ai learning optimization models model drift autonomous vehicles adaptive decision making

in the data and then projecting those forward. And that works marvelous as long as the world you operate in never changes. But the minute your world changes from what it looked like historically, those models, they drift. You've heard data scientists talk about drift management. Models drift. So the problem is basing your models on historical trends and patterns means the minute those models are put into production, They're out of date. They're wrong. And you need constant tinkering by the data science team. It's called the Data Science Full-time Employment Act to constantly keep those models up to date. But AI is fundamentally different. It doesn't optimize. It learns and adapts. It's constantly learning and adapting. Think about how an autonomous vehicle works. It's not trying to make the optimization. It's trying to make the best decision in the context of the situation it's in.

autonomous vehicles context-aware decisions real-time adaptation variables and metrics situational decision making

So autonomous vehicle is making, by the way, just a handful of decisions, probably 2018 decision it makes around braking and turning and windshield wipers. It uses 30,000 to 40,000 variables to help them make those decisions, but it only makes a handful of decisions. And the decision it makes right now, based on the context it's in, The decision might be very different 5 seconds later. A light turns red, it starts to rain, a ball rolls across the street, it sees a car pulling out of a parking spot, right? It sees a clown riding backwards on a unicycle, right? Everything's changing in the model. So what it does is it tries to make the right decision in the context of that moment. Constantly learning and adapting based on the context of the moment. This makes this technology very different and it makes it incredibly powerful because it's not held captive to what's happened in the past. It's using in the past trying to help it build the models and the variables and metrics, but it's using the context of the current decision, the current situation, the context of the current situation to make the right decision in that moment.

desired outcomes stakeholder workshops kpis metrics intentions

In order to do that, What the model has to do, two key things up front. What are my intentions? What is it I'm trying to accomplish? And what are my desired outcomes? We run these workshops, we teach in the class, and we bring together a bunch of diverse stakeholders. When I was at Dell, this was a methodology we used at Dell to engage with customers. We bring together a broad range of stakeholders in a room about this size with all kinds of flip charts and post-it notes trying to dive down into the desired outcomes. We wouldn't have two or three. We'd have 50, 70, 80 desired outcomes across a wide range of stakeholders. All those diverse things. That's how we start. What are we trying to accomplish and what do we think is good look like? And then we want to start for each of those desired outcomes, what are the KPIs and metrics around which we're going to measure the desired outcomes? By the way, you don't want one. You want at least three. If you have one metric for desired outcome, you get a bias. You need 3 to triangulate.

neural networks thousands of variables context-driven decisions multi-metric bias avoidance

And so all of a sudden you go from 80, 60, 90 desired outcomes to 1,000 variables and metrics. Thousands. And that's okay, because what's happened is these models are designed using a, this is a formula for a neural network, to process all those different variables and metrics, knowing what your desired outcomes, to make the right decision in this moment, given the context. Again, lots of work before we ever start putting science of the data. And in order for us to make certain that we have thought holistically about how we're going to make decisions, how we're going to look across our broad and diverse range of stakeholders and constituents to make decisions, you do need to think like an economist. And this is how an economist thinks. Economics and finance are not the same thing. AI models do a crappy job of optimizing on financial metrics because most financial metrics are lagging indicators.

economics mindset lagging indicators value creation financial metrics stakeholders

They measure what's happened. And pardon my bluntness, but AI does a shit job of optimizing around things that have already happened. So you need to think like an economist and start bracing across all these different variables and metrics. Think about how your organization creates value from a customer perspective, from an employee, stakeholders, community, ecosystem. partner, society, environment, workforce, ethical. Broad range of how value is created. By the way, in my 40 plus years of doing this, probably closer to 50 at this point, of doing this with thousands of companies, every company I've wrestled with, I've worked with, has wrestled with trying to define how they create value. We've got a finance mindset, and that mindset is the antithesis of what AI can do. So we're going to start by thinking like an economist. Again, you're going to get these slides, but you need to think more broadly.

group exercise restaurant selection decision variables brainstorming human decision making

