Jonathan Schaeffer of Kind by Synsira

All right. Welcome back to another episode of Modern Cyber. I am really excited about today's episode, because it is not very often that you get to talk to somebody who has been working in AI for as long as today's guest, forty years. Can you imagine that forty years in AI, having seen all the evolution? We're going to get into that and a lot more on today's episode. I'm delighted to be joined today by Jonathan Schaeffer. Jonathan is the inventor of Synsira, or the founder of Synsira and the inventor of Kind and one of the pioneers of artificial intelligence with, like I said, more than forty years of experience as an AI researcher, entrepreneur and innovator. Kind is a privacy first AI platform that helps individuals organize, search, and interact with their personal knowledge without surrendering control of their data. We're definitely going to get into that in today's episode. Jonathan's work also includes creating AI systems that achieved Guinness World Record recognition in both Checkers and Poker. I'm sure that's going to be a really interesting story. And he has co-founded the Alberta Machine Intelligence Institute, or Amii, and launched multiple successful technology ventures. Jonathan is a former professor and a leading voice in trustworthy AI, and is focused on building practical AI systems that enhance human intelligence while protecting privacy, ownership, and intellectual property. Jonathan, that is quite a career. Thank you so much for taking some time with us today.  

Thank you very much. I'm delighted to be here.  

Awesome. I would love to kind of start not at the beginning necessarily, but let's go through a little bit of this because that really is a really amazing time period to have seen AI evolve. And we talk today about, oh, you know, we had AI ten years ago, but we just called it machine learning. But what was AI at the beginning of your journey? What was it like? What was actually available?  

Okay, I'm going to sound old. So maybe I'm an AI dinosaur that has evolved into AI Velociraptor. Well, okay. Without the fangs, perhaps. Okay, I got into AI, believe it or not, in in the late nineteen seventies. And the, the mantra there was expert systems. Expert systems were founded on the basis that intelligent systems needed human knowledge. And so where do you get human knowledge? You get it from humans. You talk to people, they give you knowledge. You write it all down and you turn it into hard coded rules. If you're a building an AI system to be doctor, you say, hey, this person's got a cough. Uh, what is the appropriate treatment? The doctor gives you the information, but the information isn't always accurate in the sense that it might be just a cough, but it might actually be a symptom of some other disease. And so the knowledge would be written out, and then you would put beside it a probability as to how good that diagnosis was. Oh, he's just got an ordinary cough. We'll give him some cough medicine that works. And maybe that's the answer ninety five percent of the time. But one or two or three percent of the time, it might be some symptom of a disease or something worse. You had the knowledge and you'd have probabilities. And these programs would just analyze a certain situation and come up with an answer.  

The problem is, way back when, people quickly discovered it was really hard to extract knowledge from humans. Doctors would tell you rules of thumb or heuristics, and that was even harder to get the numbers out, how somebody has a cough and it's just a common cold, is it ninety five percent of the time? Is it ninety nine percent? Is it eighty four point three seven percent of the time? Really hard? Yeah. And what happened over time as AI evolved? First, we developed systems that were good at learning what those probabilities were. You could analyze data and discover, oh, you know, when people say have a cough, it's ninety six point eight seven three percent of the time, that cough medicine would be the answer. So we could learn that part just by analyzing data. Get the statistics. And that was great because it made it easier to acquire knowledge. Yeah. And then people started discovering that, you know, human knowledge, it's hard to extract it out of experts, but you could start extracting knowledge out of data. And so, for example, when they started building systems that translated between two languages, how do you come up with a set of rules to translate, say, English into French or French into English? They started analyzing data sets, and they would find those relationships and be able to build something that was pretty good just using that. And so over time, we've seen the evolution of going from a state where the human had to specify every piece of knowledge to today, where we're either extracting it from data, large sets of data, or we're having the computer do all sorts of analysis and discover on its own what that knowledge is. And so it's been this incredible transition from AI that was knowledge, human knowledge intensive to today, where it's no human knowledge or at least no explicit human knowledge. It's all implicit in the data that we give to these computers.  

