The Defense Advanced Research Projects Agency (DARPA) has divided AI’s evolution into 3 distinct waves. Currently, we find ourselves in the 2nd wave, dominated by machine learning and big data. The 3rd wave however, is nearly upon us, and will allow AI to go from learning and perceiving to reasoning and possibly even generalizing. The ability to generalize is AI’s holy grail. Though rarely mentioned by its name – Artificial General Intelligence (AGI) – it’s often depicted in SciFi movies and books. DARPA predicts that next-generation methods will be required in order to achieve AGI.
Peter Voss is the man who coined the term AGI and is one of the field’s foremost thought leaders. Peter joins us on this episode to discuss cognitive architecture, a theory of computational structure he advocates for, and which he believes is our best path to an AGI future. We chat about a number of interesting subjects, and along the way learn why the very nature of cognitive architecture may eliminate the problem of bias in AI, why conversational AI is the killer app for cognitive architecture, and why the Turing Test isn’t very useful for appraising a machine’s intelligence.
Guy Nadivi: Welcome everyone. My name is Guy Nadivi, and I’m the host of Intelligent Automation Radio. Our guest on today’s episode is Peter Voss, founder, CEO, and chief scientist Aigo.ai. Peter is one of the world’s luminaries in the field of AI, whose stated mission in life is, “studying and understanding all aspects of intelligence, and actually creating an AI system with general intelligence that can learn, think, understand, and reason more like the way we do.” Peter believes that something called cognitive architectures is the true path to artificial general intelligence. And we’re going to be speaking with him about that today. Peter, welcome to Intelligent Automation Radio.
Peter Voss: Thanks for having me, Guy.
Guy Nadivi: Peter, you started your professional career as an electrician, a number of years ago in South Africa. Please tell us a bit about how you ended up in the field of artificial intelligence. .
Peter Voss: Yeah, certainly. One of my first jobs was as an auto electrician, then I progressed to electronics engineer. I started my own electronics company. Then I fell in love with software and my company turned into a software company. I’ve developed several technology platforms, including programming language and a database system, and also an ERP software system. That became quite successful, company grew very rapidly. We actually did an IPO, so that was super exciting. But that also allowed me, when I exited the company, to take off five years to study intelligence because what struck me, as proud as I was off the software that I developed, software today doesn’t really have any intelligence. If the programmer didn’t think of some particular scenario, it will just give you an error message or do something silly. So I really wanted to figure out how we can make software more intelligent. .
Peter Voss: I had the opportunity to actually take off five years to study intelligence from many different angles, from starting with philosophy, epistemology, how do we know anything? What is reality? How do we know it? What is certainty? Then from cognitive psychology, what are IQ tests? What do they measure? How do children learn? How does our intelligence differ from animal intelligence? And those kinds of questions. And then of course also finding out what had already been done in the field of artificial intelligence. .
Peter Voss: Over that five-year period, I came up with a design for a cognitive engine, a sort of a thinking machine. And then in 2001, I launched my first AI company, hired about 12 people, and for several years we were just in R&D mode, basically turning my ideas into actual prototypes and code. So that’s sort of my journey. By 2008, then actually had our first commercial product. And since then, I’ve basically been switching my time between getting commercial systems off the ground as a CEO, and also continuing to increase the intelligence of our system. .
Guy Nadivi: So let’s talk about Artificial General Intelligence, or AGI for listeners unfamiliar, which is for simplicity sake, the ability for a machine to understand, learn, and basically mimic human intelligence. Now we’re nowhere near achieving AGI, but there’s great expectations that we’ll get there before the end of the century, perhaps even by its midpoint or possibly sooner. Peter, you’re on record is advocating for cognitive architectures as the path to AGI. Can you please help our audience understand what differentiates cognitive architectures from the other approaches used up until now, and why you’re advocating for it? .
Peter Voss: Yeah, certainly. Another useful way of looking at the sort of artificial intelligence area and the approaches is DARPA talk about the three waves of AI. And what they mean by that is the first wave of AI is sort of logics, formal logic systems, expert systems and that. And that really dominated artificial intelligence for several decades and Deep Blue, the chess champion that IBM built, is an example of that. But then about eight, nine years ago, the second wave hit like a tsunami. And that is basically all to do with deep learning, machine learning, statistical systems, big data. So it’s when the big companies that had a lot of data and had a lot of computing power figured out how they could build neural networks or connections systems that could do really useful things. .