And so let me walk you through a really simple exercise about how the human mind makes decisions. Because when you're making a decision consciously or subconsciously, you are making some of these trade-off decisions. When you go to a restaurant, for example, you might choose local grown versus not local. You might choose a local-based restaurant versus a chain. You might choose a restaurant that pays your employee more. Each of you have a different set of variables and metrics you're going to do. In fact, here's what we're going to do. I want you to pair up with somebody, and I want you to think about what are the variables and metrics that you might want to consider when trying to decide where you're going to go eat tonight after the conference here. So pair up, get a piece of paper. We're going to take a couple of minutes. Get buddy, buddy. And I want you to identify what variables and metrics you might.

group activity brainstorming timing energy building

Oh, I love the energy. We've got two minutes on this. Write them down. I love it. Right, right, right. Right, right, right. One minute, one minute. 30 seconds, 30 seconds.

human mind creativity brainstorming energy idea generation collective thinking

15. All right, 10, 9, 8, 7, 6, 5, 4, 3, 2, 1. Okay, can I get your attention? I want you to pause a second. I want you to pause. Did you feel the energy in this room? You see how everybody's sort of geared up? The human mind is a phenomenal machine. It's A phenomenal machine. When you allow the human mind to start to do its thing, the ideas that come out are phenomenal. Phenomenal. So let's go through, I'm going to point to a couple of folks and tell me, give me two or three variables and metrics that you thought were important.

decision variables restaurant metrics location price quality

How many other people are there? Okay. Chain or independent? Good. Somebody else. Location. Location. Labor. Labor. Type of food. Type of food. Price. Price. Quality. Quality. Time. Time. They can be repeats. Yeah, time. Time, good. Reputation. Reputation. Proximity. Proximity. Who you're with? Who you're with? Convenience and time. Convenience and time. The atmosphere. Okay, one more. Cost. Cost. All right. So when I do this with high school students, I do this at the Waukee Innovation Learning Center where I get students together and we brainstorm. We have more time in that situation. They come up with about 120. Now think about this. Now I bet you went across the room. If I had given you two more minutes, fair minutes, you were to come up with 100. Your mind is weighing off all these different factors. You're weighing off factors about time, cost. quality, local, community events, maybe ethical, organizations that have ethical behavior.

variable weighting decision context subconscious processing trade-offs situational priorities

Your mind weighs these things off. Now you may not write them down as you do the process, but you're processing your head. So you've got 120 variables and metrics trying to figure out where to go eat. So how is it going to figure out which of these variables and metrics are most important? Remember, intentions, in desired outcomes. So for me, when I get done here, I've got to go over to the university and teach some faculty members how to build their own digital assistant. Then when I get done with that, I went some places quick and convenient. I'm going to be tired by the end of the day. And these variables here, the weights in these variables increase. I'm going to want affordability and convenience, and I don't want to have to pay for parking. So there's a number of these variables whose weights go up. Remember, a utility neural network looks at variables and weights. And uses those weights with those variables to help make a decision. So, I'm going to Chipotle, baby, right? Uh-oh, Friday night, date night, and my wife don't want to go to Chipotle. Right? So what's important to her?

collinearity principal component analysis thousands of variables granular outputs conflict in variables

Like, I have any say in this whatsoever. Right? Well, she wants some place nice, got a nice ambiance. Somebody's going to, they're going to wait on her. You know, she can dress up. It's got, you know, maybe she can be seen by somebody or she can see friends, right? We're going some place way, way too expensive. This is how the human mind works. A large number of variables and metrics, many of them conflict with each other. This is what we want this is not. not traditional machine learning where you're trying to get down using principal component analysis down to what are those four or five most variable metrics. No, I want thousands of them. And I want the ones that conflict with each other. I don't avoid collinearity. I embrace it. Because the more variables and metrics I have, the more granular, the more relevant, the more meaningful outputs I'm going to get. So #1, If you want to become an AI solution engineer, step number one, learn to think like an economist.

yoda questioning systems thinking systems generative ai digital assistant

And think about the broad range around how organizations, companies, people, society measure value. Okay, got a good start. Let's talk about Yoda. Yoda. And I love this quote. You cannot build a system that thinks well without building one that questions well, which that was my quote. That was given to me one of the classes I had taught. And we're going to build a thinking system with Yoda. But we need to understand before we dive into this the realities of generative AI. Now when people throw the word around AI today, they almost always mean generative AI. And generative AI has a fundamental problem. It's based on correlations or statistical averaging across a wide volume of data. So think about your average large language model having the equivalent of 440 million books in it.

