Debbie Reynolds “The Data Diva” talks to Jim Barnebee, CEO, AIM-E (Artificial Intelligence Made Easy). We discuss our joint work on next generation connected systems with IEEE, his background in programming and development including work with IBM on Watson, bad data in and bad data out, AI does what you program it to do, in depth analysis of input data required for successful language models for AI projects, difficulty of adapting generalized AI to problems, the need for correct analysis of the problems to be solved by AI, why data can affect AI performance and his hope for Data Privacy in the future.
Debbie Reynolds “The Data Diva” talks to Jim Barnebee, CEO, AIM-E (Artificial Intelligence Made Easy). We discuss our joint work on next generation connected systems with IEEE, his background in programming and development including work with IBM on Watson, bad data in and bad data out, AI does what you program it to do, in depth analysis of input data required for successful language models for AI projects, difficulty of adapting generalized AI to problems, the need for correct analysis of the problems to be solved by AI, why data can affect AI performance and his hope for Data Privacy in the future.
52:37
SUMMARY KEYWORDS
people, data, ai, system, query, math, problem, run, artificial intelligence, area, ontologies, sentiment analysis, ontological, individuals, algorithm, determine, talk, person, natural language, build
SPEAKERS
Debbie Reynolds, Jim Barnebee
Debbie Reynolds 00:00
Personal views and opinions expressed by our podcast guests are their own and are not legal advice or official statements by their organizations. Hello, my name is Debbie Reynolds; they call me "The Data Diva". This is "The Data Diva" Talks Privacy podcast where we discuss Data Privacy issues with industry leaders around the world with information that businesses need to know now. Our special guest on the show is Jim Barnebee; he is the Chief Executive Officer of AIM-E, Artificial Intelligence Made Easy. Welcome.
Jim Barnebee 00:41
Thank you. Glad to be here.
Debbie Reynolds 00:43
So you and I met; we actually have a mutual friend, Vikas Malhotra, who is the head of WOPPLI Technologies. And I know you're a member of his board of advisors. But we got together and we collaborated with IEEE with Vikas. And that's how we met. So I think we were on a call one day and you were just brilliant. I thought oh my god, I have to have him on the show. Because I feel like you have such a deep knowledge and you're easy to talk to. And I thought that my audience would really love to hear about you and the things that you work on in Artificial Intelligence. So why don't you start out by giving us an idea of your career journey in AI and the things that you do at Artificial Intelligence Made Easy.
Jim Barnebee 01:39
Of course. So I'll try to be brief about my career history. I've been doing this for, like 40 years. So I got my first professional programming job when I was 16. So yeah, I've been doing this for a long time; I started out doing development as most people do. I was a beta tester. For Java, I helped create some of the Java specifications. I was on the committees that Sun built in order to continue Java as a growing form. Once Java became super stable, I moved on to architecture work. And I was a UML evangelist for First at first because that was a great way to do software engineering. And then I became an ontological evangelist. And ontologies are not a subject most people are familiar with. So let me explain it. An ontology is an overlay of data. So if you have a bunch of heterogeneous data sources around the world, and you want to be able to run queries on those data sources as a data set, there are a couple of different ways to do that. If you can take all of the data and conglomerate it into one location into a data lake, you can obviously run predictive analytics on it and move forward from there. If however, you can't move the data, right, you can't move it; you can't change it; it's got to stay in place. What do you do? Well, what you can do is create what they call an ontological overlay, which is essentially where you define the business concepts that are in your data. And then you map those concepts to all the different databases, then you can use one query structure to query all of the databases in the world at once. So what it does is produce a graph. So it's just grab, just set math, right? Set mathematics, it produces a graph, and you are able to query that graph. And then that graph goes out to the data sources and pulls the data back that is relevant to the query that you've presented. So that's how ontologies work. So I was an ontology evangelist for the Federal division in Unisys for a while, and I became an IBM person to work on Watson. And the reason I joined IBM to work on Watson was because they were utilizing ontologies on the back end. And there's a different set of Artificial Intelligence areas and capabilities. My specialization is in natural language processing and data manipulation using natural language processing, which was ontological. So it made sense for me to work on that on the back end. However, they changed their approach to Watson about a year and a half after I joined IBM, and it was a good change. It's not anything bad about what they did. What the way they did was changed from a pipeline that used ontologies to query data across the data set that they wanted to access and change it to a neural network approach. Neural networks are not my specialty. So I went from helping build it to helping sell it. And I joined Global Business Services at IBM, with the Watson team doing financial technology. And I spent five years or so building Artificial Intelligence systems that process natural language for Fortune 500 companies. So if you've ever called up a call center and gotten that message that annoyed you, that said, oh, press one for this, and two for this and three for this. That's not what we built, what we built as a system where you can call up the call center, and you could talk to the chatbot, just talk to it; I want to do this. And it would say, okay, you need to do this, this