Peter Voss - What’s missing in today’s artificial intelligence
Peter Voss is the CEO and chief scientist of the advanced chatbot platform Aigo, as well as one of the people who came up with the term artificial general intelligence (AGI).
In the second conversation with Peter (our first one revolving around AGI), we discuss what's missing in modern-day AI technologies and how we can responsibly unlock their potential. Peter also shares insights and lessons learned from the development of Aigo, the chatbot with a brain.
Links & mentions:
"This is a part of the problem with the current technology, that you can't rely on it, that it has the limitations. It doesn't know what it doesn't know. It can't really think about its own thinking very well at all. And so the lack of intelligence or the limitations of its intelligence actually is what make it dangerous."
Welcome to the Agile Digital Transformation Podcast, where we explore different aspects of digital transformation and digital experience with your host, Tim Butara, content and community manager at Agiledrop.
Tim Butara: Hello, everyone. Thanks for tuning in. I'm joined today by returning guest, Peter Voss, CEO and chief scientist of the advanced chatbot platform, Aigo, with whom we already had an awesome conversation about artificial general intelligence, AGI, and whether or not humanity is ready for it.
And in today's episode, we'll be talking about what's missing in modern AI systems. And Peter will share his key lessons learned from actually creating a chatbot. So Peter, welcome back to the show. It's great having you back. I remember that our conversation last time was particularly insightful, particularly engaging. And I think we're in for another great one today.
Peter Voss: Yes, thanks for having me again, definitely was very was fun.
Tim Butara: Yeah, you know, as we said last time, the field of AI is developing so quickly that, you know, there's always something to discuss. There's always something new to talk about. And yeah, let's get right down to business, Peter.
And before we discuss what's actually missing AI today, I want to first devote a little bit of time to discussing what AI is getting right today. So what are some examples of good AI uses and and implementations that you'd really highlight here?
Peter Voss: Yes, as you say that, you know, the field is moving very fast. And I mean, it's only been a few months since we've last spoken. So much has happened, you know, the explosion of what's called large language models, you know, like Chat-GPT. You know, it's really quite revolutionary. I don't think I haven't come across a single AI researcher who wasn't really amazed and blown away by just how much these relatively simple systems can do.
I mean, they're simple from an engineering point of view in terms that they don't have a very complex architecture. It's just masses and masses amounts of data and computing power. I mean, these models can cost a hundred million dollars to train and just what it costs to do the actual computation, but the actual underlying technology and architectures is not that complex.
So we really didn't expect them to be able to do as much as they can, you know, to have such fluid conversations and language, and to be able to solve math problems and help with programming and, you know, all sorts of things. So the world of large language models has now been expanded to visual systems, you know, creating images and those are amazing and even starting to show create videos and voice.
So I think from a creative point of view, this is really phenomenal. And you know, there are many, many use cases to give you ideas to create copy, to create images, have poems, help you with email, with assays and so on. So there's definitely a lot going on now. Well, we can talk about the limitations of them, but I think really everybody should should start to use these tools and see where they are useful to them and how they can, can use them. So that, that has really been a big change in the last, you know, 6 to 12 months.
Tim Butara: So we're really, as we also discussed in our first conversation, where we discussed the sparks of artificial general intelligence paper from Microsoft researchers, they highlighted a lot of areas of progress that they really didn't expect and couldn't anticipate, such as using tools, kind of using calculators and stuff like that to kind of help with the computation and everything. And it really seems like we have, obviously, on the one hand, we have this unexpected progress in the context of AI getting things right.
But you just said that we can also discuss the limitations of AI and well, obviously, the main topic of our conversation today is what's missing in today's artificial intelligence. So obviously we need to discuss this too, because it's not just unexpected progress, but there's also probably the other side.
So where do modern AI systems fail and what's the missing piece to kind of unlocking this power of artificial intelligence in a responsible way?
Peter Voss: Yes, I think, you know, because we're all so blown away, everybody's so blown away by what it can do, that it sort of exceeded expectations so much that I think there's a general sense that people overshoot now and, you know, expect that, well, wow, this, this feels like human level intelligence.
And I think that's starting to sort of, reality is starting to set in as people try to use them for more. Advanced or more critical applications, and so there are a number of inherent limitations in this technology, and I'll be talking more as we as we go forward. I'll be talking more about the difference between what's now called generative AI, you know, which is a subset of statistical AI.
