Latest / The Adults In The AI Room / Peter Voss on the Third Wave of AI - Episode #22

Transcript

Peter Voss on the Third Wave of AI - Episode #22 · The Adults In The AI Room

• 0:36
  SPEAKER_02: Well, good afternoon, good evening, good morning, depending upon your time zone.
• 0:40
  SPEAKER_02: Welcome to our podcast.
• 0:41
  SPEAKER_04: I'm Seth Early.
• 0:42
  SPEAKER_04: And I'm Chris Featherstone.
• 0:44
  SPEAKER_04: It's good to be with you.
• 0:45
  SPEAKER_02: Nice to see you.
• 0:46
  SPEAKER_02: How's uh how's uh how's everything in Montana?
• 0:48
  SPEAKER_02: I know that you just uh were just chatting, you said it was freezing, snowing, and uh uh making winter uh uh winter like uh signals out out there for you.
• 0:58
  SPEAKER_04: Well, as they as they say, you know, as the con the the common phrase now, winter is coming, right?
• 1:03
  SPEAKER_04: Not to be you know popularized by uh by the you know pop culture, but winter is here.
• 1:10
  SPEAKER_02: Yeah, right, right.
• 1:11
  SPEAKER_04: And it is cold, yeah.
• 1:12
  SPEAKER_02: It is not here yet.
• 1:13
  SPEAKER_02: And I'm happy to say that it's a spectacular autumn, and I'm very happy with that.
• 1:17
  SPEAKER_02: It's a very nice fire season.
• 1:19
  SPEAKER_02: So I'm gonna have some folks over this weekend for fires.
• 1:21
  SPEAKER_02: Well, let's um today our guest is a serial entrepreneur, engineer, inventor, and a pioneer in artificial intelligence.
• 1:29
  SPEAKER_02: He's parlayed his passion for studying intelligence and how it develops in humans into the creation of a natural language engine and artificial intelligence applications.
• 1:38
  SPEAKER_02: Please welcome founder and CEO and chief scientist at IGO, Peter Ross.
• 1:42
  SPEAKER_02: Yeah, thanks for having me.
• 1:44
  SPEAKER_02: So, Peter, uh, you know, one of the things that we've uh talked about in our prior conversations is, you know, your path to what you're doing.
• 1:51
  SPEAKER_02: Do you want to just give us the thumbnail of kind of how you got to where you are and especially interested in your thinking around artificial uh general intelligence as opposed to narrow intelligence?
• 2:01
  SPEAKER_02: Why don't you talk a little bit about um your career and how you got to doing what you're doing?
• 2:06
  SPEAKER_00: Yeah, certainly.
• 2:07
  SPEAKER_00: So I started out as an electronics engineer, started my own company building electronic equipment for industrial applications.
• 2:15
  SPEAKER_00: And then I fell in love with software, and my company changed very quickly into a software company.
• 2:20
  SPEAKER_00: Well, actually, a systems, hardware, and software company.
• 2:24
  SPEAKER_00: And I ended up designing an ERP software system for small to medium-sized businesses.
• 2:30
  SPEAKER_00: That company was quite successful.
• 2:32
  SPEAKER_00: We went from the garage to 400 people and an IPO.
• 2:36
  SPEAKER_00: Uh, so that was really exciting.
• 2:39
  SPEAKER_00: Yeah, love to do that again.
• 2:41
  SPEAKER_00: And um, but it's when I sold my interest in that company, I sort of had the freedom to think about, you know, what big, exciting project do I want to tackle next?
• 2:51
  SPEAKER_00: And what really struck me is that software is inherently quite dumb.
• 2:56
  SPEAKER_00: You know, and I'm saying that we'll being quite proud of the software that I wrote um and that you know, that we built the business around.
• 3:05
  SPEAKER_00: But still, if the programmer didn't think of some situation, um, the program would just, you know, crash or do something nonsensical.
• 3:14
  SPEAKER_00: There's really no common sense, there's no thinking.
• 3:16
  SPEAKER_00: So that started me on the journey of how can we build software that actually has some intelligence, software that can actually think and reason and have common sense and can learn.
• 3:28
  SPEAKER_00: And so that's the journey I've been on for the last 25 years or so.
• 3:33
  SPEAKER_00: And uh initially took off five years just to study intelligence and all things related to intelligence, you know, starting with philosophy, epistemology, theory of knowledge.
• 3:44
  SPEAKER_00: You know, how do we know anything?
• 3:46
  SPEAKER_00: What is real?
• 3:47
  SPEAKER_00: What is reality, you know, and and so really the the the I still ask myself that question what is what is reality?
• 3:54
  SPEAKER_00: Yeah, well, pinch yourself and then you know, you know what's real.
• 3:58
  SPEAKER_00: Um so you know, I found that it's really important for me to understand the fundamentals of knowledge and intelligence and reasoning and how we think and learn.
• 4:09
  SPEAKER_00: And so I studied how children learn and how our intelligence differs from animals, and what IQ tests measure or don't measure, you know, stuff like that.
• 4:19
  SPEAKER_00: And of course, also uh learning about what else had been done in the field of artificial intelligence over the last 60 years, 50 or 60 years.
• 4:29
  SPEAKER_00: Um so that culminated in 2001.
• 4:34
  SPEAKER_00: I then started a company to actually turn my ideas into code, into actual prototypes and a framework.
• 4:42
  SPEAKER_00: Uh, around about that time, I also got together with some other people interested in sort of a similar pursuit of real artificial intelligence.
• 4:51
  SPEAKER_00: Uh, and I'll talk more about you know what the difference is there.
• 4:54
  SPEAKER_00: And so we ended up writing a book and coming up with the the title and the term artificial general intelligence or AGI, which you know has since in the last 20 years actually been adopted quite wide widely.
• 5:10
  SPEAKER_02: So in it, you actually coined you actually coined that phrase.
• 5:14
  SPEAKER_00: Yeah, I coined that together with two other people.
• 5:16
  SPEAKER_00: There were three of three of us uh that came up, you know, we were brainstorming the title for the book, and we came up with this this term.
• 5:24
  SPEAKER_00: And uh I really like it.
• 5:26
  SPEAKER_00: The artificial general intelligence in general sort of has the the meaning of you know that it's very broad, uh, can be applied very broadly, it's not narrow.
• 5:36
  SPEAKER_00: Uh and also it has that little G, which is you know a symbol used in uh psychology to denote uh IQ or general intelligence.
• 5:46
  SPEAKER_00: So um, yeah, that that's a term.
• 5:49
  SPEAKER_00: And then since then I've basically been uh developing this through various companies and commercializing this technology that that we've um you know developed and keep improving to get closer and closer to human human-level intelligence.
• 6:05
  SPEAKER_02: And tell me when you did your studying, sorry, Chris, when you did your study of psychology and learning and knowledge and epistemology and so on, what what were the big takeaways?
• 6:16
  SPEAKER_02: I mean, what was the insight?
• 6:17
  SPEAKER_02: Was there an aha moment that you got out of that, or is it more of a a range of insights and a range of understanding?