Yeah. And it also seems to me that just from hearing you kind of describe that, we've gone from a system where it was much more deterministic at the beginning with what you described as those expert systems where maybe you had a probability distribution, but that was like pre-programmed almost into I think of it as the code, but maybe the data set is the right way to think about it. And now we've kind of gone through then, like the next phase where we could measure empirically from data sets. Oh, it's not ninety five percent of the time. It's actually ninety six point eight seven three or whatever the number was that you gave based on the observed data sets and the ability to kind of crunch larger data sets to now today where we're doing a combination of that, you know, probability extraction from unstructured data sets like the entire internet. If I had to kind of summarize it, is that kind of a fair assessment of that evolution?  

Hey, you're sounding like a professor. You're using all the right jargon, boy. No, that's that's excellent. That's, that's a great summary of what's going on.  

One of the things, though, that has been most interesting to me is like, there used to be the idiom when I started my career in technology in the late nineties was, you know, I hate computers because they do exactly what you tell them to do. But that's like not quite the case anymore. And that's one of the things that I find. On the one hand, okay, it's kind of quirky and interesting, but on the other hand, I see it as a great cause for frustration or potential cause for frustration as people start to use, you know, modern AI systems, mostly generative AI and large language models, because those are kind of the flavor du jour, if you will. And those are not deterministic. Those do have a little bit more not randomness, but a little bit less human predictability embedded in them based on how they work today. How do you feel about this transition?  

Well, first of all, you can't argue with the results. Clearly, these systems are performing at a phenomenal level. But as a scientist and somebody who has built systems and then tried to fix them or debug them, it's becoming incredibly difficult because these systems are so complex. They've been trained on not just gigabytes or terabytes or petabytes of data. I mean, just massive amounts of data. But they've also required incredible amounts of processing where there's no such thing as an exact answer. Because in many ways, how you take the computing and data and you throw them into this little pot and the way it gets stirred can come up with different answers. And you add in another piece of data and everything can percolate and change. It makes it incredibly difficult to test, debug, authenticate these systems. And from my point of view, it's challenging because you want a system that's doing the right things for the right reasons. And in the beginning, to your point, it was actually easy. You could get a trace of everything that the program did and followed through the logic and validate manually that everything was correct. Now you're just doing statistical analysis as to whether it's correct. And quite frankly, um, these systems are so large and complicated right now, uh, that there are likely problems and bugs. And when these systems produce an answer that you know is wrong, or maybe we can jokingly call it a hallucination. Yeah. Diagnosing what happened turns out to be incredibly challenging.  

Yeah. We we experienced this the other day. And I mean, just to take a little, you know, 30s for a tangent here. We train. I mean, we help organizations with their governance, their own AI adoption. And so part of how we do that is we run a lot of AI internally. So we understand not so much the actual algorithms because to your point, it's really, really difficult to understand what's going on kind of behind the scenes in the model. But for instance, just like what the interactions look like, how could you incorporate them into a corporate workflow or a business process or whatever? And so we have one that we have very limited to a specific Slack channel, and we give it certain documents for analysis purpose. And we've got an upcoming conference that we're all going to. And we gave it a bunch of our travel plans, and it kept consistently getting the date and the day of the week alignment wrong. And so we're looking at and we're saying, hey, we're arriving on Sunday, August second and we're leaving on Sunday, August ninth. And it kept telling us that it was Saturday, August ninth. And so we, you know, continuously correct. And we say, no, no, August ninth is a Sunday. And then it says, okay, great. You're arriving on Monday, August second and leaving on Sunday, August ninth. And we're like, no, August second is a Sunday. And so every time you would kind of correct either end of it, it would adjust by one, and you'd have that off by one error at either the arrival or the departure date. And we had to reframe the entire thing. And this is with a what is considered to be a very capable modern model. And one of the things that I kind of look at, and to your point, I kind of scratched my head and I go like, what is actually going on here? Is there an actual math equation being executed? Is there an actual calendar lookup, or is this just token prediction that's happening and it's output back to us? And I think that's a really difficult thing for most people, ourselves included, to try to, you know, kind of puzzle out the logic of what's going on in this calculation.  

