The Philosophy of AI: Examining the Ethical Implications with David Danks (Part Two)

Forbes Books presents the Startup Science Podcast with Gregory Shepard, brought to you by startupscience.io. Greg is an entrepreneur who has built and sold 12 businesses. He's a recipient of 4 private equity awards and is a featured TEDx speaker and Forbes author. Here's Gregory Shepard.

I'm back with David Dayes, the professor of data science and philosophy at UC San Diego. We've been discussing the intersection of philosophy, cognitive science, and machine learning to tackle complex problems we see around AI. David, let's dive into time series models. First, what is a time series model? And second, how can entrepreneurs utilize that for forecasting, planning, and dealing with challenges on the fly?

Yeah, it's a great question. So time series models are basically Models of things that evolve and unfold over time. So sometimes I just want to predict, do you have cancer at this moment based on a radiographic image? That's not a time series model. I've got this moment in time. Do you have cancer? But often what I want to do is predict how might your cancer unfold over time under various treatment regimes. So if I give you this chemotherapy or this surgical intervention or this radiation treatment, Can we figure out how the cancer is likely to change over time? That requires us to explicitly model time. We have to think about the ways that things change over time, that the cancer probably isn't going to change much in 1 minute, but it will change over 1 year. And so it raises lots of difficult challenges from a data science perspective in terms of how we model time. How granular are we? There's a kind of trade-off of If I measure a lot very quickly, that's great for accuracy, but computationally inefficient. If I measure every once in a while, that's computationally efficient, but I'll get less accurate. So it's these kinds of trade-offs that we worry about. Now, for entrepreneurs, the key is to think about what kinds of sort of time data you're going to have. What access are you going to have to learn about how things unfold over time? In whatever domain or problem you're looking at, and then figuring out how do you balance this trade-off of you don't want to remember all of the past because that's computationally a nightmare, but you want to remember enough of the past to inform your future predictions, to be able to help people plan out and trace out the possible trajectories of what lies in front of them. And that is. You know, there's no technical answer to that question. That trade-off of how much do I remember, how do I balance accuracy in my future predictions with the computational cost, there's no technical solution to that. That's going to depend on how important is it to get the future exactly right? How far out in the future am I trying to predict? How much do my customers care about me trying to hold on to the past data? Do I have access to really big compute and storage and all these kinds of things? So what's important for an entrepreneur, I think, is, and you're going to notice a theme in some of the things I've been saying, what's important is that they be intentional, that they stop and they think to themselves, okay, what are we trying to achieve? What kind of balance is most important or most valuable?

And am I sure that I'm getting things right from the perspective of my So when you're thinking about these time series models, how can it be used to identify capital and emerging market trends, right? And these customer behavior patterns, like it seems like, you know, the whole buying and selling of stocks will be incredibly disruptive because of these time series models. Right? So you could predict based on certain sets of data that are publicly available outcomes. And then, you know, like I said earlier, like a lot of the white-collar people, I mean, you know, what do you need a broker for if the system is, you know, better than an army of brokers? So what are your thoughts on that sort of stuff?

Yeah, so the first thing to say is we're already a long way down in that world. So the vast majority of trades on stock markets are now done by automated trading systems. That are exactly using time series models to try to predict very short-run fluctuations in where markets are going and where particular stocks are going and to try and take advantage of very small inefficiencies. Now, what that highlights is that we're using automation, we're using AI systems essentially right now for the very quick decisions. We're not using them as much for the very, for the longer-term decisions, for where is a sector going, what is a good value stock. We're still, I think, in a lot of ways relying on human expertise, in part because we don't know exactly which information about the world is most important. So we're relying on human expertise to do a lot of that information integration for us. But you're absolutely right that it's one of the, I would say, holy grails of fintech is how do we come up with models that are really good predictors at the Let's say 1-week timescale. We've got models that are very good at the quarterly predictions. We've got models that are very good at the 1-second predictions. It's actually the middle ground where I think that there's a lot of uncertainty right now and a lot of challenges are arising. One of the issues that comes up, of course, with the kinds of time series models you're talking about, say in financial technologies, is that you're trying to model the very thing that you're then trying to influence. Right. So there's a sort of self-referential problem here, which is I could get the model exactly right, but then I want to do things. I want to take advantage of inefficiencies in the market. And that activity, the act of trying to take advantage of the inefficiencies, can actually mean that I'm falsifying my own model. I'm making my own model wrong.

