Today, we talk about the variety of ways AI shows up in our lives, both in our personal lives and in our workplaces. And we address a few questions. Should we be afraid of it? Where should we be wary? How can we best use it?
Our guests are Grant Erickson, the CEO of IntelliTect, and John Shovic, director of the Center for Intelligent Industrial Robotics at the University of Idaho in Coeur d'Alene.
This is part two of a series of AI discussions. Hear part one here.
This interview edited for length and clarity.
Doug Nadvornick: I try to get my head around where artificial intelligence is used in our society. Can we just throw out sort of a list, maybe, of things that might be surprising beyond, I guess, the typical uses of AI?
Grant Erickson: Yeah, I think there's a few places that are kind of obvious. When you hold that little button down on your phone that allows you to speak to it, you're using a little language model to make that speech into text, something that a machine can then go and process.
It used to be very deterministic, you'd say a thing, and if you said four of the right words, it would do a thing for you. And now it's much more broad, and it can handle basic things, and accidentally ordering toilet paper just because you happen to be talking about it. I think those are a few kind of like really pretty simple areas where you're seeing AI stuff.
DN: Are there areas that might surprise us where AI is being used?
GE: I think there's some other ones that are kind of interesting. Think about car diagnostics, you think about the scam detection that banks are doing on your credit card. That's a really good example of, hey, this person has these spending habits, and maybe they're traveling across the United States, they can kind of figure that stuff out.
I think you look at even something as simple as the smart thermostats. Security cameras now, it's not just taking video anymore, it's identifying people and objects, things inside of that security footage, and that's all happening without kind of the big cloud language models that's oftentimes in pretty accessible, off-the-shelf kind of security systems.
DN: Do you have other examples, John?
John Shovic: Well, the area of AI I work in is more involved in manufacturing, in making machines smarter to do better things, hence the Center for Intelligent Industrial Robotics.
Now, what we consider a robot certainly isn't just a six-arm thing. It's also autonomous cars. It's manufacturing machines, CNC machines that are cutting things, making things, and all of these assembly line things, the autonomous robots running around. They're all using something we call physical AI and physical AI is just a different name for AI being put to a different application. The application in this case is moving atoms around. In other words, controlling a robot to pick up something and move something somewhere else.
But I guess the first thing I'd want to say about all this is we always tend to think of AI as being a mono block and it's not.
DN: Mono block as in?
JS: Well, Chat GPT, what you mentioned earlier. Those are large language models. Those are very useful things. There's probably a dozen out there that are commercially and economically viable. But that's only one part of AI.
If you look at the other aspects of AI, they're entirely different kinds of algorithms which don't involve predicting the next word, whether it's in a piece of code or the next token or synthesizing pictures and these are AI things that are set up to work on data.
What I mean by data, it's stuff coming off of machines, sensing what's going on in the environment, and then modifying your programs via a variety of different AI techniques to find the answer and then to be able to explain the answer back to an engineer so they can improve something, whether it's in a manufacturing line or whether it's in a classroom. Either of those things and all those kinds of AI algorithms out there really complicate things in the real world because so many people say, ‘oh, we're using AI in this program, in this product, right?’ But in somewhere around 60%-70% of the cases, they aren't. They're just using some data analytical technique. Or they may just be putting a wraparound Chat GPT to make it look like they're really doing something. So you need to be a little suspicious when you hear, well, this product has AI in it. Everybody says it. But in many, many cases, they're not really using AI and they're not really doing something different than the next guy down the street.
DN: How long has AI been used in the industrial setting as you're talking about?
JS: A variety of techniques have been used for many years but what has happened in the last, I would say six years, is that large language models have come on the scene and that provides us tools we can use in part of manufacturing, being able to identify what the problem is with the machine by doing a very deep and complicated search of the documentation for the machine. That's turning out to be a useful part of AI and a useful use of these large language models like Chat GPT.