Since you're able to repeat that bad behavior over and over again, it's something that's factual in the program. It's either a bug or it's a flawed data set that this analysis is relying on. But in either case, somebody has to diagnose it and it's going to be challenging.  

Interesting. How do you talk to people about this today? Because I'm sure you've worked with other organizations that are telling you the same thing. They're like, yeah, the results are you can't argue against the results. But then at the same token, I can't guarantee you one hundred percent accuracy. So what's, what's the advice that you give to people today about how they should think about this?  

The, the acronym AI is a terrible acronym, in my opinion, because artificial intelligence, the field is not artificial and it's not intelligence. You just talked about, um, you know, token prediction just a moment ago. Yeah. These systems are really completing sentences. They're finding words. What is the next word? Predicting it and stringing it together. And it's amazing how well it works. I personally would not have predicted it would work as well as it actually does, because it does not understand the semantics. You're going to ask a question and it comes back with a beautiful, eloquent answer, but it doesn't understand what it's saying. There's no cause and effect. There's no meaning. There's nothing behind what it's saying. It's just a sequence of tokens or words to the computer. AI really needs to be thought of as augmented intelligence. Okay, we need these programs to help augment or inform us so that we can do better decisions or write better documents, etc..  

The reason I'm emphasizing that is because we need to have the human in charge at all points. These programs, because of their non-deterministic nature, they're flipped. There's a lot of coin flipping that goes on in an LLM, a large language models like ChatGPT or Claude or Gemini, and the answers that can come back may be wrong. And sometimes they're wrong in a trivial way. In your example, hey, it's not Saturday, it's Sunday, right? And other times they can come back and they're wrong, but in a way which is convincing to you that it's right when it's actually wrong. AI as augmented intelligence is important because the human has to take responsibility for these AI's. Think of the AI as being I'm hey, I'm a professor. It's my graduate student or it's my. If I'm a doctor, it's my intern or if I'm a lawyer, it's my articling student. It's your apprentice. The point here is that you are the master and you've got an underling, and you're asking that underling for advice, and they'll give you the advice. They'll give you an answer. But you're the professional. You're in charge. It's your job to validate that the information is correct, and if so, use it appropriately. And when it's not. Coax your apprentice, your graduate student, your intern to try and get them to come up with something that is correct that you can use. The human has to be in the loop. The human must be in charge. The human must take responsibility for anything that they decide to use. That comes back from one of these programs.  

There's three points of what you said that I'd love to dig in a little bit more on. Number one is just on the responsibility side. There's already legal precedent now that organizations are responsible for the results of AI systems that they use. If you put a customer support chatbot back by an LLM out into the world, and you expect your customers to interact with it, and then that chatbot, quote unquote hallucinates a discount to your customer or a free license. Most times that's going to that's going to be upheld in court. The one exception to that that I saw was the people who kind of tricked the sales agent into selling them a car for ten dollars that apparently was not upheld in court. But other things like bereavement fares and whatnot have been upheld in court. So do bear in mind like you have organizational liability on it.  

Second, I loved your analogy about this. Our VP of product, um, we've been talking about kind of the quote unquote, let's call it the maturity level, or let's maybe call it the capability level of these systems over the last little while. And about two years ago, he said that he said that in his opinion, the right way to think about these were you have a toddler who's an assistant but has access to every book ever written. And now it's gotten better in terms of, let's say, like reasoning and quality and so on. And I think that's more a result of training and some of the token prediction algorithms. To your point about flipping a coin a little bit better, better, relatively speaking, to the desired outcome. But the other thing that I think is really important to understand here around your human in the loop point is I see a lot of desire from organizations to automate business processes in the name of cost cutting, but not necessarily in the name of quality. And what do you see as that kind of like as the tension there? Or do you think that is a false choice? Do you think that's the wrong way to think about like where, where and why and when to implement kind of an AI driven or an agentic process, if you will.  