Oh, interesting.

Another big challenge is, uh, how do we have models that explicitly include Here's the things that I might do so that we have a model that can be a bit more stable over time.

So it seems that if you're looking at the startup ecosystem, right, it seems like you could do, you know, you would have investors in private equity using this, right? So you submit something to the investor that gives them a bunch of data, they put it into a series, and then it spits out, here's the possible outcomes or probable outcomes For this particular business. And I'm not sure if that's good, you know, because when you're early on, you know, you're still trying to figure out your business and it takes a couple of years to understand the customer well enough and how you should, you know, work through it because, you know, you can spend years learning about the customer and build something wrong. So if an investor is looking at a business and they're looking at these models, maybe those models would drastically change if the founder made a different decision or pivoted or whatever. So what are the dangers that you see and how could a founder utilize it instead of for risk, use it for opportunity?

Yeah, so I think you're right to raise all of these kinds of risks, the risk of sort of the homogenization of startups where everybody gets directed towards the same thing, the risk of self-fulfilling prophecies. If we do this, then, then we're going to end up in this feedback loop where, where we think we're getting it right, but what we're actually doing is misreading or misinterpreting what it is that our customers want. So I think you're right about that. I think you also hit on exactly the key idea, which is you can in some sense use these models as predictions about everybody else. And that can help you to see where are the blank spots in the map, where are the parts of the map that nobody else is going to be ending up in. So, and I don't think I'm revealing anything, hopefully that isn't too big of a surprise to some of your listeners, but a lot of good VC companies are already doing this. Private equity companies, they're already doing this. They're trying to predict where's everybody else going to go so that they can figure out where the marginal advantages might lie. Where's the part of the landscape that is unexplored. And so I think it's a great idea. The trick with all of these things is the quality and quantity of data that we have. In AI, we sort of, you know, when I look back over the now over 25 years that I've been doing this, in AI, we've gone through cycles where we're constrained by not having good algorithms. Like we our algorithms and models aren't good enough. And then we'd be constrained by the data not being good enough, right? So sometimes we've got algorithms and models that are really good, but we don't have good enough data for them. And then we get good enough data and then we realize the shortcomings of the model. And so we have to build new models and we sort of oscillate back and forth between what's the sort of rate limiting factor. Is it the model, the kinds of models we have, or is it the kinds of data we have? Right now we are on the we need better data part of that cycle, I think.

Yeah, the quality is horrific right now.

Exactly. And so in fact, we you know we know from public comments that some of the big tech companies, their major advances in the last six to twelve months have not come because they've gotten better algorithms or because they've gotten more compute. They've come because they've gotten much better data, much cleaner data. They've gone through and had people label the data and make more. sense of the data. So we really are at a place right now that if you want to be able to, to get ahead, the best advice I would give an entrepreneur is figure out if you have access to data that nobody else has figured out how to use yet.

That's—

so let me ask you about this causal discovery and what sort of techniques could founders use to find opportunities, just like we were talking about the investors. You know, finding those opportunities.