The other thing that is going on is when you have a manufacturing line, you may have 25 million data points coming off that a month and, before the latest set of techniques, we didn't really have mathematical and statistical techniques that could tease out what is going on on that manufacturing line by looking at these multi-millions of data points. You can't do it by hand, it's just too many data points. You can't do it by spreadsheet, there's just too many data points. So we have new AI techniques now that search that data looking for hidden correlations and coming up with solutions saying you should do this to increase the productivity of your line. Those answers don't come out of regular AI, they come out of this other kind of this data AI analysis. What's really interesting about that is you can do it in such of a manner where it not only predicts a result, yes, this board is gonna be bad, but you can ask it why, and it'll tell you why, so the engineer, the technicians can go in and fix whatever problem is causing a mistake in the manufacturing line and that is really magical.
I know we all think Chat GPT is magical, but when you see some of these results coming out of these new data techniques, you just go, oh my gosh, we can save this company millions of dollars a year and it's been in their data all the time.
GE: You really are kind of talking about two things. There's this idea that there's a way to process data at scale that we haven't been able to process before in the past because we had to kind of know the equations, things like this.
Now AI is kind of able to figure out what are the relationships between these five different data points that have been collected every second for the past three years and be able to draw those things together, but then also be able to solve that problem and also be able to use maybe some deep documentation learning as well to kind of pull those things together and solve problems that just are beyond the scope of a human being's ability to hold the numbers in their head.
JS: He's absolutely right. But these new data techniques that we're using, these analytical data techniques like genetic algorithms and decision trees and things like that versus large language models, they do have some commonalities.
One of the things I'm sure you've heard is large language models, Chat GPT and what not, will hallucinate. Let's not anthropomorphize this too much. The software does not hallucinate. It is a billion statistical variables and it's just math. There is no reasoning, there is no thinking, there is no creativity involved. It's just reporting things back. Why does it hallucinate? It's because it's got data that points it in the wrong direction and we call that a hallucination, but it's not. But it can be very confident too. And once again, it's not a person, it's not a thinking machine, it's just data.
On the other side, when you start looking at these analytical, these 23 million data points, you can come up with answers that are also kind of hallucinations. We generally just call them nonsensical answers. An engineer looks at that and says, no, that can't be the problem. That makes no sense whatsoever. So there's always the human and the algorithm going back and forth. You see the same thing when people are using like Chat GPT to write code. You go back and forth, trying to get the machine to do what you want it to do. But it is a machine. It's not a general AI or anything. It is truly garbage in, garbage out.
DN: You've each used the term multiple times, large language model. So let's isolate that and define it so that people understand. What is a large language model?
GE: So there's several different types of AI. The general idea with AI is you feed a whole bunch of data into a system. You say, now figure out, based on the data that I have, what is the next most likely thing given some input? If I give it a bunch of data that's just counting like one, two, three, four, five, and then I say seven, eight, nine, and it's going to come back with the number 10 because that's the kind of sequences that it has. Now, it's incredibly simple.
When you talk about large language models, those are ones that we fed a whole bunch of human language into in order to get them to predict. They do that with really complicated math and matrices and vectors and all kinds of crazy stuff. But at the end of the day, what it's trying to do is it's trying to take human language and by looking at a large corpus of data, which is books and the internet, everything they can think to feed into it, to then ask some kind of a question, the next most likely words that are going to come out are going to be the answer to that question.
JS: Let me just add one more thing on top of what Grant said, the magic of some of these new large language models. Yes, we're predicting the next token, but we're predicting the next token with context. So it's not just what's the most likely next word, it's what the most likely next word, given all the context you've given it. You might be talking about the Spokane parks. Okay, that gives it context. So when it goes and predicts what the next word and the next token and what the next token is, it's closer to what it's paying attention to the right things.
GE: And let's just push that maybe one step farther. For those folks who have tried out Google, you've now Googled within the last, six, eight, 10 months, and you'll see the very first answer is now an AI-generated answer by default. You're seeing that, you're putting in context of your question and sometimes it's just a question, like, how old is Tom Cruise? You can just ask that question in regular text, and it's going to come back with some answers there. It's only going to know things that existed at the date when you cut off all of the information that it got, so it doesn't know anything beyond that and so what it has to do is it has to find things.
So when you look at something like Google, what it's doing is it's grounding itself in hopefully things that are correct. By the first question it asks is, based on this prompt, what are web searches that I should do? And then it goes out and searches the web for those things. It takes the results that are coming, the top results coming back from Google, probably, and then it feeds those into what John was talking about, it's context.