• 6:25
  SPEAKER_00: I I think that's sort of an accumulation of things, sort of getting a better and better understanding, you know, initially started out with saying, you know, well, what is consciousness?
• 6:36
  SPEAKER_00: What is intelligence, you know, and what is free will?
• 6:39
  SPEAKER_00: How does that all fit in?
• 6:41
  SPEAKER_00: You know, how how can we how can machines think?
• 6:45
  SPEAKER_00: And um over time that just became clearer and clearer of what the answers are to those to those questions.
• 6:52
  SPEAKER_00: But if I have to say what are maybe two of the biggest insights that, you know, I I think are not commonly known about AI and AGI.
• 7:03
  SPEAKER_00: The the the one is the importance of concepts and concept formation, and to really understand how concepts work.
• 7:12
  SPEAKER_00: Because it's high-level concepts or abstract concepts that are really the key to human intelligence, that we can uh we can create uh really an infinite number of abstract concepts and concepts of concepts.
• 7:30
  SPEAKER_00: So we can create this hierarchy of very, very abstract uh concepts like you know, loyalty or government or you know, things like that, that are based on lower level concepts all the way down to sort of the grounding of reality of you know our perception.
• 7:47
  SPEAKER_00: So I think the importance of concepts and concept formation is one.
• 7:52
  SPEAKER_00: Uh the other thing that I learned is in the study of IQ, I was actually involved with um helping to develop a new type of um IQ test.
• 8:03
  SPEAKER_00: It's not really an IQ test, it's more a cognitive process profile.
• 8:06
  SPEAKER_00: So the the strengths and weaknesses of a person's cognitive processes.
• 8:12
  SPEAKER_00: And what I learned through that is that there's one dimension that's particularly important, and that dimension is called metacognition.
• 8:21
  SPEAKER_00: So that's sort of your ability to consciously or subconsciously apply apply the right strategy to problem solving.
• 8:32
  SPEAKER_00: You know, some problems require uh logical thinking, uh, or other other problems really require you to have a more intuitive solution, get a more intuitive solution because they don't have a logic, strictly logical path, and and so on.
• 8:48
  SPEAKER_00: So those are probably two of the two of the sort of many elements, but those are the two that stick out.
• 8:54
  SPEAKER_04: What what led you down the route of actually you know studying animals and humans and children?
• 9:00
  SPEAKER_04: And what what was the outcome of that?
• 9:03
  SPEAKER_00: Well, you know, it was really important for me to um to understand cognition in in its broadest terms and uh and and to understand it deeply.
• 9:16
  SPEAKER_00: And one of the key elements of intelligence is learning, the ability to learn.
• 9:22
  SPEAKER_00: And you know, the one kind of puzzle when you start going into the field is to say, what makes humans so special?
• 9:28
  SPEAKER_00: What is makes human intelligence special?
• 9:32
  SPEAKER_00: And so understanding how children learn and how animals learn, and why, you know, why there's such a big difference?
• 9:42
  SPEAKER_00: And so it was really uh interesting for me to understand, for example, um, chimpanzees and their early development are actually ahead of human children in their cognitive development, but then they kind of top out, you know.
• 9:57
  SPEAKER_00: So why is that?
• 9:58
  SPEAKER_00: And you know, even chimpanzees uh or you know, um other or the higher apes raised within a human family simply cannot get past a certain threshold.
• 10:11
  SPEAKER_00: Um, and and sort of trying to understand that, and this is kind of where the concept formation and abstract thinking, you know, that really uh uh was an answer.
• 10:21
  SPEAKER_00: But I've also just generally had a um innate interest in philosophy and ethics in psychology, you know, what makes humans tick.
• 10:31
  SPEAKER_00: So there was there was that element just that I was inherently interested, but it fitted in nicely with understanding better how can we uh build machines that can think and reason and learn the way humans do.
• 10:44
  SPEAKER_00: And then also like dealing with the puzzle of what is free will, you know, the sort of more philosophical aspects, and how can how can machines have free will, or what kind of free will do we have?
• 10:56
  SPEAKER_00: What kind of free will will machines have?
• 10:58
  SPEAKER_03: So awesome.
• 10:59
  SPEAKER_04: I'd love to be able to at some point get, you know, maybe your thoughts on at what point do you know do we get to a machine that actually has a conscience, right?
• 11:10
  SPEAKER_04: And what is consciousness as opposed to what we're talking about with abstract layers of concepts that actually create context and can actually make decisions and you know how that you know kind of comes together.
• 11:20
  SPEAKER_04: But it's fascinating.
• 11:22
  SPEAKER_00: Well, I can I can give you I can give you a quick answer on that, but of course it'll you know open up a whole lot of additional questions.
• 11:29
  SPEAKER_00: So uh the the the quick answer is that um consciousness is an essential component of higher level intelligence.
• 11:40
  SPEAKER_00: So you cannot have high-level intelligence, human-level intelligence without consciousness.
• 11:45
  SPEAKER_00: And the sort of quickest way I could justify that statement would be to say the system has to be aware of itself as a thinking, acting entity.
• 11:56
  SPEAKER_00: So it has to have a self-concept that you know it so basically says, you know, I took this action, not something else, and I can take this action, and by thinking about things, or you know, I may come to a different conclusion, and my actions have an effect on the world.
• 12:15
  SPEAKER_00: So it's it it and a high-level intelligence absolutely has to have that kind of self-awareness, and in in in in roughly speaking, that is really what consciousness is.
• 12:28
  SPEAKER_02: I think that um when you get into that realm, it feels a little a little bit of uh uh there's some challenges to being able to verify, validate, right?
• 12:39
  SPEAKER_02: Because I think as much you know, machines will probably be pretty good at emulating uh some type of behavior or awareness, or it may seem that it's the you know, the emulation game, right?
• 12:50
  SPEAKER_02: It's the it's the Turing test, but but will it will how would we ever know whether that's actually a reality of of uh of a machine?
• 13:00
  SPEAKER_02: And and I I have a little bit of a hard time kind of re-broaching that or you know, drawing that conclusion because I do think about the complexity of the human brain, and I think about you know, uh one neuron can connect to 10,000 others, you have hundreds of neurotransmitters, those are all analog, there's a billion of those uh neurons.
• 13:20
  SPEAKER_02: And so I think orders and orders and orders of magnitude, but but it does come back to you know, uh, what does that really mean?
• 13:28
  SPEAKER_02: And and is there a way of understanding and defining it?
• 13:31
  SPEAKER_02: I think that's kind of a tough area to explore because again, it is really um uh it's it's it's a lot of um it's it's more philosophical than it is scientific, right?
• 13:43
  SPEAKER_00: Yeah, well, there's sort of the the standard uh comeback of saying, you know, how will we know whether machines are conscious?
• 13:50
  SPEAKER_00: I mean, the standard comeback is well, how do you know that I'm conscious and not just a machine, you know, flapping my lips.
• 13:59
  SPEAKER_00: Acting like it's conscious.
• 14:01
  SPEAKER_00: Acting conscious.