So statistical AI or generative AI, which is really what what this is all about versus cognitive AI which is ultimately what we need to have human level intelligence. So the current technology, I mean, Chat-GPT, the word GPT already tells you some, gives you some important clues. G means generative which basically means they make up stuff.
You know, for better and for worse, they can be very creative and creating, you know, poems and songs and, you know, email and, and so on, but they can also make up stuff that is just completely wrong, completely bogus. And the problem is they do it with such confidence that it's very hard to tell.
So, you know, if I've been using it for programming, for example, and I might ask it, how does this work? Or can I do this? Or can I do that? And it might give me an answer, might give me like ten times when I use it, eight times, it might be really useful information and twice it might be information that sounds extremely plausible, but it's just plain wrong.
So the generative part is an inherent feature of it. It has masses of amounts of information, but that information wasn't curated. It's just scraped from all over the internet, you know, good and bad. So there's a lot of garbage in the training and the way it puts together, even good information can just be wrong, that the combination of things may just not be true. So the, the generative part is both the strength and the weakness, but it's inherent in the approach. So you can never really rely on the information, there always needs to be a human in the loop and the human level intelligence to say better double check this. This doesn't sound quite right, you know.
So the second part, the P in GPT is pre-trained and that's the other huge limitation of the system that all of the capabilities that it has, that are inherent in it, were trained, as I mentioned, it can cost a hundred million or more to train this model, but it's read only. You know, once the model is trained, the model itself is not actually adjusted, so it cannot learn additional things or it cannot correct its knowledge except for some sort of short term tricks that they that they've worked out. But the inherent model is pre-trained. And it, you know, it cannot learn and adjust and adapt to new things.
That's why Chat-GPT will tell you, you know, the information is up to date, up to the training date of, you know, September 2022, or whatever date it gives now. There are ways of using short term buffers that you can give it additional information, but that information that you give it in the prompts that you, you know, you might stack up is not fully integrated into the system and it's not permanent.
It doesn't change the actual core training of the system. So the pre-training is, basically it makes it that it's not adaptive and it cannot really learn incrementally as you go on. Now that's crucial for any- you know, if you had a human assistant, you have a new business partner, you have a new product, you have, you know, things happen in the world. You need to basically take that information and integrate it. And it becomes part of your current knowledge and it may completely reverse knowledge that you had previously. So I think those are two critical components.
The other one is the sheer cost of running these systems, that they they require so much processing power is, you know, is and require so much data. I mean, I don't know if we spoke about that last time, but a child, you can show a child a single picture of an elephant. And it will be able to recognize, you know, pink elephants and sideways and upside down elephants and so on was just, you know, one one example. Whereas with statistical AI and generative AI, you basically need hundreds or thousands of instances for it to be able to learn.
So the sheer cost of giving it new information is very, very substantial and that also limits its application. You know how quickly it can potentially learn something new or how many examples it needs to be able to learn. So those are- and those limitations are inherent. That's why I say there's, you know, DARPA, I think we spoke about it the last time where DARPA identified three waves of AI, you know, the first wave being sort of good old fashioned AI, what all the expert systems and logic based systems in the seventies, eighties, nineties.
The second wave is really that we're riding right now is an big data, big compute, statistical systems, generative systems, but we ultimately need to move to cognitive AI that can think and learn and reason more the way humans do. And, you know, that are not pre-trained or not only pre-trained and that are not just generative in the sense of just making up stuff from statistical correlations, but can really think about and understand what they're doing and saying.
Tim Butara: So, yeah, this is the human level intelligence that we also kind of spoke about last time, we basically, if i remember correctly we said that in order for AGI to to be possible and to actually be beneficial to humanity, human level intelligence would basically be a necessity.
Peter Voss: Correct. Because, and in fact, you know, this is a part of the problem with the current technology, that you can't rely on it, that it has the limitations. It doesn't know what it doesn't know. It can't really think about its own thinking very well at all. And so the lack of intelligence or the limitations of its intelligence actually is what make it dangerous, you know, or more dangerous.
You really want, you want more intelligence for the system not to make mistakes that, you know, a human wouldn't make or a human wouldn't make if they thought things through properly. I mean, we obviously make a lot of mistakes, but many of the mistakes we make is because we don't think about things carefully enough, you know, don't take enough data into consideration or just have emotional reactions or so on, but an AI that really can think very clearly will not make silly mistakes.
