the company-specific knowledge APIs and capabilities.

15:12And then the third layer is for each individual user, what unique information you learn about them. Yeah. How large can – so the core model doesn't grow in size, but the knowledge graph grows in size. Correct. and that's housed in a graph database. How large can that get? Is there a limit? Yeah. So, we, you know, currently all of the applications we do are text-based, so they're not imaged, they're no images. And you can really have a lifetime worth of conversation pretty much, you know, in a few megabytes. I mean, it's really not a problem. You know, we currently, I mean, compared to large language modules, it's completely trivial, you know, whereas they're talking trillion parameters.

16:12We're talking maximum a few million. You know, we simply don't need more than that for these particular applications. Now, for a more general AGI, of course, that will become much bigger. It will be, you know, in the tens of millions, hundreds of millions range. Yeah. Well, how do you get then from this to AGI? I mean, it's not simply a matter of scale. You mentioned this right now is text-based. You know, we talked last time about Jan LeCun's JEPA architecture, And he's right now, he's working in still images, but with an expectation to move on to video where this architecture will build a model of the world based on its input, unstructured data input, no labeling required or anything like that.

17:22And to me, his argument against the pre-trained transformer architectures is that the knowledge that they're dealing with is what's contained in human knowledge, in language, but written human language. Human language contains a lot of, you know, lies and falsehoods and, you know, rhetorical devices and things. so it'll never understand the underlying reality whereas his architecture is learning from in much the way as a human would from what it's experiencing and it sounds like your cognitive architecture in IGO has that potential I mean can you talk about are there parallels with what Lacoon is doing and what you're doing?

18:24Yes, absolutely. In fact, our early development, our early prototypes that we built had multiple sensors. In fact, we originally started off with a virtual critter, a virtual mouse in a virtual environment that had whiskers and ears and could smell things and virtual smell and all that. So I think it's very important. So I would definitely agree with that sentiment that you want some grounding. You want grounding in senses. And also you really do need for general intelligence, you do need some kind of a vision. It doesn't need to have the same sense acuity as humans necessarily. But you do need vision, I think, to have a sense of 2D, 3D world and also action movement.

19:15so the current project that we're working on that is directly aimed at going all the way to human level intelligence is a version of our technology that again has vision included in it because I do believe that's important as you know mouse you can manipulate a mouse and keyboard and see you know take in a desk whatever is shown on a desktop so I would agree with that a for the grounding and as you say for really the tie to reality the grounding. Now I'm not totally up to date with his approach and I'm not sure if he's published everything but typically what everybody else is doing is basically bulk training and I think one of the keys of intelligence is it has to absolutely be real-time incremental it has to be able to learn real-time incremental and sort of much of it has to be one-shot learning you know for example you can show a child a photograph a single photograph of an elephant for the first time they've never seen one before and they'll be able to recognize you know pink elephants and upside down elephants and elephants in the wild and so on and the systems really need to be capable of doing it and all of the the big data efforts right now are talking about plowing more data into training and it's all bulk training you know that's why they cost spend a hundred million dollars or more and have to use chips that cost $30 ,000 per GPU chip or training chip.

21:07So I think that's an important element of being able to learn incrementally in real time. Yeah. And you mentioned earlier in how condensed this data is, if only contained in language, And I had heard, and I'm trying to think whether it was you or someone had mentioned to me that they had spoken to someone who's developing a watch, a device that would run 24-7. Was it you that told me about this? No, it doesn't ring a bell, no. Right. And so you would wear it throughout your life and it would continually pick up data and encode it into a model so that over time you would end up with a virtual twin.

22:10I've heard you talk about something similar with Igo. Yes. Yeah. yeah it's actually um i i had some friends that oh maybe as long as 15 years ago were doing life logging where they basically have a camera um you know permanently attached recording everything and um but that seems to have fizzled you know i think one of the problems is also is how do you ever get back to it it's like people taking photographs of their their lunch and dinner and all sorts of things, you know, and they have tens of thousands of photographs. Do you ever look at them again? So, yes, I think what we are doing, one of the aims is what we call a personal, personal assistant.

