1 00:00:01,518 --> 00:00:03,098 Welcome, Scott. 2 00:00:03,458 --> 00:00:08,718 FICO was in some ways one of the original AI companies, at least in the sense of 3 00:00:08,718 --> 00:00:12,458 business analytics, going back to the invention of the credit score back in the 4 00:00:12,458 --> 00:00:13,338 50s. 5 00:00:13,338 --> 00:00:18,498 How does that long history influence the way that today you think about deploying 6 00:00:18,498 --> 00:00:19,352 AI? 7 00:00:19,439 --> 00:00:21,399 Yeah, thank you, Kevin. 8 00:00:21,399 --> 00:00:24,799 You know, it's part of our legacy, right? 9 00:00:24,799 --> 00:00:32,599 We've always focused on having a responsible view on how we use analytics. 10 00:00:32,599 --> 00:00:36,779 Originally, back in the 50s, right, we decided that, you know, decisions could be 11 00:00:36,779 --> 00:00:40,479 made more ethically and more fairly through the use of mathematics, right? 12 00:00:40,479 --> 00:00:45,759 And that started what was, you know, the use of analytics and credit scoring, which 13 00:00:45,759 --> 00:00:47,887 evolved to the PICO credit score. 14 00:00:47,887 --> 00:00:53,647 which later on in the nineties, we also branched into AI with our fraud solutions. 15 00:00:53,647 --> 00:00:58,507 And so we've always had at the center of our mind a focus on the fact that we 16 00:00:58,507 --> 00:01:01,367 develop models that impact people. 17 00:01:01,367 --> 00:01:04,807 And therefore those models, whether they be AI or something else, right? 18 00:01:04,807 --> 00:01:08,807 Or machine learning or something else needs to be built carefully and 19 00:01:08,807 --> 00:01:13,917 responsibly and reproducibly and with a level of accountability around it. 20 00:01:13,917 --> 00:01:16,547 And so that's been our big focus for us. 21 00:01:16,687 --> 00:01:19,767 given the gravitas to some of the decisions that get made with these types 22 00:01:19,767 --> 00:01:22,455 of analytics and models that we develop at FICO. 23 00:01:23,406 --> 00:01:28,566 So is it really from the consequences because many technologists for a long time 24 00:01:28,566 --> 00:01:31,496 said, well, this is just math, it's neutral, it's not biased. 25 00:01:31,496 --> 00:01:32,816 And we'll get into that. 26 00:01:32,816 --> 00:01:36,506 I think nowadays there's a lot more recognition of the problems. 27 00:01:36,506 --> 00:01:41,986 But what got the organization, what got you personally early on to put such an 28 00:01:41,986 --> 00:01:44,582 emphasis on these questions of ethics and responsibility? 29 00:01:45,007 --> 00:01:50,187 So in the early days of credit risk, and it still prevails today, there are a lot 30 00:01:50,187 --> 00:01:55,107 of challenges with respect to discriminatory usages of models and of 31 00:01:55,107 --> 00:01:55,567 lending. 32 00:01:55,567 --> 00:02:01,947 And so back in the, before we used math and models and making decisions about 33 00:02:01,947 --> 00:02:07,587 credit, whether you got a loan or not could be based on whether or not we went 34 00:02:07,587 --> 00:02:11,627 to the same church, did we go to the right, live in the right part of town, 35 00:02:11,627 --> 00:02:12,399 right? 36 00:02:12,399 --> 00:02:15,239 And these are kind of very subjective sort of measures. 37 00:02:15,239 --> 00:02:18,839 And so it was very, you know, from the very get -go and the launch of the way we 38 00:02:18,839 --> 00:02:23,199 developed analytics, we understood that we were actually combating a problem, which 39 00:02:23,199 --> 00:02:27,789 is, you know, the data back then and in the data even today is full of bias. 40 00:02:27,789 --> 00:02:31,679 And therefore, you know, we go in with this view that data is dirty, data is 41 00:02:31,679 --> 00:02:32,359 bias. 42 00:02:32,359 --> 00:02:37,747 And so when we develop a model, we have to do that in such a way that we can, you 43 00:02:37,747 --> 00:02:37,967 know, 44 00:02:37,967 --> 00:02:43,127 fared out where those biases could be, remove them and build the most fair and 45 00:02:43,127 --> 00:02:44,077 ethical model possible. 46 00:02:44,077 --> 00:02:48,067 So I think, you know, that was really part of our origin story, if you will, in terms 47 00:02:48,067 --> 00:02:52,227 of why math would be more empirical, more fair, right? 48 00:02:52,227 --> 00:02:55,847 We still have challenges today and how math gets used, right? 49 00:02:55,847 --> 00:02:56,957 How these scores get used. 50 00:02:56,957 --> 00:02:59,117 And that's a big part of the infrastructure, right? 51 00:02:59,117 --> 00:03:00,707 Of how decisions get made. 52 00:03:00,707 --> 00:03:04,847 But we wanted to make sure that the models or scores that are used, 53 00:03:04,847 --> 00:03:08,827 to help aid in making fair decisions are built to the highest level of standard 54 00:03:08,827 --> 00:03:11,853 possible from an ethics and fairness perspective. 55 00:03:12,718 --> 00:03:17,318 challenge today as the models have gotten so much more sophisticated and we moved to 56 00:03:17,318 --> 00:03:22,958 AI and deep learning and it's not as clear even how you get from the inputs to the 57 00:03:22,958 --> 00:03:23,786 outputs. 58 00:03:24,143 --> 00:03:26,923 So I think you've hit on it exactly right. 59 00:03:26,923 --> 00:03:31,863 I mean, one of the biggest sort of challenges we have today that wasn't the 60 00:03:31,863 --> 00:03:37,583 case when I started here at the company 25 years ago is that we somehow all accept 61 00:03:37,583 --> 00:03:41,823 that, you know, AI is used to make all these decisions and the models have become 62 00:03:41,823 --> 00:03:43,723 more and more complex. 63 00:03:43,723 --> 00:03:45,763 There've been more and more parameters. 64 00:03:45,763 --> 00:03:49,263 There've been people enamored with how quickly they can build models and how 65 00:03:49,263 --> 00:03:50,927 large these models can be. 66 00:03:50,927 --> 00:03:53,047 And they don't focus on the physics. 67 00:03:53,047 --> 00:03:56,007 Like I'm a PhD physicist and I think about physics, right? 68 00:03:56,007 --> 00:04:00,187 And I think the world is defined by, you know, laws of physics and types of 69 00:04:00,187 --> 00:04:02,107 behaviors that are relatively simple, right? 70 00:04:02,107 --> 00:04:07,127 So, you know, we don't need models with hundreds of thousands or millions of 71 00:04:07,127 --> 00:04:08,947 parameters to make it a credit decision, right? 72 00:04:08,947 --> 00:04:13,047 What we need is we need level of transparency in terms of the types of 73 00:04:13,047 --> 00:04:13,927 models. 74 00:04:13,927 --> 00:04:14,319 But... 75 00:04:14,319 --> 00:04:16,879 You know, as an, as an industry, right. 