EP 036 Artificial Intelligence with Peter Voss

Joining us today is Peter Voss, a Pioneer in AI who coined the term ‘Artificial
General Intelligence’ and the CEO and Chief Scientist at Aigo.ai. For the
past 15 years, Voss and his team at Aigo have been perfecting an industry
disruptive, highly intelligent and hyper-personalized Chatbot, with a brain,
for large enterprise customers.

A transcript of our conversation Artificial Intelligence:

Mike:
Peter, welcome. Thank you so much for being on the show. I really appreciate you taking the time out of your day. You’re a really accomplished guy. You’ve done a lot of amazing stuff, I’m really impressed, building companies and doing research in artificial intelligence. So just to give you an idea of the people listening right now, we’re thousands of software developers who are working at the junction of electronics and programming. So people working in embedded devices and that kind of thing. Some people that are listening now are doing the garage thing. They’re in their spare time, they’re working on building things that they think are great. Some people are software developers, embedded software developers, that’s their career, and that’s what they work on. And here we are talking about artificial intelligence and maybe for some people that’s like, oh, that’s a really heavy technology that I might not be able to incorporate into what I’m doing right now. The work that I’m doing.And for me, the most exciting part of the past decade has been that I feel as though artificial intelligence, it’s been an enabling technology to incorporate somewhat simply anymore into devices. For example, like image recognition is just a good example. I can make an API call, use somebody else’s model and I can categorize or classify an image with lots of issues, somewhat. But generally, I can do it okay. And that’s exciting. But that’s an example of a Narrow AI. It sounds to me like you’re more into artificial general intelligence and that’s where I’d like to start this conversation because some people might not know the difference between Narrow AI and General AI. So I’m curious, just if you could maybe tell the audience what the differentiators are there.Peter:
Yeah, certainly I’ll be happy to talk about that. Just before I go into it, I just want to say thanks for inviting me. And my own background is actually electronics and building electronic stuff. So I can very much relate to your audience. That’s really what got me excited about getting into programming ultimately, is really falling in love, was being able to program electronic chips. And so the intersection between hardware and software, certainly something I can relate to. And as you say nowadays hardware has become so powerful that you can actually do some pretty exciting things on the hardware itself in addition to being able to just have the vast cloud capabilities of all sorts of fancy algorithms that are just on tap.
But let me answer your question. So artificial general intelligence is really the same as what the original meaning of AI was all about. When the term AI was coined some 60 odd years ago, it was really about building thinking machines, machines that can think and learn and reason the way humans do. And at the time they thought, we’ll probably crack this in a few years. Well, it turned out to be a hell of a lot harder than that. So what happened over the decades? The term AI has really morphed into what is practically Narrow AI. So instead of building a thinking machine, basically creating some software that can think and reason and learn the way humans do, what people ended up doing is to tackle one problem at a time and say, “How can we write software where to solve this problem?” And there’s actually a profound difference that is underappreciated.
And with Barrow AI, the intelligence really resides in the programmer or in the data scientist and not so much in the algorithm. So it’s the data scientist that figures out, how can we solve this particular problem by writing some algorithms or training some model? So for example, we have the first wave of AI, DARPA talk about three waves of AI. And the first wave of AI is logic approaches to AI. And Deep Blue, IBM’s chess world champion program is a good example of that. And it’s not that they built a machine that could figure out how to play chess. It was the programmers that basically thought, “How can we use the power of computers, the specific capabilities of a computer,” which was largely brute force, but not entirely brute force, but basically, “how can we be smart about using what computers can do to solve the problem of playing chess really well?” And they did that.
So you have the same in other applications, whether it’s container optimization or medical diagnosis or some expert system where basically a flow chart is created and with the knowledge of people. So that’s Narrow AI and the current wave of machine learning, deep learning is also Narrow AI and I’ll expand on that a little later.
So when I entered the field of AI, I was disappointed to find out this is what AI was doing. And I said, “Well, I’m really interested in getting back to the original dream.” And I think that hardware and software capabilities have progressed enough over the decades that we can now have a shot at this, that we can actually make progress towards building thinking machines. So in 2001, I got together with a few other people who had similar ideas and we actually coined the term, three of us coined the term artificial general intelligence or AGI to get us back to this original dream of building thinking machines. And we wrote a book with that title on that topic. So really that’s what I’ve been focusing on, and I think it’s quite feasible. It’s quite possible for us to actually build thinking machines in the very near future.