generative ai correlations statistical averaging social media bias outliers

And 90% of those books, 95%, 98 are coming from social media. It means you're getting a whole lot of books about the Kardashians and Taylor Swift. And so your answers are averaging across all those. The outputs from a generative AI represent the most common data trends. And most of our data is highly questionable, skeptical, biased. And so what it does, it's a correlation-based tool. It's not a causation. It can't tell you why. It can only correlate about what it's observing. gives you very average generic answers. This is because it struggles with outliers. It immediately takes second and third standard deviation items that are the items that might be most valuable and averages them right out. And finally, it has a bias towards historical data patterns.

regression to mean mediocrity average results generative ai limitations aspirational goals

It's wed to that and it's averaging across that. And so what happens, you get this regressive model, this collapse We're getting averages of averages of averages of averages. And pretty soon you're getting results that at best are average. Remember, if you're making your decisions based on averages at best, you're going to get average results. I can guarantee you everybody in this room here is better than average. If you were below average, getting to average is probably great. But everybody in this room, all of my students, they don't have aspirations of being average. And I can guarantee that people who've come to this session and who are spending time are not here to be average. So what do we do? What do we do with a tool that's got value to it? It's not causal. We're going to talk about causal in a bit. But it is correlation-based. How do I get this to work for me? Well, we're going to turn it into Yoda, your own digital system, by going through a five-step process. Now I'm going to go through this process fairly quickly.

yoda training five-step process problem focus nurse retention claims processing

Again, you're going to get the slides. If you follow me on LinkedIn, I'm constantly publishing more content about Yoda. The first day of class, we spend the entire class setting up and building Yoda. And then throughout the entire 13 weeks of the class, the students pick a company and a problem to go after. So this year we picked nurse retention, we picked claims processing, we picked childcare, selection optimization at the university. Previous cases, we've picked things like product rationalization and such. We pick a company, we pick a problem, and we immediately start training Yoda on that problem. We're going to do the same thing here in a second. But then we go through a five-step process. Throughout the semester, they're training Yoda on the problem they're going after. Throughout the semester, they're taking, and what it does By doing this problem, oh, I don't think I have the slide here anymore. Bummer.

problem-specific training filtering irrelevant data domain focus yoda customization relevant sources

We basically take that 550 million books out there and we condense it down to the 30 or 20 or 15 books that are relevant to the problem I'm going after. I'm training it. I'm training to say I care about nurse retention. Give me all the research on nurse retention. Give me all the information on nurse retention. Give me all the comments on... Kardashians go bye-bye. Taylor Swift go bye-bye, right? Chicago Cubs, sorry, you always go bye-bye, right? So I'm telling it what's important. And so I'm tricking this tool to think like me, to understand what's important to me. So here's how we do this. We're going to have a simple example that I did recently with a farming co-op. about how do we use a tool like generative AI to help us figure out what crops to plant. So here's a scenario where step one is to define our intentions, our desired outcomes, our boundaries, our constraints, give it as much information as possible about what it is we're trying to achieve about who I am.

crop selection farming co-op intentions desired outcomes profit maximization

what kind of farm I am, what size farm, how long has it been in the family, where is it located? As much detail as possible. I want to train it to define my intent and my context. So I'm A 10,000 acre farm located outside Charles City, Iowa, my hometown. And these are my objectives. So we summarized it, right? So in reality, it's probably a lot longer than this, but I need to balance as a farmer selecting what crops I'm going to plant, but I've got to look at profit maximization and risk mitigation and resource efficiency and blah, blah, blah. Again, these things all conflict with each other. AI is a marvelous tool for providing transparency and making these balancing decisions. That's its real power. So I'm going to immediately set up conflict. I want profitability. I want water quality. I want environmental, I want long-term sustainability, right? I want it all. Let's make a little commercial, right? I want it all. But how do I make those trade-off decisions? So that's step one. Define what it is I'm trying to go after.

stakeholder workshops desired outcomes dell methodology diverse perspectives rationale discovery

Takes time. We do this, we did this at Dell. This was a two-day workshop. just making sure that everybody's on the same page. And by the way, we brought in people, diverse people, even people who didn't like each other. In fact, I preferred to have people who didn't like each other because there's probably a basis for why one person had a position and another person had a position. And once you get by opinions with people and you get down to rationale, now you got meat. Now you got meat. All right, so we've defined our problem. Now we need to start training this tool. We've got it focused, and I'm going to do two things here. I'm going to leverage outside trusted validated research, and I'm also going to start mining that critical domain knowledge organizations have. So in the farming example, we went out and found 16 different research studies. There are four of them, right? We went out and found validated research studies. Use Google search as the best tool for this, by the way. I can see it.