23:01And as I mentioned, it would really be called a personal, personal, personal assistant because three different aspects of the word personal, you know, are really meaningful here. The first personal is that you own it. It's your property. You control it. It serves your purpose and not some mega corporations. It's not like Siri or Alexa. Clearly, you get them for free, but they don't serve your purpose first and foremost. So that's the first thing, that you own it. And the second personal is it's hyper-personalized to you. It gets to know your history, your likes, what type of things you buy, who your friends are, and everything like that.

23:47So it can do things for you. And the third personal is that the issue of privacy, that you decide what it shares with whom. So certain things that you may share with your spouse, other things with your co-workers, and some other things you share with Amazon. So that's a personal assistant. And so I don't see it so much as a twin recording everything you do, but more in an extension or an exocortex eventually that, you know, you can rely on it so much to give you advice to do things for you. It's like an expansion of your own mind, of your own cortex. Yeah. And how do you do that if you're this personal, personal, personal assistant?

24:33How would you gather that data? Is it that, you know, someone would spend a certain amount of time every day, you know, through natural language, just inputting data? Or would it be, as I described, some device that captures data as you move around? well it could be any number of ways i mean we are talking about something that really is at human level capability so it if you wanted to it could you know read your email uh past past email go through it and you know get to know things about you or anything else that you may have that that it can find out but to a large extent it would be through interaction and things you ask it to do for you that it would pick up and ask for clarification, you know, do you always want to fly with this airline?

25:29You know, is that something you prefer? Or, you know, what kind of hotels do you want to stay in? You know, and whatever it might be that it would learn through interaction, but as I say, potentially could, you know, read all of your Twitter or Facebook or email and gather more information from that. Yeah. But again, is that a model that you're working on? And that would be, we're not talking about AGI there. We're talking about just a really useful tool. No, I think we are talking about AGI there. I mean, there may be early versions of it which, you know, wouldn't qualify. But the work that we're doing right now is to expand our system in the level of, in a number of levels.

26:23A, in the sort of adaptability, meaning what it can learn. Obviously, the capacity of how much it can learn, we are expanding, but that's more of an engineering challenge. And we're expanding the ability to reason. And, you know, for example, theory of mind, our current system can't really reason about what the other person is thinking. So, you know, that's something we need to expand the capability of. It's higher levels of abstract thinking. So it's really the cognitive abilities need to be cranked up. And then we need to teach our system a lot more general knowledge. at the moment, it's not a lot because, you know, for the applications we do, it doesn't need to know about, you know, baby showers and different sporting events and traveling the world and stuff like that.

27:19It simply doesn't need to know that. But once you get to a personal assistant, you really want it to have that kind of general knowledge, that knowledge about the world, about things. So it's really scaling up the knowledge, scaling up the ability to learn autonomously at the moment, there's still quite a bit of human in the loop for when the system learns. So that will decrease as its cognitive ability cranks up. But it's really the same architecture that we are using currently. It's just a matter of expanding it to be more general, more adaptive, and more autonomous. That's fascinating.

28:06I'm sorry, I'm just seeing my batteries low.

28:16I'll go for a few more minutes, and then I'm going to have to run down and get my cord.

28:27So it's the same architecture. How do you encode that knowledge? if you're I mean there's a certain amount of label data at the very beginning of what is a verb, what is a noun and that sort of thing. But as you move into increasingly abstract space, how do you encode that knowledge? Yeah, so it's really the The important thing is that the system needs to be able to learn this similar to the way a human does. So it will learn about different instances, for example, whether it's examples of people or places or activities. And then the cognitive ability is able to do concept formation, is able to conceptualize this automatically and say, well these things fit into this category and other things fit into a different category and part of that is where we will give it a curriculum to encourage that kind of learning similar to the way you would you would a human to basically feed it information give it examples give it exercises to be able to build up this knowledge structure layer by layer so the more fundamental things are, you know, really solidly embedded, you know, like time and space, for example, you know, and comparisons of bigger and smaller and all of the different kinds of relationships.

30:06And then as it gets better at being able to learn autonomously, then it's basically a matter of hitting external data sources, like Wikipedia is an obvious one. And, you know, we've already done quite a bit of work in that area but this is also where large language models can actually help us where we can you know extract information from them as long as you have a system that's smart enough to be able to double check the information and see if it makes sense if it makes sense to integrate it yeah would would would you be able to avoid uh the bias that exists in in, you know, human databases like on the internet generally.