Yeah, so the basic idea of causal discovery is, um, uh, kind of one that we're all familiar with, which is we can look out in the world and we can see things that are correlated with one another, but that doesn't mean that they're causal. So if I want to predict or figure out whether it's raining outside— and granted, I'm in San Diego, it doesn't rain very much— but if I want to figure that out, I can look to see if people are carrying umbrellas. But if I want to make it rain, it does no good to tell people to start carrying umbrellas. That's not an effective strategy to make it rain, to say, hey everybody, start carrying an umbrella around. And so things that are good predictors are often not good causes. They're not the things that we can use to control or shape our world. And but in fact, when we're doing policy, when we're creating new systems, we often want to control our world. I don't just want to diagnose that somebody has cancer. I want to be able to do something to fix the cancer, but that means I need causal knowledge. So over the last 30 years, there's been a, a whole lot of folks, myself included, who have developed new algorithms to figure out something about the way the world works causally from data that we collect where we don't intervene on the world. We don't mess with the world, but nonetheless, we can figure out how the world works. So the advantage that I think this holds for entrepreneurs is that many businesses depend on having some kinds of causal knowledge. They depend on being able to do things in the world to make people's lives easier, to help people find new kinds of information or new opportunities. And so we often need causal knowledge, but most AI systems are only going to give you good predictions. They're going to tell you, well, you might try having everyone carry around an umbrella and see if that creates rain. So there's a real opportunity, I think, for entrepreneurs to think carefully about where and when they might need causal knowledge, to think about, you know, what does it matter that they're going to be trying to wiggle something in the world to bring about a good outcome for their customer? That, you know, this new ad campaign, is it going to raise sales? Well, That requires causal knowledge. And so I actually think a lot of entrepreneurs should be taking on board these causal discovery algorithms and using them to be able to better predict what's going to happen when they make real changes. And it's not just in terms of using them for your customers, it's also using them for yourself. If we bring in this new market, how is that going to impact our business? That requires doing causal time series modeling. It means putting together the 2 different kinds of techniques that we've been talking about so far. It's hard, but there are some really good introductions available online and a lot of open source code. So there's a lot of code that's out there that makes it very easy to build these kinds of causal models and to learn them from the data that you might have.

So you could use that then for, on a broader basis, you could do the, I did this huge research on starting in 2016 on why, when, and how founders fail. Which is what I wrote the book about. And I, it was so manual, right? You know, you're sitting there looking at the transcriptions from the interviews and looking at case use cases and all these papers that people drafted and all these different reports. You could, it sounds like you could use that exact system to do that in a fraction of the time. Just point at it, collect all the data to make sure it's good data. and then send the machine in there to go look at the data and come out and say, well, you know, if a founder does this, they have a 10% more likely chance of failing versus if they do this, they have a 20% chance of being successful. I mean, that sounds like it— would that be a good use of this?

I think that'd be a fantastic use of this, uh, these methods, and I'd be happy to talk about it with you offline. Yeah, I mean, I will say, so we, you know, one of the projects that I'm working on right now, one of the things we're looking at is trying to understand the causes of blood pressure crash in surgery. So drops in blood pressure are a massive predict— a very important predictor of post-surgical complications. And so what we're trying to do is figure out why do some people have these rapid unexpected drops in blood pressure and others don't? But we don't just wanna predict it because even if we can predict it, that leaves the anesthesiologist wondering, okay, so what do I do now, right? They have a limited set of tools. Which of their tools should they use? Well, it's going to depend on why the blood pressure drop happens, what the cause of it is. And so that's what we're trying to figure out. And of course, we're doing this in a context where we can't just go in and do A/B testing, right? We can't just randomly start doing things to people when they're in surgery. And so these kinds of methods are invaluable for those. Similarly, you want to be able to tell founders, here are ways to increase your chances of success that actually are the causes and drivers of success. But you don't want to just do a bunch of A/B tests on entrepreneurs. I think people would be annoyed if you said, here, just try this and see what happens.

Yeah, the investors wouldn't be happy either. So let me ask you this too. So when you look at this, you know, AI, ML, the whole, the whole set, right? The whole thing that we categorize as AI, do you feel like there are some more practical applications that, that you could point to that founders that are listening or investors that are listening could say, this is a direction we should look at specifically in AI? Because, you know, right now you go to webinars, you read papers, all this emails that you get and everything, and everybody's saying everything is AI, you know, even though they're not really AI, right? And so you have these founders that are sort of scrambling saying, I need to get into this AI thing. Where do I point myself? Do you have any recommendations for what categories or verticals that they should be looking at specifically?

