Camille Morhardt 00:39

Hi, and welcome to today’s episode of InTechnology. I’m your host, Camille Morehardt. And we have an episode on What That Means: artificial general intelligence and cognitive architecture. I’ve got with me today Peter Voss, who is CEO and Chief Scientist of Aigo.ai. And he’s been working on artificial general intelligence for decades; he actually helped coined the term along with two other people for AGI. And his company now, Aigo.ai, is working toward AGI, but in the meantime, it’s also got a commercial chatbot platform, as he calls it a chat bot with a brain, that’s out there being used already by companies like 1-800-Flowers. So welcome to the podcast, Peter.

Peter Voss 01:28

Yes, thanks for having me.

Camille Morhardt 01:29

How did you initially become interested in artificial intelligence or AGI?

Peter Voss 01:35

So I started out as an electronics engineer, started my own company doing electronics. And then I fell in love with software, and my company turned into a software company. We went from the garage to 400 people, entered an IPO. So that was exciting.

But when I left the company, it occurred to me that software really is quite dumb, you know? And I was proud of the software we’ve created. But still, it doesn’t have common sense, it can’t reason, it can’t learn the way you know humans can. So that is really what interested me then to pursue, is to figure out how to write software that can think, learn, and reason the way humans do. So I spent five years studying all different aspects of intelligence to really get a grounding and then actually coined the term AGI together with two other people in 2002. And launched a company to ultimately develop AGI.

Camille Morhardt 02:30

So this is kind of a big claim. So I’m going to pause on it. So you said you and two other people coined the term “artificial general intelligence,” or AGI as we’ll refer to it. How did you coin it?

Peter Voss 02:41

When I finished my research, I was ready to actually start building some prototypes and exploring building and actually achieving the technology. And I came across a few other people who were also interested in pursuing the original dream of AI to build thinking machines. And so we decided to write a book together, and three of us decided on the book title. And we tried, you know, different ideas and so on. And then the three of us, Ben Goertzel, Shane Lake, and myself, we decided on artificial general intelligence. You know, we didn’t know that the term would ultimately take off. And so yeah, that book was written in 2002, was published in, I think, 2005.

Camille Morhardt 03:26

Well, we’ll put a link down, too,  if anybody’s interested in taking a closer look. Let’s start by actually defining AGI. I mean, what is general intelligence versus specific intelligence or intelligence to begin with at all?

Peter Voss 03:39

So the original vision of the founders of the field of AI, when they originally coined the term AI some 60 odd years ago, was to build thinking machines, machines that can think, learn, reason the way humans do. And they actually thought this could be achieved in like a year or two. Now, of course, it turned out to be a much, much harder problem than that. So what happened over the years, and over the decades, the field of AI really turned into a field of narrow AI, where you take one particular problem that you want to solve that seems to require intelligence, and you figure out how to automate that. And so a very good example here is IBM’s Deep Blue, you know, the world chess champion in the 90s. It’s really not that the software has the intelligence to figure out how to play chess. It’s the ingenuity of the engineers to figure out how to use computers and put different algorithms together to play a good game of chess. And they can’t even play a game of checkers. So it’s really the external intelligence, the intelligence of the engineers, or today the data scientist, that achieved these particular narrow objectives.

Now, of course, with ChatGPT and the massive amount of data that it’s trained with, it has a much, much broader range of capabilities. But I’ll come back and we, you know, we can talk about the limitations and why that isn’t really AGI. But really the field of AI turned into narrow AI and as coining the term of AGI was to get back to this ideal to this vision of having thinking machines. And really the most important thing here is a machine that can learn a very broad range of capabilities by itself, that would imply that it could become a scientist, a doctor, researcher, you know, a call center operator, or any of things by simply learning the way we use. So that’s really what artificial general intelligence implies–the ability to have the core intelligence to be able to learn pretty much any skill that humans can, and, of course, many more.

Camille Morhardt 05:52

So one of the differences–well, there’s an infinite number of differences between a human mind and a machine, or I’ll say, computer machine, because people might argue the human mind is a machine, I’ll leave that one for a moment; but the amount of energy that we consume to run our brains is pretty small. To train a computer, or to train a model, you’ve got to have, you know, an incredible amount of compute power and then you have to have an incredible amount of data. Whereas a human can be shown sort of one picture and then kind of understand that picture in a variety of different contexts until some other thing happens that causes you to focus or narrow your perception of something. Whereas a computer might have to have millions of millions and millions of pictures of cats, right, before it really understands. Is that something that computers need to get to before they can achieve AGI–sort of a similar power consumption and similar single shot learning? Or is that not required, they can learn differently?

