ai.bythebay.io: AI: from an Idea to the Customer Panel
Recording: ai.bythebay.io: AI: from an Idea to the Customer Panel
So welcome everybody to the closing panel of the first day of AI by the Bay. So this panel is called AI from the idea to the customer and the idea of this panel is essentially there are multiple businesses which are driven with data. So AI is a pretty broad concept. So we have an evolution of data science, machine learning, AI, but in the end, businesses have very specific needs. They have customers who pay money and they need to make money and they need to serve specific customer needs. So it's a kind of very good way to give me to this concept of data and AI in the wild. And here we have folks from different companies at different stages in their data driven journey. Not only that, each of these folks have history with similar companies before
So I kind of invite our panelists to tell us about their current roles, but maybe the history with data driven business, where their current company is in this data driven journey, what kind of data it uses. And based on this we can zoom in onto specific AI use cases. So the way it's gonna run, everybody's gonna do a three minute introduction when they can kind of share the key aspects of their roles and their users of data and their companies and where they are in this data journey. And then essentially I can pick up on this, folks can pick up on each other's answers. And we just will have a fairly informal conversation. And about halfway into the panel we'll open the floor to audience questions. So we'll have the mic there and so our track manager, Michael, will essentially let folks line up there and then we can just take questions. Right? So about halfway into the panel you can basically start line up next to that mic if you have questions
So this is how we're gonna run. Let's get started. So my name is Alexey Krabarov. I'm the founder of By the Bay Conferences. Before that I used to be a chief scientist. I used to be a director of analytics. So I think I held jobs pretty similar to some of the folks here. And so I understand the space
And in my mind AI is a technosocial problem. It's not just technology. It's the way you build teams. It's also revolves around data engineering because data does not grow on trees. And it is really important to have a business use case. Right? So what I let other folks tell about their current users of AI. Should I start? Yeah. So hi everyone
I'm Xinwei. I work at a company called Grab. And I think maybe some of you won't be so familiar with that company. So let me tell you a little bit about that. We're a ride sharing company that operates in Southeast Asia. We're the leading ride sharing company there. And in contrast to our competition there, we offer not just taxis and cars, but also like motorbike taxis, shuttles, social ride sharing, social carpooling and commercial carpooling. So we consider ourselves kind of like an all service platform for transport for Southeast Asia
Right now we're doing a few million bookings a day. About 29 million users. So it's a space that we're really excited to be in. So we didn't always start that way. And I would say that right now we consider ourselves very much a data driven, AI enabled company. I mean, if you imagine millions of bookings having to be optimized, matched with drivers and passengers, pricing, pooling, like all those things. There's a lot of stuff that goes on behind that to make that work efficiently. But when we started, you know, four years ago in Malaysia, that really wasn't in our minds
I think we started to kind of serve a market need, which was safety. And, you know, for the first few months when we were doing hundreds of thousands of transactions, lots of stuff we could solve through simple rules. And as we grew and as we scaled, we've really kind of dug our heels into AI and machine learning and built capabilities there. So today we have three R&D centers. One's in Singapore, one's in Seattle, where we can draw on great talent there, and one's in Beijing. So I'd love to talk more about that journey that we've come along. Thank you. Hi, my name is Sarah Asher, and I'm director of product at Salesforce Einstein
And I don't know if you guys already heard a little bit about Einstein from Vitali earlier today. But basically, what Einstein is, is AI integrated into all the various pieces of the Salesforce platform. And this means a bunch of different things. It can mean, for example, for a particular client, they have potential customers, and they want to know which customers are the ones who are most likely to purchase. That's one aspect. Another aspect is given emails, how do you know the mood? So to use NLP to figure out, you know, whether an email, someone's happy or someone's sad. And that could be used for determining whether a case needs to be high priority, low priority. A third use case, for example, for marketing is you have images, and you want to be able to tell whether the image has the name of your brand inside the image or not
So it's actually very all-encompassing. It covers all sorts of various different pieces of machine learning, and it's all under this umbrella of Salesforce. And so Salesforce has some advantages to our data, for example. So we are a cloud service, which means all our data is already in the cloud. It means that it's already relatively well-structured. We have access to it. And the other thing we have a lot of is we have data scientists who can do interesting research and create very innovative models. One of the challenges that Salesforce faces is when we create a model like predictive lead scoring, it needs to work for all our different companies
And our companies range from very, very small to very, very huge. And we need to have models that work for each individual customer based on their data and not anyone else's data. So those are, like, some of the challenges that Salesforce, in particular, has to solve. Thank you. Hello. My name is Gabor Maly. I'm senior director at Sony PlayStation, a fairly new role for me. I knew Alexi from other roles in the past