validated research domain knowledge peer-reviewed studies google search credible resources

I can validate it. I can go out there and I get a bunch of credible resources that talk about crop selection, soil conditions, irrigation, all the factors that go along with that. I'm loading it with credentialed information and telling it, find me more research like this. Credible, peer-reviewed, in some cases legally liable resources. Easy to do. There's a lot of great resources out there. And so whatever problem I'm going after, whether it's nurse retention or customer retention or inventory optimization, get a bunch of research I can get about that problem in that industry. But now this is where things get really interesting. The real secret sauce of this process is organizational domain knowledge. And so there's a book, a textbook we use in our class called The Art of Thinking Like a Data Scientist. It's an eight-step process where we take through And we bring stakeholders together and we brainstorm across eight different steps, think design, thinking kind of concepts.

domain expertise design thinking stakeholder brainstorming desired outcomes impediments

We're starting to mine all that incredible domain knowledge. So for example, here, summary, here's improved crop selection effectiveness. We've got our desired outcomes, maximize crop yield, optimize timing of planting, blah, blah, blah. So you start, remember, if you get a lot of people together, you don't have 12, you have 80. What are the benefits? What are the potential impediments? What are the failure ramifications? What are the potential unintended consequences, and how are we going to measure that? And then we basically take a page out of the design thinking journey book. Journey maps. We create journey maps in each of our stakeholders, understanding who's involved. By the way, more stakeholders, the better. The more granular the stakeholders, the better. The more desired outcomes I have, better. The more KPIs and metrics I have, better. I want more, more, more, more around the problem I'm trying to solve. This is a lengthy process. We spend one week on each one of those eight steps to really make sure we've really well defined the problem and captured all that valuable domain expertise.

domain expertise employee layoffs productivity gains differentiation ai collaboration

Side note, companies that are laying off their employees are watching their domain expertise walk out the door. What a huge failure. Short-sighted, because there's a big difference between productivity gains by laying people off Productivity gains are not differentiated because anybody can copy them. What is differentiated is taking that domain expertise of your people who have been in industry for two years, 20 years, 40 years, and turning that into insights that AI can collaborate with. Inside note. All right, drink of water here. So we've got a great foundation. We have a lot of external data to give us perspective and validation. We're now starting to mine all that internal information, right? We're starting to capture all that domain knowledge. Now, I need to teach this tool how to think. And there are four pillars around which I want it, when it has a conversation with me, I don't want it to give me answers.

socratic method ethical foundation decision flaws un sustainability conversational ai

I want a friggin' conversation. So there's four things I'm going to do. Number one, I'm going to upload the Socratic method. The 7 questions that Socrates asked his students, he only asked 6. We had to add a 7th one because Socrates didn't care about sources of data. I do. I do care about sources of data. Very much care about sources of data. Some sources where I don't want to get data from. So we uploaded the Socratic method. So immediately, instead of giving me answers, it's answering me with questions. What are the perspectives? What are the rationales? What are the different diversity? I'm immediately getting, it's not a tool that gives me answers, it's a tool that starts to think better. If I want a tool that gives me better answers, I need to have a tool that thinks better. Number 2, I'm going to upload an ethical foundation. I upload the books of Matthew and Luke. Those books cover the parable of the Good Samaritan, which is, we can love the higher Bible if you want. But I like Luke and Matthew. I get the parable of the Good Samaritan, right? A huge difference between do no harm and do good.

ethical ai good samaritan confirmation bias recency bias decision flaws

One's passive, one's active. And the parable and the lessons around how to help others, right? One of my students loaded the Quran. gives it an ethical foundation. So it's making decisions, making recommendations, having a conversation based on ethical foundations. When I upload it, I say, tell me, think like Jesus, think like Buddha, think like whoever you embrace, right? Number 3, humans make bad decisions. We're horrible decision makers. If you want any proof of that, just go to Las Vegas. And so we upload the 19 decision flaws that humans have and we tell it, help me to avoid making confirmation bias, make it recency bias, making fear of missing out, missing some, right? We train it to say, don't let me fall trap to that. And finally, we upload the UN sustainability framework so we have a sustainability perspective as well. So now when I'm having a conversation, With Yoda, I'm having a conversation with Socrates and Jesus and guys who don't go to Vegas.