76 00:04:16,899 --> 00:04:20,999 we, we've kind of adopted these big models and somehow bought into bigger is better 77 00:04:20,999 --> 00:04:25,579 without that and sacrificing the ability to actually understand what drives those 78 00:04:25,579 --> 00:04:26,519 outcomes. 79 00:04:26,519 --> 00:04:30,259 Now, if you look at things like, you know, the FICO score that we produce, we still 80 00:04:30,259 --> 00:04:33,259 use a generative additive models for that. 81 00:04:33,259 --> 00:04:37,399 And primarily we do that because when we, let's say, Kevin, give you a decision, 82 00:04:37,399 --> 00:04:41,679 you, you, a decision gets made based on a score and you say, well, I don't 83 00:04:41,679 --> 00:04:43,983 understand why that decision was made. 84 00:04:43,983 --> 00:04:49,923 We need to be tremendously factual in terms of what are the reasons or why you 85 00:04:49,923 --> 00:04:51,683 got the better score or the worst score. 86 00:04:51,683 --> 00:04:54,433 Cause you, you're going to need to act on that, right? 87 00:04:54,433 --> 00:04:56,683 Because you're still going to need to get that loan. 88 00:04:56,683 --> 00:05:01,643 And so for us, it's not just like, okay, yes, let's use explainable AI, which I 89 00:05:01,643 --> 00:05:05,763 like to talk about as being good, bad, and ugly with most of it being bad and ugly, 90 00:05:05,763 --> 00:05:06,143 right? 91 00:05:06,143 --> 00:05:09,443 Let's focus on interpretability as job one. 92 00:05:09,443 --> 00:05:13,071 Because if I tell you, these are the three reasons why you didn't get the great. 93 00:05:13,071 --> 00:05:15,711 the best score, then you could go and act on them. 94 00:05:15,711 --> 00:05:20,031 And if you did that, right, we would see that consequential sort of improvement in 95 00:05:20,031 --> 00:05:20,491 the score. 96 00:05:20,491 --> 00:05:23,791 And that's the level of sort of visibility, which is these scores are not 97 00:05:23,791 --> 00:05:26,161 just meant to block, right? 98 00:05:26,161 --> 00:05:30,571 They're meant to assist and provide understanding of what the outcome was and 99 00:05:30,571 --> 00:05:32,737 maybe how you improve that outcome in the future. 100 00:05:33,742 --> 00:05:37,962 Clearly that matters in credit scoring where as you're talking about their human 101 00:05:37,962 --> 00:05:41,942 consequences and it's in many cases regulated. 102 00:05:42,122 --> 00:05:46,862 Do you think that in general that companies have gone too far with the big 103 00:05:46,862 --> 00:05:48,262 non -explainable models? 104 00:05:48,262 --> 00:05:51,102 What if you're using AI for something like sales forecasting? 105 00:05:51,102 --> 00:05:55,060 Say, does it really matter if you can explain what the rationale was? 106 00:05:55,151 --> 00:05:58,071 I think it still does. 107 00:05:58,271 --> 00:06:02,191 I think you have to look at the consequence of the decision. 108 00:06:02,191 --> 00:06:07,451 So if we're going to produce a picture of a five -legged elephant with generative 109 00:06:07,451 --> 00:06:08,751 AI, it probably doesn't matter. 110 00:06:08,751 --> 00:06:09,551 That's art. 111 00:06:09,551 --> 00:06:11,091 And if you don't like the picture, we throw it away. 112 00:06:11,091 --> 00:06:13,071 There's no consequence there. 113 00:06:13,111 --> 00:06:17,481 But if I'm going to do a sales forecast, and I have shareholders that are impacted, 114 00:06:17,481 --> 00:06:20,411 and I have employees that are impacted, I have bonuses that are impacted, then 115 00:06:20,411 --> 00:06:25,071 that's a very consequential decision and one where, at the very least, 116 00:06:25,071 --> 00:06:29,551 you know, I should have understanding and confidence in, you know, how the analytic 117 00:06:29,551 --> 00:06:32,261 reached an outcome, right, or a recommendation. 118 00:06:32,261 --> 00:06:36,871 And so I'd say, you know, that this sort of concept on models that are more 119 00:06:36,871 --> 00:06:41,211 interpretable and maybe simpler so that we understand and we feel comfortable in the 120 00:06:41,211 --> 00:06:45,431 assumptions that went into those models or what was learned by those models to make 121 00:06:45,431 --> 00:06:46,811 the decision. 122 00:06:46,811 --> 00:06:50,311 But we understand that so we can have confidence in the models, right? 123 00:06:50,311 --> 00:06:54,011 One of the things that I think about, Kevin, is like over the last year, right, 124 00:06:54,011 --> 00:06:54,895 a customer, 125 00:06:54,895 --> 00:06:59,435 confidence in models, AI in particular, has gone down considerably. 126 00:06:59,435 --> 00:07:04,615 Consumers, like 60 % roughly, depending on which survey, have a lot of concern about 127 00:07:04,615 --> 00:07:05,435 use of AI. 128 00:07:05,435 --> 00:07:09,035 Businesses themselves have concerns about using AI because they don't feel like they 129 00:07:09,035 --> 00:07:10,275 have it under control, right? 130 00:07:10,275 --> 00:07:12,665 And even in the financial sector, same thing. 131 00:07:12,665 --> 00:07:17,435 So I think, you know, we are at a sort of like tipping point where folks are going 132 00:07:17,435 --> 00:07:21,515 to say, well, that's all cool that we can, you know, use, you know, use a thousand 133 00:07:21,515 --> 00:07:23,439 GPUs to create a model, but... 134 00:07:23,439 --> 00:07:25,749 you know, do we need to and what do we lose? 135 00:07:25,749 --> 00:07:28,809 Do we lose our ability to actually interpret or have trust in the model? 136 00:07:28,809 --> 00:07:32,459 And I think a lot of people run into that, even if it's not a regulated environment 137 00:07:32,459 --> 00:07:35,015 that they are sort of contemplating the use of. 138 00:07:36,398 --> 00:07:42,018 So how should an organization think about the trade -offs in terms of, you know, 139 00:07:42,018 --> 00:07:46,478 they adopt a large language model or something that's not today particularly 140 00:07:46,478 --> 00:07:49,644 interpretable versus the alternatives? 141 00:07:49,679 --> 00:07:50,259 Yeah. 142 00:07:50,259 --> 00:07:56,059 So, you know, I think one needs to figure out what is the best tool for the job. 143 00:07:56,059 --> 00:07:56,219 Right. 144 00:07:56,219 --> 00:07:59,759 And so, you know, I've had this conversation with many executives where, 145 00:07:59,759 --> 00:08:03,259 you know, I try to explain to them that, you know, there are different AIs for 146 00:08:03,259 --> 00:08:03,999 different purposes. 147 00:08:03,999 --> 00:08:04,619 Right. 148 00:08:04,619 --> 00:08:09,959 And, you know, in general, right, we should use the simplest and minimalistic 149 00:08:09,959 --> 00:08:13,223 solution, right, to, to. 150 00:08:13,903 --> 00:08:14,943 to solve a problem, right? 