So now it has all this information that is probably newer than whatever it had when it started, and then it's going to put in your prompt and it's going to answer your question based on these things. That way it can give things like references and it's not just guessing based on whatever it thinks it knows internally.
DN: So does that mean you could do the same prompt a week apart and you're going to get a different answer in the second week?
GE: Even minutes apart. You will get a different answer from most of the LLMs because there's a certain amount of randomness that's inserted into those.
DN: We hear these stories about people having conversation with Claude or with ChatGPT. What allows that to happen?
JS: Conversational style. It's read a whole bunch of books that have conversations between humans and then novels and all that sort of stuff. I don't know if you remember 30 years ago, a program called Eliza came out and it basically would take what you're asking and then turn around and ask a different question of you, that particular type of therapeutic style. What we're doing today is just the same sort of thing with more context. There is not a person. You are not talking to a person in there. It is not a human being. It does not think like you.
One of the best examples I give people is say, go ahead and try asking ChatGPT this: I'm 15 minutes away from my car wash. Should I walk there? Okay, now, how would you as a person answer that?
DN: You could say, if I'm walking, why do I need a car wash?
JS: Exactly. The large language models will answer, well, if it's 15 minutes, you should just walk rather than drive your car there. Okay, context, missing context. You as a human being immediately pick that up. LLMs don't deal with that very well yet. And yes, they're wonderful tools, but they aren't human beings. They don't think like human beings.
DN: What's a good way for somebody who wants to sit down with a ChatGPT for the first time to learn about it without feeling overwhelmed?
GE: I get this question all the time.
We met with a friend this morning who was doing some things and trying to create some stuff. I would encourage you, if you haven't tried it before, honestly, just go Google some stuff.
In the old days, it was like, hey, you gotta put in just the right terms to get the right Google answer. Now, you're actually better off if you just ask a plain English question, including all of the detail, because anything that you don't specify, it's going to guess and so you want to make sure you put the right stuff in there so you're getting the right answer back. I would just try that using Google's Gemini, doing something like that. If you're running on a Windows PC, you can click that little copilot button that's now down if you have a Windows 11 machine. Click that button. That's a pretty good one. Find something that you really enjoy and go in there with the assumption that it's going to be wrong a certain percentage of the time. Keep those expectations pretty level and start asking questions and then it's really kind of fun just to go discover things and let the human creativity and let the human curiosity kind of guide that conversation.
JS: One of the first things I tell people to do is ask it, who is [your name]? We actually know the answer, right, when you're asking that question. But even there, you'll see these come up and hallucinate.
Did you know I had a two-year-long NFL career that was ended by a bad concussion?
GE: And I played hockey in Canada.
JS: And, you know, these are amazing facts. I guess the concussion was so bad that I just don't remember anything about playing for the NFL.
DN: What do you think about when you see something that's so obviously wrong?
JS: Those are almost laughable. What isn't laughable is when it makes up a citation and it's not real. You click on it, it doesn't go there. Or it's stating facts very confidently that aren't correct. And remember, it's reading a tremendous amount of stuff, some of which is on the internet, which has no controls whatsoever. The answers it comes up with are based on what it has learned and what it has learned is not necessarily correct.
But there's something more, not nefarious, that's not the right word, but something more complicated going on here.
When you apply these reasoning rules that the programmers built on top of the large language model, they can come to the wrong conclusions. I mean, humans come to the wrong conclusions all the time for very bad reasons. But these things, they'll come to the wrong conclusion and then the next time you run it, it'll have a different conclusion. It'll have a different set of citations. The problem is how confidently it states them.
We ran a study here at the end of the last semester at the U of I, where I had a bunch of people use a large language model tool, Claude Code, and we had them do some of their homework assignments with this. There was a small sample set, but what we found, the people that knew the least about software didn't check their work and so the stuff didn't work. The code provided was off and it had problems in it, where the other students that did know what was going on worked with it and got it running. You can see the same sort of thing when you sit down, if you believe these things. You've got to be very skeptical of these answers.