and this. And then if it couldn't do it, get the right person on the phone now to help you move forward. The idea behind that was to make people's jobs easier. So they didn't have to completely redo the same steps over and over and over again, 100 times a day, such as password resets, right? Now, there are automatic systems that can do password resets. But again, you have to go to the system; you have to log into the system, you have to put the stuff in, and you have to run the password reset. There are people in the business environment who will outsource that to a second person. So now you have two people involved in one operation, just trying to change one password; the average cost of changing a password per person per incident is about 70 bucks. If you can reduce that to say $5. And you have 10,000 people changing their password a month; you can save considerable money. Right? So those are the kinds of systems we're building for IBM's financial clients. Now, as part of that, you have to deal with protected data. Because it's financial transactions, you have to deal with whether or not the users that are accessing the platform want to access it. How do we make it better? How do we make it so that the users want to use it as opposed to typing the number? So that's where we start talking about things like sentiment analysis, which is a big deal right now in AI systems because people feel that when they're speaking to a bot, and they're having this discussion with this natural language process, that analyzing how they feel about the process in real-time, is somehow invasive. If the system is constructed correctly, what that data should do is let the developers know that this particular interaction is frustrating to a user. That's why they do it is to determine what the person likes and what the person doesn't like on a holistic basis. So if you have 10,000 users, how do you determine which part of the natural language interface makes those users comfortable and happy and they want to use and which part really annoys them, and they're going to start cursing? That's why they use sentiment analysis. Now, can that like any other technology, it can be vastly abused? Sort of like PCI? Yeah, PCI data can be vastly abused, straight data access, anything like that, that is a technology that is progressive, can obviously be misused. So for example, I wrote a program quite some time ago that you could just go out and it would pull every corporate mentioned from Twitter in the past month, run it through analysis and tell you how everybody feels about your company. And for the people who don't we can isolate their demographic group and try to improve that. So we have those kinds of systems in place now. Is that engaging in a privacy issue, or the perspective of the business using it? There are it's a delta. So if you are not collecting PCI data or PII data, and what you're doing is creating sentiment analysis data or a particular area of the system, then you can use that, that there's no ethical or legal implication to doing that, because what you're essentially saying is, this is the part of the system I need to improve. You know, 100 people 200 people used it. 25 of them didn't like this. I don't know which 25. But I know 25 didn't; that's a perfectly legitimate use for a system for sentiment analysis. What is not a good ethical use for sentiment? And this is my opinion, is to determine and you can do this easily with the same data to determine the exact demographic group that is either happy or unhappy with your product, your brand, your messaging, whichever piece and be able to target. And so you're going after individuals in an individual group; you are identifying individual people and using their data to change your marketing approach or your delivery approach so that you reach a specific demographic. If you're just using demographic data, great if you're identifying individuals to build that demographic data, it's a failure because you've crossed that line of ethical behavior into targeting specific people for specific reasons. This is very similar to what Facebook does which everyone's so upset about. Whatever you interact with on the platform is recorded and tracked against you personally. So the personalization engines that we use actually direct results based on your personal choices previously; let's talk about how that can be good and bad. The good part of it is you have Amazon and it can tell you, Hey, you bought this thing; you can buy this other thing for it for half price. Fantastic. That's something that people want. The downside of that is that we now have a system in the background that knows everything you want and everything you've ever done. And everybody you've ever talked to you within that ecosystem. Going back to ontologies. There are actually methodologies for querying all those ecosystems at once. So if Facebook has some data on you, Twitter has some data on you, and WhatsApp has some data on you, you can use a social aggregator that is built on an ontological platform to go out and grab specific data about specific people out of all those platforms, and collate most people don't realize that that's available. And I can give an example. It's kind of funny. About 10 years ago, I was doing a presentation for the GSA, which is a government group, and its role is to buy things. So if the Army wants a tank, they tell the GSA and they go pay for it. It's basically the accounting system. We're doing a presentation for him. I was talking about ontologies and, you know, overlays for their financial data and stuff like that. And one of the guys said, so can you show me what this is used for? Outside of you know, the financial information? I said sure, give me your email address. He gave me his email address. And I pulled up a social aggregator program that was running at the time, and it's quite good. And I put in his email address, and I was showing this during the meeting. And I said, okay, here's your house, here's your pictures. Here's your family. By the way, here's what your daughter's boyfriend wants for Christmas. And half of them want to buy it and half are terrified.