Peter Voss 06:57

Yes, you actually put your finger right on the issue of why current approaches are not AGI,  and in fact, will not get to AGI. There’s a very common misunderstanding about what intelligence entails or what human intelligence entails. A lot of people make the mistake of believing that having a lot of knowledge makes you intelligent, but that’s not really the case. It’s the ability to learn to conceptualize, to generalize and to think, conceptually. That is really the hallmark of intelligence. So you know, you can have a person that has very, very little knowledge about really anything–children would have that–but they had the ability to learn. And that is really the hallmark of intelligence to be able to learn and to be able to learn largely by yourself, and LLMs can’t do it.

So when you’re saying the current computers have massive requirements of data and computing power and energy, that’s because they tried to brute force it. It’s all about having trillions of pieces of information to try and simulate something that looks like intelligence, but really isn’t intelligent. So a truly intelligent AI wouldn’t have that requirement. And exactly as you say, it could see one picture of a giraffe, the first time they’ve ever seen a giraffe, and would be able to recognize it the same way a child can. So this one short learning was very little data. And it’s not about how much knowledge you have. But it’s the ability to really deeply understand, to reason, and to be able to acquire knowledge. So the kinds of systems that we are working on require a tiny, tiny fraction of training data and computing power and energy. And that really is ultimately what you want.

Camille Morhardt 08:45

Humans have, I don’t know, five or six senses. And obviously, we’re aware of others that we use in machinery. But do you feel like computers are going to need a similar kind of sensory input? I know that there’s arguments about autonomous driving, and whether it can use vision, or it needs, you know, other kinds of technology, in addition to it, but that would just be one example. What kinds of sensors do we need? Obviously, large language models are using, you know, writing and possibly that’ll extend to video, but what other kinds of things?

Peter Voss 09:16

Right, it’s a very interesting question, and I don’t know that we know for sure how little senses you can get away with. But I often talk about my theory of AGI as the “Helen Hawking theory of AGI”–Helen Keller and Stephen Hawking. So, if you take a Helen Keller with very limited sense acuity and you take a Stephen Hawking with very limited dexterity, obviously, they were both very intelligent people. So clearly, you can be an intelligent being with very limited sense acuity and dexterity. And that’s actually the approach we are following, because it’s easier that we don’t have to deal with robotics and all of the complications. So exactly how little sense acuity you can get away with and still have the level of learning ability that you want is not totally clear yet. But it is clear that you don’t need to have the full sense acuity that we have, for example.

Camille Morhardt 10:17

Will machines be learning, then, always from observations that humans have already made? Or, you know, if it was one of the interesting ways people learn is, you know, they experience it themselves. You know, the machine itself as the microscope, or rather than reading something that a human observed, you know, or that I guess, would be my curiosity around why you might need a sensor.

Peter Voss 10:38

Yes, so a commercial system is text only, but our AGI development system has vision, and of course, text input, as well. And I think there’s something like either vision or tactile is required to really internalize a 3D world. I mean, Helen Keller could clearly feel things around her. So I do believe that some form of vision is really essential for grounding the sort of three dimensional/four dimensional experience that we have, or that an intelligent entity needs to have.

It will be able to learn. And what we are doing now is we’re training it through interacting with a virtual world, but also potentially with a real world through, you know, cameras and other inputs. But it can also learn a lot through reading. But there’s a very important difference in how you read. With large language models. their approach is basically to dump gigabytes of PDFs, or whatever, and statistically build a model from that and there really isn’t understanding the way we would read a text. So you know, if we read a technical paper, for example, or even a novel, we will read one sentence and immediately we ingest that sentence and we try to make sense of it, you know?  Is this something new? Is this something that confirms something we already know or does it contradict it ? What does it make you think of? Do you need to go off and do some other research? So we process the information as we go along and this is really how understanding an intelligent learning has to work. And perhaps ask for help, ask for clarification, if you don’t understand that you have a teacher. So that is the way we see an AGI learning by interacting with a world to some extent, but also, by hitting the books.

Camille Morhardt 12:26

I think what you’re talking about is essentially comparing new information or new input with existing patterns of thought or context. So a lot of that seems to me based on memory, for at least humans; I’m retrieving files or understandings of things that I’ve come across before. So how does AGI, or how do I suppose gen AI today, how are they approaching, like, short-term, long-term memory or understanding the context of a situation?