DN: So we hear these horror stories about students using this as a way to do their homework so they don't have to do it. What are your thoughts about students, what are your experiences with students using AI when it comes to doing their work?
JS: Well, we have a term now in the industry called AI slop, and you know, you can just tell.
If I looked at a research paper that's been generated by AI, I'd know it in the first couple of paragraphs. How do I quantify that? Why do I know? I don't know exactly, but it's obvious. It really is. And what you find when a student uses a tool to build a complex program moving robots around, when it goes wrong, which it invariably does, they don't know how to fix it. They can't sit there and explain the code.
When you're doing something in physical AI where you're using generated code to move things around, there's a certain amount of danger involved in that. You can break things, you can hit people and stuff like that. So it's just not ethical in my mind to take a program from a large language model and run it on one of these physical things, moving robots around without doing a code review and really looking at what the code is doing. Because of that, people who just generate this code lack the critical thinking skills to be able to analyze and to debug when something's gone wrong.
So how do we do it? How do we handle that? Well, people need to use these AI tools earlier in their career. Maybe they're around, they're going to use them and we still have to make them think. We still have to make them understand what they're doing because you just can't have people believe these things blindly.
What we're finding is it's the people who understand code who are the ones that get the productivity increase, not the people that are trying to program an app in the weekend.
GE: I do a fair amount of speaking to students at the high school, junior high, even elementary age level, and the temptation to be able to, 'hey, I have to do whatever the paper is.' I read a book and now I have to write a little paper with this prompt. I mean, you can literally put that in and that is gonna write a better paper than any of those students for the most part. Maybe they're seniors or something like that. They're gonna be fine. They can go and play video games and I can press the button and I get the paper and I'm gonna get a good grade and don't use too many EM dashes and don't make it sound too flowery and maybe you give it a couple of things you've written in the past to make it sound like it's in your voice, the teacher doesn't know. That works and they may get better grades. The problem is that there's this massive easy button that they just pressed. And when you start pressing that, you stop learning how to think, right?
I've taken a lot of math in my life, lots of calculus and things like that. I've used calculus once in the last 25 years. It's not the subject. The purpose of learning is not to know the thing necessarily. It's the ability to get the mental pathway set up.
We need to think more like going to the gym. You can't get nice abs without going and doing work. You can't get the ability to think without working hard at thinking and unfortunately, it's hard. It's a really hard process and I think that that to me is where for students, it's so important. Yeah, it can give you a better job. You get a better result, all that, but it's this idea of AI as a solution to allow me to do something that I really wanna do over here that maybe it doesn't require as much thinking. It's, I'm gonna think about this and think about how we wanna structure this and then I'm gonna use AI to help me augment and do a better job. I'm gonna write something down. I'm gonna go through the process of creating those words, building those synaptic pathways of creativity and then I'm gonna have it review for me and say, 'hey, how can I do this?' Or maybe I'm gonna take the outline and have it review the outline and help me with, oh, did you think about these points? But I'm not gonna have it do the whole thing for me. Use it as an assistant as opposed to a wheelchair.
DN: Do you think that's the central challenge, the central issue that we have with AI?
GE: Yeah, for young people, it totally is. We have this massive temptation of I can just press the easy button.
JS: So what we have to do is teach the students how to use the tools while preserving their ability to think.
DN: So we've made an example of young people. But young people aren't the only ones who are learning how to use this technology. I mean, even folks like me in their mid-60s are using it too. What's the best way for somebody like me who's older, who has a lot more experience to use the technology?
JS: It's a lot easier than you think. Go and log up for a Chat GPT account. Start asking questions.
One of the things my wife did, she put in a picture of our front door and said, please change the color of this door to red. I could have gone in and used one of these photo editing tools and whatnot, but it came back with a very good example and we could see what that looked like with the door painted red. Wasn't a good look. But she can do that. She can sit down there and do those sorts of things. That gives people abilities.
Now, if you say, 'okay, I want you to create a recipe for me for the next three days. Give me a meal plan.' That's great usage of that sort of stuff. Is it going to come up with something that's a five star from a fancy chef? Yeah, maybe if it read a five-star review and you didn't give it criteria like it has to be under $10 a meal or whatever. Yeah, those are all great uses.