Mike:
Oh, that’s fantastic. I’m thinking about deep learning and in my mind, so I studied neuroscience back in the early 2000s before the deep learning, at least it seemed to me at the time before… I like to say, I learned about back propagation before it was cool. But in my mind there was really a phase shift, I guess it seems there was a phase shift in image recognition capabilities that happened between like, I don’t know, early 2000s and after 2000s where that narrow, deep learning algorithm at least created a capability that we didn’t have before. And as I reflect on that, part of me begins to think that maybe that use of it is something that doesn’t happen very often. The discovery or creation of that is like a once in a generation discovery, for example. Our ability to now use and understand this algorithm.
And it’s not like it’s gotten us obviously artificial general intelligence by any stretch. Anybody who uses these deep learning algorithms know that they’re flawed in many different ways. But if I understand correctly, what you’re saying is that you don’t think deep learning, and I think a lot of people would agree with this, deep learning is not going to get us to AGI, artificial general intelligence. We need a different algorithmic approach to get there. Is that correct? Because part of me thinks, well, let’s say one if statement isn’t going to get me lots of choices. But what if I had lots of if statements that all were interdependent? Could I then get to something more complex? And so part of me thinks, “Well, is deep learning the building block to AGI? Or do we just need to really rethink the whole structure of that?” And I guess I’m curious with your company, Aigo, is that the approach you’re taking where you’re rethinking the entire approach for creating a more general intelligence?