• 14:02
  SPEAKER_04: Well, you know, I was I was thinking about when you were when you're talking about, you know, diving into um, you know, the base foundations of how children and animals learn and what that looks like in terms of of pulling together abstract concepts to actually then make decisions on them and things.
• 14:18
  SPEAKER_04: And then, you know, like you were talking about the layer of concepts that need to um that could be mutually exclusive, they could be interrelated, they could be hierarchical.
• 14:28
  SPEAKER_04: What does that look like when it starts to come together?
• 14:31
  SPEAKER_04: You know, because then when we start to bridge the gap between what this looks like for maybe specific artificial are, you know, artificial intelligence and then maybe generalized artificial intelligence, how that actually then provides value to you know to the market.
• 14:47
  SPEAKER_04: It's one thing to do research and do all that, but then you know, and this is probably the plague of you being not being able to turn your brain off, is then applying those concepts to actually that provide value in today's world, right?
• 15:00
  SPEAKER_04: Whether it be business value or you know, or augmented value, right?
• 15:06
  SPEAKER_04: And where humans come in.
• 15:07
  SPEAKER_04: So I think the the key, I'd love to get your take, Peter, too, in terms of, you know, maybe for folks listening, how you define artificial intelligence and then how you define artificial general intelligence and the difference between the two.
• 15:23
  SPEAKER_00: Yeah, absolutely.
• 15:24
  SPEAKER_00: Very good, very good question.
• 15:26
  SPEAKER_00: So when the term uh artificial intelligence was coined some 60 plus years ago, uh, it was really about building thinking machines, machines that can think, learn, and reason the way humans do.
• 15:39
  SPEAKER_00: That was the original vision of AI.
• 15:43
  SPEAKER_00: Now, originally they thought they could crack this in like, you know, a few years.
• 15:48
  SPEAKER_00: Now, it it turned out to, of course, be much, much harder than that.
• 15:52
  SPEAKER_00: And so what happened over the decades is really that that vision completely drifted to saying, well, if we can't achieve the grand vision, let's at least um build systems that seem to exhibit something that we call that we consider intelligent.
• 16:13
  SPEAKER_00: So a perfect example of that, I think, is um playing chess, you know.
• 16:18
  SPEAKER_00: So I think somewhere in the late 80s around about there, IBM uh built uh deep blue and became the world chess champion.
• 16:26
  SPEAKER_00: And you know, generally people thought that, well, if you can play chess, then surely that's something intelligent.
• 16:33
  SPEAKER_00: But basically the the vision of AI got lost over those decades where it turned into narrow AI, where you are solving one specific problem at a time, whether it's playing chess or container packing optimization, you know, traffic optimization, or uh some expert systems for medical diagnosis or so.
• 16:56
  SPEAKER_00: And what is uh a very, very important point that actually is generally not at all appreciated is what happened to AI is the intelligence that we're seeing in things like deep blue or expert systems is actually not the intelligence in the AI, it's the intelligence of the programmer or the data scientist or whatever.
• 17:21
  SPEAKER_00: So the intelligence is actually external, somebody figuring out how can we build, how can we build a computer, what algorithms do we need, what tricks can we use to play a good game of chess?
• 17:33
  SPEAKER_00: And even today, with the big data approaches and statistical approaches, it's still the same thing, you know, protein folding and so there isn't intelligence that figured out how to do the protein folding thing that is inherent in there.
• 17:48
  SPEAKER_00: It's people figured out if we prepare the data in a certain way, and if we design the neural network in a certain way, and we experiment with this and experiment with that, then we can have a machine that can actually spit out, you know, can play Go or do protein folding or whatever, but it's solving one problem at a time, and the problem is really solved by the programmer or the data scientist.
• 18:13
  SPEAKER_00: And that realization of AGI, uh, AGI is getting back to the original dream of AI, is to have a machine that can figure these things out.
• 18:25
  SPEAKER_00: So it that's why learning is such an important thing and reasoning is so important.
• 18:30
  SPEAKER_00: Uh, so learning and reasoning are absolutely in which implies you know, memory, short-term memory, long-term memory, being able to use context, concept conf uh concept creation formation.
• 18:43
  SPEAKER_00: Um now, today's AI systems don't typically have any of those component components that are actually essential for real intelligence.
• 18:53
  SPEAKER_00: So there is it's really like night and day.
• 18:57
  SPEAKER_00: Almost everything that's done in the field of AI today is narrow AI.
• 19:02
  SPEAKER_02: Right, right.
• 19:04
  SPEAKER_02: So when you start thinking about you know applying those lessons uh that you learn from your study of intelligence and knowledge and and uh consciousness and psychology and learning and cognition, and you think about uh you know the history of AI, you know, where symbolic AI, knowledge representation was really the main game, you know, 50s, 60s, 70s, 80s, 90s, right?
• 19:31
  SPEAKER_02: And then statistical approaches came about uh and and started to uh eclipse uh knowledge representation and symbolic AI.
• 19:40
  SPEAKER_02: And now we're seeing people thinking, well, we can't forget symbolic, right?
• 19:44
  SPEAKER_02: We have to have knowledge representation.
• 19:46
  SPEAKER_02: Uh how how have you begun to uh uh incorporate in it sounds like the way you you're thinking about this, because you have to have knowledge representation, concept representation, and within those concept representations, the ability to infer other relationships or do some type of reasoning.
• 20:05
  SPEAKER_02: Talk a little bit about that evolution and and how you look at things like ontologies and uh symbolic AI versus uh uh machine learning uh and statistical types of approaches.
• 20:16
  SPEAKER_00: Do you want to speak to that kind of realm?
• 20:19
  SPEAKER_00: Sure.
• 20:20
  SPEAKER_00: So um DAPA actually a few years ago came up with a presentation that they called the three waves of AI or the third wave of AI.
• 20:31
  SPEAKER_00: And I found that a quite a useful model where they refer the first wave of AI, they're they they basically reference um uh sort of logic-based systems, I think is the way to think about it, which would include actually rules-based logic and expert systems and so on.
• 20:52
  SPEAKER_00: Yes, symbolic, uh, symbolic driven uh systems, you know, which deep blue again would be uh the good the good example of that.
• 20:59
  SPEAKER_00: And that dominated AI for for many decades, even though neural networks had also been studied, they they never really quite made it uh in comparison to the symbolic approaches.
• 21:13
  SPEAKER_00: So that's the first wave.
• 21:14
  SPEAKER_00: And then, of course, we've we had the second wave hit us like a tsunami about 10 years ago, when the big data statistical approaches really started to work, you know, the neural neural networks of of of a certain type.
• 21:30
  SPEAKER_00: And it was really the the breakthrough of uh deep pocketed companies having a lot of data, having a lot of computing power, that they could get really very significant, significantly better results uh with neural networks, with statistical systems rather than um um symbolics.
• 21:51
  SPEAKER_02: And that was a part that was part due to data sources, computational power, plus yeah, very much so.
• 21:57
  SPEAKER_00: They uh you know there were uh some.
