In this Digital CxO Leadership Insights series video, Mike Vizard talks with Arize AI CEO, Jason Lopatecki, about why the key to artificial intelligence (AI) success stems directly from the ability to observe machine learning models both before and after they are deployed.
Mike Vizard: Hello, and welcome to the latest edition of the Digital CxO Leadership Insights Series videos. I’m your host Mike Vizard. Today we’re with Jason Lopatecki. He’s the CEO for Arize. And we’re talking about the need to monitor and observe machine learning models. Jason, welcome the show.
Jason Lopatecki: Yeah, thanks for having me.
Mike Vizard: I think everybody often has it in their head and says, “Wow, we’re going to build an AI model. And we’re just going to invoke it through an API, and we’re done. This is awesome. It’s just great.” And then there is this nasty little thing called drift that occurs, but I’m not sure if people know exactly how drift happens. So maybe you want to explain to folks why drift is something I gotta pay attention to when I’m creating my models?
Jason Lopatecki: Yeah, so I’ll start from a high level; you know, the reason I started this company, this is my second time, as an entrepreneur. I built a company previously from early stage to a public company, you know, rented headquarters as a public company. And I was the founder who built our data science team. This is kind of the early days of data science, we didn’t have all the tools we have today. And put lots of models in production, the models essentially ran our business; we had programmatic advertising, the software would decide what ads to buy. And when there were problems, when there were challenges, customers would come to me and in my team, and I’d be on the front lines. I’m trying to answer why, why the models did what they did. And when money’s on the line, your business is on the line, you’ve put these models into your company to do things that are important. It, for me, was eye opening, that there was no solutions to help and it was eye opening how complex the models were, when trying to get answers of what the problems were. And so when I, you know, I exited that company and wanted to go start my next one, this was the area I was passionate about; I knew the pain. I’d felt it myself, and wanting to go build a company, to help people really do build software to help people understand what models are doing when there’s problems, help you fix them. And what data is causing those problems, do it quickly. And really, for us, it was really managing our investment in ROI, our ROI, making sure that ROI worked, and we can get to problems quickly. So that’s kind of how I got into it myself. Why, why it’s important if you’re investing and putting models into anything important in your business. And I saw I’ve seen, you know, Michael, with your interviews with Capital One, you’re interviewing a lot of people who are doing that today, putting models into important things in the business. You need software to help you tell you help to tell you when something’s wrong, and get you to the answer around the data causing that problem. So that’s how I’ve got into this space myself; a little bit of a big high level picture around why it’s important. And then happy to dive into more of the details.
Mike Vizard: What do you think are some of the root causes of drift that are out there that people aren’t paying attention to? It does seem like the data changes, but more importantly, the business changes? So when do I know that the model needs to be tweaked?
Jason Lopatecki: Yeah, so people are building, we see as a broad spectrum of ways in which in time periods, people build models. I have some customers who are building models every day, I have some customers who are building models weekly or monthly. And depending upon that period, that data moves underneath you; the actions your customers were taking or the training data that your model was built on last week might not represent what’s happening in the real world today, I would say what we see as well is, it’s so easy to introduce a problem into a data pipeline, we know how easy it is. How easy is it to get two underscores or to just mess up a name that was fine in one week and destroyed the next? The models are built on this data and they’re looking for the data they’re built on when those mistakes occur, very easy to cause, you know, company costly problems, not just to your bottom line based upon the decisions it’s making, but also the time spent in trying to track down through arguably your most complex systems; what data is causing that issue, that really choking – breaking or causing problems in your model. So what we see is data changing all the time – we see data changing in between periods of model building. We see simple mistakes in pipelines causing the big, big outcomes and really tough troubleshooting problems. All coming up to this thing where people call it model drift, which is really – is the data I built my model on different than what it is acting on? And is the way it’s acting problematic? Did it not extend to this data naturally?
Mike Vizard: Do you think senior leaders understand the risk? As one once said to me, it’s one thing to be wrong, it’s another thing to be wrong at scale. But do the business leaders kind of get that notion and that idea, and they have a intelligent conversation with the data science team about it?