30:57Yes, I mean, ultimately, you know, biases like that can, the best remedy for biases, handling biases like it, is ultimately rationality that you say, is this really a relevant factor? You know, is race a relevant factor, you know, or, you know, whatever it might be, is the background irrelevant factor to achieve a particular objective. And, you know, humans aren't that good at it, but basically the more rational you are or the more rationality you apply to it, the better you'll be at basically eliminating bias in your decision-making. Yeah. And that idea of combining large language models with this cognitive architecture approach I mean, do you think AGI, not everybody, but there are several different initiatives to try and get there.

31:58And you wrote a big chunk of the original artificial general intelligence book with a number of other, I mean, Ben Gertzel and Shane Legge. They've gone off in their own directions working on AGI. Do you think ultimately all of these different threads or the most promising will come back and intertwine? Or do you think there is only one path to AGI and one of you guys will find it? Or do you think that there are several paths and there will be different flavors of general intelligence? Yes. So I'm quite certain that it has to be something along the lines of cognitive AI, though there are different ways of achieving it.

32:52For example, one could start with robotics, you know, a much more grounded, embedded approach that the system learns through that. It certainly has its advantages, but also has big disadvantages of having to deal with the robotic aspects of it. um so I think there are different ways of getting there but you know even Sam Altman and Demis of deep mind have said quite clearly that large language models are not the way to AGI they're a dead end yeah so and but they don't really have an alternative you know it's more at the moment well let's build bigger systems let's kind of see what happens you know but on the one hand they recognize that it's a dead end but on the other hand they don't seem to really know how to solve the problem and you know the the biggest impediment by far well there's several but i would say the biggest one is that they are pre-trained you know gpt they're generative they make up stuff which is not a good thing uh rather than to reason about things and and have a solid knowledge base.

34:00Uh, they are P pre-trained, which means they cannot really learn, uh, once the model is built. So you to build a new model, you're talking about a hundred million dollars and I don't know how many weeks, um, and additional training doesn't really work well. I mean, you have catastrophic forgetting, um, the buffer is not integrated to context buffer is not integrated. So the fact that they're pre-trained, uh, they're bulk trained, is really a killer. You know, I mean, imagine hiring somebody as an assistant and they come to you and they can do Excel and they can do QuickBooks and, you know, they know a bunch of things.

34:41Okay, they sometimes make some really bad mistakes, you know, but let's ignore that for now. But so you tell them, you know, we're just taking on some new products. I've got a new partnership here. We're shutting down one of our branches and, you know, a few other things happen in our business. next day they come in they don't remember a thing about that you know not that's not intelligence so um you know that problem really has you know cannot be addressed by building bigger models and the gpt the t the transformer um really locks them into this but the problem is transformers have been so incredibly successful in so many areas.

35:25I mean, what GPT can do is amazing. You know, it really is amazing. So they're kind of locked into by their own success and they really need to somehow tear themselves away. And the other issue why I say cognitive AI is really the way to go and why I think the current leaders in the field with statistical AI are unlikely to actually achieve AGI is because their backgrounds are inherently statistics. They're mathematicians, statisticians, logicians. You know, that's the way they see intelligence. They basically see, hey, we've got all this data. We've got this computing power. We've got computers that can do certain things.

36:13What can we do with that? How can we use our human intelligence to make these things do amazing things like become world chess champions or go champions or do protein folding or whatever? But these are all really narrow AIs that rely on the ongoing external intelligence of humans to make them work. So you really need to start with cognition, with cognitive understanding intelligence and say, we absolutely need these. These are the following requirements to build an intelligent system. And if your system doesn't have that, it's not going to ultimately be intelligent. Do you follow how Shane Legg and Ben Goertzel are pursuing AGI?

37:04yes and somewhat I was just at the AGI conference the annual AGI conference in Stockholm and you know caught up with with Ben his current focus seems to be more on a sort of society of mind not in the Marvin Minsky in sense or at least not in my opinion but that a lot of narrow AIs is, you know, a society of narrow AIs will give you AGI, which I disagree with. I don't believe that's the right path, but that's the path he seems to be pursuing. Shane Legg has, you know, has actually not published much over the last 10 years that I'm aware of, but he recently did give an interview and published something that seems to be relying very much on reinforcement learning, which to me does not, you know, it's kind of the opposite of learning with, you know, one-shot learning, basically.