• 22:00
  SPEAKER_00: Some sort of minor breakthroughs, I would say, in terms of uh neural network architecture, but really not that revolutionary.
• 22:07
  SPEAKER_00: It was more a matter of, you know, hey, suddenly this stuff is starting to work when we throw, you know, a thousand times or million times more data at it than we had ever done before.
• 22:19
  SPEAKER_00: So, you know, speech recognition and image recognition suddenly became, you know, they beat all the benchmarks that had previously been set up by, you know, symbolic uh approaches or logic approaches.
• 22:31
  SPEAKER_00: Um, so and and we're still in that era in the last 10 years is basically big data.
• 22:37
  SPEAKER_00: And it's it's now sort of become self-perpetuating in the sense that uh it's been so incredibly successful that you know more virtually unlimited amounts of money have flown into this this area.
• 22:52
  SPEAKER_00: Uh I mean, Microsoft just okay, we'll write a check for billion dollars to open AI, you know, and um, and now they're gonna increase their investment, you know.
• 23:01
  SPEAKER_00: So uh the it's it's really the only game in town uh because it can do so many things so um impressively.
• 23:10
  SPEAKER_00: Um, however, it's still narrow AI, and it's still there's no thinking, there's no learning.
• 23:19
  SPEAKER_00: Once you build the model, the model is essentially a read-only model.
• 23:23
  SPEAKER_02: Uh so there is there's learning in that there's optimization and maybe during the training process, absolutely.
• 23:32
  SPEAKER_00: Yeah, absolutely.
• 23:33
  SPEAKER_00: But once but once you you deploy the model, it's essentially a read-only model.
• 23:38
  SPEAKER_00: There isn't really uh significant learning, there's no no reasoning.
• 23:42
  SPEAKER_00: So that's the second wave that we're in.
• 23:44
  SPEAKER_00: And then DARPA talked about the third wave of AI, which are you can describe more as cognitive systems.
• 23:50
  SPEAKER_00: You know, there we are getting back to answering the question of what does intelligence require?
• 23:57
  SPEAKER_00: You know, intelligence requires reasoning, uh, adaptation, you know, which means learning, contextual um interpretations, and and so on.
• 24:07
  SPEAKER_00: So, you know, the things that are uh I listed before, which really is if your starting point is what does intelligence require?
• 24:15
  SPEAKER_00: The kind of human-level intelligence that we're interested in, not narrow intelligence of playing a good game of go.
• 24:23
  SPEAKER_00: Uh, but what does general intelligence require?
• 24:26
  SPEAKER_00: Um, really, the second wave doesn't have any of those, and the third wave addresses addresses that.
• 24:32
  SPEAKER_00: And a good way to uh uh to refer to them, I think, are cognitive architectures is to have an architecture that inherently addresses the requirements of uh intelligence.
• 24:44
  SPEAKER_02: And in that cognitive uh architecture, we're really talking about you know a hybridization of the symbolic and machine learning.
• 24:52
  SPEAKER_02: Yes.
• 24:54
  SPEAKER_00: You are referring to hybrid AI these days, and it seems that yeah, and and I I want to be a little bit careful about the hybrid because um you know people are definitely talking about the limitations of the second wave of of statistical systems, you know, stochastic parrots, uh, you know, and and and terms like that.
• 25:16
  SPEAKER_00: So uh at the moment in the mainstream AI world, there isn't any consensus of how to overcome them, overcome these limitations, but there's a general sense that we need we need some we need reasoning, you know, and um we somehow need to be able to handle language symbolically, we need to be able to do mathematics, you know, some we need symbol manipulation.
• 25:43
  SPEAKER_00: So they're talking about uh combining first wave and second wave.
• 25:48
  SPEAKER_00: But as long as that combination is just sort of, you know, we have this system first wave and the system second wave, and we kind of glue them together in some way, that's not really gonna work.
• 26:00
  SPEAKER_00: Right, right.
• 26:01
  SPEAKER_00: It's it's so yeah.
• 26:02
  SPEAKER_00: So it has to be if you're going to use the term hybrid, it has to be deeply integrated at every level, the the technologies.
• 26:11
  SPEAKER_00: But it's to me, it's it's the wrong way of thinking about it.
• 26:14
  SPEAKER_00: Your starting point is not we need a bit of this and it need a bit of that.
• 26:18
  SPEAKER_00: Your starting point really needs to be what does intelligence require?
• 26:22
  SPEAKER_00: You know, start more with a clean slate, not hey, we have these super powerful um models that we've built with machine learning, deep learning.
• 26:30
  SPEAKER_00: How can we tweak tweak them to get to where we want to go?
• 26:35
  SPEAKER_02: Interesting.
• 26:35
  SPEAKER_02: Uh, I do want to dig deeper on that, but I we're about halfway through, and I wanted to remind folks that uh today we're speaking with uh Peter Voss, who is an AI pioneer and uh coined the uh the term our artificial general intelligence, uh AGI, and we are just talking about the third wave of artificial intelligence.
• 26:56
  SPEAKER_02: We'll be right back.
• 26:58
  SPEAKER_02: We'd like to thank Marketing Artificial Intelligence Institute for their support of this podcast.
• 27:03
  SPEAKER_02: Marketing AI Institute is a media, event, and online education business that makes AI approachable, actionable, and accessible to marketing leaders around the world.
• 27:12
  SPEAKER_02: They run the world's leading marketing AI conference, MACOM, which brings together hundreds of marketing and business leaders to better understand, pilot, and adopt AI.
• 27:22
  SPEAKER_02: Today's podcast is brought to you by Simpler Media Group, owner of the Customer Experience Decoded Podcast, covering the latest in customer experience and marketing.
• 27:30
  SPEAKER_02: CX Decoded is a semi-monthly podcast designed to deliver insight from leaders driving the evolution of customer experience with an emphasis on digital touch points, customer journeys, respectful data management, and voice of the customer.
• 27:59
  SPEAKER_04: Their information architecture expertise has allowed clients from all over the world to efficiently organize, access, and utilize their data, ultimately resulting in better business outcomes.
• 28:08
  SPEAKER_04: They deliver customer-driven solutions that scale and adapt to the needs of a business as it evolves.
• 28:14
  SPEAKER_04: Founded in 1994, Early Information Science has proudly delivered result-driven solutions for more than 100 brands across the top industries.
• 28:22
  SPEAKER_04: You can get more information about early information science at early.com.
• 28:27
  SPEAKER_04: E-R-L-E-Y.com.
• 28:31
  SPEAKER_04: Working with customers in AI, the number one question that I get usually is where do I start?
• 28:36
  SPEAKER_04: What data do I need?
• 28:37
  SPEAKER_04: How do I gather it?
• 28:38
  SPEAKER_04: And it's great questions.
• 28:40
  SPEAKER_04: Well, if you want to catch the wave of AI, not be crushed by it, you really do need to read a book written by Seth Early called The AI Powered Enterprise.
• 28:48
  SPEAKER_04: It gives you the ability to look and understand how to harness the power of ontologies to make your business smarter, faster, and more profitable.