Well, I think the most important thing is to think about, uh, what are the tasks that are often repeated and where you know what counts as success. So if you know when you've been successful, It's the kind of thing that you can build AI to do. So that's one of the sort of hallmarks of AI systems is they really perform best when we know whether we've succeeded. And so thinking about the overall goal and workflow and structure of your organization and asking, where are the pieces that we keep doing this thing over and over and we know what success looks like? So we'll know that we've succeeded when we get there. Those are the pieces that you want to look at and say, that's where we should be using AI. The parts that require judgment or that are really context sensitive or that require somebody to go, you know, this case is just a little bit different than that other one. Those are the places that you want to make sure to reserve the judgment for the human in your organization, because those are the parts that AI is still struggling with. And so in a kind of roundabout way, I can come back to, you know, your question, which is when people ask, well, what vertical is, you know, most ripe for disruption by AI or these kinds of things? My response is all of them and none of them. Because in every vertical, in every place that people have looked, some parts of the jobs, some parts of that sector are clear, repeated, easy to turn, relatively speaking, easy to turn into AI, and other parts are not. I mean, we're already seeing this, for example, with generative AI systems where there was sort of this idea that DALL-E and ChatGPT were going to put all advertising workers out of jobs because we'll just have advertisements created by the generative AI systems. And what people have very quickly discovered is generative AI systems can do a lot of the tasks, that somebody does in an advertising office, but not all of them. And there have been multiple studies that have found exactly this, that almost every job has some parts of it that AI should not be trying to do because AI is not going to do it well. And almost every job has some parts that AI should be doing and can do very well. So to an entrepreneur, to a founder, I guess my advice is, Think about the parts of people's lives that are boring, repetitive, easy to turn into just a you know the system takes care of that for me, and that's going to be the sweet spot in terms of rapid adoption and happy customers.

So more like a practical versus like a sentiment, right? Because it exactly does a terrible job with sentiment. It doesn't know what you're actually trying to do. You have to be. Extreme, you know, it's a machine, right? So you're saying like the practical, like, hey, go find this, or tell me what these numbers mean, or tell me if my numbers are off compared to these numbers, or that sort of stuff. Is that—

Exactly. So, you know, plan 5 itineraries for an upcoming vacation. Actually, it would be really good at that. Picking which of those itineraries we want to do, I don't want to delegate that to an AI system because That's going to depend on the context and knowing the details of, you know, what it is that I like to do on vacation, but also what my wife likes to do on vacation. That's a pretty important consideration as well. And so I think that that's really the key is to understand that almost nothing can be completely turned over to AI, but almost everything can be partially assisted by or augmented by AI. And finding those niches, finding those places for augmentation, both within an organization and in dealing with customers is really, I think, you know, whoever figures that part out first is going to be a rapid winner in this whole AI race.

Wow. You know, I really appreciate your time. I could talk to you for hours. You're just fascinating. I really appreciate your time. I appreciate you joining the podcast. Thank you so much, David.

Thanks so much for having me on, Greg.

And that's it for another episode of the Startup Science Podcast. If you can please leave a 5-star review on Apple or Spotify, I'd really appreciate it. And founders, as a thank you for leading the changes in our world, I'm offering free access to the Startup Science platform, Visionaries by Gregory Shepard. It's a one-stop shop for founders and has everything that startups need, from academy, To investors, to grant finders and pitch finders, and a whole lot more. Go to GregoryShepard.com to get Visionaries access. Use the code FOUNDERS in all capitals for free access to the platform and check it out for yourself. I'll see you next time. Namaste, my friends.

The Startup Science Podcast with Gregory Shepard is brought to you by StartupScience.io. Founders can log on to startupscience.io and use code Forbes to get free access. To connect and find out more about Greg, go to gregoryshepard.com. The Startup Science Podcast is a production of Forbes Books.