So I think the field right now is too immature for people to be trusting AI to do the jobs that many companies are asking the AI to do, and I think customer service is a perfect example. These AI are making mistakes. To your point, there is liability. Some of these mistakes have made companies look bad, terrible. In fact, in the media, companies have looked at the AI as a way of cutting costs and in many cases, sorry in some cases have fired or let go of all the people in their customer service and brought in AI agents. And as you as your listeners obviously know, there's a reverse trend there. Customers are not happy with dealing with these bots. It's very frustrating. Most of the time, I never get the answer that I'm looking for. If I'm actually calling or going online to discuss a problem with one of these bots, it's by definition something unusual, and therefore they just haven't got the training or the knowledge or the the AI has the sophistication to fully understand what I want. I just don't get the solutions. I just don't get the solutions that I'm looking for.  

The problem here is that AI has been sold or the impression has been created, unfortunately, that AI is going to replace humans. And a lot of people have talked about it. And you have people who are forecasting the future and saying, a lot of these jobs are going away. And the reality is that we have just incredible abilities. The human the human brain is unbelievably amazing. You know, we can tell instinctively when the AI is right and when it's wrong. We have empathy. We have understanding of society and how to work with other people. We understand cause and effect. The AI has none of this. You said it's a toddler. Emotionally, it's a toddler with this vast amount of knowledge. It's a correct analogy, but it's an idiot savant, really, because it can do one thing. And one thing well, is take this vast amount of knowledge and synthesize it, but it can't do all the other things that we want it to do. Yeah. To have a proper interaction.  

And as an aside. Although November thirtieth, twenty twenty two is a landmark day in my life, people say November thirtieth, twenty twenty two. Who the hell cares? What's. What's that mean? I've seen earlier versions of ChatGPT and okay. Unimpressive, but cool. And then November thirtieth comes along and they roll out this new version and it was like, wow, okay. As an AI person, I was wowed because it did things I had not expected, and I was impressed for quite a while. And then things started to nosedive a little bit. And we can talk about hallucinations and other problems with this technology. But the one thing that really bothers me about ChatGPT is its name ChatGPT. Who cares? Well, the GPT nobody cares about because like, it's a geeky technical acronym. Okay, forget it. It's the chat word. That's a very clever and it's a dangerous word because what it's trying to do is put a human face on this technology. You're not querying it. You're having a chat. Who do I have a chat with? I have a chat with my friends. I have a chat with my wife. I phone my daughter up for a chat. Chat is a word that implies a relationship, right? Right. And so it's an attempt to anthropomorphize these programs and look at the word hallucination. These programs are getting answers that are wrong, sometimes blatantly wrong, sometimes subtly wrong. But they're happening often enough that it's a problem. And if this was any other field, we would say it's lying. It's fabricating. As a computer scientist, I would say it's got an error rate. What is the error rate? Okay, five percent. Right? We don't say that. We say it hallucinates. What does hallucinate mean? Okay. It's an attempt to anthropomorphize anthropomorphize it again, it's like a cute way of making an excuse. Okay. It had a hallucination. Yeah, we all occasionally have hallucinations. The point here is that many of these companies design their interfaces in a way that tries to create a relationship between you and the program. There is no relationship. You can tell the program I'm in love. You can tell the program I'm dying, and the program can generate responses that you may interpret as being sympathetic or whatever, but it does not understand. Yeah. And the human aspects that are critical to us, we're not computers talking to each other. We are humans talking to each other. And there's a lot that gets said and unsaid in any kind of interaction. And a lot of what AI is being sold to people and to corporations is misleading and in my view, potentially dangerous.  