• 28:56
  SPEAKER_04: Seth goes into the intricate details of how to build an ontology, develop and employ an ontology as a secret weapon for the basis for how to actually utilize AI as a powerful competitive advantage.
• 29:08
  SPEAKER_04: This book is really the first one to combine sophisticated explanation of how AI works with a practical approach of applying AI to the problems of the business.
• 29:17
  SPEAKER_04: And it's applied to, you know, various use cases from customer experiences to business operations to product development.
• 29:23
  SPEAKER_04: It's a powerful, powerful tool in the arsenal of understanding how to deploy AI to make your business better.
• 29:29
  SPEAKER_04: The AI Powered Enterprise by Seth Early.
• 29:31
  SPEAKER_04: Harness the power of ontologies to make your business smarter, faster, more profitable.
• 29:35
  SPEAKER_04: Out on Amazon.com today.
• 29:37
  SPEAKER_04: And now back to the show.
• 29:39
  SPEAKER_02: Let's talk, let's keep going.
• 29:40
  SPEAKER_02: You mentioned cognitive architecture.
• 29:42
  SPEAKER_02: So talk about what a cognitive architecture looks like and how that level of integration, you know, and and uh uh hybridization.
• 29:52
  SPEAKER_02: I'm using just a term to kind of talk about that.
• 29:55
  SPEAKER_02: You're talking about a clean sheet approach, but you know, talk a little bit about what a cognitive architecture is and uh what is that path to to getting there today.
• 30:05
  SPEAKER_02: So maybe you can expand on the components of it.
• 30:08
  unknown: Right.
• 30:08
  SPEAKER_00: So uh cognitive architectures have also been around for a while.
• 30:13
  SPEAKER_00: And uh, you know, often the criticism now would be well, we've tried this for decades and it hasn't worked.
• 30:19
  SPEAKER_00: Though I'd like to just remind people that the same thing was said about neural networks 10 years ago.
• 30:25
  SPEAKER_00: You know, we've tried that for decades and it hasn't worked.
• 30:27
  SPEAKER_00: Well, it didn't work until it worked, you know, until somebody figured out how to get it to work.
• 30:33
  SPEAKER_00: So I think cognitive architectures uh yes, it uh are the right approach to the third wave and to getting to human-level intelligence.
• 30:42
  SPEAKER_00: And and really, uh, you know, as I've mentioned already a few times, your starting point has to be what does intelligence require?
• 30:50
  SPEAKER_00: Now, an intelligence system um requires long-term memory.
• 30:56
  SPEAKER_00: So now our memory, human memory is is certainly opaque.
• 31:03
  SPEAKER_00: I mean, we can't pinpoint, you know, I mean, there's uh it's a it's a complex topic, you know, people say there's a grandmother neuron or you know, that that kind of thing.
• 31:13
  SPEAKER_00: But uh by and large, you can say that our our neural network is is really not scrutable, it's pretty opaque, you know, of what's happening where, when.
• 31:24
  SPEAKER_00: Um, but from an engineering point of view, you you really would much prefer to have something that's scrutable.
• 31:32
  SPEAKER_00: So to me, the starting point is if we can have largely a system that's essentially um uh a semantic network of some sort, a knowledge graph, that that is the foundation of it.
• 31:45
  SPEAKER_00: Because computers, you know, we are talking about computers, one of their strengths is they can have access to massive amounts of data, databases and structured data and so on.
• 31:56
  SPEAKER_00: So, why throw that out?
• 31:58
  SPEAKER_00: So I think you do really want to have uh a knowledge graph as the basis, the foundation of it.
• 32:05
  SPEAKER_00: But that knowledge graph can't be based on traditional uh Boolean logic, because the knowledge we have about the world um is often uh fuzzy, you know, there it's contextual, it's um uh to some degree statistical.
• 32:25
  SPEAKER_00: You may have seen X number of versions of this and the Y number of versions of you know, the opposite.
• 32:32
  SPEAKER_00: Uh, it needs to be able to hold contradictions, and you know, traditional logic systems can't do that.
• 32:39
  SPEAKER_00: Um so you may learn two contradictory uh uh versions of something, and you need to be able to hold both of them, you know, pending resolution.
• 32:50
  SPEAKER_00: Or maybe this is true in this situation, that's true in that situation, uh, or maybe you simply haven't resolved it, you know, and you but you need to be able to capture that knowledge.
• 33:00
  SPEAKER_00: So you need a very uh flexible and and sort of uh fuzzy contextual uh knowledge representation as the key.
• 33:10
  SPEAKER_00: Um, you also need short-term memory, which is sort of the activation.
• 33:14
  SPEAKER_00: What is currently, what is the context currently?
• 33:18
  SPEAKER_00: Who are you talking to, or what are you doing?
• 33:21
  SPEAKER_00: Uh what do you already know if you're having a conversation?
• 33:25
  SPEAKER_00: What do you know about that person?
• 33:27
  SPEAKER_00: Uh, even theory of mind, what do they know?
• 33:30
  SPEAKER_00: What do you expect them to know?
• 33:32
  SPEAKER_00: What don't you expect them to know?
• 33:34
  SPEAKER_00: You know, what are the goals?
• 33:36
  SPEAKER_00: So you you need a uh uh a reasoning system that can take into account your long-term permanent short, you know, uh long-term uh memory or uh knowledge.
• 33:48
  SPEAKER_00: Um but combine that with the current context, short-term uh and and goals uh to be able to reason.
• 33:57
  SPEAKER_04: So, you know, I read it, I I read an article that you had you wrote.
• 34:01
  SPEAKER_04: You made it's they might have been quoting you, or you were you know speaking ours and article whereas it was literally around the the foundations of AGI are really centered in quality data, not quantity of data.
• 34:13
  SPEAKER_04: Right?
• 34:14
  SPEAKER_04: I'd love to get your thoughts on that and how that fits into a reasoning system, like we're talking about, where you have short-term and long-term memory, because one denotes maybe more data for long-term, or what that looks like.
• 34:30
  SPEAKER_04: But I'd love your thoughts on the quality of this quantity discussion.
• 34:33
  SPEAKER_00: Right.
• 34:34
  SPEAKER_00: So the distinction that I make here is actually also a very practical one.
• 34:40
  SPEAKER_00: And and I mean, you know, we've been talking about this sort of really sort of from a theoretical and research perspective, but there's also the practicality of building systems that can actually do things in the real world.
• 34:54
  SPEAKER_00: So um you have these systems that have, you know, billions and billions, or maybe even a trillion parameters that have like all the knowledge that's on on the internet and books and whatever, you know, GPT3 has all of that information, uh, all the possible information you could have.
• 35:13
  SPEAKER_00: But it's statistical, you know, it's basically what is the most likely thing that the person will say next.
• 35:19
  SPEAKER_00: Now, you you try and implement this in a business scenario.
• 35:24
  SPEAKER_00: You know, we we provide um a chatbot, conversational AI for um for enterprise customers.
• 35:33
  SPEAKER_00: Now, you don't you can't have a system that has all the knowledge of of the world in there and might respond in some totally inappropriate way, not you know, that that may be the most common way somebody has responded to that, but it's totally uh out of context, totally inappropriate to the current situation.