Jason Lopatecki: I think we’re kind of in the first inning still in this space. I think most, you know, that first inning, it’s about excitement. It’s about the buzz. It’s about the ability to solve these tough problems, which can move your business. And so you throw data scientists at it, you throw people out at it, you throw teams at it. And the last year or two, we’re really about building that next batch model, building that thing to move that needle in the business. But once you deploy it, you end up with a whole host of other problems. And I think we’re kind of in that coming away from that first inning, which is we’re in the second phase where some people know I’ve talked to some business leaders who’ve had big problems, lost a lot of money and they can point to specifically what that issue is; I think many are still in this this, you know, inning zero of like, the buzz and excitement is there and you’re solving these problems, but you don’t know the issues that they’re causing. Because you don’t have visibility into them. You don’t know what it’s doing. So I’d say it’s a mixed bag but we’re early and I don’t think everyone knows what the potential risks are as you put these models and important business decisions that you can have without knowing what’s going on.
Mike Vizard: We’ve had the idea of continuous monitoring and observability and DevOps and application development forever. Are we kind of replicating some of the best practices that we use for application development for the building and deployment of the models? And they’re slightly different terminologies and slightly different use cases, but the concepts are similar. And can people follow along in that regard?
Jason Lopatecki: Yeah, I think the growth and maturity of infrastructure, observability, which is I think you’re talking about the data dogs, new relics, really has made the conversation about the need for observability to be easier. People when people know they need it; they don’t know necessarily what it is, or why it’s different. So there’s, I would say, three categories I see coming out today. There’s your traditional infra observability, which is about timing, your app timing, timing problems, and how it’s related to software that’s causing those timing problems – tracing, logging, all relate to that. You have data observability, which a growing set of like Monte Carlos or big eyes. And they’re doing like table monitoring. So let’s make sure your warehouse, your SQL update that you did to your warehouse didn’t break anything. And let’s make sure the tables freshness is great. And I would say there’s this third pillar of monitoring, kind of observability coming up, which we are the leader in, which is model and AI observability. And in our case, what we’re looking at is the model entity instead of data as the table infrastructures like app and app timing models; we’re looking at models and we’re looking at model performance. And a lot of what we do is outcomes. So a model is predicting an ETA – when will your driver come to you? Well, am I off? Am I off by 20 minutes? Am I off by 10 minutes? What data caused it to be off by a very statistical nature? So we’re looking at statistical measures of difference in data, we’re looking at that by feature, we’re looking at and tracing through performance. So very model centric, very outcome centric. I think the big idea observability people get; I think those are how each of these three are different. Different teams have different understandings of it and are further along their journey.
Mike Vizard: We hear a lot more interest from regulators and all things AI these days. Do you think that at some point organizations are going to have to produce an observability report that kind of see tells people this is how the model behaves? This is what happened and, you know, letting them actually understand what the AI model is doing versus treating it as a black box.
Jason Lopatecki: Yeah, I think it’s coming. I absolutely think it’s coming. I think that there’s probably arguments on whether it’s good or bad and whether the regulatory stuff is good or bad. I probably won’t have a position on that. But I think it’s certain that it’s coming. It might be coming faster than we all think. I think most of us, I don’t know if you’ve interacted with ChatGPT recently, but wow. Talk about a step up in closeness to AGI, meaning like a real something that might pass the Turing test, or something might feel as slightly more human than we think. And I’m sure it’s gonna scare a lot of people. I’m sure that what they, what the exponential pace in this industry right now, and the breakthroughs we see, I think will scare a lot of people. And you will get regulation in some way and whether and what that looks like and how it happens. And is it simply a report? Is it required for monitoring? I don’t know. And I don’t also want to take a position on if the regulation is good or bad. But I think the pace, and what people are going to see is probably going to force that, especially with what we see kind of going on in the different either early stages of incarnations of those laws in different regions – it just feels slightly inevitable. Are they going to get it right? I don’t know about that.
Mike Vizard: Yeah, all right. Well, just for the record, we can validate that this is actually me talking to an actual you?
Jason Lopatecki: Yeah, yeah. Maybe not actually.
Mike Vizard: That said, you know, there’s an old maxim that says that which you can name becomes less scary. So, do you think observability will help make AI less threatening because people better understand exactly what is going on here?
Jason Lopatecki: I think so. I think, hopefully, we’re a component that helps. You know, I deeply believe that software that helps the humans understand what AI is doing, when it’s off the rails, how to fix it, is absolutely needed. So I think we’re part of that, which hopefully eventually brings comfort to those that are concerned. Realistically, when you look at what these do, and models do, I don’t think we have this. You know, those of us in the details don’t feel like this is general intelligence. Don’t feel like it’s out there. But it can feel like it as you’re interacting with it, it can start to build fear in people that don’t understand it. And hopefully software that helps humans troubleshoot, monitor it, fix it brings one other level of comfort to the mix.