So, you know, you mentioned the term earlier augmented. I think you didn't call it augmented intelligence. You called it augmented. No, I said augmented intelligence. AI augmented intelligence. So if chat is the wrong way to think about it and, you know, I totally understand what you're saying from the Anthropomorphization perspective. And I think I think I only speculate that the argument around that is very often that AI is a scary thing for a lot of people. And putting more of a friendly, quote unquote, friendly face on it is a way to make it less scary and make it more approachable, especially for non-technical people who maybe don't understand the inner workings. But put that aside for a second. I don't want to make a judgment call as to whether that's right or wrong, but what is a productive way for somebody in a modern organization to think about it? What is a productive way to think about it? Is it that augmented intelligence approach? Is it that, hey, this is a thing that can be useful for certain specific tasks where you either have low stakes or you have, or you don't have to have a one hundred percent accuracy necessarily, or it's going to go through a human in the loop evaluation, because I'm going to use it to write an email template, but then I'm going to spot check that email template before it goes out the door. Like what's the right way to think about it to get those productive results?  

All right. So this is going to make me sound old. But I'm, I'm young enough to remember Star Trek the first series, which was nineteen sixty six to nineteen sixty nine. Okay. And I don't claim that I actually saw it in nineteen sixty six to sixty nine. I think I saw, you know, first or second reruns. That's fine. But when I got into the field of AI, I used the computer in Star Trek as a model, if you will, of what I thought AI should be. And to this day, actually, in most situations, I believe that's the model that we should be following. What did the computer do? Captain Kirk Spock would ask the computer a question. It would answer if it had the factual information, and if it didn't, it would go off and compute for a few seconds, a few minutes, sometimes longer. Analyze all the data and come up with its best recommendation. There was no personality. There was no dialogue. Spock. Kirk had to make a decision. They would ask the computer. The computer would inform them. Sometimes Kirk followed the computer's suggestions. Other times, especially when he factored in the human element, he ignored it. But the point here is that he was in charge, and the computer was informing him with information that was useful, and it was his call to to decide whether to use it or not.  

Right now, that's not the situation we're in. And this is not the situation that some companies are landing themselves in, because computer humans, because of our empathetic and our social nature, seem to build a relationship or treat many of these programs as a social interaction. And therefore, when who whatever entity you're engaging with gives you some information back, you tend to trust it. Mhm. And that's where the danger happens, because you may rely on information that you have not vetted because these programs make mistakes. And I'll talk to people and I'll talk to companies about this. And it's sort of sometimes they take it seriously and sometimes it goes in one ear and out the other because they they're taking the path of least effort. It's a computer, it's AI, it's trained on the whole internet. It's got this massive computing. It's got to be right. And people are often too lazy to check. And then there can be consequences. So I am an advocate for removing the anthropomorphization and treating the programs, these programs, what they should be, which is giving you fact based information that helps you do whatever task you're you're trying to achieve.  

That's interesting. Especially interesting because right now, if I think across the landscape of AI services and, you know, here at Firetail, we support about roughly one hundred different AI services from the standpoint of like discovery, visibility and, um, kind of organizing telemetry data for our customers. Um, I don't think a single one of them is built with this kind of fact, you know, just the facts, ma'am model of...  

I'm glad you said that. Hey, yeah. You just dated yourself. That comes from Dragnet, a nineteen fifties television series.  

In my case, I only saw the movie remake in the late eighties, but, uh, with Dan Aykroyd, yes. Dan Aykroyd, Tom Hanks. Great movie. Yeah. But anyway, but but none of them are built that way. They're all built to have a little bit more kind of, if not personality, at least a kind of simulation of a human analyzing the data. You know, if you look at, if you watch the reasoning and even the ones that, in my view, have the least personality. So let's take something like a DeepSeek, which doesn't give a lot of kind of more opinionated responses, but gives you that if you watch the reasoning, it's very much the user has just asked me for this, so now I must go perform this process. I've just tried to analyze this data. It didn't succeed. So now I must go do this. And it uses all these first person pronouns and things like that that are anthropomorphizing itself, even if the output is not super like, hey Jonathan, you should think about the problem this way and giving you kind of like leading or opinionated type of responses. So I'm curious, like, do you think that a just the facts, ma'am model could actually succeed? Would it stand out? Do you think it would be hard for people to adopt?  