Debbie Reynolds 13:04
Right.
Jim Barnebee 13:07
So that kind of information exists. And it's being collected. The question is, is it being used appropriately? And how do we safeguard that usage? Right, yeah. That's why the data becomes important when you start talking about analytics, right? So there's been a big thing with facial recognition.
Debbie Reynolds 13:31
Right.
Jim Barnebee 13:34
How that works is that you have a system that builds a model. And that's just math; the model is math. Well, we use math to describe events that either are happening or we think are going to happen. And then we can test that mathematical model against the real world. That's great. But how do you build the model? So if what you do is you take the system and you give it training data, and you say, take this information? This is the 100,000 things I have free to look at. Here are 100 that are correct. So x matches y. I'm going to give you 200,000 x's, tell you x matches y, and see how many y's you get. So and then you just try to make that a little better. So you get you know, 80-85% y out of that match. Here's the problem. If your 200,000 x state training set has just a bunch of white guys, guess what? It's never going to find anything but white guys. I don't care if you put on a green frog, right? Like the green frog is a white guy because it doesn't know the difference. It hasn't been trained on anything except that one dataset.
Debbie Reynolds 14:54
Right?
Jim Barnebee 14:54
There's a lot of problems with AI in the way that the models are created because of the training dataset. So when people get upset about this model is making bad decisions? Yeah. Not surprising. If you train it with bad data, you get bad models. So garbage in, garbage out. That's our problem. So when people get very concerned and upset about face facial recognition algorithms that don't work, most of them just don't right? Just don't. Right? It's not because the system is bad, right? It's not the math; that's your problem. Right? It's the input, right?
Debbie Reynolds 15:36
It's the data you give the math. Well, I think too the problem is, and you're right about this. So the problem is, okay, you have a training set, you, you know, solve your problem, or get the data back, the why's, as you say, but then you turned around and tried to apply that to a broader group of people, and it was never trained that way. And so then now you get people who don't really fit into the box, right? They don't fit in there. So you're kind of an anomaly. If you're people like me, you know, I'm always not fitting into things. So that, to me, that's a problem. That's more of a human problem.
Jim Barnebee 16:23
It's straight-up a human problem with the data. Yeah. I mean, if I've got an algorithm, and I tell it, my training set, here's my match, that's a white guy. Anything else in my training set is a green frog. It's going to take anything else and put it in the green frog category. Because it's, it doesn't know any better. It's just math, one and one equals two. Okay? And it gives you back to if you want to make it so that actually deals with the population that you're working on. Your training data set has to absolutely reflect that population.
Debbie Reynolds 17:01
That's right.
Jim Barnebee 17:03
Because if you don't, it can never figure out the right answer.
Debbie Reynolds 17:07
Exactly.
Jim Barnebee 17:08
People think AI is smart. AI is incredibly stupid.
Debbie Reynolds 17:14
Well it's doing what you tell it to do, right?
Jim Barnebee 17:16
It's doing what you tell it to do. People ask me about my job. And I tell them I tried to make computers less stupid. But they are. It's a math engine. Right? The fundamental level of computer science, it's a math engine. So it's just equations. It's just math; it's predictive analytics, which you know, is curve-fitting stuff. It's just math. If we feed really bad data into our math, we get a really bad model out the back end, or you end up with a flat earth it doesn't.
Debbie Reynolds 17:48
Right. And then too one thing that I'm concerned about is the way people use statistics to try to tell the story about humans, right? So let's say you have an algorithm saying, okay, let's say who's guilty or innocent of a crime. So we think 90% of people are innocent. But who are those 10%? And how did you decide? How did the algorithm decide those 10%? Those are not numbers. So those aren't people, right? So if something bad happens to you, the statistics don't mean anything.
Jim Barnebee 18:24
You're absolutely right. So this goes back to how you design the system that finds the x equals y. So if you're talking about an image recognition algorithm, they take a whole bunch of data, give it a match and tell it to just go learn, just trial through the data keep going and over and over. And when they get the results and say this was right, this was wrong, they send it back and run another cycle. Which is great. But again, if you're doing something like these individuals are guilty, these individuals are innocent facial rec isn't going to do that. And the fact that they tried to apply it to that problem, it's taking a hammer, and trying to I don't know, loosen up a screw, it's not going to work, right? wrong approach. What you need to do for that kind of situation where you're trying to solve a holistic issue, rather than a precise one is to create a larger structure of the data that you can actually analyze. So this is a great place where ontologies come in.