Peter Voss: So that’s currently dominating the AI space, machine learning, deep learning. But when DARPA talk about the third wave, they indicate that something more is needed, that you basically need interactive learning, one-shot learning. You need to get away from just statistical systems. You need adaptive systems, systems that can basically learn immediately, adapt their behavior, and also to be able to reason. And they call that the third wave of AI. And I identify that really as what is needed, the architecture needed to implement that as a cognitive architecture, something that is inherently designed to have all of the capabilities that intelligence require, such as immediate learning and reasoning and deep understanding and so on. .
Guy Nadivi: What are some of the more interesting use cases you’ve applied cognitive architecture based AI to, and what kind of results did you get? .
Peter Voss: Right. There’s a relationship between cognitive architectures and artificial general intelligence. So just to briefly talk about that, in 2001 I actually coined the term AGI, Artificial General Intelligence, together with two other people when we wrote a book on the topic. And the idea behind that was really to get back to the original dream of AI, the term that was coined some 60 years ago. That was to build thinking machines. So we felt in 2001 that the time was ripe to refocus on that original objective, where up to that point AI really had focused on narrow AI, solving one particular problem at a time. And so cognitive architecture to tie it back to artificial general intelligence, is in a much better position to cover a very wide range of different applications and to be able to adapt and learn to changing circumstances. And that’s really what AGI requires. .
Peter Voss: So my focus over the last 15 plus years has really been on natural language conversation, rather than robotics or vision or any of those other fields of AI. And the applications in this area for conversational AI are tremendous, are really large. On the one hand, obvious use cases are to have it as an assistant that can help you with customer service, whether that is for a retailer, or for a bank, or financial institution, or for a phone company, a cable company, that kind of thing. Or internally in the company as well. There are also other applications, many medical applications, as a medical coach for example, that can help you manage diabetes or some other condition. As an elder companion. In the car, you want to be able to talk to your car and you want it to understand you and to remember you and to learn your preferences. .
Peter Voss: In robotics as well, if you have a robot in a hospital or a hotel, you want to be able to talk to the robot and tell it what you want to achieve. In hospital, go to the dispensary, pick up this order and deliver it to that room. Or in a hotel, bring me a shower cap, and tomorrow morning I want two eggs over easy. That kind of thing. So you need conversational AI that is adaptive to those areas. Gaming is another application. So really anywhere where you want an intelligent conversation that is personalized to the individual user, that can learn your individual requirements. .
Peter Voss: We’ve done work with, for example, a sales assistant as a front end to Salesforce, because salespeople are notoriously bad at using Salesforce. So if they have a conversational AI that they can just talk to, they’re much more likely to actually be able to use it, where they can just say, “Tell me about my next appointment. What are their hobbies? What product are they interested in? Do they have any kids?” And can tell you that. And then when you’re done with your appointment, you can say, “Remind me next Tuesday to follow up. Set this to high priority. Send them brochure X, and let my boss know what’s going on.” So just myriad of applications. .
Peter Voss: We’ve having some great success as a hyper-personalized concierge for a gifting company, where our agent can basically learn your individual preferences, who you buy gifts for, who are the important people in your life, what kind of gifts they prefer, when you want them, and so on. So just many, many applications are possible with this general conversational AI. .
Guy Nadivi: Now, speaking of conversational AI, there are some concerns about biases creeping into the AI that powers things like chatbots. How do cognitive architectures address the issue of biases differently than current machine learning methodologies? .
Peter Voss: Yeah, it’s a good question. So a big source of the biases that you get in machine learning applications is basically that you just feed massive amounts of data, and it’s not really curated very much. So whatever bias is in your data is going to be reflected in the outcome of the AI. Whereas with a cognitive architecture, you typically have an ontology that is specifically trained where you have a human in the loop. So it’s not the quantity of data that matters, but rather the quality of data. And so that allows you to look for potential biases and eliminate them as you build your ontology, the knowledge base and the business rules that you have in there. .