Peter Voss 12:55

Yeah, I’d like to take a little detour and talk about a model that DARPA published a few years ago, and they call it the “Three Waves of AI.” And according to DARPA, the first wave of AI is what we know also called good old-fashioned AI. It’s basically the work that was done in the 70s, 80s, 90s, which is largely logic-based approaches, some statistical, but really mathematical/logical approaches like expert systems, and so on.

And the second wave is basically, hit us like a tsunami about 12 years ago. And that is all about deep learning, machine learning and now generative AI, so statistical systems. And what happened there is obviously some big companies had a lot of data, had a lot of computing power, so it was sort of what can we do with that? In a way, it’s just a hammer, we’ve got, everything looks like a nail; and they managed to do a lot with this big data approach.

Now, the third wave of AI, according to DARPA, really starts with understanding cognition–what cognition requires, what intelligence requires. And memory is one of the requirements. But it’s not just short-term memory and long-term memory. Short-term memory is very important, but it’s really how you use the short-term memory. Is that it forms part of the context of understanding: Who are you talking to? What is the subject you’re talking? Theory of mind, what does the other person already know? What was said earlier? What assumptions? What is the goal that you’re trying to achieve? So all of those things are partly short-term memory, partly called from long-term memory. And that serves as a context for being able to really make sense of what you’re doing and to have the appropriate response in the appropriate thought patterns. So that is really what the cognitive AI approach is.

Now current large language models, whatever memory they have is really external to the model. The model itself is read-only. I mean, if you Just look at GPT, “G “means generative, which means it makes up stuff, which, you know, for better or for worse, it has trillions of pieces of information, good, bad and ugly. And it basically statistically generates things from them. So that’s the generative part. But the “P” stands for pre-trained; it’s trained at the factory, $400 million dollars or more. And the model itself is read-only–it does not change as it comes across new information or whatever. The new information that you have is either stored in an external database that is indexed, but it’s still external, it doesn’t change the model. Or it’s part of the context window, which means its short-term memory that is kind of fit in to the model as you do inference. But again, it’s not integrated. And there are also severe limitations with how much information you can actually put into this short-term memory. So that’s really a very fundamental limitation of large language models that the model itself cannot change, cannot learn in real time.

Camille Morhardt 16:00

Is that in perpetuity? Or is the assumption that it will be doing that in the future, but that we’re kind of in early generations right now?

Peter Voss 16:08

Uh, I didn’t get to the third letter of GPT. I mean, the “P” already tells you it pre-trained; but the “T” is it’s transformer technology. So as long as we using transformer technology–and that seems to be pretty deeply embedded in the whole LLM scene, I mean, people are now building chips based on transformer technology. So it’s getting pretty much baked into the approach. And transformer technology inherently requires some kind of back-propagation, some bulk training, and it cannot learn in real time; there really aren’t the algorithms that allow you to update the model. You know, there’s a little bit of fine tuning you can do at the output layer, and so on. But really, the core of the model cannot be updated. In fact, if somebody could figure out how you can train a system in bulk, and then later update it in real time, the model, that would change my mind about the path we might take to AGI.

I think there are some other problems with large language models with GPT that also prevent it from getting to AGI and that has to do with the way concepts are formed. So statistically forming concept is not really the same way that humans conceptualize. I’ll give a simple example, which might not exactly explain it. But if you look at word-to-vec encoding, vector encoding, you find that “dog” and “cat” are closer in vector space than “dog” and “puppy.” And you know, for us, “dog” and “puppy” seem to be closer related than “dog” and “cat.” But because in text, you’re far more frequently find “dog” and “cat” close together or mentioned together. So it’s the way the vectors are encoded in these large language models. So I think there are really some fundamental limitations that are not fixable, that will prevent large language models from getting to AGI.

Camille Morhardt 18:08

How does AGI achieve what I guess humans would commonly call critical thinking?–Which is kind of tying it back to the LLMs because hallucinations, you know, how do you know or how does it know or how does whatever is interpreting it know that it’s gone off the rails? How are you finding the truth, if the truth is what you want?