Debbie Reynolds 19:29
Right.
Jim Barnebee 19:31
So if I go into an ontology, and I say, here are the 25 characteristics that we're going to look for to determine guilt, innocence and guilt we're talking about. So these are 25 characteristics. And then I put in the dataset that runs on top of 10,000 people, and it's got to be a cross-section of every court case that's been run in the past four years, right? Do you need that kind of data? Who has been innocent? Who's been guilty? And what are all their attributes? Not just, you know, height and weight and race, but where do they live? What are their socio-economic conditions? What is the rate of crime in their area? What are the extenuating factors? Like, is this person a minor? Or do they have something else going on? We have to take all of that data into consideration in order to make a cohesive decision holistically. So you can put all that information and ontology, then you could query it to find the people who matched all those conditions, then you can do a statistical analysis on that. Because then you're actually comparing apples to apples. The biggest problem we have with AI in training data is that people compare apples to oranges; it doesn't work.
Debbie Reynolds 20:56
Or like I say, apples to grapes.
Jim Barnebee 20:58
It just doesn't work. And it's not the fault of the math because it's just, yeah, of people who don't spend the time to properly curate the data and align it so they get the right answers out of the models. Half of that is people don't know what data they have, or how to cleanse it, or how to align it. And if you get somebody and it's just I'm gonna build a model on whatever data I've got. That's what you get. There's a model on whatever data you have if you have a group that can go in, and we used to do this work for these other guys for IBM, we'd go in and look at the data set and actually analyze it and say, does this data have work? Does it make sense? Is it correct? Is it organized in a way that we can get the right answer, as opposed to just any answer? That's a non-trivial process that takes time.
Debbie Reynolds 21:55
Absolutely.
Jim Barnebee 21:56
And it's very complex, and the more datasets you have, and the bigger they are, and the more distributed they are, the harder that is to do. So one of the ways to fix that is to create a big overlay of all the data and then pull the specific things you want out of that giant data set conditions that you want to impose. If you can't do something like that and say, this is good data that I'm trying to use for my training set; this is good data that I know matches. And then run that through the algorithm. If you can't do that and start with good data, you're never going to get good results. And that's why I see a lot of AI projects fall; they don't have good requirements coming out of the front end, and they don't have a good understanding of what the data is that they're feeding into it. And because of those two issues, they're never going to get good data out the other side.
Debbie Reynolds 22:52
Yeah, tell me about, you know, I feel like people maybe from movies, they think AI is magical, right? Do all these crazy things. And I don't know why you I'm like screaming at the screen like, oh, my God, that doesn't actually happen. But I think when you're trying to educate people, you have to sort of talk them down from this magical theory about artificial intelligence and what it actually does. Tell me a little bit about that process for you.
Jim Barnebee 23:20
Sure. So the number of times I've told people that there is no magic bullet. There is no magic. It's just math. It's a lot of really cool math. And it makes it look really good on the backside. So people can use it easily and understand it and work with it. At the core, it's just math. So if you look at any area of Artificial Intelligence, there is no Terminator, and there's no General Artificial Intelligence. When you talk to a bot, it's not thinking. It's running math to try to come up with an answer that answers the question you asked. So how do we do that? Well, we have different areas of Artificial Intelligence. And within each of those areas, we have specializations. So for example, if we take AI holistically, and we look at it as all the different pieces we can do, we have stuff like natural language processing, which is involved with text-to-speech and speech-to-text, and analysis of speech and putting together speech. It's essentially natural language; how do we talk to a machine? Have it understand what we're saying and talk back to us in a way we can understand. So that's what natural language processing is and does. That's what my company does is create those kinds of interfaces on to anything else. Alright, so any other system you have in the back end? We create this nice way you can talk to it, it'll talk back and it can actually understand what you're asking. Now, there's a trick to that. It's in that it's just math. So we can't ask it to understand anything that you say because we'd be there forever. So what we do is say, okay, for your domain, right? You are Fred's bank. Okay, Fred's bank. Let's go look at all the Fred's bank questions that you get, break them down, figure out how to answer them by querying the back-end systems, and how you want to hear the results back, then we can build that interface. But the reason you can't just go by an AI and turn it on, and it's magic, and it works, is because you haven't taught in anything. So the one I used to use, and I'm calling IBM a lot on this, because they're, they're good at it. We would go in and we say things like, okay, we want to buy it, we want to just use it. Okay, well, we have one that won Jeopardy, would you like us to come in and win Jeopardy for you? That's, that's no problem we have that it can go in jeopardy all day long. What I can't do is set up Fred's bank by telling you to go play Jeopardy. Alright, so we have to take the thing that learns how to play Jeopardy and teach it how to play Fred's bank. That's a natural language area of Artificial Intelligence. Then we have things like expert systems, which are basically rule-based algorithms that have a bunch of flexibility that let you do specific things for knowledge capture. And dissemination is mostly what expert systems are used for; for example, if I have 100 people in my organization, and we're all architects, and we need to store all these plans, how they operate on the plans, how they store them, what they get out of them. Those are all knowledge that individuals have in the company. Well, if that person leaves, I don't want to lose that knowledge. I can codify that knowledge into an expert system where people can go query for that and say, Hey, I need to solve this very detailed technical problem in my particular area, like I'm building a plane, and I need to know how this part goes. And the manual says this. But the guy has been doing it for 20 years. It's 20 degrees off, but which one do you want? I want the guy with 20 years of experience telling my people to do the manual, then turn it 20 degrees. That's the kind of thing we can capture with an expert system. So when you ask the question, you get back the answer that contains all that information. Does that make sense?