• 35:55
  SPEAKER_00: So, you know, for example, one of our uh one of our clients is 1 800 flowers, a group of companies was gifting.
• 36:02
  SPEAKER_00: So when you're having when a customer is having a conversation with with our AI, you really need to make sure that the information you provide is accurate, that you can rely on on the information, not you know, just sort of a a random um statistical um uh you know draw draw of luck.
• 36:24
  SPEAKER_02: Um so fitness to purpose of the information.
• 36:28
  SPEAKER_00: Correct.
• 36:29
  SPEAKER_00: And and so this is why I say the quality of the information you have is much more important for critical applications and the quantity.
• 36:37
  SPEAKER_00: Now, if you watch this want to be entertained, then maybe having a you know a lot of different things that you can talk about and get some funny, funny responses, then maybe quantity is uh the the right measure.
• 36:51
  SPEAKER_00: But if you actually have if you're trying to achieve an a particular objective, then quality of data is is uh much much more critical.
• 37:01
  SPEAKER_02: Right, especially when you're talking about um you know things that have a regulatory or compliance implication or things where you're you know you're advising a customer or you're helping them make a choice or a decision.
• 37:13
  SPEAKER_02: So all of that does need to be uh intentionally.
• 37:17
  SPEAKER_00: Well, it in in fact, I mean, even if you're not in you know banking or insurance or in one of these regulate environments, even if even if in normal any kind of business, uh it has to pass legal review, it has to pass uh marketing review, and it has to, you know, pass uh you know various reviews of of how you know business business rules that you have and you know how you want to interact with the customer.
• 37:46
  SPEAKER_00: So yeah, that's true really for any any business.
• 37:50
  SPEAKER_02: Sure.
• 37:51
  SPEAKER_02: So quick question when you start thinking about long-term memory, short-term memory, and when I think of uh you know some of the things that we work on around making predictions, orchestration, personalization, a lot of that is based on uh reading that digital body language of the customer, right?
• 38:08
  SPEAKER_02: What are they doing?
• 38:09
  SPEAKER_02: What was their last click?
• 38:10
  SPEAKER_02: What was their purchase?
• 38:11
  SPEAKER_02: Did they have a help ticket?
• 38:13
  SPEAKER_02: And and part of it is also understanding are they technical uh in terms of background, or are they a lay person or a business customer, consumer?
• 38:21
  SPEAKER_02: And some of that is uh analogous to your short-term memory, right?
• 38:24
  SPEAKER_02: What do you expect them to know?
• 38:25
  SPEAKER_02: What do you what are you anticipating?
• 38:27
  SPEAKER_02: What do you know about them and what do you expect them to be able to know?
• 38:30
  SPEAKER_02: So you can give them and a more appropriate answer, a more contextual answer, more protect, you know, next best product, next best action, and so on.
• 38:38
  SPEAKER_02: So when you and that can be represented in a customer identity graph and a knowledge graph and those types of things.
• 38:44
  SPEAKER_02: When you start thinking about longer-term uh memory, that's really looking at past purchases, perhaps, what they you know bought for a uh wife uh last birthday, whatever it might be.
• 38:57
  SPEAKER_02: Right.
• 38:57
  SPEAKER_02: So so where do you see those uh that type of short and long-term memory when you start looking at really building these cognitive assistants?
• 39:05
  SPEAKER_02: Is that a matter of connecting to corporate knowledge bases, uh uh uh uh transactional systems, ERP systems, order management systems?
• 39:14
  SPEAKER_02: Talk a little bit about the kind of a practical ecosystem in a typical corporate ecosystem and how those things are are uh are managed.
• 39:24
  SPEAKER_00: Yeah, you bring up bring up some good points here that um it really all comes under the the in my mind under the heading of context when you're having a conversation.
• 39:35
  SPEAKER_00: What what do you need uh to establish the right context for your conversation?
• 39:40
  SPEAKER_00: And that would include long-term memory of interactions that the aid the AI has had with that customer before, where you may have learned something that's not in a corporate database, but just you know, their their preferences or their their manner of speaking or whatever, you know.
• 39:58
  SPEAKER_00: So the long-term memory, short-term memory is is really the same thing.
• 40:02
  SPEAKER_00: And in fact, in our architecture, uh the the only distinction between long and short-term memory is that certain um certain parts of short-term memory you don't need to permanently remember.
• 40:15
  SPEAKER_00: You know, if they're like just courtesies and um, you know, things that are not relevant, you you wouldn't like we don't remember remember them.
• 40:24
  SPEAKER_00: There's no point in remembering them, kind of clogging up your long long-term memory.
• 40:28
  SPEAKER_00: So short-term, long-term memory are really just what you've learned from prior interactions versus what you learned from the current interaction in terms of their goals and preferences and and and so on.
• 40:40
  SPEAKER_00: Um now in in commercial applications, invariably, though, you also have back end systems that have information, you know, whether it's product availability or whether it's customer history or customer preferences or whatever it might be.
• 40:57
  SPEAKER_00: So you want to engineer a system that it can very rapidly and seamlessly integrate that information as required.
• 41:07
  SPEAKER_00: And, you know, that I mean it's typically done through APIs that uh during the conversation, if you need some additional information that's relevant, then you'll hit an API, get the information, and becomes part of that knowledge graph, either permanently or temporarily, either either as short or long-term memory, but you can use it as part of the context.
• 41:28
  SPEAKER_00: Um, and of course, in the same way, things that you learn, conclusions that you draw, actions you want to take, will then also be actioned back, you know, through APIs back to whatever uh backend system you have.
• 41:40
  SPEAKER_00: But I I think one one of the really important distinctions here again between um statistical systems or general systems that are mostly in use, they tend to treat you as a demographic.
• 41:55
  SPEAKER_00: That basically you are put in a certain bucket.
• 41:58
  SPEAKER_00: Oh, you're this kind of customer or you're that kind of customer.
• 42:01
  SPEAKER_02: Collection of attributes.
• 42:03
  SPEAKER_00: Yeah, and and that's ultimately really not what you want as a customer or uh as as a um company, if you really want to provide high-quality service, that every individual customer is a unique individual with a unique history, you know, and unique requirements, even if they overlap largely with what other customers, but still there's a very specific context for this customer at this time.
• 42:30
  SPEAKER_02: Right.
• 42:31
  SPEAKER_02: And that does at the end of the day, it still does have to be represented by some type of attribute model, right?
• 42:37
  SPEAKER_02: Because you're such a responding to.
• 42:39
  SPEAKER_02: And that but that attribute model is very dynamic and very specific to that, uh, to that individual customer.
• 42:47
  SPEAKER_00: Correct.
• 42:47
  SPEAKER_00: I mean, the way we approach that, uh, the the the knowledge graph, uh, the complex knowledge graph that we have, there is actually a part of that graph that is absolutely unique for every individual customer.
• 43:00
  SPEAKER_02: Yep, I can see that.