151 00:08:14,943 --> 00:08:18,233 So if I can solve the problem with a relatively simple interpretable model, 152 00:08:18,233 --> 00:08:18,523 right? 153 00:08:18,523 --> 00:08:22,643 Or even a linear model, maybe that's the best thing because I want to provide a lot 154 00:08:22,643 --> 00:08:26,503 of confidence in that decision and I need that level of transparency, right? 155 00:08:26,503 --> 00:08:29,703 If you really need to goose up the performance, right? 156 00:08:29,703 --> 00:08:33,203 The detection capabilities, then maybe you get more complicated and you'll have like 157 00:08:33,203 --> 00:08:37,563 an interpretable neural network, which is something that we focus on, which allows 158 00:08:37,563 --> 00:08:42,063 that level of learning those imputed relationships that drive. 159 00:08:42,063 --> 00:08:44,603 decisions but can still be interpreted. 160 00:08:44,683 --> 00:08:51,203 I would argue there's very few use cases probably where, at least in business 161 00:08:51,203 --> 00:08:54,943 context, that you have to go to completely unexplainable models. 162 00:08:54,943 --> 00:09:01,303 Even if you look at interpreting text and classifying text, you can do that without 163 00:09:01,303 --> 00:09:07,753 full generative AI models that hallucinate and that are unexplainable. 164 00:09:07,753 --> 00:09:11,951 And so I think just stepping back and making sure that... 165 00:09:11,951 --> 00:09:16,171 you say, what is the minimalistic sort of solution that would solve this business 166 00:09:16,171 --> 00:09:21,191 problem and treat it as a business problem, not as a science problem where 167 00:09:21,191 --> 00:09:22,421 you're interested in science. 168 00:09:22,421 --> 00:09:25,391 We're all interested in cool science. 169 00:09:25,391 --> 00:09:31,341 But if we're asking a business problem, people will say, where can I get to 85, 170 00:09:31,341 --> 00:09:36,771 90, 95 % of a solution and still have an ability to rationalize this, explain it to 171 00:09:36,771 --> 00:09:39,451 my board, explain it to my manager and my governance team. 172 00:09:39,451 --> 00:09:41,551 And as soon as we put on our business hats, 173 00:09:41,551 --> 00:09:45,911 You guys very, you know, the worst versus reward is very much like a typical CRO 174 00:09:45,911 --> 00:09:49,131 sort of conversation that you'd have in terms of what risks you're willing to take 175 00:09:49,131 --> 00:09:53,491 and, you know, getting that extra, you know, a little bit of detection, but 176 00:09:53,491 --> 00:09:57,171 potentially, let's say, discriminating against a group of individuals, not a 177 00:09:57,171 --> 00:09:58,121 great trade off, right? 178 00:09:58,121 --> 00:10:01,191 And that's where I think some of those decisions come back down to just making 179 00:10:01,191 --> 00:10:03,081 those solid business decisions. 180 00:10:04,014 --> 00:10:08,654 What are some of the things that you've done at FICO to make your AI systems more 181 00:10:08,654 --> 00:10:09,580 accountable? 182 00:10:10,287 --> 00:10:17,267 So what we focused on is establishing what is our model governance standards at the 183 00:10:17,267 --> 00:10:17,887 company. 184 00:10:17,887 --> 00:10:22,967 So we have a firm belief that, and we use the term responsible AI versus 185 00:10:22,967 --> 00:10:23,827 accountable. 186 00:10:23,827 --> 00:10:29,247 I like the term accountable AI, but in my mind, accountable AI kind of implies that 187 00:10:29,247 --> 00:10:31,407 I'm accountable for the wrongs that I've done. 188 00:10:31,407 --> 00:10:32,907 And that's true, right? 189 00:10:32,907 --> 00:10:36,267 But I would rather have a responsible strategy that says this is the right way 190 00:10:36,267 --> 00:10:38,207 or this is the company way of doing it. 191 00:10:38,207 --> 00:10:39,663 So at the very... 192 00:10:39,663 --> 00:10:46,143 start of it is a model governance document that model development governance standard 193 00:10:46,143 --> 00:10:48,063 for how all models get developed, right? 194 00:10:48,063 --> 00:10:52,543 Because if I have, you know, 200 or 300, 400 data scientists, I can't have them all 195 00:10:52,543 --> 00:10:56,543 making decisions around how to, you know, choose an algorithm or what's the right 196 00:10:56,543 --> 00:11:01,483 way to explain a model or how do we deal with ethics, right? 197 00:11:01,483 --> 00:11:04,303 So we have these are the models you're allowed to use. 198 00:11:04,303 --> 00:11:05,683 This is how we explain the models. 199 00:11:05,683 --> 00:11:06,883 This is how we do the ethics. 200 00:11:06,883 --> 00:11:07,439 And so, 201 00:11:07,439 --> 00:11:11,039 The first step is having that model development standard that everyone 202 00:11:11,039 --> 00:11:12,199 follows. 203 00:11:12,499 --> 00:11:15,939 The second major step then is having an enforcement tool. 204 00:11:15,939 --> 00:11:20,779 And it sounds a little bit like we're restricting creativity of data scientists. 205 00:11:20,779 --> 00:11:26,519 And we are to a certain extent because customers and our customers' customers are 206 00:11:26,519 --> 00:11:29,359 not lab experiments. 207 00:11:29,359 --> 00:11:33,967 We need to make sure that we are using technology that is proven and understood. 208 00:11:33,967 --> 00:11:37,867 And so we enforce that today through the use of blockchain, right? 209 00:11:37,867 --> 00:11:42,907 Where decisions get made that where we show that like these are the 27 things you 210 00:11:42,907 --> 00:11:46,667 need to show done as part of the model development standard, you know, and then 211 00:11:46,667 --> 00:11:51,827 have a proof of record that each were done, tested, and then approved. 212 00:11:51,827 --> 00:11:56,147 And that allows us to release models that meet a model development standard. 213 00:11:56,687 --> 00:11:59,077 And, you know, the consequences are great, right? 214 00:11:59,077 --> 00:11:59,791 You know, we'll... 215 00:11:59,791 --> 00:12:03,931 One major sort of advantage is that everyone follows the same standard, right? 216 00:12:03,931 --> 00:12:06,501 And so you have less errors and issues when models get released. 217 00:12:06,501 --> 00:12:11,951 The other major piece is all that energy that people put into like governing 218 00:12:11,951 --> 00:12:13,211 models, right? 219 00:12:13,211 --> 00:12:16,951 Where there might be thousands of models in a thousand different conversations all 220 00:12:16,951 --> 00:12:21,571 get funneled down into let's improve the model development standard, right? 