Peter:
Yeah, certainly. So a couple of things, certainly deep learning, machine learning, we’ve seen a revolution. It’s really a phase shift of what’s possible to do. Now, the interesting thing is largely the technology that’s used in deep learning, machine learning has been around for decades. As you say, neural networks, back propagation and so on. And really the breakthrough was that we suddenly had some very large companies that had massive amounts of data, massive amounts of compute. And they put that together and tweaking the algorithms a little bit and suddenly by just scaling up the operation, really that was the major breakthrough, suddenly things were possible that we just weren’t able to do before. Like speech recognition, significantly improved with deep learning, image recognition we can use for autonomous cars.
And then of course, what happened then is because of these early successes, massive amounts of money went into the field and people came into the field. So more and more people were working on this because of the excitement and the money. So this has a virtuous cycle. And I’ll talk about the downside of that, where these deep learning machine learning technologies and reinforcement learning suddenly saw so much more development than they’d ever seen in the past, and people came up with clever tweaks for this and different ways of combining and modifying them to give us ultimately AlphaGo and GPT-3 and things like that, that are very, very impressive. But as you said that they are not AGI, not by a long shot.
And the interesting thing is even the head of DeepMind, which is a giant organization, I don’t know, they have six, seven, 800 PhD level people working on deep learning machine learning. And the founder of the company said deep learning is not going to get us to intelligence, not by a long shot. Now, coming from him. That’s a pretty strong statement, and Jeff Hinton considered often as the father of deep learning a few years ago said we should throw it all out and start over. I think he’s changed his mind a little bit. He’s I think more bullish on deep learning more recently. But I think that, yes, I think there’s a general appreciation that deep learning machine learning is not AGI. And also I think more and more people believe that it can’t get us to AGI. And there are some very fundamental and straightforward reasons of what the limitations are.
For example, they don’t have what is generally accepted is they, don’t reason. There’s no reasoning ability, but there also isn’t real time learning. You can’t teach the system new things in real time, which is clearly an essential part of intelligence. So they don’t reason they really don’t learn. And I could go on, I’ve written extensively about that in my articles on medium.com. You can find them there.
So some people still stick to the idea like OpenAI claim that if they just had a million times more data or whatever the number is, a thousand times more data that GPT-X will actually be general intelligence. But I think that’s becoming the minority view and maybe the CEO of OpenMind said that they can get another couple of billion dollars from Microsoft. I don’t know. I don’t know if he really believes it or not.
And then of course there is also the point that yes, in the limit, if you develop deep learning machine learning, if you stretch the definition of it far enough of a general neural net, you could say, “Well, our brains are neural net, so therefore we can ultimately use it.” But it wouldn’t really be called deep learning machine learning. It certainly wouldn’t be relying on back propagation and so on. So I think it’s a bit of a semantic trick saying that these models can get us to human level intelligence when really fundamentally a different technology is required.
And so this is yes, what we are doing. I mentioned that DARPA have a model where they talk about three waves of AI, and the first was basically good old fashioned AI, logic based system essentially. And the second wave of course, is the tsunami that hit us with deep learning machine learning. And that’s the wave we’re in right now. And the third wave though, is what would probably be best described as a cognitive architecture where the adaptability, the ability to learn and reason is the key requirement. So you start off by saying, what does intelligence require? And you build an architecture that addresses those requirements. That’s fundamentally different from deep learning machine learning.
Now there have been some cognitive architectures similarly to neural nets have been around for decades. And people used to say, well, they don’t work. We’ve tried for decades. Well, they don’t work until they do. And one could argue similarly that cognitive architectures have been around for decades. And well, people say, “Well, we’ve tried them. They didn’t work.” Well, yeah, again, we would argue that they don’t work until they do, until you figure out how you can make them work. So I believe cognitive architectures are the third wave and is a technology that will get us to AGI.
Now, one other comment you made was about if you have one if and statement, it’s not going to get you very far, but maybe if you have billions, you’ll get there. And of course, that’s very true. If you believe that basically a current computer technology from a hardware point of view can get us to human level intelligence, if one has that belief then well, it’s NAND gates all the way down, basically. I mean, we know that basically any logic algorithm, anything you can do on a computer basically can be assembled with just a lot of NAND gates. So yes, ultimately it is programming and it’s, yeah, it’s microcode, it’s machine code, whatever it is. It’s code that’s embedded in the chip. So yes, at the very low level, it’s all if statements. But that doesn’t help you when you’re designing high level algorithms, you really have to have a model that is either more like good old fashioned AI or more neural net ish or a cognitive architecture.
And one of the things that’s actually always troubled me when I started studying artificial intelligence is this schism between good old fashioned AI and neural nets. They have their own conferences, they talk with different language. It is very unfortunate because pretty much everything that you can do in logic AI, you can also do in neural nets and vice versa. It’s just often not the best model to be thinking about the problem. Some problems are just much, much better understood and solved by using neural networks and others more by logic. So choosing the right model and the right tool, the right way of thinking about the problem can make it much more efficient to solve a problem. But ultimately, yes, these are all, first, second, third wave are all ultimately if statements or NAND gates when it comes down to it.

Mike:
This is fascinating. So my current understanding of deep learning from the experience I’ve had, is that it is very, like you said, data intensive. That is a big piece of the puzzle. Having an enough data to train a network in order to produce results that are reasonable. Do you see data intensity as being the same way for artificial general intelligence? Would a cognitive architecture also rely on lots and lots of data?

Peter:
No, quite the opposite. And in fact, let me talk a little bit about what we are doing. So I started out in 2001, put an R&D company together. We had about a dozen people working on it to turn my ideas that I’d been working on for several years into actual prototypes, of an early AGI prototype. And over many years, we developed that into something that we could actually commercialize. We first out with building a virtual critter in a virtual world and experimenting around with that. But over the years, we decided that we could actually make more progress by focusing on natural language understanding rather than sensing vision and dexterity and robotics. Robotics is really hard and also not that easy to commercialize.
So we shifted our focus on a natural language understanding, and that’s really what my companies, both in terms of development and commercialization, what I’ve been doing over the last 20 years. So aigo.ai, we have a chat bot with a brain. That’s how we describe it. Now, there are lots of chat bots. There are thousands of chat bots out there. None of them have a brain. They don’t have a cognitive architecture. They can’t think, they can’t learn interactively. They can’t reason. And we’re painfully-

Mike:
Oh, I know Peter.