• 43:01
  SPEAKER_02: I can see that.
• 43:02
  SPEAKER_02: It's a kind of the metadata cloud that we can all walk around with, right?
• 43:06
  SPEAKER_02: Our preferences, our history, our knowledge, our expectations, our thinking style, our communication style, all of those things that can be derived or applied.
• 43:16
  SPEAKER_02: Yeah.
• 43:17
  SPEAKER_02: And and so when you look at your uh uh uh engagements and your deployments of uh virtual assistants, you know, what what needs to be done?
• 43:26
  SPEAKER_02: You know, how do organizations need to prepare to build these types of more advanced um virtual assistants and bots?
• 43:36
  SPEAKER_02: Uh, you know, what kinds of what do they have to do to get their house in order in order to do this, right?
• 43:40
  SPEAKER_02: You can't start from scratch.
• 43:42
  SPEAKER_02: You got to have certain preconditions.
• 43:44
  SPEAKER_00: Right.
• 43:44
  SPEAKER_00: I mean, this the same really is true for any systems you implement in a company or automation you do.
• 43:52
  SPEAKER_00: Um the better organized the company is, the clearer it is in terms of what they do and what their rules are and how.
• 44:00
  SPEAKER_00: They want to operate, the easier it's going to be to implement.
• 44:04
  SPEAKER_00: Of course, in reality, most companies aren't that well organized.
• 44:10
  SPEAKER_00: They're always in the process of, you know, changing something, replacing systems, and you know, uh, and and so on.
• 44:18
  SPEAKER_00: Now, the way we often cut through that is if there is uh a live customer service, which you know most companies have, you talk to an agent, the question is how does the agent solve, you know, answer the answer questions?
• 44:34
  SPEAKER_00: What training material do you have?
• 44:37
  SPEAKER_00: Uh, what systems do they have access to?
• 44:40
  SPEAKER_00: And that's kind of our baseline because you know, we're talking about trying to uh provide service similar to what uh human agents are doing, but in many ways better than a human can.
• 44:53
  SPEAKER_00: And I'd like to talk about that that too.
• 44:55
  SPEAKER_00: But that's kind of often our starting point, is simply to say, how do your agents do it?
• 45:00
  SPEAKER_00: Okay.
• 45:00
  SPEAKER_00: Right, right.
• 45:01
  SPEAKER_00: You know, how do you train them?
• 45:02
  SPEAKER_00: Do you have training material?
• 45:04
  SPEAKER_02: You know, yeah, I like to say the same things you need to train train a bot or a virtual assistant.
• 45:09
  SPEAKER_02: You need to train a human, right?
• 45:10
  SPEAKER_02: So you can uh build you can build these capabilities and knowledge bases.
• 45:15
  SPEAKER_02: Right.
• 45:15
  SPEAKER_00: Except we don't we we don't yeah, we don't need hundreds of thousands or millions of examples to train a statistical model.
• 45:23
  SPEAKER_00: Uh it's more like a human, you know.
• 45:26
  SPEAKER_00: Here, read read this manual here, you know.
• 45:28
  SPEAKER_00: Here are here are the answers.
• 45:30
  SPEAKER_00: Right.
• 45:30
  SPEAKER_04: You know, here also you're you're also, Peter, you're also getting into a scenario.
• 45:35
  SPEAKER_04: I find a lot of um customers that I've worked with in the past believe that there's this notion of one bot or one digital assistant to rule them all.
• 45:46
  SPEAKER_04: And you know, and that's it's just not feasible in, you know, for let's say an organization who's looking for, like you said, you can't train everything, you can't, you know, have the world's knowledge in a single bot.
• 45:58
  SPEAKER_04: How do you advise customers that are looking for these scenarios and saying, hey, listen, I need a bot, I need a digital assistant.
• 46:05
  SPEAKER_04: They're in customer experience, right?
• 46:07
  SPEAKER_04: Or they're doing field service, or they're doing in you know, customer self-service.
• 46:13
  SPEAKER_04: How do you advise them around, let's say, the notion of having multiple bots that can potentially learn from each other or like a network of them, or do you do you have actually you know paint them or point them to a picture of a one size fits all type of perspective?
• 46:29
  SPEAKER_00: How do you help them understand?
• 46:31
  SPEAKER_00: Yeah, so uh actually I have a uh the opposite approach of what what you're talking about is in fact, I think within an organization, you should have one bot.
• 46:44
  SPEAKER_00: Uh and um the reason for that is we call ours a chat bot with a brain, is that you build out this corporate brain that you know across different things, whether it's sales or service, you know, or something else, certainly customer facing, you need the same knowledge base.
• 47:03
  SPEAKER_00: So if a customer goes from sales to service or service back to sales, you really want to be able to share that information.
• 47:10
  SPEAKER_00: You want to be able to seamlessly go from one to the other.
• 47:14
  SPEAKER_00: You know, you might be at service and then says, well, actually, let me place another, let me place a new order, or what products do you have?
• 47:22
  SPEAKER_00: You want the same system.
• 47:25
  SPEAKER_00: In fact, this is again one of the advantages of uh automation as opposed to human, because chances are you get put, I mean, we all know this experience with dealing with banks and insurance companies.
• 47:36
  SPEAKER_00: You get to one department, they can only do this one narrow function.
• 47:41
  SPEAKER_00: And if you want something else, or let me transfer you, you know, and then another 30-minute wait and you have to tell your whole story again from the beginning.
• 47:49
  SPEAKER_00: So you really want to have that corporate brain that can know as much as possible.
• 47:54
  SPEAKER_00: Now, you know, there might there might be less overlap between customer facing and internal functions, you know, if you're also using the bot for HR and internal uh you know tech support or something.
• 48:09
  SPEAKER_00: But even there, you you would still have uh quite a lot of the same APIs and the background information and you know uh things that are shared.
• 48:18
  SPEAKER_00: So you really do want to have uh one bot that can uh learn learn all of that.
• 48:25
  SPEAKER_00: And you know, we uh again going back to the 1-800 flowers example, we started the implementation on uh they have uh 12 companies are in the group.
• 48:35
  SPEAKER_00: We started with one company, one channel, and one application, you know, with 1-800 flowers.
• 48:41
  SPEAKER_00: Where's where's my order on ABC chat?
• 48:45
  SPEAKER_00: And once we'd implemented that, it became a lot easier to implement other channels.
• 48:50
  SPEAKER_00: We now have five channels, all 12 companies, and you know, about a hundred different applications, but they could build on each other and you can seamlessly go from one to the other and carry the context across.
• 49:03
  SPEAKER_00: And you can even have a true omni-channel where you can have start a phone conversation, transition to chat seamlessly, carrying the context over, um, might have an SMS interchange, you know, or even email, all handled by the same by the same brain.
• 49:22
  SPEAKER_00: So I think that's actually the ideal uh scenario rather than having all these siloed bots that don't talk to each other.
• 49:30
  SPEAKER_02: Right.
• 49:30
  SPEAKER_02: So you may so my my perspective on that uh I think is kind of a uh maybe maybe a uh compromise uh because when you start saying uh 100 applications or multiple applications, you're really talking about a framework, you know, bot framework.