So, you know, I can talk a little bit more about the limitations. We are, you know, as we're talking to corporate clients, large corporations and universities and so on, and how they use Chat-GPT and large language models, is many of them have actually decided that they cannot use this for any critical applications, that the human always has to be in the loop.
So, you know, for FAQs to help you find things on your website, you know, from all the different documents and things like that to obviously create ideas to maybe fine tune an article or an email that you're writing, you know, that's all okay, but anything where there's potentially legal liability or marketing issues involved, marketing, this department won't sign off on it without a human being in the loop and the legal department won't sign off. So you cannot rely on the output ofth ese systems. And that's sort of the general consensus that for critical, anything critical, you really cannot just use Chat-GPT without a human in the loop.
Tim Butara: But if we had human level intelligence,would this change, so specifically, if we had human level intelligence, would this change make us do away with the need for a human in the loop if the AI kind of checking other AI tools such as Chat-GPT and other LLM tools would actually have the same kind of cognitive functioning as a human?
Peter Voss: Yes, absolutely. I mean, that's really what AGI means, is that you have human level intelligence. In fact, in many ways it will be superhuman because, you know, it'll have photographic memory. It will be able to actually reason better than humans can. But in terms of when we trust them and how we trust them, that really wouldn't be that different from what humans do you trust.
I mean, you know, when you have relationships, whether it's personal relationships or business relationships, you know, you interact with somebody, you know, if it's an employee, you might give them certain tasks or a vendor, a supplier of something, you know, somebody that, I don't know, to build something for you, you might give them a small project and then you kind of build up confidence, you know, and reputation. Somebody tells you, yes, this person is really good. You look at the resume.
So it's all of these kind of things that you have to build up a reputation for saying, yes, I can rely on thisp erson or on this AI to do this particular job and AIs can obviously cross check each other as well. So it will be a lot easier to do that kind of quality assurance, really, on AIs. But, you know, we need to get there. We need to get to that human level AI, which really requires a fundamentally different approach from, you know, the big data generative AI to actually achieve that kind of intelligence.
Tim Butara: And what kind of approach would that be? Can you tell us a little bit more about this other approach?
Peter Voss: Yes, certainly. So that field or that approach is called cognitive AI. And what that really means is, you know, cognition, thinking, intelligence, the things to do with how the mind works and how the human mind works in particular. So the starting point for that is not how much data do we have or how much computing power do we have? And, you know, what particular problem can we solve? The starting point is to say, what does intelligence require? What are the components? What are the attributes that you need in intelligence? And, you know, one of the things, I mean, there are quite a number of this sort of a checklist, you know, I've written a number of papers on that, but there's a checklist.
And, you know, one of them is that the system absolutely needs to be able to learn incrementally, in real time. It needs to be able to, you know, take in information, new information and integrate that with whatever it already knows. And that has to be a process that can happen in real time because things can happen very quickly.
You may need to, you know, you need to make decisions, if you're learning a new tool, you need to experiment with that. Maybe try a few different things and okay, you get the hang of it. You can now use the tool, you know. You might have a conversation, you might get an email, you might talk to somebody that gives you new information that changes direction.
So being able to learn, interactively, in real time without having unlimited computing power or something, because, you know, there's no such thing as unlimited. So the system has to inherently be designed to do that. It also needs to have short term memory. It needs to have long term memory. It needs to have context and needs to have reasoning, deep reasoning ability and the reasoning ability needs to include being able to think about thinking, or what's called metacognition.
So it needs to be able to reflect on its own thought process and say, you know, am I approaching this in the right way? Am I thinking about this the right way to sort of, to direct the thought processes, so the architecture of the system, the way it's designed needs to be designed to be able to do that.
Another element is to be able to form abstractions, which is really the key thing that separates us from animals, is that we are able to form very abstract concepts. We can form abstractions on top of abstractions on top of abstractions. And that is unique in humans where, you know, we can learn obviously the instances of different people and different animals that we come across, but then we can put them into categories, you know, saying this, this is a human, this is a cat, this is a dog, you know, and this is a pet and this is a wild animal. And then we can have levels of abstractions that can go as far as, you know, concepts like honor or marriage or government or, you know, things like that, that are quite abstract.