expert simulation jonathan ive steve jobs david kelley aristotle

All right, now we're getting to the fun stuff. Experts. I can have this tool act as an expert. So for example, if I'm doing a product design conversation, I can say to Yoda, bring me together a panel of Jonathan Ivey, from Apple, Steven Jobs, David Kelly from IDEO, and Aristotle. Why not Aristotle? Throw them into the mix, right? I can say bring them together and I can start posting questions and getting answers from those experts, right? I can start adding expert data here. There's one gotcha here. I don't think it's in here. Sorry. This is a problem in only 45 minutes. One of the untruths about these Gen. AI tools. People say, well, the more conversations you have with this tool, the smarter it gets. That ain't true.

ai memory memory limitations contextual knowledge base conversation persistence knowledge upload

That ain't true. You have a conversation with Steven Jobs and Jonathan Ivey and Aristotle and David Kelly and you get lots of great information, it's going to forget that. It's got a memory cap and it gets flushed out every couple of days, couple of weeks. In order to sustain that, You're going to copy that conversation, that narrative, their responses into a Word document or a PDF and upload it into Yoda with a prompt that says, factor these insights going forward. So it doesn't learn from the conversation, but we can turn the conversation, the feedback we get into knowledge that we upload back into Yoda. Pretty simple process, but don't think for a minute. I mean, I know my tool. I always like my slides in landscape format with a transparent background. And by the second or third day, it's all forgotten about transparent background. I'm like, why do you keep forgetting? And they'll say, well, I don't have a very big memory. Well, that's true, right?

crop recommendations synthesis validation profit optimization soil health

So I moved away from memory to storage. I'm creating a contextual knowledge base by loading more and more of these conversations up. So there's a way around this. So finally now, I want to synthesize, I want to validate, and I want to get an answer. So I say, okay, based on your research, I'm A 10,000 acre farm in Northeast Iowa, what should I plant? And it comes up with an answer and some rationale, some recommendations as far as why it came up with that. Right, and if you had a balance, it says, hey, I'm balancing profit and soil health. I'm improving input efficiency and building resilience or diversification. So it gave me an answer. I can start having a conversation with it now. I can say, okay, this is great. I can ask for more details. What about this? So I can say, well, what happens if we get into a trade war? How does that change? So I've got a prompt that says, what's the potential impact of a 50% trade on With Mexico and Canada, it gives me some feedback, it gives me a recommendation and a tariff strategy.

scenario analysis what-if questions tariff impact continuous training contextual knowledge

I can turn this into a conversation piece. And of course, when I get this feedback, I copy it into a slide or a Word, and what am I doing? I'm uploading it back into Yoda. I am constantly training it by loading it back into my contextual knowledge base. So now I have a vehicle for having all these kind of what-if conversations. But I can have this with a high degree of confidence Because I have trained it on the problem I'm trying to go after. I have trained it on my intent. I have trained it on my desired outcomes. I've trained it on thinking like a Socrates and having ethical behaviors. I've trained it like thinking like an expert. I've done all this training. All this training to get here. A lot of work. A lot of work setting this up. But this is the difference between a tool that's going to give you the average answer versus the tool that's going to give you the average of experts. because I've trained it on the problem and the experts who I seek guidance from. I got 10 minutes, okay? Last section.

causal ai personalized healthcare doctor-patient matching behavioral propensities treatment optimization

I promised causal AI, and we're going to talk about causal AI. We're going to talk about a healthcare example. And here's the example we're going to walk through. We're going to talk about how putting AI in the middle, using causal factors, allows me to drive a better relationship between my doctor, nurse, patient and treatments. I'm getting down to the ultimate of 1 to 1 on my engagement. I'm not making decisions based on correlations, based on generalities. I'm going to make decisions based on the individual behavioral performance propensities of the doctor, the nurse, the patient, and the treatment effect. We're going to put AI in the middle to help us, not replace us. Here's how it works. It's a concept I developed back when I was at Yahoo many, many, many, many, many years ago called nanoeconomics. And that is, when you think about my challenge at Yahoo back 20-some years ago, I had 500 million visitors a month coming to my site.