Debbie Reynolds 27:44
Absolutely. Also, one thing that I think is interesting with this is that this kind of goes to the magical thinking, you know, in order for you to use a system like this, you have to anticipate what you think the questions will be right? The things that people are trying to answer. And the I think some people, they're asking questions that the model was never trained to answer, right. And sometimes, to me, those models. Just like you go to Google, if you search for something, you have a million results, which is not true, right? Like no one will use Google if nothing came up, they still are; I don't know, I can't understand your query. So in some ways, I feel like some of these Artificial Intelligence systems are built to try to make people get a result, even though it's not a very good result, just to make it seem like it's working.
Jim Barnebee 28:41
And if you look at the search engine algorithms, yeah, they're going to get the nearest next match from their dataset. If they can match one word, they'll pull that up. If they can match five words, they'll pull that up, and they'll put it in order of precedence, we match five words, four words, three words, whatever. That's a keyword search. There's also cognitive search. And Google's gotten pretty good at that. That's where if you type in a phrase and says, you know, how much do alligators eat, it can go pull up information from a zoological database about alligators because it understands that you're asking about alligators, but you're asking specific questions. Now, if you just put in the words, alligator feed, you're going to get a bunch of pictures of alligators eating things, not what to feed. So that's the difference between a keyword search and a cognitive search, a cognitive search tries to take the context of the data into account when making the query, ie what does this mean? If I put these 10 words together, does it mean the same thing as if I put these three words apart? So there's that aspect of it for generic search engines. They've got some amazing stuff; they do a great job. It's not a trivial problem, particularly with the Internet. Because indexing and understanding the context of that much data that changes constantly is an incredibly difficult problem, which is why they dedicate giant buildings of servers to it.
Debbie Reynolds 30:14
So tell me what's happening in the news right now or in the world right now related to AI that concerns you most? And what are the news stories you heard about and just made? Like, oh, my god, I can't believe this, or, you know, what was concerning to you right now?