Well, it wouldn't be hard for people to adopt, but I don't know that people will want that, unfortunately, because again, we're we're social in nature, but the reality is, is in the vested interests of these companies to make to anthropomorphize these programs, because you can build a relationship with, um, these chatbots. I have friends and I always correct them when they say, well, he said that when they're referring to ChatGPT or my, my advisor says I should do that and they're talking about Claude or something like that. And this is part of the vendor lock in. Yeah. These companies want you to use their product and not look at the others. And so it's quite subtle and maybe it's insidious by building a quote relationship unquote, with that particular program, you're more likely to do most of your work with them. You're more likely to put your files in their file system up in the cloud, you're more likely to pay a twenty dollars, thirty dollars, forty dollars monthly fee or more, or whatever it is they want to lock you in. And so part of it isn't just the quality of the program, it's also the more subtle, human like things of exploiting our social nature is.  

So I'm curious what surprises you most about, you know, at the beginning of the conversation, we talked about kind of the evolution of AI and the kind of the current state where we're at, where, you know, to, to quote yourself, like, you can't argue with some of the results. Some of the results are super impressive and they can be very assistive for people trying to get tasks done, etc. What do you find most surprising about the current AI? Is it that that the technology has gotten to where it has in a relatively short time period? Is it surprised at the way that humans interact with it and do anthropomorphize it? And, you know, talk about having a relationship or call it their friend, their assistant, their buddy, or is it something else?  

Uh, the thing that impresses me most will sound rather strange. It's my inability to predict what's happening. I've made lots of predictions in the past. People have asked me, hey, Jonathan, when's this going to happen? And I've been right most of the time. And now that sounds egotistical, but the caveat is I have been wrong, wrong, terribly wrong, way off. When it comes to the time frame. And so. When ChatGPT came out on November thirtieth, twenty twenty two, it's not wasn't a surprise to me that programs could have those kind of level of abilities, but I didn't expect it then, and I is still a long way off. And what's happened since then is like this hockey stick phenomenon where it's just gone up, up and up in terms of, of the abilities. There's the, um, I call it the, the, the, the triumvirate of technology. We've had this tipping point where three things have just come together, massive amounts of data, incredible computing resources to analyze that data and sophisticated algorithms that can run on these computers to analyze that data. And, you know, data came along, but nothing was happening. Fast computing came along and more interesting thing happens. And then you throw in things like deep, deep learning or reinforcement learning or the, the transformer, and all of a sudden things just take off. So what has astounded me is it's really unprecedented in human history, which is the incredible advances that are happening literally daily or weekly in the field of AI over the last three, three and a half years. And of course, with a trillion dollars or more of money being pumped into AI in terms of infrastructure like data centers and fundamental research and people and education, uh, I don't see this slowing down. And so what has really stunned me is things that it didn't take much work to gaze into a crystal ball to see would eventually happen many years or decades into the future is now happening literally. To use the Star Trek analogy at warp speed. Yeah. Right now. And I don't see an end in sight. It's just truly staggering.  