221 00:12:21,571 --> 00:12:23,241 And then use things like blockchain to enforce. 222 00:12:23,241 --> 00:12:26,351 So it's really gotten us much more tactical. 223 00:12:26,351 --> 00:12:31,531 And actually much more efficient in how we develop high quality models at scale in an 224 00:12:31,531 --> 00:12:32,951 understandable, predictable way. 225 00:12:32,951 --> 00:12:36,591 And then we focus our, you know, back to the data science creativity, we focus our 226 00:12:36,591 --> 00:12:38,811 innovation on improving the standard, right? 227 00:12:38,811 --> 00:12:42,671 And then, you know, then we treat your productions, developing models as a 228 00:12:42,671 --> 00:12:45,341 production process off of that standard. 229 00:12:46,606 --> 00:12:48,686 So let me unpack those two pieces. 230 00:12:48,686 --> 00:12:51,086 You talked about, you start with the standards and the ethics. 231 00:12:51,086 --> 00:12:52,486 So where do those come from? 232 00:12:52,751 --> 00:12:57,591 Yeah, so the standards are ones that are typically defined by someone like myself 233 00:12:57,591 --> 00:13:03,651 within an organization that has like a broad purview of the problems that we 234 00:13:03,651 --> 00:13:07,521 solve, that the pluses and minuses of different algorithms and then, you know, 235 00:13:07,521 --> 00:13:08,861 making them hard decisions, right? 236 00:13:08,861 --> 00:13:12,091 There's certain algorithms that we just do not allow to be used at FICO, right? 237 00:13:12,091 --> 00:13:16,231 Because, you know, they're not deemed explainable or they're not deemed ethical 238 00:13:16,231 --> 00:13:19,855 or, you know, or they have so many parameters that just... 239 00:13:19,855 --> 00:13:23,375 science would dictate that you can't collect enough data to actually specify 240 00:13:23,375 --> 00:13:26,375 all those parameters properly and therefore you have low confidence in 241 00:13:26,375 --> 00:13:26,715 models. 242 00:13:26,715 --> 00:13:32,275 And so, I'm part of that, part of that would be our legal teams and ethics people 243 00:13:32,275 --> 00:13:33,935 and also business owners, right? 244 00:13:33,935 --> 00:13:40,955 That it might have very, very keen views on how regulated AI specifies that certain 245 00:13:40,955 --> 00:13:42,835 regulation needs to be met. 246 00:13:42,835 --> 00:13:47,735 And so that all gets folded in into that model development standard. 247 00:13:47,975 --> 00:13:49,423 It gets kind of... 248 00:13:49,423 --> 00:13:55,203 through that sort of set of stakeholders and then wants to find, right, then it's 249 00:13:55,203 --> 00:13:56,413 operationalized. 250 00:13:57,486 --> 00:14:00,376 And then on the second piece, why blockchain? 251 00:14:00,376 --> 00:14:03,736 And I actually do a lot of work on blockchain and digital assets. 252 00:14:03,736 --> 00:14:07,386 So I was very excited to hear about the work you're doing there and understand 253 00:14:07,386 --> 00:14:11,746 you've gotten patents and awards for it, but can you expand a little bit more about 254 00:14:11,746 --> 00:14:15,855 what the value is of using a distributed ledger for that kind of audit trail? 255 00:14:15,855 --> 00:14:20,345 Yeah, so we started this responsible AI journey a very long time ago. 256 00:14:20,345 --> 00:14:24,755 And we started with this concept of what's we call analytic tracking documents. 257 00:14:24,835 --> 00:14:27,595 And I treated it as a contract back then. 258 00:14:27,595 --> 00:14:32,425 And it was a contract between myself and the business owners and my scientists. 259 00:14:32,425 --> 00:14:36,875 And there were named resources and tasks that needed to be done. 260 00:14:36,875 --> 00:14:43,375 And we would enforce it through weekly stand -ups, sprint reviews, agile 261 00:14:43,375 --> 00:14:45,115 development processes. 262 00:14:45,487 --> 00:14:50,477 And we made a lot of progress there, but things could slip through the cracks, 263 00:14:50,477 --> 00:14:50,747 right? 264 00:14:50,747 --> 00:14:53,567 Because if I didn't ask the right questions or if something wasn't presented 265 00:14:53,567 --> 00:14:55,587 sufficiently, right? 266 00:14:55,587 --> 00:15:01,447 We would basically go through that process in a human way to ensure that it's 267 00:15:01,447 --> 00:15:02,487 enforced. 268 00:15:02,627 --> 00:15:04,787 With the blockchain though, right? 269 00:15:04,787 --> 00:15:07,567 We're able to automate all of that, right? 270 00:15:07,567 --> 00:15:12,087 And the reason why we use the blockchain is number one, 271 00:15:12,239 --> 00:15:13,859 we have that immutable chain, right? 272 00:15:13,859 --> 00:15:17,379 Cause I want to see the history of what we did well and what we struggled with, 273 00:15:17,379 --> 00:15:17,829 right? 274 00:15:17,829 --> 00:15:23,259 And maybe, you know, I completed a task and one of my scientists tested it and 275 00:15:23,259 --> 00:15:24,539 said, no, Scott, you're wrong. 276 00:15:24,539 --> 00:15:26,379 You didn't meet that task. 277 00:15:26,379 --> 00:15:28,019 I want that on the blockchain. 278 00:15:28,019 --> 00:15:32,079 And so the immutability means that people take things seriously. 279 00:15:32,079 --> 00:15:36,299 So back 15 years ago, when they actually had a sign, a piece of paper with me, like 280 00:15:36,299 --> 00:15:40,159 a little mini mock contract, when you put your signature on a piece of paper still 281 00:15:40,159 --> 00:15:41,519 means something, right? 282 00:15:41,519 --> 00:15:42,479 Same with the blockchain. 283 00:15:42,479 --> 00:15:47,269 As soon as I commit it as done on the blockchain, it's there, it's immutable. 284 00:15:47,269 --> 00:15:49,299 I take it seriously, right? 285 00:15:49,299 --> 00:15:52,919 And then, you know, the people after me, they're checking that work and verifying 286 00:15:52,919 --> 00:15:53,699 that work. 287 00:15:53,699 --> 00:15:55,159 Their names are on the ledger, right? 288 00:15:55,159 --> 00:15:56,579 And they're taking it seriously. 289 00:15:56,579 --> 00:16:01,319 And so it provides a sort of level of accountability because we know that it's 290 00:16:01,319 --> 00:16:02,369 going to be on that ledger. 291 00:16:02,369 --> 00:16:07,759 We know that it's immutable and that we're not going to be able to change it. 292 00:16:07,759 --> 00:16:10,759 And it also establishes tremendous trust, right? 293 00:16:10,759 --> 00:16:17,279 Because I can open that distributed ledger or private ledger to regulators, to 294 00:16:17,279 --> 00:16:20,169 governance teams where they see exactly what was there, right? 