So what I can say about Sony PlayStation is, in a way, we're also software service. We have tens of millions of users that log into the network. And we're looking to consolidate the artificial intelligence and the kind of recommendations that happen. I don't know if you've fired up a PlayStation recently, but not only are there games, but there's music, there's movies, there's lots of options. It's one of the more powerful consoles you can get for a home. And that is a growing device and platform. But I've been in the business now for about 25 years. Even in the 90s, we were doing artificial intelligence
It was called one-to-one marketing. And you could send coupons out. And then since then, credit scoring. What I've seen over the years is the shrinkage in the cost per model. Before, it would cost us tens of thousands of dollars to put a model together. Now, it costs $500 and less. So that's one of the dimensions, the real-time of it. The architecture is available
So there's an incredible time. Predictive modeling is still the core, but there's lots of other recommendations. My focus is on NLP, for example, and I have lots to say on that story. Thank you. Hi, everyone. My name is Leon Katznelson. I work at IBM. Up until about two weeks ago, my position was a director and CTO for emerging technologies in the IBM Analytics group
So IBM Analytics is where we work on database technologies, big data, Hadoop, Spark, technologies that deal with data governance and so on. And I've been doing that job for quite some time. So I'm not a data scientist. I'm more on the data engineering side. That's my background. As of last two weeks, I do have a new position. And in that position, my job is to really go out into the community and work with partners, data scientists like yourself, to help out, build the community, build the awareness, build the skill, and get more people involved. I work a lot with large enterprises, but I work even more with startups
And I call them grow-ups, not to be confused with grow-ups. So companies that are getting to apply AI and machine learning in general to their problems, I mean, one of my favorite quotes is from a guy from Wired who said that, you know, the next 10,000 business cases for startups are going to be, you know, we have X and we're going to apply AI to it. So I feel my job is to help them do exactly that. So just like Alexis said, I view the problem of AI not to be a technology problem. It is a social problem. And in my view, the most social part of that is not the fact that we have a lack of technology. I think we actually have overabundance of technology. It's a pretty strange thing for a vendor to say, I really believe that our biggest problem in AI is a lack of skill and ability to harness the technology to achieve the business results
And I see this in spades. You know, I see project after project, you know, hundreds of thousands, if not millions of dollars being poured in, and people not being able to capitalize on that. So I think that's, to me, that's the problem to solve. Hi, everyone. I'm Daniel Golden. I'm the machine learning team lead at Arteris. And Arteris is a company that makes post-processing software for radiological images. So right now, our main product that we have out right now is for post-processing of cardiac MRI
So how that will work is a patient will receive an MRI if they have a known or suspected heart problem. And then clinicians, such as radiologists, will use our cloud-based software to view the images after they've been acquired, and also to make measurements on the images in order to determine the best care for that patient. So as one component of our software that we have out right now, we have a deep learning-based system to automatically contour the ventricles of the heart. And I'm proud to say that as of January, we are the only company to have FDA clearance for a cloud and deep learning-based system that's being used in the clinic today. So there are a lot of challenges for AI, for medical software in general, and specifically for AI, including ensuring that it's safe and effective, in working with regulations that mandate that it be safe and effective, and ensure that the burden of proof is on the companies to show that. And then also in the messaging that we have for our software. So we have an AI component, but we do not say as a company that our goal is to replace a radiologist. Our goal right now is to help the radiologist by automating components of their job that are tedious, that are prone to error, and that machines can do much better than humans, such as automating measurements, while allowing the radiologist to work on the aspects of the case that are most interesting to them, such as dealing with complex cases and diagnoses, and communicating results to patients
So things that I'm very concerned about in AI and healthcare, and my company specifically, are how we ensure that the regulations are clear and that we're meeting them, and then also how we ensure that when we're selling our product, we emphasize that it's to help the radiologist do their job, and it's not to replace radiologists. And then sort of how we prepare for the future when maybe 10 or 20 years from now, maybe radiologists will have a different role, because some components of their work will have been taken over by automated software, and how we can kind of prepare for that maybe longer-term future. Thank you. So I think we have now kind of a spectrum of companies, right? And maybe I'd like to dig in a little bit into kind of how can we define AI in this context, right? So I was thinking hard, you know, what AI means in preparing for this conference, because everybody has a different definition, and I consult my advisor, who basically started my PhD in, you know, in 96. Right? I was an internal student. I went to work for industry. I worked for Amazon. I went back when I had my first son, and then I finished it