From my point of view, it's you make an interesting point. And as you were as you were making that point, I wanted to check something because, you know, the trillion dollars, I don't know if that's the exact number, but it sounds right when I see the if I think about the total of the, you know, one hundred billion into Anthropic, two hundred billion into OpenAI and all of these kinds of things. So, you know, trillion sounds very reasonable. And what's really interesting is, just to your point, the amount of progress being made in basically a four year time frame. And I know there was a lot of groundwork and foundational work that happened well before that, that a lot of people don't appreciate. But let's just say like it's a trillion dollars over what, ten year time frame, fifteen year time frame, whatever the case may be. Put that in context of the tens of trillions that have been spent on the quote unquote, war on drugs over like a forty year time frame, or the well over a trillion that has been spent on eradicating poverty worldwide over a forty or fifty or sixty year time frame. And the results on those side, they really pale in comparison to the progress made on the quality of this. And I don't want to comment on what that says about, let's say, the quality of the institutions or anything like that that's been behind those. But it does show to me that, you know, if incentives are properly aligned and there is, you know, resources available to pursue these programs, we can make big structural changes in a very, very short period of time. So that's just the interesting observation from my side. As you kind of rattled off that number, and I started thinking about the time frame as you were going through it. It is it is like mind boggling how quickly and how far this has come in that time period.  

It is. But that's also scary because we have a number of well-financed, very, very large companies who are in a race and money seems to be no object because investors are pouring massive amounts of money into these companies. You know, SpaceX briefly, there was, uh, at an incredible evaluation. We know that Anthropic and OpenAI want to go public and they'll be somewhere, I don't know, close to a trillion dollar valuation. Um, so there's companies and money out there and they see a pot of gold at the end. It's a race. They're all racing to the finish line. And there's this, I think it's mythical notion of a what's called AGI, artificial general intelligence, or ASI, some call it artificial super intelligence. And there's this notion that whoever gets there first, wins the pot of gold, and the pot of gold is worth just a fabulous amount of money. And the problem that concerns me is it's more important to win than it is to fix the problems that you create along the way. Mhm. And they're racing forward at breakneck speed, and they're leaving problems in their wake like hallucinations. Yeah. Concerns about people using or abusing these tools, creating an unsafe environment or unsafe. Then there's security problems because, well, these programs, they execute text and you can inject into text security vulnerabilities. You can create problems, uh, with, uh, um, malicious, uh, prompts. And then there's the data center problem, which they're sprouting up like mushrooms all over the place, generating a, sorry, not generating, using an enormous amount of electricity and water, potentially creating environmental impacts.  

So these are some of the problems that are being created, but they're not necessarily being addressed because it's from a corporate point of view. They're in a race. And the only thing that matters to these companies is who gets there first. I don't have a good answer to this, but it's a unique problem. It probably won't be the first time we see these kinds of, uh, heavyweight sumo wrestler companies battling for supremacy. We may see the same thing in quantum technology or some other technologies down the road. But we as a society need to step in and slow things down. You can't keep running at a breakneck speed towards this finish line without accidents happening. And accidents are happening. Yeah, and let's fix the problems.  

The reason why I I'm going off on that tangent. Is I believe with all my heart and everything I've worked for, for forty plus years. And that's true for almost and probably every single colleague I've ever worked with is we all believe that AI can be a force of good in this world, that we can use AI to improve the quality of life of everybody on this planet. But if you want AI to be adopted by businesses, by consumers in a way that adds value and quality to everything we do. Then there has to be trust. And right now, trust in AI is at a low point. I'm in Canada, so I have some recent data from literally a couple weeks ago, a report came out and sixty seven percent of Canadians don't have trust in AI. Seventeen percent don't know. They don't have enough information to decide, and only sixteen percent trust AI. A United States might be a little bit better or higher numbers, but it's still a low number in the United States. I saw one from last year that put it at forty percent. The point here is this technology to be widely deployed, you need to build trust. I go out and I buy a car, I research the car, and when I buy it, I have trust in that vehicle. I trust the manufacturer. They've done all their due diligence. I trust the government who's got regulators that do inspections, that ensure that it meets all the standard safety requirements. There's trust. There is a lack of trust right now with AI, and that, in the long run, is going to be a real serious limiting factor towards the deployment of this technology in the ways that that I would eventually like to see.  