So the system needs to inherently be able to form these concepts automatically, they can't just be force fed from sort of external sources. The system needs to be able to form new concepts. And this is how really deep understanding happens. That's how innovation happens is when, you know, we put a number of different things together and a new insight and a new kind of abstract representation. So, you know, there's sort of a very different approach saying, starting with what does intelligence require? And then you build a system that creates, that allows you to create, that has all of the capabilities that you need for intelligence.
Tim Butara: And I'm betting that you implemented all of these or the ones that you didn't implement while creating Aigo, you probably learned while creating Aigo. So can you tell us a little bit more about that, about how you created Aigo, so the advanced chatbot and what were your key findings, key lessons learned from this entire process?
Peter Voss: Yes, absolutely. So, you know, Aigo, our current product, we call it a chatbot with a brain. And that's really to make the distinction between every other chatbot out there, which doesn't have a brain. And, you know, we often can, as we interact with chatbots or call into a company and deal with an automated system, more often than not, they are pretty bad, you know, they really don't understand what you want. And then if you, you know, if the conversation changes, they don't remember what you said earlier. So what our approach is, is to have cognitive AI, to have a brain, to have a system, a computer system that has memory, that has reasoning, that has deep understanding, that uses context to be able to hold a conversation, a much, much more elaborate conversation and with deeper understanding.
And we've been working on this for 20 years. So we've gone through the long stretches of just pure development, because it's not a trivial problem to solve, to build an artificial brain, but we now have a product that's still quite a long way from human level understanding and knowledge. One of the big problems is giving the system enough common sense knowledge, the knowledge that we automatically pick up just by interacting with the world, you know, growing up in the real world. That's quite challenging.
So right now the applications we have tend to be focused on solving a particular, as we train the system, to solve a particular set of problems. So an example is, one of our clients is 1-800 Flowers, and it's a group of companies where Harry & David, and Popcorn Factory and so on, you know, very big company, and our chatbot basically gives a hyper personalized concierge type service to their 20 million customers, where it remembers the kind of things that they like to buy, what gifts they buy, who they buy it for, it remembers previous conversations you've had, so it can be more proactive about that.
So those are the kinds of applications we do with our technology right now. But we really, we just, in fact, since we last spoke in the last few months, we have launched a new project to build a separate team that will concentrate on taking our technology to the next level, and having a sort of a big concentrated drive to get much closer to human level intelligence, which is, you know, that development is happening in parallel with our commercial implementations, which are using the sort of proven technology that we have, but on the development side, we are, you know, doing more radical changes, to really rapidly get closer to human level intelligence. So we are currently in the process of expanding that and building that team. So to finally get to true human level intelligence.
Tim Butara: And so what is your, what has been kind of your number one finding, maybe something that you didn't expect when you started developing this, when you maybe, I don't know, also, I assume that you must have had some special reasons for forming this more radical development team. Can we talk a little bit more about this also?
Peter Voss: Yes, certainly. In fact, it's exactly the success of Chat-GPT that has sort of been the catalyst in that. And there are a number of factors there. The one is a lot of people now feel much more optimistic about achieving human level AI.
We can talk about the risks and the concerns about it. But you know, there's a general sense, hey, we may not be that far off. You know, that, just what Chat-GPT is showing us that maybe we can build these systems after all, you know, and not in a hundred years or a thousand years, but maybe more in like five or ten years.
So I think there's a general, a large amount of optimism in terms of being able to achieve human level AI sooner rather than later. And that of course also has an influx of money. Pouring billions, tens, hundreds of billions of dollars into the technology. So there is this sort of, the environment is very much pro AI development right now and getting closer to human level AI.
So that has changed from say a year ago where there's, you know, more optimism and more people are interested in getting into the field, into development. But the second thing specifically for us that has motivated our concerted push now is, as I mentioned earlier, one of the difficulties is getting enough common sense knowledge into our brain.
Previously, I always estimated that it would take, you know, thousands of people to gather this information and teach Aigo that. Just as a reference, Amazon had 15, 000 people working on Alexa; now, they've actually not quite shut down Alexa, but they very significantly scaled down the whole Alexa effort. I think they realized they're kind of not really making progress on the approach they were using.
But anyway, that was my concern that it would take thousands of people. And we simply don't have the resources in our company right now to do that. Now with large language models being around, we can combine the reasoning ability that we have and sort of the inherent quality assurance that we can, that our system can do, but utilize the knowledge that is embedded in large language models to train our system. So the cost of training our system has reduced very, very significantly with these large language models and sources that are available now. So the combination of, on the one hand, the, the much bigger interest and momentum in the field of AI combined with the ability to train our system much, much cheaper has really, well, we've got to, you know, jump on that.