Peter:
… aware of that.

Mike:
Oh, painfully. I remember when Alexa first came out, my hope was like, “Oh, these are going to be great in no time.” And honestly, sometimes I think they’ve gotten worse. I don’t know. It’s yeah. But anyway, I didn’t mean to interrupt, but wow. I think everybody listening knows that natural language processing abilities of just any of these are just subpar, not living up to our expectations.

Peter:
Right. They are getting worse in one sense, and that is they try to have more and more functionality that they add, but that makes it easier for them to get confused. Just let me quickly talk about the chat bots and the difference having a brain and not having a brain. So all the chat bots, other than the one we provide, as far as I’m aware of use basically first and second wave technology, they use second wave technology as a categorizer. So if you say, blah, blah, blah, weather, it basically says, “Oh, okay. Weather.” So there’s a model that basically just as a categorization model, it takes whatever you say and guesses, categorizes from that statistically, basically, what is the most likely thing out of the hundreds of things that Alexa can do that you want to do? So that’s a categorizer and if you say, “I hate Uber don’t ever give me Uber again,” chances are it’ll trigger the Uber app. So that’s the one part, is the categorizer.
And then the second part is basically a little flow charty type program that somebody writes where it says, okay, where do you want to go? How many people are going and do you want UberX? But it’s a simple flow chart for program. And that’s how all chat bots that are out there are done. There is no reasoning engine. They don’t use context unless somebody very carefully wrote some program to maybe use the previous utterance that you did, or the previous tasks that you did, but there’s basically no intelligence be behind it.
So our approach is to have this cognitive engine, this Proto-AGI engine, that actually has deep understanding of what you say. So it does a deep pass, but it does have past contextually. What is a conversation about? Who are you talking about? What is the topic? Are you having a business discussion? Are you talking to your spouse? So it takes context into account. It takes into account what was said earlier, and it learns interactively. And we have some demos on our website that show how we do that.
Now we’re still a long way from human level understanding, but we believe have the core architecture to do that. And to get back to your original question here, in terms of, does that approach require a lot of data? No, not at all. It’s the opposite. It’s the quality of the data that matters, not the quantity. So we basically have an ontology, the background on knowledge that you need to have to have understanding of it. Now, the deeper your understanding, the more background knowledge you need, but it’s measured in tens and hundreds of thousands of facts, maybe millions, not zillions as deep learning or GPT-3 for example, they brag about, I don’t know what the number is now, but it’s basically gazillions of words that have been fed in.

Mike:
So is that ontology then, provided? You are creating that ontology or is that context being extracted from the interface? So I’m talking to my wife or I’m having a business discussion. Is that context right there is pre ingrained in wherever the domain that the intelligence might be at? For example, let’s say I’m on my, let’s say I’m running a business and I’ve got a chat bot that is helping people find customer answers or answers to issues. So then that chat bot knows, okay, the domain I am operating in is X, Y, Z. I’ve told it ahead of time, “This is the domain.” And that’s where it’s getting that context then?

Peter:
Yeah. Yeah. Very, very good question. And I can use an example. One of our big customers is 1-800-FLOWERS and they use Aigo basically as a concierge service that can help you with your gift buying and be hyper personalized. So there are essentially, if you think of three concentric circles of ontology and knowledge that we have, the inner circle is the common knowledge that is shared-

Speaker 3:
[crosstalk 00:24:18].