Peter Voss: But there’s actually another angle to this that helps. With the second wave of machine learning, deep learning, the system is inherently a black box. If it gives a certain response, you can’t really pinpoint why it gives that response. If you see bias in it, the only remedy is really to retrain the system, train it with a different data set and hope that that fixes it and doesn’t break something else, or doesn’t create some other bias. Whereas with the cognitive architecture, at least in principle, it’s not opaque, you can actually figure out exactly why it’s giving you a certain response and you can then remedy it, and you can give it the extra knowledge that it may be missing or the extra business rules. So it’s much more manageable to be able to eliminate undesirable biases. .
Guy Nadivi: Harvard Business School published an article not long ago, calling for the auditing of algorithms, the same way companies are required to issue audited financial statements. Peter, what do you think about AI algorithms being audited for bias in the same way? .
Peter Voss: I think it sounds pretty impossible actually in most cases. I think it’s a really, really hard problem, because almost every situation is unique. There’s even a more fundamental thing of, what is a good bias and what is a bad bias? I mean, the word bias has a negative connotation attached to it, but there is sort of experience that you have, there are statistical facts that you have. So it really I think comes down to more having the business itself audit itself, and having the right moral structure in place in the company itself. I think external auditing is very, very difficult except maybe for very certain narrow industries. So one would hope that the leadership in companies care enough about the issue to basically eliminate the bad kind of biases. But that’s hard, as with any kind of business ethics, because on the one hand you have huge push towards maximizing profitability, and any sort of moral imperatives that would undermine that just requires really strong leadership. .
Guy Nadivi: Historically in the software business, a killer application was needed to help a hardware platform achieve commercial success. Is there a killer app, or short list of killer apps, for cognitive architecture based AI that will help it achieve breakthrough commercial success? .
Peter Voss: Yes. I think it’s a very obvious that conversational AI is sort of the killer app for cognitive architectures. Cognitive architectures themselves don’t limit themselves to conversational AI, because they also need it for robotics and vision, so basically sense acuity and dexterity aspects of AI. But conversational AI just has such a huge potential market. There’s such a demand for having hyper-personalized conversational assistance. Whether this is an elder companion, whether it’s a personal assistant that helps you, or whether it’s something for large companies where they’re trying to provide better, more personalized hyper-personalized service, consistent service to their customers at a much lower cost. I mean, the promise there is essentially as if you had a dedicated service representative or sales representative, whatever the case may be, allocated to you, remembers you, remembers your previous conversations, remembers what you said, what your preferences are. So I think that’s a killer app. In this case, hardware doesn’t really come into it that much. But for cognitive architectures, yes, I think conversational AI seems to be the obvious choice. .
Guy Nadivi: Most people have heard of the Turing Test, which basically states that if a human can’t tell if they’re communicating with another human or a machine, then that machine or computer has passed the Turing Test. I recently learned that Steve Wozniak of Apple fame proposed an alternative called the coffee test, and this tests some machine’s intelligence by seeing if it can enter an average American home and figure out how to make coffee, which if you think about it is not entirely straightforward. It has to find the coffee machine, find the coffee, find a coffee cup, add water into the coffee machine, and then brew some coffee by pushing the correct buttons. I know some humans who would have trouble passing that test, myself included since I don’t drink coffee. Peter, what’s your personal favorite test when it comes to appraising machine intelligence?
Peter Voss: Yes, I think the coffee test is certainly a good one, if you ultimately want to confirm that you have reached human level AI. But as you said, I would also personally fail the test because I’m a tea drinker. So the Turing Test itself is not actually very useful. In some ways it asks too much, in other ways it asks too little. In the way it asks too much, it basically expects you to be able to fool people that you’re a human when you really aren’t. So if, for example, you want to divide a nine digit number by another nine digit number and come up with a result to 10 decimal places, it would have to say, “Oh, I don’t know how to do that” when it really could. But it also asks too little, in that you just have to be good enough to fool the human judges. So then it depends on the rules of the game, you can kind of game the system.