295 00:16:20,169 --> 00:16:24,719 And if Scott was particularly bad on a task and he messed it up seven times and 296 00:16:24,719 --> 00:16:27,589 seven times his data scientists rejected it, right? 297 00:16:27,589 --> 00:16:31,559 And maybe eventually Scott was taken off that task and was replaced by Julie, 298 00:16:31,559 --> 00:16:33,189 that's fine, right? 299 00:16:33,189 --> 00:16:35,215 But that is the level of sort of... 300 00:16:35,215 --> 00:16:39,895 clarity and transparency and seriousness around meeting that standard. 301 00:16:39,895 --> 00:16:43,274 And if you look at regulators, they don't have that level of transparency. 302 00:16:43,274 --> 00:16:47,455 So it drives the right behaviors in a data science program. 303 00:16:47,875 --> 00:16:52,475 It also provides that level of trust when it comes to explaining how the model is 304 00:16:52,475 --> 00:16:54,815 built, because there's no disagreement. 305 00:16:54,815 --> 00:16:57,155 There's no like, well, I think it does this. 306 00:16:57,155 --> 00:17:01,915 If we say, for example, and I say, well, these are the latent features that are 307 00:17:01,915 --> 00:17:04,025 learned, they're persistent in the blockchain. 308 00:17:04,025 --> 00:17:04,815 There's no... 309 00:17:04,815 --> 00:17:09,175 I think it is, or I thought it might be, no, if the requirement is I have to tell 310 00:17:09,175 --> 00:17:12,455 you what all the latent features are and what the relationships that were learned, 311 00:17:12,455 --> 00:17:13,585 that's on the blockchain. 312 00:17:13,585 --> 00:17:17,955 And if later on, or a governance person or ethics person, lawyer, regulator wants to 313 00:17:17,955 --> 00:17:23,375 look at that and they say, ooh, we don't like that latent feature, that's fine. 314 00:17:23,375 --> 00:17:25,755 That's the conversation that needs to be had, right? 315 00:17:25,755 --> 00:17:30,255 And so that's one of the reasons, but it's enforcement, it's accountability. 316 00:17:30,655 --> 00:17:33,647 And frankly, you know, it's... 317 00:17:33,647 --> 00:17:36,887 you'd think data scientists wouldn't like it, but it actually provides tremendous 318 00:17:36,887 --> 00:17:40,427 comfort because unfortunately most data scientists are put in situations where 319 00:17:40,427 --> 00:17:43,187 they're ill -equipped to take on that level of responsibility. 320 00:17:43,187 --> 00:17:44,587 So it also protects them. 321 00:17:44,587 --> 00:17:48,647 So it's a sort of happy harmony of kind of benefits. 322 00:17:48,647 --> 00:17:52,807 And then, you know, frankly, after that, when the model gets released, it only gets 323 00:17:52,807 --> 00:17:57,207 released when all those requirements are met and accepted and you have that whole 324 00:17:57,207 --> 00:17:58,217 history. 325 00:17:58,543 --> 00:18:01,173 That blockchain goes with the model to where it gets deployed. 326 00:18:01,173 --> 00:18:04,763 And it's also the thing that lives with the model so that later on a year from 327 00:18:04,763 --> 00:18:08,163 now, two years from now, when there's questions about what that model does and 328 00:18:08,163 --> 00:18:11,443 who built it, why they built it, why is it behaving this way? 329 00:18:11,443 --> 00:18:15,183 You can go look into that blockchain, understand what's inside the model. 330 00:18:15,183 --> 00:18:20,033 And moreover, today we actually put in the entire requirements on how you monitor it, 331 00:18:20,033 --> 00:18:20,203 right? 332 00:18:20,203 --> 00:18:22,143 Which is another big issue in AI, right? 333 00:18:22,143 --> 00:18:23,383 How do you monitor these models? 334 00:18:23,383 --> 00:18:27,183 And so the only person that really knows how to do that is the person who built it. 335 00:18:27,183 --> 00:18:30,243 Might as well codify that at the time it was built and say, these are the operating 336 00:18:30,243 --> 00:18:31,323 parameters for my model. 337 00:18:31,323 --> 00:18:33,267 And then you run it in production. 338 00:18:34,446 --> 00:18:35,896 It sounds like a lot of overhead though. 339 00:18:35,896 --> 00:18:39,146 How do you build all that into the workflows? 340 00:18:39,471 --> 00:18:42,171 So there is work. 341 00:18:42,171 --> 00:18:47,731 And so we have a UI that we essentially codify all those requirements. 342 00:18:48,291 --> 00:18:52,751 And so as part of the model development process, they have to list that in. 343 00:18:53,911 --> 00:18:56,031 I don't think it's any incremental effort. 344 00:18:56,031 --> 00:18:58,131 It's more about organizing. 345 00:18:58,231 --> 00:19:03,041 So you think about my desk is actually, you can't see it, but it's a big mess. 346 00:19:03,041 --> 00:19:03,887 And so like, 347 00:19:03,887 --> 00:19:05,327 It's very cluttered, right? 348 00:19:05,327 --> 00:19:08,907 And there's order that I understand, but I don't think you'd understand necessarily 349 00:19:08,907 --> 00:19:11,327 unless you're as cluttered as I am. 350 00:19:11,327 --> 00:19:16,727 So we're structuring the work in that, that the requirements get met. 351 00:19:16,827 --> 00:19:20,707 There is overhead, but without that, we'd go have to go back and maybe answer 352 00:19:20,707 --> 00:19:25,207 questions that should have been answered, you know, a month ago or six weeks ago. 353 00:19:25,207 --> 00:19:30,307 And so we actually find that it's more efficient because, you know, we've 354 00:19:30,307 --> 00:19:32,399 actually codified exactly what. 355 00:19:32,399 --> 00:19:35,839 you know, a governance team will be looking for exactly how we're going to 356 00:19:35,839 --> 00:19:39,039 report the reasons for how a model behaved. 357 00:19:39,039 --> 00:19:43,139 And so, you know, there's more work upfront to organize that work into, you 358 00:19:43,139 --> 00:19:46,739 know, a structured way or on that model governance standard, but there's so much 359 00:19:46,739 --> 00:19:50,819 less work after, right, or during the model build because you've already 360 00:19:50,819 --> 00:19:52,009 satisfied all the requirements. 361 00:19:52,009 --> 00:19:55,619 You know, one of the things that's interesting to me is there are still a lot 362 00:19:55,619 --> 00:19:58,419 of organizations that have model governance teams, right? 363 00:19:58,419 --> 00:20:00,079 And, you know, model governance team, 364 00:20:00,079 --> 00:20:03,159 generally operates that I build the model, I throw it over to you, you don't know 365 00:20:03,159 --> 00:20:05,479 anything about the model, you start asking questions, right? 366 00:20:05,479 --> 00:20:08,619 The questions may or may not be relevant, you may or may not have the same level of 367 00:20:08,619 --> 00:20:10,719 expertise in what I've done. 368 00:20:10,719 --> 00:20:13,659 These governance processes can take nine months, right? 