Jim Barnebee 30:30
What really concerns me most? Well, there's two things that concern me most of the implementations. The first is that Apple, the general public, isn't clear on what different types of AR, and AI are and what they do. There are many people who still believe or think from the movies, oh, it's Terminator, and it's going to take over and kill everybody. Or, you know, it's going to do all my laundry for me; it's going to be the AI that, you know, runs the robot that runs my house. Neither of those things, right? Maybe eventually, we'll get a robot that works. And then they're coming up with those, and they're getting better. But they're still just doing what you tell them to do. There is no AI system anywhere in the world that thinks for itself, doesn't happen. If we ever get that. Fantastic. I'd love to see it. You know, people say to you like you want general art. That's what they call general artificial intelligence, as opposed to specific. People say, Would you like to have general artificial intelligence? Yeah, we don't have much of the other kind; it can't hurt, right? But the question there is, what happens after? So people call the idea that an AI becomes self-aware the singularity because we have no idea how that will affect everything. Right? Not something we have to worry about. You got 10-20 years before anybody starts getting close to that. Some of the quantum stuff is being started starting to be looked at in that regard. But it's going to take decades for us to hit the singularity. Until then, what we have are specific math, right, specific types of artificial intelligence algorithms that work on specific types of problems. So if your problem, as I mentioned before, is trying to determine what somebody says and what you should say back? Well, that's natural language processing. Expert systems are stuff like teach me how, you know, teach my system about clinical, medical, clinical stuff. And then the doctors can go back and pull that information easily. Right. So we have a bunch of doctors put some information in, and then another doctor can go look that information up. Medical is a great area for that because you've got a vast amount of data that comes at medical professionals constantly. There's no way they can keep up with they're not going to read all that can't be done. Right. Even if they spent 24 hours a day, they're not going to be able to get that all that information brought in. If we use an expert system, as that data comes in, we codify it, classify it, stick it in, stick it in a data set, and then use an expert system to allow us to query that data set. Then when the doctor comes in and says, okay, I have this problem, and they can put it in and they can see from the data the last 15 people who've had that problem. And what the most recent research is regarding. So that's kind of an area for an expert system, stuff like taxes, stocks, flight tracking, things that people do that we want to enable others to share that knowledge. So that's more of the expert system area. We have fuzzy logic, your self-driving car. Great example. It's got to all the time figure out where it is where everything else is how to get around it what to do. So it's not doing if I see a car stop, it's doing if I car these conditions and these conditions and these conditions, and these are the different choices I want to make, right? It's taking all that into account to make a fuzzy decision. And fuzzy decision. We're talking about probabilistic. So you can't predict the outcome because it's a probabilistic problem. Exactly. We have things like robotics, that robot room in your house, hey, we can do that. Driverless, the actual driving of the car. Manufacturing stuff, right? So you have the robot that puts together the Toyota. That's an entire area of artificial intelligence. Fun, that's kind of interesting. So if you have a grid, and you have a robot that can do neural network visual record ignition and expert systems and fuzzy logic and you put it into a robot, and you give it a pegboard. And you show it once how to put the pegs in the board. It can do that forever and never make a mistake. You teach it again; it's just an AI system. So you teach it what you want. And it goes and does it quickly, repetitively forever. What it doesn't do is make independent decisions about whether or not to change it. It just does what you're telling it to do, right? The math enables you to make that fuzzy and say, well, if these things happen, do this; if these things happen, do this. And if these things happen, do some other things. You can do that. But again, it's just saying, well, here's what happened. What's my probability? It's 20 of this 30 of this 40 of this, okay, I want to make that decision. It's not actually making an independent, thought-based decision. They just they can't do that.
Debbie Reynolds 35:59
Right.
Jim Barnebee 36:00
Neural networks are another area of AI. So I was talking about, you know, facial recognition, noise recognition. pathfinding, so neural networks are great for stuff like how do I solve the traveling salesman problem. For those who aren't familiar with the call traveling salesman, a fairly well-known medical problem where you have 50 cities, and you have a person going to each city to sell X, what's the most efficient route between the cities? Right, and how you calculate that that's a fairly well-known problem. It gets interesting when you've got moving targets, so you've got a person who's delivering food, and they have four deliveries on their plate, well, then one drops off, one comes on, he changed your target profile, you change where they have to go. We have a traffic jam; we have to route around it. So we changed how they have to go. So you have to monitor all the conditions around that path and determine how to restructure and route that path in real-time. That's a very non-trivial problem. That's difficult. So it's sort of an extension of the traveling salesman problem that we keep running into. And we can use neural networks to find the best path through those if that makes sense. It's the same tech we use on facial recognition; you give it to data, you tell it what the right answer is, and you see what it comes up with you optimize. And all AI basically works that way. And there's a lot of precision here, and I'm sure I'm gonna get pushback from AI people on this. And that's fine. So the point I'm trying to make is AI is math, its algorithms, it doesn't think for itself, there is no Terminator, you just feed it data, tell it what you want, and see what comes out the other end, and then go back and fix it and optimize it until you're getting a result. That's mostly correct. You're never going to see, and again, I'm probably going to have AI people yell at me, but I have never seen a system that was 100% accurate in AI, ever. I don't even think it's possible. Now, there's a lot of fault-tolerant systems and a lot of good systems that function like your driverless cars, right? But those algorithms have internal mistakes, and things go wrong. They just have other stuff on top of it. This is if this goes wrong, go over there and do that.
Debbie Reynolds 38:40
Right.