nanoeconomics yahoo propensity models personalized ads behavioral data

And I had a fraction of a second to figure out what ad to show them. I had to know not only what ad to show them, but what the value of that ad was, what the recency or the urgency of the need was for them, because in some cases I was bidding for those views. Now think about what I knew about you on Yahoo. I knew every site you went to. I knew every click you made. I knew every ad you clicked on. I knew every ad you didn't click on. I knew what you put on social media. I knew what you put in searches. I knew what you wrote on your e-mail. Sorry, you should have read the fine print. So I had this great level of detail about what I thought you were interested in. And we built a detailed propensity model on each and every one of those 200 million people. So when they came to my site, I had a fraction of a second to figure out what's their value, what's their urgency so I can make the right bid. For example, if I knew you were interested in vacations or car, getting you to click on an ad for a vacation or car gave me 19 bucks. If I knew you were interested in a cup of coffee, that paid less than a penny.

patient propensity models treatment adherence risk scores behavioral responses social support

So differentiation was built by my knowing more about each of these 200 million visitors. And by the way, these propensity models, these scores, I kept them on your cookies. Pretty simple game. You erase your cookie and I was blind. But most people didn't erase cookies. So this is a concept we're going to embrace. And so we're going to build very detailed profiles on each and every one of the players involved in a situation where we have a woman who's got cancer for the first time. She's older, she's very nervous about it, she's very scared. We're going to try to understand her situation from a causal factor in order to make sure we give her the best chance for the right kind of treatment. So the first thing we do, much like we did at Yahoo, we're going to build a very detailed model on her propensities. Treatment adherence and risk scores, behavioral response, social support. There's a whole bunch of variables and metrics we're going to capture.

doctor profiling treatment profiling matching algorithms empathy experience

These are basically psycho scores, lack of a better word. We're building very detailed scores that are predictive of her performance and behavioral propensities. And we're going to turn them into a score. The more we have, the more detail we have about what makes this particular patient unique. We're not going to make decisions based on average. Based on average, women this age in this area get this kind of treatment. That's bullshit. That's not what AI can do. I want as much detail about her so I can personalize care, I can match, I can optimize, I can orchestrate real-time guidance. This is the nanoeconomics in reality. This is working to make sure we're providing the right kind of care. And guess what? I do the same thing for doctors. Not all doctors are created the same way. Some are more experienced, some are more empathy, more, right? So I'm going to build very detailed profile, quantifiable causal based profile on each and every one of my doctors and my nurses.

human-centric ai matching service doctor-patient conversations outcome monitoring human strengths

So I can match the right doctor to the right patient. Given the patient's propensities and tendencies, I can match the right doctor. And same thing for my treatment methods. Very detailed profiles and side effects and efficiency and et cetera, et cetera. This is not hard to do. just means you have to have a different approach of understanding what are the variables and factors that are most important for care. What are the variables and factors that are most important deciding where you're going to go eat? A lot of similarities here. It just means that you... what I think the power of AI is. AI is ultimately going to make us more human because these are human conversations with the nurse, with the doctor to really understand where they are, what the propensities and behaviors are. This is where nurses and doctors, we excel. This is not what AI excels at. AI is great at matching. And so once we've got this, AI sits in the middle And it acts as this matching service, provides guidance, provides feedback as it goes around.

patient scoreboard propensity scores pain tolerance emotional resilience dynamic updates

We've got these very detailed profiles, and it's not only matching who the right doctor is for the right treatment and the right patient, but it's monitoring what works, what doesn't work. So it's going to give you a scoreboard. Let's see what the scoreboard looks like, shall we? Try to make it bigger, seat in the back. So this is a model that we're going to use to understand our patient in more detail. So you can see here's Sarah, her age, her situation. Propensity scores we built on her. We built and we update them all the time. She may become more tolerant to pain. She became less tolerant to pain. We want to know those changes. We can use language models to help us track that, but we've got to always turn it back into causal factors.

doctor adjustments causal interventions counterfactuals outcome projections expertise integration

We have to understand causal factors give us the why. And we need to understand why. If we understand why, then we can do interventions and counterfactuals. Interventions mean we can project before we make a decision what the likely outcome is. If we know the why, we can understand what kind of outcomes we think it's going to derive. So Here is Sarah. You see you got a whole bunch of information about her. You can kind of scroll down here. The doctor, as they have conversations, can change. Her pain sensitivity is not as high as we thought, and her emotional resilience is actually, where's the, she's got a really strong network, much stronger than we thought. She's got family nearby, et cetera, et cetera. So we can, as a doctor who's having these conversations, the doctor can use their expertise to help not necessarily override, but train this tool. All right, now we're going to match to a doctor. is to take a look at these propensities and say, across these doctors, here are the different factors we look at in doctors. Let me kind of scroll down here. And we think that Dr.