369 00:20:13,659 --> 00:20:17,179 I would rather say when the model is released, you and I, if you're the 370 00:20:17,179 --> 00:20:21,039 governance person Kevin, you and I are agreed on what you need to see. 371 00:20:21,039 --> 00:20:24,179 And as proof of work, when that model is released, it's shown. 372 00:20:24,179 --> 00:20:25,479 You can go look at the blockchain. 373 00:20:25,479 --> 00:20:28,799 And so in some sense, it drives all that governance into the model development 374 00:20:28,799 --> 00:20:29,219 process. 375 00:20:29,219 --> 00:20:34,119 So it's not an entirely different process that oftentimes takes twice as long or 376 00:20:34,119 --> 00:20:37,165 three times as long as actually developing the model in the first place. 377 00:20:37,870 --> 00:20:41,050 Yeah, do you think there will be more pressure for that kind of approach? 378 00:20:41,050 --> 00:20:45,730 Certainly in banking, we have model risk management standards and those kinds of 379 00:20:45,730 --> 00:20:46,120 requirements. 380 00:20:46,120 --> 00:20:50,570 But as we talk now, especially with generative AI, where you've got companies 381 00:20:50,570 --> 00:20:54,130 that are building the models, other companies that are adapting them to 382 00:20:54,130 --> 00:20:57,210 applications, and then companies are taking those, are doing the rag chain and 383 00:20:57,210 --> 00:20:58,050 so forth. 384 00:20:58,050 --> 00:21:02,350 Now you have a lot more organizations that are somehow touching or dependent on the 385 00:21:02,350 --> 00:21:04,047 model, and then something goes wrong. 386 00:21:04,047 --> 00:21:04,842 Yeah. 387 00:21:04,842 --> 00:21:08,882 is that this kind of audit trail you think going to be ultimately what organizations 388 00:21:08,882 --> 00:21:09,711 need to go to? 389 00:21:09,711 --> 00:21:13,111 I think it most absolutely will be. 390 00:21:13,111 --> 00:21:17,551 And so you look at the work that people do in LLMs. 391 00:21:17,730 --> 00:21:21,391 I look at different proof of concepts that people want to do. 392 00:21:21,451 --> 00:21:24,321 And we obviously do proof of concepts here at FICO. 393 00:21:24,321 --> 00:21:26,621 We also do core research in this area. 394 00:21:26,621 --> 00:21:30,031 We've been doing it for 20 years in generative AI. 395 00:21:30,031 --> 00:21:34,661 But there's a lot of imprecision in what people understand. 396 00:21:34,661 --> 00:21:36,691 For example, like RAG. 397 00:21:36,691 --> 00:21:38,611 You talk about this augmentation. 398 00:21:38,611 --> 00:21:39,527 But like, 399 00:21:39,599 --> 00:21:42,479 you know, what, how, what's the parameter on what gets used, what doesn't get used? 400 00:21:42,479 --> 00:21:43,579 How do you do the embedding? 401 00:21:43,579 --> 00:21:45,199 What, you know, what's the chunking? 402 00:21:45,199 --> 00:21:45,949 What's the overlap? 403 00:21:45,949 --> 00:21:50,019 These all really, really matter from a business context perspective, but a lot of 404 00:21:50,019 --> 00:21:52,719 it's just kind of, you know, glossed over, right. 405 00:21:52,719 --> 00:21:54,139 In terms of looking at the outputs. 406 00:21:54,139 --> 00:21:58,019 And I think, you know, if people are going to go leverage some of these technologies 407 00:21:58,019 --> 00:22:02,735 that are a little bit more on the fringe and in terms of responsible AI or, then. 408 00:22:02,735 --> 00:22:06,175 for sure they're gonna have to keep track of these as decisions, right? 409 00:22:06,175 --> 00:22:11,595 So they can revisit it down the road if they find unsatisfactory results in the 410 00:22:11,595 --> 00:22:12,315 use of it. 411 00:22:12,315 --> 00:22:16,005 And I think that'll be a big part of where people need to get to. 412 00:22:16,005 --> 00:22:20,375 Otherwise, you'll have a model and even simple models are really hard to 413 00:22:20,375 --> 00:22:20,835 understand. 414 00:22:20,835 --> 00:22:25,615 So these large applications of generative AI most definitely will have to have a lot 415 00:22:25,615 --> 00:22:28,623 of structure and just picking apart things like RAG. 416 00:22:28,623 --> 00:22:30,723 There's a lot, a lot of decisions. 417 00:22:30,723 --> 00:22:36,403 And then there's a lot of things that, you know, people take for granted that they 418 00:22:36,403 --> 00:22:38,163 don't truly understand how it works. 419 00:22:38,163 --> 00:22:38,393 Right. 420 00:22:38,393 --> 00:22:44,643 And so from that perspective, just codifying that as a risk or delving in to 421 00:22:44,643 --> 00:22:51,803 where they will use, let's say, you know, a private LLM sort of capabilities versus 422 00:22:51,803 --> 00:22:55,083 building their own will be critically important in terms of where they can, you 423 00:22:55,083 --> 00:22:56,655 know, faithfully say, listen, 424 00:22:56,655 --> 00:23:00,085 Yeah, we only use our documentation answering the question or maybe we don't. 425 00:23:00,085 --> 00:23:02,255 Maybe there is hallucination going on there. 426 00:23:02,255 --> 00:23:05,155 And I think some of those things have to be codified so that there's some 427 00:23:05,155 --> 00:23:06,119 governance around. 428 00:23:07,502 --> 00:23:12,342 Are there areas where you've changed your mind or you've realized that you were 429 00:23:12,342 --> 00:23:16,302 going in the wrong direction in terms of governance and responsibility in recent 430 00:23:16,302 --> 00:23:17,030 years? 431 00:23:17,487 --> 00:23:20,217 Well, that's a great question. 432 00:23:20,217 --> 00:23:22,967 I think it's an evolving process. 433 00:23:22,967 --> 00:23:24,627 I'll put it that way. 434 00:23:24,627 --> 00:23:29,207 I think with blockchain, I'm very confident that's the direction forward. 435 00:23:29,207 --> 00:23:36,087 In fact, we're so confident at FICO, we have this AI decisioning platform, which 436 00:23:36,087 --> 00:23:39,607 was a leader in a Forrester wave from 2023. 437 00:23:39,607 --> 00:23:42,287 And we're using blockchain beyond just models. 438 00:23:42,287 --> 00:23:45,987 We're using it to kind of keep track of the entire decision process. 439 00:23:45,987 --> 00:23:46,511 And so, 440 00:23:46,511 --> 00:23:50,311 I think this blockchain was a great sort of approach. 441 00:23:50,531 --> 00:23:57,151 I think, to your point about what is the right sort of heftiness of a process, this 442 00:23:57,151 --> 00:24:01,861 is still a refinement for us in terms of the pluses and minuses. 443 00:24:01,861 --> 00:24:05,271 So I don't think there's anything where we backed away and said, let's take an 444 00:24:05,271 --> 00:24:07,151 entirely different direction. 