Jim Barnebee 38:40
So but some, there's never 100%. So if you're talking about language, right, you get 80%. If you get 85% in a language model, it's considered really good. 15 percent flows through. We don't know what to do with it. Obviously, expert systems are a little simpler because you're recording knowledge and sending knowledge back through the system. Right. So whether or not you're using that knowledge to track flights and help airport personnel, or you're using it to help financial people with stock trading, or doctors, right, those that your knowledge base that's being queried, you're still going to have issues. Right. And I've seen this with like medical programs where it's mostly right. But occasionally it grabs an article or a piece of data that's not related to the question you asked. And in a medical situation that can be serious. That's why there are no automatic doctors. Right? That's the problem. We can't fix with an automatic doctor to have a doc in a box, which gives you advice and information. But it should not be making decisions, right? So we're talking about autonomous systems, systems that are working by themselves, you're still going to have that 85-15 rule, that 15%, where the autonomous system has to go back and run its error systems. Oh, that doesn't fit the curve that I normally have. What do I do with it? Okay, I have to run through my fault tolerance, figure out what's going on, and then fix it. So that's why you see things like automatic cars that 99.9999% of the time don't have an issue. And occasionally, somebody, my personal favorite was they took an automatic car and they painted a circle around it. Good move. It was told not to go across that line. The line was all around it, right? So it just stopped and went, okay. I'm where I'm supposed to be. I'm doing the right thing. I'm sitting here. And it would refuse to move. Because the rule said you can't cross that line. Okay.
Debbie Reynolds 41:01
Right.
Jim Barnebee 41:02
They had to kind of figure out a way around that. Because yeah, that was funny. Yeah. And automatic parking spot, draw a circle. There you go. Right. Those and that's, that's way out. And that 15%, right? How many people who have an automatic car are going to make it drive into a circle of the white line where it can go cross?
Debbie Reynolds 41:25
Right.
Jim Barnebee 41:26
Probably not gonna happen. But it could absolutely be the error category. What do we do now? Absolute thing I would really want to convey to people about Artificial Intelligence is, it's not bad. It's not dangerous. It's just a tool. It's a chunk of math. We use it to make things faster, better, and easier. Sometimes people use it in other ways. That's not as productive for the society, shall we say?
Debbie Reynolds 41:59
Yeah, I like to say, people think of AI as a teddy bear. And I think it's a grizzly bear. So I think you need to know what it's doing, why it's doing it, and they can't like abdicate your human judgment to a machine, right? You're telling it what to do. It shouldn't be telling you what to do.
Jim Barnebee 42:19
Exactly. AI systems are fantastic helpers, helpers. Absolutely.
Debbie Reynolds 42:25
It's not leaders.
Jim Barnebee 42:26
I had a professor used to say to me that if a human being has to do it more than once, get a system to do it because you're wasting brainpower. Human beings should use AI systems to accomplish tasks faster, easier and better. They should not be using them to take advantage of other people. Right, or to do other things of that nature. But the system itself is simply a useful chunk of math modeling.
Debbie Reynolds 42:58
Exactly.
Jim Barnebee 43:00
Yeah, it's a bear. You know, you can train it to work in the circus and juggle balls, or you can train it to go eat somebody.
Debbie Reynolds 43:07
That's right.
Jim Barnebee 43:11
And if it's a bear, you just train it, and it does what it does. Yeah, we have to just make sure that when we use these systems, we use them. Well, we use them correctly, and most important that people know what to expect out of them. I keep saying it. But one of the main reason reasons any AI project fails in any industry at any time is poor requirements.
Debbie Reynolds 43:35
Yes.
Jim Barnebee 43:36
If they don't know what to expect on the other end of that project, and you don't have methodologies to test and make sure that the users are getting what they want as you go through it. You're going to have issues when you get to the other end because it's not going to meet the need that was originally submitted for the project. So right, take off my developer hat for a second, put on my manager hat for a second. We need to know what the outcome is that the individuals want. If they want to ask a question and queries from this BI system. Fantastic. Then we know what we expect. We expect those queries to run this way at the end. If you have somebody who comes in and says I want some AI, okay, great. What exactly do you want it to do? make things better. Okay. Which exact thing do you want to make better? In what way? If you can answer the question of what you want to make better? How, what do you want to do to it to make it better? And how do you measure the result? You can have a successful AI project?
Debbie Reynolds 44:46
Yeah, I think too. I was incensed. There was a situation where someone had a medical situation they couldn't figure it out, I guess, and was looking at X-rays and they could tell from the X-rays the race of the person. And some of the doctors that were using it. They said, well, we don't know why it did that. It was like, what do you mean? You don't know why it's like, you should be very clear on what the AI is doing and what the result is. And if the result is not what you expect me to go back and look at it?