Mike Vizard: So do you think we’ll be in a world soon, where we’ll be able to identify when something is using, say, GPT algorithms? And therefore, the monitoring tool helps surface that so we’ll know what we’re engaged with at the time?
Jason Lopatecki: I think you will. I think I think you absolutely will. And like all this stuff, there could be cat and mouse games, can someone you try to get around the detection of if you’re using AI or not? Yes. So I think there’ll be systems in place to help you understand if, if what you’re interacting with is, is AI and there’ll be systems to try to get around those systems. So expect a bit of an interesting world over the over the next five years as the space develops on. I definitely think absolutely. There will be many instances where people will be interacting with AI and not know that over the next five years. That is a you know, that is a given.
Mike Vizard: Old school – that which I can observe I can manage. So the next question becomes if I have observability, can I start setting metrics that create bounds for performance and allow certain things to happen so that things don’t go off the rails?
Jason Lopatecki: Yeah, so absolutely; that’s the core of what we do is try to make that consumable and easy. And take what is one of the more complicated technologies out there and try to help you put the bounds on it and create automatic checks and tell you where things are off. So that’s the core of what we do. I would say bigger picture as an enterprise, you know, you again, bring it back to this maybe the Capital One interview that you did. There’s another angle of this, which is, enterprises are investing a lot in AI and ML, and it’s just going to grow. But I don’t know if they all know what the ROI is on that. Like, it’s very hard to understand, like if you’re not tracking what it’s doing. Like, if I put it in a model into production that’s supposed to improve something, but I don’t know how much it’s improved and how good is it versus random – what is my investment like? That investment, that time it took to build it – you know, should I have done it? So I think there’s this other enterprise level problem I see, which is, in this first inning, we don’t care about return on investment, we’re just investing. But there’s another phase here, especially in this environment, where people are cost centric and cost conscious that the return should matter, and eventually well, and to know that return, you need to know how well it’s doing. You need to know how it’s doing versus not doing it at all. You need to know if it’s affecting certain things, you know, important outcomes worse than others. And so, there’s a lot around observability, that isn’t just, you know, but isn’t just purely technical, but it’s also like, you know, business wise, is it like getting what I want out of what I put into it.
Mike Vizard: So with that in mind, what do you know now that you kind of wish you knew two years ago, because to your point, a lot of people are just investing now because they’re afraid they’re going to be left behind. But you’ve been out here early, and you know, what some of those things you can share that might lower the quote unquote “idiot tax” for everybody else.
Jason Lopatecki: Yeah, um, my view is start early with the right components around your investment. So what I see is a lot of teams building models and starting with data science. But do it in a way that’s like, maybe the data science, maybe those data scientists have never done MLOps before, or they’ve never put a lot of models into production. Or maybe you’ve not even been in a production engineering team. And so you can get, you can start early, and not think about the tools, the great tools that are out there that help you, and you end up just building everything. So there’s a lot of companies that built kind of everything, every every tool, and what you end up with is a sprawling kind of unit with a bunch of technical debt, which you’re kind of cleaning up. And it makes sense in the early stages of a market when you don’t have tools, you don’t have things. My advice would be choose the right pillars, have your tool sets in the beginning, so that you’re not building these and replacing and building a lot of tool sets. You know, outside of us, I would say like Weights and Biases is an amazing company for experiment tracking. So check out them. And what that does is the foundation, it’s the foundation so that, you know, with us, the people were kind of like, well, do I need model one and model your model – wondering do I need it after five models or six models? And our answers typically, well, if the thing you’re putting in production does anything important you probably should have something in place. So I think it’s – get the foundation of tooling, right. Don’t try to build everything. If you try to build everything, you’re not going to do that. And yeah, and that’s probably what I would tell people from what I’ve seen here is get the foundation right in the beginning. Think about MLOps in the beginning. Choose your right set of pillars or partners that enable you to go from one model to 20 models without rebuilding everything.
Mike Vizard: All right, folks. You heard it here, keep your friends, enemies and AI models as close together as possible because you don’t know who’s gonna do what next. Jason, thanks for being on the show.
Jason Lopatecki: Yeah, thanks for having me.
Mike Vizard: All right. And thank you all for watching this latest episode. You can find this one and others on the Digital CxO website, including that Capital One interview that Jason so kindly referred to, and we thank you all once again for spending some time with us.