445 00:24:07,151 --> 00:24:12,753 But what we have tried to do is try to figure out how to make it as... 446 00:24:13,263 --> 00:24:14,683 less burdensome as possible. 447 00:24:14,683 --> 00:24:17,943 That's why, you know, a big part of it, the challenge wasn't a model development 448 00:24:17,943 --> 00:24:22,063 center for FICO, because that's part of our business and our challenge wasn't 449 00:24:22,063 --> 00:24:27,203 blockchain technology, you know, that's one you can get your hands around. 450 00:24:27,203 --> 00:24:30,843 You know, our challenge was like, you know, how to come up with a UI so the data 451 00:24:30,843 --> 00:24:33,543 scientists don't spend a lot of time having to worry about that. 452 00:24:33,543 --> 00:24:34,513 And we did that, right? 453 00:24:34,513 --> 00:24:38,287 And so that's one of my core focuses is like, 454 00:24:38,287 --> 00:24:41,827 Okay, you know, making sure that that user experience for a data scientist is good. 455 00:24:41,827 --> 00:24:45,307 I don't like if I have a world -class data scientist, I don't need her being a world 456 00:24:45,307 --> 00:24:48,797 -class blockchain sort of, you know, whisperer either, right? 457 00:24:48,797 --> 00:24:53,287 I mean, and this is where, you know, most machine learning, you know, teams and AI 458 00:24:53,287 --> 00:24:56,407 teams today that build models have these areas of specialty. 459 00:24:56,407 --> 00:25:01,167 And I think that's been one of the major learnings is just no one human being can. 460 00:25:01,167 --> 00:25:04,167 can encompass all that in one head, right? 461 00:25:04,167 --> 00:25:10,027 And so figuring out what that MLOps needs to look like, and nowadays maybe, 462 00:25:10,146 --> 00:25:14,113 regulated AI ops, if you will, right, is really important. 463 00:25:15,438 --> 00:25:20,178 What are you hearing from customers or when you go out there and talk to people 464 00:25:20,178 --> 00:25:22,991 about their thoughts on responsibility in AI? 465 00:25:22,991 --> 00:25:24,191 Yeah. 466 00:25:24,331 --> 00:25:30,251 So the customers are really intrigued by the use of, let's say, the audible AI 467 00:25:30,251 --> 00:25:32,971 concepts that we have and blockchain. 468 00:25:33,651 --> 00:25:36,111 They're under pressure. 469 00:25:37,851 --> 00:25:43,831 More than 50 % of, in one of our survey we did a couple of years back, 50 % of 470 00:25:43,831 --> 00:25:48,351 financial institutions in the United States don't have model standards. 471 00:25:48,751 --> 00:25:51,535 And roughly 60 % of banks, 472 00:25:51,535 --> 00:25:54,555 you know, say that they're very concerned, they don't have a handle on how to control 473 00:25:54,555 --> 00:25:59,175 bias and, you know, and fairness and development of models when it comes to AI. 474 00:25:59,175 --> 00:26:03,875 And so, you know, we see customers that are a little bit nervous about their 475 00:26:03,875 --> 00:26:10,075 abilities to do this in a very tactical way and in a very certain way. 476 00:26:10,075 --> 00:26:14,175 In addition to that, right, they see this as one way that they're going to have to, 477 00:26:14,175 --> 00:26:17,285 you know, respond to upcoming AI regulation, right? 478 00:26:17,285 --> 00:26:19,343 We saw the EU AI Act. 479 00:26:19,343 --> 00:26:21,443 is marching ahead, right? 480 00:26:21,443 --> 00:26:26,423 Many organizations are starting to already codify how they're going to respond to 481 00:26:26,423 --> 00:26:28,683 that AI regulation, right? 482 00:26:28,683 --> 00:26:32,063 Which comes down to, let's go through all the little bullet points of things that we 483 00:26:32,063 --> 00:26:33,343 need to demonstrate, right? 484 00:26:33,343 --> 00:26:35,223 And then how do we do that, right? 485 00:26:35,223 --> 00:26:38,603 And I think they recognize for themselves, particularly if they don't have a model 486 00:26:38,603 --> 00:26:41,283 development standard, that it's all over the place. 487 00:26:41,283 --> 00:26:47,603 And so they find this as one way to kind of, you know, herd cats, if you will, and 488 00:26:47,603 --> 00:26:48,847 get more efficient. 489 00:26:48,847 --> 00:26:53,887 And I think organizations that do that will find benefits. 490 00:26:53,887 --> 00:26:59,147 It might be a little bit of time to do all that work, but frankly, where AI 491 00:26:59,147 --> 00:27:05,027 regulation is going today and consumer trust is heading right now, it's a 492 00:27:05,027 --> 00:27:07,907 necessary sort of task that people need to do. 493 00:27:07,907 --> 00:27:12,087 But once they do it, you'll have teams that are supercharged at doing it and 494 00:27:12,087 --> 00:27:13,231 doing it very efficiently. 495 00:27:13,231 --> 00:27:13,807 And... 496 00:27:13,807 --> 00:27:16,907 You know, I think that's where you start to get these multiples of, you know, 497 00:27:16,907 --> 00:27:20,647 productivity, just because people are so used to doing that versus you and I 498 00:27:20,647 --> 00:27:24,707 sitting there and looking at a particular type of model and figuring out how the 499 00:27:24,707 --> 00:27:25,997 hell we're going to explain this thing. 500 00:27:25,997 --> 00:27:26,267 Right. 501 00:27:26,267 --> 00:27:29,387 And then, you know, being able to sleep at night and make sure that we were able to 502 00:27:29,387 --> 00:27:29,977 do it, right. 503 00:27:29,977 --> 00:27:34,267 Is a challenge versus, you know, having a tried and true proven method. 504 00:27:34,267 --> 00:27:34,847 Right. 505 00:27:34,847 --> 00:27:36,247 Again, right. 506 00:27:36,247 --> 00:27:40,127 When we saw business problems and that's what a lot of our customers do, right. 507 00:27:40,127 --> 00:27:41,447 It's not about. 508 00:27:41,551 --> 00:27:44,251 algorithm A versus algorithm B, it's about outcomes. 509 00:27:44,251 --> 00:27:48,111 And so, you know, them just having that path is really important. 510 00:27:48,111 --> 00:27:51,071 And then being able to show that they follow the path appropriately and defend 511 00:27:51,071 --> 00:27:53,031 it, super important to them. 512 00:27:54,094 --> 00:27:57,318 What do you think is the biggest challenge going forward? 513 00:27:58,223 --> 00:28:03,423 I think one of the biggest challenges is the AI regulation that's coming. 514 00:28:03,423 --> 00:28:10,823 So when generative AI, well, the large language model thing popped up, it took 515 00:28:10,823 --> 00:28:13,043 almost all the oxygen out of the room. 516 00:28:13,043 --> 00:28:16,743 And I was involved in so many conversations that were just utter 517 00:28:16,743 --> 00:28:21,403 nonsense about, we don't need data scientists anymore. 