Jim Barnebee 45:26
Well, you definitely need to go back and look at it. But what you're talking about there as ancillary data, they put it in all the conditions for here, the x-rays, and here, here's the matching set, right, and then go run it, run it, run it until the neural net actually understands that, oh, that's what I'm supposed to look for. If the data that is matched with what they're supposed to look for in the X-rays, and in that X-ray data, they say, the race is actually included in that data, then the system can make an inference, and it's actually just probability, right? It's just running numbers. And the inference is going to make is individuals who are listed as this race have this more often than individuals of this race have this; you want to fix that and take race out of it. That's right, remove race from the data set. And then it won't find it anymore because it's not there. Right. And so that's the kind of the and they didn't know I'm sure they didn't know ahead of time that they were even feeding race into that characteristic set of 100, 200 or 300 different characteristics, that they're feeding into that program to tell it, how to look at those x rays. If just one of those sets of characteristics happens to include race from the data they polled, then it'll show up on the other end; it goes back to that data cleansing I talked about right? Got to make sure your data is good; it makes sense. And it does only what you want it to do before you feed it to the AI. Otherwise, you get results on the back end. I don't know where that came from. Right? You put data in on it.
Debbie Reynolds 47:07
Exactly. So if it were the world according to you, Jim, and we did everything you said what would be your wish for either AI, privacy or data stuff in the future? What are your thoughts?
Jim Barnebee 47:20
Well, privacy is a whole different subject. I think in the privacy constraint privacy arena, what we need is for individuals to be able to control their own data. So I can decide whether or not the bank wants to use my data in their algorithm to determine pricing on savings accounts, right? I get to decide if my personal data goes into that dataset. That would be the first thing. And there are some people working on things like that right now with sovereign identity and digital identity. So if we can get something like that, where individuals can actually control the usage of their data, that's step one. Step two is to look at the algorithms. Right? So how many of those are producing data that wasn't expected and why? And that goes back to checking out the data, making sure it's cleansed; you and I've talked about specifically image recognition, and they're all showing up as white guys or green frogs, right? So that's a problem with the dataset. And if we don't put good data in, we get garbage out the other side. What I would love to see is AI helping everyone. I would love to see it easy and simple. So you could just walk into a room and tell the room what you wanted, and it would go figure it out and bring it back, I want to see Star Trek computers. I want to see where you can just talk to the thing, and it goes out and figures everything you want out and brings it back. That and if it has to use neural networks or fuzzy recognitions or make a robot, you know, run the screen up and down, whatever it has to do to accomplish what you want. One of the biggest issues I think we have with computer science as a whole is that most people can't use it. Right? Most people are not going to go in and create a data model. Most people are not going to go in and build cloud infrastructure, right? Just examples. You have to go hire an IT person to go do that. I've been trying to work myself out of a job for decades. I want people to be able to just work with machines flawlessly and seamlessly with no effort. It shouldn't be hard for people to get assistance from Artificial Intelligence; it should be easy. Easy, right? So that's what we try to do is just make it easy. And what I would love to see is every computer, every system everywhere in the world, so easy to use that a four-year-old can walk in and use it without ever having touched a machine.
Debbie Reynolds 50:09
That's what I want. That's amazing. I love that. Oh, wow. Thank you so much. This is a very instructive episode. I love it. I'm so excited about being able to collaborate with you on actually, some of the interesting things. And we're working on it. Man, this is amazing. I love this episode. Thank you so much, Jim, for being on the episode. This was great. I think I've never heard anyone explain AI in the way that you did it, you can do it. It's such a simple way that anybody can understand. So I think you live up to your domain of your company AIM-E, because you definitely made it much, much simpler for me and other people to understand.
Jim Barnebee 50:53
Well, that is the goal. To make it easy for everyone to use this and get the benefits out of these types of systems.
Debbie Reynolds 51:02
Excellent, excellent. Well, I'm sure we'll be in touch soon.
Jim Barnebee 51:07
As William Gibson used to say the future has arrived, it's just not very evenly distributed.
Debbie Reynolds 51:14
That's very apt. That's a very apt analogy. We need to get it more evenly distributed for sure.
Jim Barnebee 51:21
If we could take all of the magnificent things we can do now and put them in everyone's hands and an easy way to do things, we can build a utopian society. We have the technology; we have the ability; we have the resources. People can't get access to them and use them. Right. That's what we have to fix.
Debbie Reynolds 51:43
I agree. I agree. Wonderful. Wonderful. Thank you so much. I really appreciate it and I look forward to chatting with you soon.
Jim Barnebee 51:51
Me too. Thank you very much. Thank you