Peter Voss: As far as the work we’re doing in conversational AI, it’s actually very clear when you give customer service in any different domain, or whether you have an AI in a car or in a VR experience, or wherever the conversation is, and you want to hold an ongoing conversation. It very, very difficult to actually do that well. So it’s easy to see improvements in that area. Basically, how many conversations can the system hold in the real world, and for how long can it maintain? How well is the quality of the conversation? And I think these are not really generalized benchmarks. In fact, I am not a fan of academic benchmarks at all, because you’re then optimizing to the benchmark rather than optimizing to intelligence. So I think in conversational AI, it becomes pretty obvious how well it handles a wide variety of real conversations with real humans.
Guy Nadivi: Whenever we have an AI expert like yourself on the show, I always like to ask them the following question. Over the long-term, do you think that AI machine learning or cognitive architectures will ultimately augment more people or replace more people?
Peter Voss: So I think it will free us up in many ways and improve our lives. Ultimately AI will be able to do pretty much everything that humans are paid for doing right now. That’s in the longer run, and that will free humans up to do the things that they want to do, whether it’s creativity or learning or human relationships, raising a family, whatever it might be, without having to work. That I think is a long-term view. But it will also augment us, in the way that we will essentially have, I like to call it an XL cortex, an extension to our brain, an extension to our mind, that will allow us to make better decisions in life. It will be almost like an angel on your shoulder that can help you make better decisions in life, provide you with more information, better information to make a decision. It will also maybe prevent you or slow you down from reacting at the spur of the moment emotionally about something that you might regret later on.
Peter Voss: Having that personal assistant that becomes part of you, part of your life, I think will just enhance our lives and make us better humans. So I think both are true. It will replace human labor in many ways, but will also enhance our lives.
Guy Nadivi: Peter, for the CIOs, CTOs and other IT executives listening in, what is the one big must-have piece of advice you’d like them to take away from our discussion, with regards to implementing cognitive architecture based AI?
Peter Voss: I think there are quite a few different things but first of all, cognitive architectures are still very new. We’ve been working on it for more than 15 years, but deep learning machine learning is really dominating the field right now. So I think it’s important for management to really understand the limitations of machine learning approaches. They are inherently static. You basically collect a whole lot of data, you train a model, and that model is then deployed in the field. And it cannot learn interactively, it cannot adapt to each individual. You basically have a one size fits all. So if they really want to have a better conversational AI, and again that’s my area of expertise, is to say, well, will the technology that you’re buying, will it actually be able to do that? Will it be able to learn interactively? Will it be able to do one shot learning? Somebody tells you a fact, I’m going to Oregon next week. Will it be able to learn that and use that information without having to be specifically programmed for it? Does it have deep understanding? Can it reason? To really ask those hard questions.
Peter Voss: Now to implement cognitive architecture well, you need to really deeply integrate it into the enterprise or into the application that you want, for it to be effective. It needs to have access to the backend information, to your business rules and so on. And it’s non-trivial, it needs to have commitment. It’s not something you just have a fancy looking tool that you can simply quickly put together some flow chart of a conversation and expect that to work. And therefore it requires really a commitment from top levels of management to have a successful implementation of a cognitive architecture that you understand what you’re trying to achieve, the steps you need to go through to implement it and integrate it. But the rewards of course to doing it well, are just very, very significant. Most large companies, or most companies implementing chatbots right now, are deeply disappointed in the performance that they get, for the reasons I mentioned. They actually don’t incorporate a cognitive architecture and they’re not deeply integrated into the company’s ontology and business rules and so on.
Guy Nadivi: Interesting food for thought for the many executives now budgeting for future investments in this field. All right. It looks like that’s all the time we have for, on this episode of Intelligent Automation Radio. Peter it is a real honor to have someone so highly regarded in the field as our show’s first expert guest on artificial general intelligence. You’ve certainly enlightened me on the topic and I suspect our listeners got great insights from you as well, and it’s something they’ll be thinking about more than they previously did. Thank you so much for coming onto the podcast.
Peter Voss: Yeah. Thank you again for having me, Guy.
Guy Nadivi: Peter Voss, founder, CEO, and chief scientist at aigo.ai. Thank you for listening everyone. And remember, don’t hesitate, automate.