518 00:28:21,403 --> 00:28:24,803 And these things are super intelligent. 519 00:28:25,323 --> 00:28:27,791 And I think now, right? 520 00:28:27,791 --> 00:28:31,931 we have gotten to a more sensible place where people understand that, you know, 521 00:28:31,931 --> 00:28:34,871 these large language models are cool and they're fun and they're interesting, 522 00:28:34,871 --> 00:28:35,611 right? 523 00:28:35,611 --> 00:28:38,511 They're also not meant for many regulated environments. 524 00:28:38,511 --> 00:28:41,441 They also predict the next token in a sequence of tokens. 525 00:28:41,441 --> 00:28:43,371 They're not some Uber intelligence, right? 526 00:28:43,371 --> 00:28:47,331 Even if we feel that they are based on how they talk to us, communicate. 527 00:28:47,431 --> 00:28:52,091 So we are all smartening up and letting the hype cycle die down. 528 00:28:52,091 --> 00:28:56,091 But we lost like a year and a half or two years on responsible AI journeys. 529 00:28:56,091 --> 00:28:57,935 So like when I mentioned earlier, 530 00:28:57,935 --> 00:29:02,155 you know, about 50 % of organizations don't have an AI governance standard at 531 00:29:02,155 --> 00:29:02,655 all, right? 532 00:29:02,655 --> 00:29:05,495 Which means Jack and Jill are building models entirely differently at their 533 00:29:05,495 --> 00:29:06,155 organizations. 534 00:29:06,155 --> 00:29:09,475 And there might be hundreds or thousands of Jacks and Jill's, right? 535 00:29:09,475 --> 00:29:13,195 That's a huge problem when regulation is going to insist that you write a report 536 00:29:13,195 --> 00:29:16,475 for every single model and you can justify that, you know, you follow the 537 00:29:16,475 --> 00:29:18,555 regulations, let's say. 538 00:29:18,855 --> 00:29:21,775 And I think this is our biggest sort of challenge. 539 00:29:21,775 --> 00:29:25,755 In some sense, it's a positive because responsible AI will get, have to get real 540 00:29:25,755 --> 00:29:27,239 really, really quickly. 541 00:29:27,247 --> 00:29:31,567 But for those laggards, right, in defining those standards, there's going to be a lot 542 00:29:31,567 --> 00:29:37,347 of work to do, which means they need to figure out who in the industry is leading 543 00:29:37,347 --> 00:29:42,027 in those spaces and try to mirror some of that so they can kind of catch up and 544 00:29:42,027 --> 00:29:45,327 define those governance standards themselves for how they develop and 545 00:29:45,327 --> 00:29:46,801 operationalize models. 546 00:29:47,438 --> 00:29:48,048 So that's interesting. 547 00:29:48,048 --> 00:29:54,538 You think that all the excitement around LLMs and Gen .AI got organizations to take 548 00:29:54,538 --> 00:29:56,975 their eye off the ball somewhat in terms of responsibility? 549 00:29:56,975 --> 00:29:58,295 Absolutely. 550 00:29:58,795 --> 00:29:59,045 Yeah. 551 00:29:59,045 --> 00:30:01,035 And frankly, right. 552 00:30:01,115 --> 00:30:06,815 The general, all that LM gender of AI, you know, height also caused regulation to 553 00:30:06,815 --> 00:30:07,495 speed up. 554 00:30:07,495 --> 00:30:07,645 Right. 555 00:30:07,645 --> 00:30:11,915 And so like, like the Biden administration and putting out that their AI guidance, 556 00:30:11,915 --> 00:30:12,535 right. 557 00:30:12,535 --> 00:30:16,295 A lot of it was predicated not on, you know, using AI and machine learning to 558 00:30:16,295 --> 00:30:17,955 make, let's say a credit risk decision. 559 00:30:17,955 --> 00:30:19,695 And a lot of it was based on. 560 00:30:19,695 --> 00:30:25,995 Wow, look at all these LMX, you know, live experiments on consumers, right? 561 00:30:25,995 --> 00:30:29,155 Where organizations, including the organizations that, you know, big 562 00:30:29,155 --> 00:30:31,065 organizations that build these models can't explain it. 563 00:30:31,065 --> 00:30:35,135 And I think that created a lot of that trust issue and it also created a lot of 564 00:30:35,135 --> 00:30:36,785 concern in the regulatory environment. 565 00:30:36,785 --> 00:30:40,495 So I think, you know, it most definitely people took their eye off the ball, but it 566 00:30:40,495 --> 00:30:45,235 also raised this sort of heightened concern about, whoa, which direction are 567 00:30:45,235 --> 00:30:46,075 we going here? 568 00:30:46,075 --> 00:30:46,225 Right? 569 00:30:46,225 --> 00:30:47,311 And, you know, 570 00:30:47,311 --> 00:30:50,911 That's why like if you look at the EU regulation in detail, right, you got a 571 00:30:50,911 --> 00:30:54,771 very high risk and high risk and different classifications of AI for the purposes of 572 00:30:54,771 --> 00:30:58,791 just saying, okay, you know, there are some types of models where we're going to 573 00:30:58,791 --> 00:31:03,291 have to be much more pragmatic and it might take three, four months to build 574 00:31:03,291 --> 00:31:06,211 that model with two or three people, because we're going to build it very, very 575 00:31:06,211 --> 00:31:07,001 carefully, right? 576 00:31:07,001 --> 00:31:11,251 It's not about spin the model on the set of GPUs and be done with it, right? 577 00:31:11,251 --> 00:31:12,295 And I think... 578 00:31:12,595 --> 00:31:16,435 I think that's one of the challenges that we've had with, with generative AI, but I 579 00:31:16,435 --> 00:31:20,715 do think it's all coming around and they'll, you know, if regulation can use 580 00:31:20,715 --> 00:31:23,675 good direction, it is, there'll be no choice, right? 581 00:31:23,675 --> 00:31:25,775 You'll have to kind of catch up. 582 00:31:25,775 --> 00:31:27,245 but, but most definitely, right. 583 00:31:27,245 --> 00:31:30,485 Those were not the conversations for, for many different organizations. 584 00:31:30,485 --> 00:31:35,015 And it has been frustrating for some where they see that, you know, people are more 585 00:31:35,015 --> 00:31:38,287 enamored with what a generative AI will produce than trying to, you know, 586 00:31:38,287 --> 00:31:42,167 properly developed models in a trustworthy and accountable way and having some of 587 00:31:42,167 --> 00:31:45,241 those hard decisions about a model of illness standard. 588 00:31:46,446 --> 00:31:50,206 But I'm glad to hear that after all of that, you're still optimistic, it sounds 589 00:31:50,206 --> 00:31:53,886 like, that organizations are going to get to where they need to go. 590 00:31:53,886 --> 00:31:55,446 So many important issues. 591 00:31:55,446 --> 00:31:56,966 It's really been a pleasure speaking with you, Scott. 592 00:31:56,966 --> 00:31:57,907 Thank you so much. 593 00:31:57,907 --> 00:31:59,605 pleasure Kevin, thank you.