So we're in the process of building a team, a separate team right now using, you know, some of our existing people that have been with us for many years and have a lot of core knowledge, of course, but then also expanding the team to get to human level AI as soon as possible.
Tim Butara: And yeah, since you mentioned, what would be some main concerns of achieving human level intelligence, that you just said that we can also discuss this a little bit?
Peter Voss: Yes, so, I don't really buy into the whole thing that AI wants, you know, may want to kill us or something, take over the world. I see absolutely no evidence that the kind of systems that we're building that are designed specifically to help humans. I mean, that's what they are, assistants to us. And, and that's an inherent design and motivation. Now it's obviously a long debate and you know, a lot of very smart people are quite concerned about AI because of having a mind of their own and their own agenda. And you know, that humans will just be treated like we treat ants.
As I say, I think mainly that's coming from people who have just a theoretical approach from it, and they've sort of thought themselves into a particular corner, I think when you actually work with these systems and how you design them, it's, you know... anyway, for both theoretical and practical reasons, I don't have that concern now, of course. It will be very, very disruptive generally in terms of, you know, human society, but mainly in a very positive way.
But let me give you the one sort of typical example that I give is imagine training up one AGI as a cancer researcher, PhD level cancer researcher. So you now have this AGI, this Aigo or whatever, that can do cancer research. Now you make a million copies of that AI. You have a million PhD level cancer researchers chipping away at that problem. And you take that same idea now and apply it to problems with designing better batteries, nanotechnology, dealing with pollution, you know, clean up, with poverty. I mean, every problem that faces humanity, we will be able to throw massive amounts of intelligence at helping us solve the problems, including governance, you know, and how we manage our affairs as a society to come up with better ideas that will just improve human flourishing.
So I see massive advantages in having AI. Now, of course, when AI can really, truly do many human jobs, those jobs will be replaced. Now people, you know, there's one part of us that says, well, that's really terrible. Let's flip that on his head and say, how many people do you know that would not want to win the lottery?
You know, I mean, almost everybody wants to win the lottery. And AI, AGI, really is like everybody winning the lottery because there will be so much wealth created that we will not have to work. So you can do the things ,you can work on the things that you want to work on, not the things you have to work on to to make a living.
So there's just tremendous upside, but the adjustment for people to win the lottery, of course, will be traumatic for many people, you know. I mean, we know a lot of people don't cope that well with suddenly not having to work anymore. So again, AI can help us get through that period, you know, in terms of helping us to find the best way that we can motivate ourselves and enjoy the freedom that we'll have.
Tim Butara: Yeah, I remember we also discussed this analogy with winning the lottery last time and kind of also highlighted the potentially negative aspect of it. I think you pointed out that a lot of lottery winners actually end up reporting a decrease in their happiness and kind of kind of comfort and well being as opposed to an increase. So, yeah, that's a lot to consider here.
Peter Voss: But that'll be a new problem we have to solve. But, you know, as one might say, well, that's that's a nice problem to have.
Tim Butara: Yeah, true. Well, Peter, thanks for another great conversation. I'm really glad we got you back for another one. Before we jump off the call, if people would like to reach out or connect with you, what's the best place for them to do that?
Peter Voss: Well, aigo.ai, you can email me, email@example.com, but I'm also on LinkedIn, on Twitter, Facebook. So not hard to get hold of me, whatever sort of medium you prefer.
And I always welcome ideas, questions, and collaborations. And we're definitely right now looking for people who want to join us on our journey to actually finally build AGI, to build human level AI. So whether they are investors or technology people or people who want to use our technology right now in their business.
Tim Butara: Well, I certainly hope someone takes you up on your offer, Peter, and gets in touch with you about that. And yes, thanks again, Peter, and have an awesome day.
Peter Voss: Yes, thank you.
Tim Butara: Well, to our listeners, that's all for this episode. Have a great day, everyone, and stay safe.
Thanks for tuning in. If you'd like to check out our other episodes, you can find all of them at agiledrop.com/podcast, as well as on all the most popular podcasting platforms. Make sure to subscribe so you don't miss any new episodes. And don't forget to share the podcast with your friends and colleagues.