Peter:
… across all different applications. It’s about knowing about people, places, time, how to hold a conversation, how to start a conversation, how to end a conversation. So that’s the common knowledge and we’re continually expanding that the system has more and more knowledge. It’s the core ontology.
Then the second layer, just next outer circle, there is the domain specific knowledge that you need to teach it. So for 1-800-FLOWERS they have 12 different brands, [crosstalk 00:24:52] chocolates and cookies and so on. So the specific terminology, their product categories, for example, whether you’re buying for an anniversary or a birthday, we need to make sure that the system has that kind of ontology, the knowledge that recognizes what these terms mean and how they relate to each other. Also, business roles. Are you a passport member that you have certain privileges? And things like that. So we teach at that ontology and either we can do that for the customer or the customer can do it with certain tools that we provide.
Now, the third layer, the outer layer is the customer specific ontology. So that’s when you talk to Aigo and you say, “I want to buy some roses for my niece, Amy, who has a birthday next Thursday.” Now, that is unique to you. So you are now teaching the system in natural language, that part of the ontology. And Aigo will remember that. And so a year from now, we might set up Aigo to be proactive and say, “Hey, your niece’s birthday is coming up. Do you want to send some roses again? She really liked them.” If we happen to know that she liked them.

Mike:
Right. No, absolutely. That’s…

Peter:
So there are basically these three levels of ontology, but as you can see, none of them are big data in the sense of deep learning machine learning.

Mike:
Right. Yeah. That’s fantastic. Oh, I’ve got so many questions. I know we’ve got limited time. I’m really curious how that ontology actually gets encoded. I’m curious, but I don’t want to go there. Maybe it’s a little too technical, but I am interested in talking about that third circle you mentioned, which is the context of the person. And I’m sure lots of people have thought about this kind of thing, but our preferences, the situations, how we act, all of those things I imagine could be captured by a machine or an algorithm and help develop each of our own personal proxies. So people would have their own proxies that would basically be a, yeah, this is what Mike prefers, or this is what Peter prefers or Amy prefers. And these are the relationships Michael has and wow, Michael seems to have a quick temper sometimes when it’s late at night and he hasn’t eaten today.
All of those things that would begin to form a proxy of the decisions I might make or the preferences I might have. And then I just think, let’s say I’m using a web browser and that web browser has access to my proxy. Then now that could act as the third circle for Aigo or any intelligence that I would be interacting with. So it’s like I’m providing, I’m allowing my data to be shared with a third party intelligence because I, darn it, I want this thing to respond to me in an appropriate way. That seems like, I mean, I’m just curious, is that the… Let’s say I go to 1-800-FLOWERS and Aigo is characterizing me in some way. If I go somewhere else where Aigo is operating, is that information coming over or does it have to relearn that about Mike?

Peter:
All right. Well, you caught straight onto our vision as a company. So yes, very much so. So we call that vision, that path that we on a personal, personal assistant, because yes, right now we are working with large enterprise companies. So they control, obviously the Aigo chat bot. And they protect the data. In fact, it runs behind their firewall. We are not a SaaS company. And people actually like that model a lot where we deploy, we provide the technology that they can integrate into their technology stack. We help them do that. You know these things all run in the cloud these days anyway, but it’s in their cloud service. It’s essentially on-

Mike:
Right. Well that’s-

Peter:
… it’s essentially on prem. So they totally control the data and that, and we don’t have to worry about outages and things like that. But the vision that you’re painting, as I say, we call that a personal, personal assistant, because yes, you might be interacting with Aigo’s 1-800-FLOWERS. And the Flowers’ Aigo knows something about, that you have a good temper, that you’re in a good mood early in the morning, whatever.

Mike:
All true. All true.

Peter:
Who you buy gifts for and for what occasions and what the relationship is to these people and where they live and whether you delivered at their office or at their home. And all of that stuff, Aigo can learn. But now, if you’re also using, if we have Aigo deployed with your bank or with AAA and you have a breakdown on the side of the road, that Aigo wouldn’t know anything about that it might have learned from other Aigos.
So the ultimate vision is what we call a personal, personal assistant. It should really be called a personal, personal, personal assistant because the word personal actually has three different meanings that are relevant here. The first meaning is personal, that you own it, you control it, it’s your agenda, not some mega corporation’s agenda, not Siri, not Alexa. So it’s yours. You control it. That’s the first personal. The second personal is it’s personalized to you. It knows your preferences, your history, your desires and your relationships and timetable and all of that. So it’s hyper personalized. It’s customized to you. It’s not a one size fits all. And the third personal is the security, the privacy issue. That things that are personal to you, that you don’t want to share with other people. So you can entrust with your deepest secrets basically, and you decide what you want to share with whom. So ultimately that becomes almost like an Excel cortex, like an extension of your personality, of your mind that you have your own personal, personal assistant.
So we are working on that as well, but as it’s more consumer product, so it’s a lot harder to get into that market where people are currently used to getting these things for free, free in quotes, you’re selling your soul. But-