Peter Voss 18:28

You know, one of the unique human features is that we have metacognition–we know what we’re talking about, we know what we’re thinking well is to some degree, not very deeply. In fact, there’s an interesting insight I got while I was studying intelligence. I actually spent quite a bit of time on a project developing a new, not exactly an IQ test but a cognitive profile test, was actually in South Africa, that doesn’t have cultural bias, which is really important in South Africa–that you could have a test that didn’t have the bias of Western backgrounds. Long story short, one of the things I learned from that exercise, I ended up computerizing this test, and I learned that metacognition is actually the single most important dimension to predict general intelligence. So when people are good at automatically using the right mental technique that matches the problem, well, they tend to be generally intelligent. So metacognition is essential.

Now Daniel Kahneman talks about System 1 Thinking and System 2 Thinking and that’s also very much related. System 1 thinking is sort of roughly speaking, most of our mental activity is just subconscious and automatic. But we have this supervisory function–this System 2 function–that kind of monitors what we are thinking, what we’re saying, and it can then redirect and correct us. And LLM simply don’t have that, whereas a cognitive AI approach clearly will need that and will have that higher level thinking, that reasoning.

Camille Morhardt 20:01

Can you say a little bit more about the metacognition and what the using the right tech- just what humans are doing when you’re using the right technique to say, I’m going to solve this problem now, how am I going to think about it? Because that’s not with a supervisor, that’s in your head? That’s just—

Peter Voss 20:17

Correct. Yeah. I mean, those are thinking skills. For humans, we actually are not particularly good at rational thinking, it’s sort of an evolutionary afterthought. For an AI, it will actually be able to think much more rationally, but it has to have the mechanisms for doing that. So these are the mental skills that have to be learned by the system, and you know, that’s part of the curriculum, the training that we do, where we teach the system, how to think critically, and how to avoid logical fallacies and things like that. So it’s really that you’re training a system to think well.

I do want to comment on something else you mentioned, that is can LLM not help with AGI? Absolutely. In fact, a few years ago, one of our biggest concerns in getting to AGI was always, how do we get the massive amount of common sense knowledge that you need to function in the real world? And to understand the real world? Where do you get that?  You know, it would take 1000s of people years to teach AGI all of these things. I believe now that we have LLMs, we can actually use LLMs to extract information as long as you have that supervisory function that basically makes sure, you know, you can cross check different LLMs, you can ask things in a different way, you could potentially also double check things on Wikipedia where appropriate, or wherever. So I think that makes the problem of training an AGI, giving a common sense knowledge that it needs, a lot more tractable by having them.

And AGI will inherently be a terrific tool user. So you know if my AGI, if I wanted to write a poem, in a certain style, I would probably could just ask ChatGPT to do it, you know?  It doesn’t have to have that ability by itself, it’s not going to have the massive amount of knowledge that is embodied in a large language model, but it will be an excellent tool user.

Peter Voss 22:30

Yes, certainly. So the company really has two divisions. The one is the AGI division, and that’s a part we’ve really been talking about is to continually get closer and closer to human level AI. And we are currently expanding that to accelerate development. But the other part of our business is existing commercial business where we have what we call “a chatbot with a brain.”

So, you know, most chatbots are not that great, and people get frustrated by them. So ours uses a cognitive architecture. And it has short-term, long-term memory, the ability to reason and so on. So we are targeting large call centers with that to really replace call center operators, because call centers have a really terrible time trying to get staff to try to staff and keep stuff. So for example, one of our big customers is 1-800-Flowers group of companies–Harry and David and Popcorn Factory, and some others. Now with our automation, we saved them 1000s of call center operators, and people didn’t have to wait–they got immediate, hyper personalized service. So that is where our technology is already being deployed right now is in call center automation. But we also are really looking at other personal assistant type of applications on the commercial side.

Camille Morhardt 23:49

And so with the hyper personalized call center, now I’m giving personal information to AI. So what kind of protections should companies look at having in place for that?

Peter Voss 24:00

So that’s the beauty with the approach of cognitive AI is, you don’t have this huge statistical model that everybody’s information gets combined. With our approach, each individual customer has their own profile that’s stored completely separately. So what you learn during a conversation with a customer that may be helpful in future conversations, like “I’m buying chocolates for my niece for her birthday,” and maybe next year, you know, you can suggest that chocolates or there might be buying something for your niece. So that information will be kept per individual. And individuals can opt out of that, but it’s not combined with other people. So that, again, is an advantage of a system that isn’t just a statistical combination of things. So it’s really hyper personalization, at scale, for 20 million people in the case of you know, 1-800-Flowers.