GE: Here's something more controversial.
I have just finished reading a book series and there's a whole bunch of books in the series. Love the book series. And I thought, you know, I wish there was another book. What I did was I wrote about how I wanted it to do it and so I said, 'okay, here's the world I want you to write in kind of a fan fiction thing, and I want you to write a four-act thing.' And so I wrote the four-act summary and then said, 'now I want you to write character bios for each one of the characters and setting bios.' So it did all that. And then I said, 'now what I want you to do is, for each one of the acts, to generate one-page summaries of each chapter all the way through the books. So these are all individual prompts moving forward and I'm building all these files that it can now go back to for context. And then I said, 'based on the character descriptions, the general plot arcs and the location descriptions for each chapter, write the prose for the chapter in this voice.' And then I created an audio book and I made pictures to put in the book and created a PDF and so it was just kind of fun to go through that process because I could make something that it's, I don't know if it's mine really, but it's this interesting thing that I hope I'm going to enjoy. And so now I'm listening through this audio book that I guess I didn't really write, but maybe it's gonna be good. Maybe it's not. The problem is I now have like a 24-hour audio book that I need to listen to in order to determine whether AI did a good job. We'll see how many hours I get through before I say, 'hey, I'm gonna throw this out' or maybe it'll turn out to be something that'll be pretty fun.
DN: Let's finish with this. Do you think that AI usage is going to become an essential skill for not only young people to get into college or something like that, but for somebody to go out and get a job or even navigate life?
GE: It already is.
JS: I agree. It already is. And it needs to be taught. We need to teach responsible use of AI and we need to continue. We still have to make people think and that's not an easy thing to do. But these are tools and we'll get better at using these tools and some people are going to be much more productive because of these tools.
GE: We're really what I call in the wild west. We saw the initial large language model and you can kind of pull up an app on your phone and ask it some questions it can kind of answer, and if it was something new, it couldn't answer it. But now it can search the internet and do all these things and so we're seeing is what I would call generation three.
So we have this first generation, which can do basic stuff. We can have the second generation, which can do some minor lookup things. But now we're getting into the place where they can do really useful things like in business. But that's harder. There's this whole thing where it's not just simple solutions anymore. It's like, 'hey, you actually have to go in and engineer a solution.'
Let's say you have your business and you want to be able to do a particular thing, optimize a process, whatever it is, like you're having to go in there and think about how you're putting those things together. Create a marketing plan to do this. ChatGPT can do that thing. Copilot can do those things.
But we're now breaking into that next kind of third wave of applications that we're building that are that next level of complication. We're really looking at how do you solve really key business problems and I would say, if you're in business and you're trying to solve some of those, if you're on the Microsoft stack looking at the Copilot subscription for Microsoft 365, it's probably a pretty good way to kind of check that out, because now it's going to have access to some of your internal data, to be able to pull some from those files and in a safe and walled garden kind of way, all kinds of good stuff you can get from there.
Then looking at really from a business perspective, what are those things that I want to make better about my business or even my personal life? It enables us to improve our lives in ways, in places where our skillset may not totally be there, it allows us to kind of leverage that to hopefully move some things in a positive direction.
DN: Any final points for either of you?
JS: Well, I would like to talk a little bit about AI hype.
Really, what we're in here is we're in an AI financial bubble and I'm not smart enough to know when it'll pop or anything like that, but it's absolutely clear. Who's going to be the winner? I'm not smart enough to tell, but I will tell you after this bubble pops, we're going to be left with some very useful tools that will help us continue to increase productivity in the United States and really take the whole human race to another level. I'm really convinced of that, but I wouldn't buy any of these stocks at the moment.
GE: I've just been chatting with people and they're talking about AI fatigue. All my conversations are about AI. It's always about AI this, AI that.
You know what? I think you should get together with some friends and just talk about life and play some board games and go for walks and do things where you're interacting with other people around human things and the things that you really care about and the things that you love and make sure that as you're kind of learning about this AI stuff, that you really keep grounded with that community that you're in. Those kinds of things are just so important because there is no substitute for meaning and purpose.
One note: IntelliTect is a financial supporter of Spokane Public Radio, but we cover it like we cover any other company.