Mike:
Selling your attention. I think is-

Peter:
Yeah. Right. But there are applications, for example, we’re using Aigo as a personal assistant for sales people. Sales people hate using Salesforce. And so if you can just have an Aigo that you can talk to and say, “Tell me about my next sales appointment. What product were they interested in? Does he have any kids or what are their hobbies?” Or whatever. So Aigo can tell you if that information is available. And then when you’re done with your sales call, you can say, “Aigo, remind me next Tuesday to follow up, send them brochure X and let my boss know what’s going on.” That kind of thing. Now, once you’re using Aigo like that at work, then obviously it’s a small step to then say, “Hey, Aigo me to pick up the kids on the way home.” And then your spouse wants an Aigo as well, of course. So then-

Mike:
Right. Yeah.

Peter:
Yeah.

Mike:
I feel like what you’re describing is just the holy grail of what everybody really wants right now, to be able to say a sentence that makes perfect sense to you and I and have an intelligence be able to just understand all the context in there and actually do it correctly.

Peter:
Right. And that’s what-

Mike:
It’s wonderful to think about.

Peter:
… we’re doing. Yeah. And we have the core technology to do that. We have the right approach. I’m confident of that, but it’s a hard problem to get all of the subtleties of common sense knowledge and common sense reasoning that we just pick up being born into this world and interacting with world and people as we grow up. So it’s non trivial to get that general background knowledge that we have.
Now, I mentioned earlier one of the downside of the success of deep learning and machine learning is that it’s really sucked the oxygen out of the air in terms of AI development. Tremendous progress has been made in developing neural nets, back propagation, reinforcement learning and tweaking these algorithms, working with massive amounts of data. The tools that are available are fantastic. But on the other hand, it’s basically nobody’s working on real intelligence because you want to do a PhD? Well, you’re not going to get a sponsor unless it’s deep learning machine learning. You want to earn the big bucks? You got to work on that field. You want to get funded in a startup? It’s got to be deep learning machine learning. So it basically has crippled other approaches and approaches like the third wave of cognitive architectures that really nobody is working on that.
And the only reason we are working on it is because I am from a theoretical perspective that I actually spent quite a few years, about five years studying intelligence, different aspect of intelligence before I even started working on this. How do children learn? How does our intelligence differ from animal intelligence? What do IQ tests measure? And really understanding deeply what intelligence requires? And once I understood that it was clear to me all through this tsunami craziness of successes of deep learning, that it cannot be the right approach. It doesn’t address fundamental requirements of intelligence. If you had a personal assistant that couldn’t reason about things or that needed to go back and be trained for 24 hours on new data, it clearly, isn’t the right approach. A child, a three year old child, you can show it one picture of a giraffe and it’ll recognize giraffes in different colors, shapes and sizes and so on. Whereas with deep learning, you need thousands of examples and number crunching forever.
So there are inherently just a lot of fundamental, it’s fundamentally the wrong approach for general intelligence. So that’s the reason we look at what can we learn from deep learning machine learning? Are there some algorithms? Are there some tricks? Are there some data models that we can use, for example, in speech recognition? Yes, that’s very useful. They’re statistical approaches and using them, but as far as the core reasoning, using context, deep understanding, they’re actually pretty useless.