• 49:47
  SPEAKER_02: And maybe there's going to be a repository in one department that uh the bot has to go to, or repository in another department.
• 49:55
  SPEAKER_02: So it's kind of where what are the boundaries of what we call a bot versus an application versus a knowledge source, right?
• 50:02
  SPEAKER_02: So the framework should be uh uh holistic framework, but there'll be multiple knowledge sources and potentially multiple applications because for any typical organization, they will have lots and lots of initiatives, right?
• 50:16
  SPEAKER_00: And the question absolutely, but you really want to have one brain, you know, so that the context can be carried across um that you know any shared business rules, any any you know, shared information truly is available.
• 50:32
  SPEAKER_00: Um, and as long as you're building individual bots, you're not going to uh to have that that kind of context carried but you know between bots.
• 50:42
  SPEAKER_00: Yeah.
• 50:43
  SPEAKER_02: Yeah, and I guess the question becomes what is what is considered an individual bot versus what is considered a framework?
• 50:49
  SPEAKER_02: Uh get into the semantics of that, but uh I think I understand your point.
• 50:54
  SPEAKER_02: Your point is to have the same reference architecture and the same uh knowledge uh structure and the same reasoning ability that can integrate across these different um compartments and silos and processes.
• 51:08
  SPEAKER_00: Yeah, correct correct.
• 51:10
  SPEAKER_00: Yeah, yeah.
• 51:11
  SPEAKER_04: And it's you know, to your point, Seth, it's that context carryover, right?
• 51:16
  SPEAKER_04: I mean, yeah, you're spot on because it's it's you're only your bot is only as good as the data that it can actually access, right?
• 51:24
  SPEAKER_04: And if you can, if you can, like you said, the having the framework there of understanding what to go access, what to go get, because for instance, you're not gonna, you know, let let your bot go and get personal information about you and carry that over to, let's say, other systems that don't need access to it, right?
• 51:41
  SPEAKER_04: So compliance, security, and across however at the same time, being able to authenticate over here in in let's say a customer service example, and then carrying that over to let's say, you know, not to be, you know, too mundane in retail, but to be able to shop and stuff too, and it has the context of both.
• 52:00
  SPEAKER_04: Now we're getting to a uh an area where we're actually integrating interacting with a personified piece of of silicone that actually is helping me self-serve and do what I need to, and at the same time solving my problems, right?
• 52:14
  SPEAKER_04: So then I can actually, you know, get stuff done without having to or get get my needs, you know, uh satisfied without having to, you know, take any, take any more or whatever else.
• 52:25
  SPEAKER_00: So right, right.
• 52:26
  SPEAKER_00: Yeah.
• 52:27
  SPEAKER_00: I mean, it's it's really the the you know, the the short-term memory, the context, everything that has to be carried through.
• 52:34
  SPEAKER_00: And you can't really do that if you're going to be handing it over to another piece of architecture that has to kind of then get booted up and configured, and and and it makes no sense from a development point of view as well.
• 52:46
  SPEAKER_00: It's it's much cheaper and easier to add functionality to an existing uh system if it's designed to work that way, you know.
• 52:55
  SPEAKER_00: I mean, we we we you know we we talked to banks, and I remember one of them, they said, you know, we've got you know, we've got 12 different bots, and we've got like 13 problems.
• 53:07
  SPEAKER_00: You know, it's like they don't talk to each other, each one has a separate development pass, customer experience is very different.
• 53:14
  SPEAKER_00: You know, you change something on one bot, and you know, then the others work differently, it's a different user interface, a different experience, they don't talk to each other, and you know, it's no, it's it's it's it's not not not a good way to go if you uh if you can avoid it.
• 53:31
  SPEAKER_02: Yeah, uh yeah, go ahead, Peter.
• 53:34
  SPEAKER_00: I I do want to just quickly mention something that's also not super obvious, and I touched on that.
• 53:40
  SPEAKER_00: And is there's a general uh belief that you know chatbots and IVRs uh can help you save money.
• 53:48
  SPEAKER_00: And it's you know, even the term containment is sort of really a negative one for one one for me, is um you really want to think about uh automation and conversational AI as offering a much superior experience to your customer.
• 54:05
  SPEAKER_00: It's not the second class, it's actually going to be the first class experience.
• 54:10
  SPEAKER_00: And the reason for that is 24-7 available, no wait time, and you can actually have a hyper-personalized uh conversation.
• 54:20
  SPEAKER_00: So, for example, if you're calling about your Wi-Fi that's not working, you know, and the first interaction says, Well, have you tried rebout rebooting your your your router?
• 54:30
  SPEAKER_00: You know, uh said, Okay, let me try that.
• 54:33
  SPEAKER_00: And then, you know, 10 minutes later you call call back and you talk to a human, they say, Have you tried to router?
• 54:41
  SPEAKER_00: Right, you know, whereas talking to to Igo, you know, to to a uh intelligent uh chatbot, it will have that history and say, Okay, did that work?
• 54:51
  SPEAKER_00: You know, and then okay, it's not working.
• 54:54
  SPEAKER_00: Okay, try moving it into the kitchen, you know, and then half an hour later you say, Well, did moving it to the kitchen help?
• 55:01
  SPEAKER_00: You know, right, right.
• 55:02
  SPEAKER_00: You cannot get that kind of experience with a human, uh, quite apart from training limitations, you know, and and all of the other limitations you have in in in uh live call centers.
• 55:13
  SPEAKER_02: That's what's cool.
• 55:14
  SPEAKER_02: Well, Peter, uh, this we've come to the end of our time today.
• 55:18
  SPEAKER_02: It's really been wonderful.
• 55:19
  SPEAKER_02: I'm so glad we had a chance to explore this.
• 55:21
  SPEAKER_02: That hour went by in a blank.
• 55:23
  SPEAKER_02: You know, we could spend hours talking about this.
• 55:26
  SPEAKER_02: It's always a pleasure to catch up with you.
• 55:28
  SPEAKER_02: So thank you so much uh for your time today.
• 55:31
  SPEAKER_02: It's really been uh it's been great.
• 55:32
  SPEAKER_02: And I can't wait to uh talk to you in six months or a year and see how things have continued to evolve because I I know you guys are making tremendous progress over there at uh at IGO.
• 55:43
  SPEAKER_02: Thank you again for your time today.
• 55:44
  SPEAKER_00: All right, thank you.
• 55:45
  SPEAKER_00: It was fun.
• 55:45
  SPEAKER_00: Thank you.
• 55:46
  SPEAKER_00: Thanks, Peter.
• 55:50
  SPEAKER_01: Thanks for listening to the Early AI podcast with your host Seth Early and Chris Featherstone.
• 55:57
  SPEAKER_01: We hope you enjoyed the program and took away some ideas that you can start to implement in your organization right now.
• 56:03
  SPEAKER_01: If you do need help, check out www.early.com for information about early information.
• 56:11
  SPEAKER_01: And of course, don't forget the link to any resources we need to do.