That's a super thoughtful kind of closing thought, but we're not quite done yet. But I do want to transition into a couple of quick topics to wrap up, but I do love the sentiment there, and I really appreciate you sharing that with our audience. Two things to wrap up. Number one, tell us about the checkers and the poker.  

Uh, in nineteen, I was first of all, in the nineteen eighties, I had one of the best chess programs in the world, tied for first in the world Computer chess championship. Uh, but then I lost out because colleagues, uh, built Deep Blue and I switched to checkers. And in nineteen eighty nine, I had this vision that I could build a world checkers champion. And I did in nineteen ninety four. The program won the Human World Checkers Championship against the world human world champion, Doctor Marion Tinsley. And that was recognized by the Guinness Book of World Records. That was nineteen ninety four. Tinsley died, unfortunately, in nineteen ninety five, and people said, oh, your program could never have beaten him when he was in his prime. And so I thought, well, uh, Tinsley was a great player, but he occasionally made mistakes. If I built a program that didn't make a mistake, then I would be the better player. And it took another thirteen years of number crunching on hundreds of computers around the world before two thousand and seven arrived, and I announced that I had solved the game of checkers. I had a program that would never lose. It would play checkers perfectly.  

Okay. And I happened to do a little bit of research. And I found that one of these things and I just, let's say fact checked this after eighteen years of computation, you found that checkers always results in a draw. If neither player makes a mistake on a solution that involves ten to the fourteenth calculations from endgame positions with fewer than ten pieces on the board. Is that accurate?  

Yeah. Interesting, I. Checkers is not a game that I've played a lot. I've played, I don't know, twenty times in my life and never really devoted any serious time or thought to it, but I did not realize that there is a, you know, optimal move for every board situation. I guess that must be also kind of the case or provably the case.  

Yeah. In every situation there is one or more best moves. But interestingly enough, checkers is a microcosm of what's going on today because we built massive data sets. Ten to the fourteenth may not be big today, but you have to remember this was in the early nineteen nineties. I started building this. Yeah. Put put your head around this. I bought my first one gigabyte disk in in, I think it was ninety or nineteen ninety one. It was five thousand dollars plus tax for one gigabyte of data. So we've come a long way and massive computing. Well, I didn't have GPUs or AI chips back then. So my solution was to use hundreds of computers around the world to do the computations. So you had, you know, this this technology trio, the triumvirate that I talked about, you had lots of computing and lots of data. And my job was to create the, the learning algorithms that combine them. And then we got a superhuman checker playing program.  

Awesome, awesome. And just to wrap up on today's episode. If people are interested in your work, where's the best place for them to find out? And then give us a quick overview on Kind, the first product coming out of Synsira.  

Right. So I talk the talk. So I decided I had to walk the walk after forty years as a professor at the University of Alberta in Canada, I left the university. I formed a company. It's called Synsira Software Solutions, and we produced our first product. It's called Kind. It's commercially available. You can see it on our website, synsira.com. And you know, all those problems that I talked about with LLMs and more, we try to solve them. So we have a product that will run on your desktop. A new version that comes out shortly is one hundred percent on your desktop. It's like having your own local LLM. Your data stays on your desktop. It's completely private. We use the resources of your computer, which is typically a powerful computer and typically idle ninety plus percent of the day. So you can do all your local, your, all your AI processing locally in a private way without having to use and encourage the building of more data centers. So private, secure, accurate AI on your data, only your data, it's not polluted by any external sources.  

Interesting stuff. And we'll have a link to that, as well as a link to your profile and your Wikipedia page. Uh, I think you're one of about five or ten guests that we've had on modern cyber that has their own Wikipedia page. We'll have all that linked from the, from today's show notes. Jonathan Schaeffer, I cannot thank you enough for taking the time to join us here. We went a little bit long, but I think it was definitely worth it for our audience. So thank you so much for taking the time to join us.  

Thank you. It's been my pleasure.  

Awesome. We will talk to you next time on Modern Cyber. Thanks so much. Bye bye.  

‍