Mike:
All right. Wow. That’s fascinating. All right. I think I got one last question here, and I want to try to bring this back to the folks listening who might be thinking to themselves, well, okay. Maybe some of the people have attempted to use a Narrow AI in their application. For example, I know a common one is that you can get a mini TensorFlow running on an Edge device and you can use a pre-trained model to recognize faces, for example. So the camera might take a shot of somebody’s face, and then it can say, okay, hey, this is this person I’m going to unlock the door because I know it’s my wife or whoever. And that’s pretty cool because I can operationalize it. You had mentioned that Aigo is not a SaaS company and that’s, for somebody listening, SaaS is just software as a service.
But I’ve heard this term thrown out there, intelligence as a service. And that’s when I envision, I was just working on an application the other day that did image recognition. And literally, it was like an API call. So I’m making an API call to something, and I’m getting a little nugget, well, call it in quote, unquote intelligence. A very narrow intelligence, but it was doing a specific task for me. And I’m just wondering, how do you see potentially Aigo, that more reasoned intelligence being operationalized for like ye old developer? Who’s boots on the ground. I’m a developer. I want to try to implement some intelligence in my device, some reasoning in my device. I don’t know. I know this is a very futurist question, I guess, but I don’t want to stick us to anything necessarily practical at the moment, but I’m just like in the future, what do you see how somebody might actually grab onto that, a developer might actually grab onto that intelligence and implement it in something?

Peter:
Yeah. So two things, the one thing I was shocked when I started working on AGI on our prototypes of how much we could achieve with relatively little CPU and memory. I started working on this 20 years ago. Now, obviously in 20 years, the amount of processing power that we have on a phone is just amazing. I mean, our current Aigo in the normal mode that it runs could easily run on a smartphone. The amount of CPU that you have on memory, you have a smartphone. So we actually, and of course it keeps growing. So basically a single Aigo can actually run, not quite on an Edge device. They don’t tend to have enough memory yet, but probably soon, I mean, we’re talking about a gigabyte of memory or so. Not a lot. You can certainly have a small, single board computer that could run Aigo.
So first of all, we can make Aigo available running on a very small device. And it’s something I’d like to do. It needs to make commercial sense to be able to do that. The other thing, if you really want it as an API call, we do expect to offer Aigo services as a service, intelligent as a service. Now, one thing I’d like to mention here is it can’t be just a normal rest API because that undermines, the whole idea of intelligence is being able to take the context of what was said earlier, the personalization, what you learn about the person. So if you want that, you really, I mean, it can still be a risk call technically, but you need to then reference a brain that you’re talking to.
And that’s actually the way Aigo is implemented that when a chat comes in or an input voice input comes in, you have to identify what brain am I talking to? And then that needs to be routed to that brain that had the conversation before. Now, when you’re done with your conversation, that brain might go to sleep, but next day or next week or next hour, whenever, when you start talking to Aigo again, we need to wake up that brain. It needs to go back to the same brain, because you wanted to remember what you said yesterday and the day before, what your preferences are.

Mike:
Oh, that’s fantastic. Awesome. Peter, thank you so much for taking the time to talk. I know this has been fascinating for me. I’m sure all the listeners are excited and pumped up and definitely want to check out Aigo. Is there any last thoughts you want to leave the listeners with? Again, these are developers out there. They’re where the rubber hits the road folks. I don’t know if you have any final thoughts to leave with them.

Peter:
Yes. I mean, I would love more people to get excited about AGI and talk about and say, “Hey, let’s do it.” I believe we have the hardware and software that we can actually get to have much higher general intelligence we have now, I would say. To get to human level intelligence, but I don’t want to scare people off. That’s a whole nother topic, but we can have the personal assistant that we’d all like to have, but more people need to be working on it.
So I think more people are talking about it and what’s possible with the third wave of AI cognitive architectures and say yes, deep learning machine learning is great, but it’s just not going to get there. We’re looking under the wrong lamp post and that’s where all light is shining right now. That’s not going to be a solution to AGI. And yeah, if you’re interested more, as I said, I have a number of articles on that, short articles on medium.com. Not very technical. You can find them. And also our website, Aigo.ai. And of course, if you know any enterprise companies that are struggling or unhappy with their chat bot, which should be pretty much every large company that should be unhappy with their chat bot, get them to talk to us and chat bot with a brain.

Mike:
Absolutely. Thank you so much for your time, Peter. Really appreciate it.

Peter:
Right. Thank you.