Camille Morhardt 24:57

So are these two pieces of information kind of bifurcating–like on the one sense, companies that can collect information will say about what you’re saying, “niece and flowers,” and then cross check all that data and say, well, it turns out that like lots of people are buying for their nieces flower, you know, who knew that chocolates were being bought for nieces? Do we need—? You know—“let’s do a promo” versus customized where data is staying in the silo? How are those two things–

Peter Voss 25:26

Right? Well, you know, that’s really a policy for each company to decide what they want to do. But obviously, they have the capability of aggregating this data, hopefully, you know, in a totally anonymized way. But it can also be used for the individual if they opt-in–you know, for, for example, for reminders or so.  The policies are really up to the individual company to decide how they want to use the data. But the technology allows them to really completely separate it, and for requirements of European Union, for example, that’s fantastic because you literally can go in and say “yes, I’m deleting All of the history for this one customer who wants to opt out.”  It can easily be done.

Camille Morhardt 26:09

What do you think some of the biggest threats are? And I’m thinking cyber threats–I’m not thinking gray goo, but you could take us in that direction if you want–as humanity gets closer, I’ll assume we’ll get closer to AGI?  What do we need to really be concerned about?

Peter Voss 26:26

So I tend to believe that a lot of the problems we have with technology and with nascent AI is that it isn’t intelligent enough. And that I think increased intelligence will actually make it safer. So I’m not really that concerned about risk from AI directly. But of course, humanity dealing with this radical abundance that AGI will create, will be a large disruption. And, again, I think AGI will help because my vision is that everybody in the world will be able to have their own personal assistant that you own, that serves your agenda and not some mega corporation’s agenda. It’s hyper personalized to you. And you control what it shares with whom. And that will be like the friend and helper but also potentially a psychologist that can help you deal with this wealth. You know, I mean, we’ve all read stories where people who win the lottery often can’t cope with it that well, you know, or who don’t need to work or can’t work anymore, because maybe personality and that is so tied up in their work. So, most people you talk to say, if you ask them, “Do you want to win the lottery?” And they say, “Of course! I would love to win the lottery,” you know? And that’s really what AGI promises in a way is that everybody, it’s like everybody winning the lottery that you will not have to work or very little in order to afford a very luxurious lifestyle, because AI will bring down the cost of goods and services tremendously.

Camille Morhardt 28:05

And then I expect that brings up just the human concern of the increased discrepancy then between people who have access to that personal assistant, and people who don’t.  What is your take on how we humans move forward in that space? Is it just policy or—

Peter Voss 28:23

Yeah, I mean, what my own view is–and I hope I have some part to play in this or that other people will come to the same conclusion is–that everybody in the world should have access to that. And I think there are some promising signs. I mean, everybody has access to ChatGPT.  Well, I mean, people who have, obviously have, at least have internet, but that’s very rapidly also, pretty much everybody in the world. Presumably there will still be a premium pricing model or something if you want to use it for research. But I think that a basic personal assistant like that will be available to everybody.

Camille Morhardt 29:02

Would that be centralized computing, then?  You’re saying that compute wouldn’t have to be so huge. And right now I know with LLMs that’s part of the situation is you’ve got massively centralized compute power that one human is probably not going to afford or have space for. So are we going to have distributed compute with AGI, then?

Peter Voss 25:32

Oh, yes, I mean, currently, the systems that we are building will run on a five year-old laptop. Now, of course, as we crank up the intelligence, that will require more computing power, but computing power constantly increases, and the prices are dropping. So I see this personal assistant being able to run quite happily on a smartphone-type device. And no, I don’t see it at all as a very centralized kind of service. In fact, that’s not desirable, from my point of view, is that really, anybody should have access to that core technology.

Camille Morhardt 29:59

Well, Peter Voss, thank you very much for the conversation on artificial general intelligence and cognitive architecture. I appreciate your time.

Peter Voss 30:07

Well, thank you.

The views and opinions expressed are those of the guests and author and do not necessarily reflect the official policy or position of Intel Corporation.

Intel technologies may require enabled hardware, software or service activation. // No product or component can be absolutely secure. // Your costs and results may vary. // Performance varies by use, configuration and other factors. // See our complete legal Notices and Disclaimers. // Intel is committed to respecting human rights and avoiding complicity in human rights abuses. See Intel’s Global Human Rights Principles. Intel’s products and software are intended only to be used in applications that do not cause or contribute to a violation of an internationally recognized human right.