doctor selection treatment selection empathy matching fear management ai recommendations

Vasquez is the right one for us for the following reasons. And you can see that you've got matches here that talk about we think it's empathy and fear. If I select a different doctor, we can see that there's a risk regarding the patient fear management. This doctor here is not really good at that. So we select a doctor. Hopefully you can see, this is really quick going through this about how understanding the doctor's tendencies and behaviors, matching that with the patient. This is what AI does. It's great and it can tell you why it's matching why. And it gives you a chance to override if you think that's not correct. And the tool is going to learn when you override. It's going to get smarter by you driving the overrides. And then next one, I'm running out of time here, so I'm going to be quick. We're going to select a treatment. So it's also going to look across, I only pick five, three treatments. It's going to recommend the right treatment. blah, blah, blah. And then I'm going to generate an advisory. And then I go into my large language model and I say, okay, generate for me a plan. And I click this button and it starts going to the internet and it starts finding current research. It starts pulling things together, right? Intelligence and external validities and research that's out there and it brings it all together, gives me some information.

personalized care plans causal insights treatment recommendations accept modify reject ai learning

I can develop a care plan. So now I have a plan that's recommended for me. I can either accept points on here We'll accept this one. We'll ignore this for a second. I can modify this one and say blah, blah, blah. And this one I'm going to basically reject, and it's going to say why. I'm going to give it a reason why I'm rejecting it. It's going to factor all that in and create a customized care plan for Sarah that talks about here's what your doctor is, here's what's going to happen, here's what we know about you, the whole treatment thing, it's all pulled together, and printed off for her. based on causal. We feed it causal insights and we leverage that matching capability of AI to deliver a personalized care package for Sarah that gives her and the doctors the best chance for success. By the way, I vibe coded that.

vibe coding trained yoda mini-me ai simple instructions problem understanding

If you know what vibe coding is, when you've gone through and built a Yoda, my students like 2 paragraphs of instructions, it'll build this for you. Why? Because you've trained Yoda. And now Yoda knows the problem that you're trying to do. Yoda become a sort of mini-me of your mind. And it creates that for you. But this is how you bring together. This is where I think the big aha moment is going to be in the last two minutes here. These generative AI tools are great at providing context. It's great at finding all this information relative to what's important to me, but it doesn't know causality. If I use entity propensity models, EPMs, these are causal driven. If I bring causal together with my generative AI, I have that ability to deliver very, not only more personalized care, but understand the whys behind why it recommends this versus that. And then I can do interventions. I can say, what if I do this instead? And it can project what sort of outcome you're going to have from that. This is where the marketplace is heading.

generative ai causal models entity propensity models epms personalized care

It's a game changer. It takes a lot more work than just throwing a bunch of random data into a generative AI tool. There's no easy button for this because the hard part about it is the domain expertise of your people. That is where the power is. It's in the people. It's not in the technology. And so anyway, I hope you found that useful, valuable. You know where the marketplace is going. You know where to find me. And let me wrap up with cool way slide. Here's how you reach me. You can find me on LinkedIn. By the way, I have a Dina Big Data GPT. If you use ChatGPT and you go to their library, it's like a library of GPTs. You can find Dina Big Data GPT. It's a mini me. It's got all my books, all my blogs. Anytime I write a blog, it's uploaded in there. It's smarter than me from a under-signed stuff, right? If I get ready to write a blog and then come back and say, geez, Mars, you wrote a blog about that about six years ago.

domain expertise people power technology limitations differentiation effort required

Like, really? Six years ago, I was writing about this stuff. Anyway, so anyway, so if you have questions, and I hope you have questions, I'm going to be around about the next hour or so. Find me, ask me questions. I learn when you ask me questions. I teach because my students are asking me questions, which allows me to learn. And I'm here to learn, and I hope you're here to learn as well. Thanks for your time. Thanks. Correlation versus causation is one of those things that we like to go lecture everybody on. We yell at our families and that has nothing to do with them. But the actual application of it in this quantitative application.