So I had ample time to think about this. And I went back to Lyle and asked him, you know, you were teaching AI for 20 years, and Lyle and we were back on Wednesday with Russell and Norvig on the panel. So he used the textbook by Russell and Norvig called AI the Modern Approach. It was 1996. Right? So we'll have all these three authorities, you know, on the panel. And so I asked him, how can we define it? And, you know, because machine learning was around for 20 years. Data science was around for two years, or three, you know, as a buzzword. And suddenly we have AI
So we kind of brainstormed it a little bit, and came up with the following rudimentary definition. That AI is different from machine learning because it's decision-making in context. Right? You need basically to have a context for decisions, and then you need to make decisions. Machine learning is fit and function, generally speaking. It's a mathematical problem. And also there's an optional component of feedback loop, right? Because the system should improve itself. After it made a decision, it should be able to judge whether the decision was good or bad. So how about I propose this as an operational definition for you guys, and I just wonder if you agree with this or not, and if you can kind of think of how your business case relates to the definition, if it bears out or not, in any order
And you can, you know, pick up on all these topics and talk among it yourselves as well. One sort of cynical definition I've heard for AI is it's whatever is five years into the future. So we never quite have AI right now, but in five years what people are thinking about, that's AI. And in five years from now we'll revise that another five years down the line. But when I think about AI, I think about the connection that a system can make with humans. And to me, one big component of that is empathy and sort of understanding and the ability to explain decisions, where a deep learning model can be engineered to have a component where there's some explanatory processes, but really explaining at the level that the human can understand and sort of empathizing, making an emotional connection is something I associate with AI that maybe we don't have yet. It's funny. I don't associate that with AI, but I think it's a necessary component for people to embrace AI
Because I think it's, you can create the most beautiful machine learning algorithms that have the greatest results, and they're gorgeous, that's 99% accuracy, though that's probably overfitted, but nevertheless. You create the perfect models. But are people actually going to believe them and trust them? And how do you get them to believe them and trust them? And so I think there are two aspects for this. One of them goes back to what you were saying in terms of a feedback loop, which is a place for people to be like, no, actually, this is wrong. I've taken the knowledge that I have and bring it into the system. And then the other piece is just trusting the model on its own. For a previous job, I remember there was one company who refused to use random forest as an algorithm because they didn't understand it. And so even though logistic regression is the one they used, because they could actually look at the formula and anyone with some math skills can look and see what's important and what's being used, and it was just too much of a black box for them to trust using like a random forest
So again, I'm not sure I'd call that as part of AI, but I think that's the only way that we can get adoption of AI and have it actually improve people's lives. Do you think it would be explainable as well? What I found, though, is for the reward, that the reward now can be financial. So it takes a while to convince people that it's going to really help them. But once they see the money coming in, they become fairly convinced. And if you can also do some simple statistics like A-B testing, so sort of experimentation is part of artificial intelligence, except that it's statistics, and a lot of us don't think as statisticians, but now we're sort of enriching what intelligence can be by doing A-B testing, proving out that artificial intelligence is actually working. Oh, yeah. In terms of the context, well, that can also be automated, though. A lot of people don't know what the context is, but if you're looking at a specific website and you think of it in one way, if it's automotive versus another one's consumer electronics, and it can be adapted, but a lot of that can be automated, and you don't know that it's actually doing a lot of fancy decision-making
But I've found that that can also be machine-learned. I mean, I think the definition of AI is something that, as you said, if you ask ten different people, you might get ten different answers. And it's been every time there's a new kind of development in the field, then that definition changes again. I mean, in general, maybe you should think of it as like, what are the best kinds of problems that we should apply these techniques to? What we found at GRAB is that as long, if it's a big, obviously the bigger the data, the more real-time that the decision has to be made, and if it's an automation and optimization problem, those kinds of problems really lend themselves to being solved well by AI techniques. And things like dynamic pricing where you have to get a price on your app within seconds. So, I mean, definition-wise, it's definitely about decision-making and context, it's definitely about learning. And I think we can also talk about whether it's good to have a human in the loop or not. But thinking about it as a class of problems that it's best, kind of best fit to solve, I think, is another way of looking at it
And I don't know what the definition is or it should be, but I certainly draw a lot of that from having conversations with people who don't have PhDs. And I think the basic view is if you say machine learning, it conjures up one image without knowing what it is. If you say AI, the view shifts towards intelligence. And I think that's something that you kind of meant by the feedback loop as being in there. And, you know, we need to kind of move away from the word artificial for a second and focus on the word intelligence. And I think the context in there is absolutely critical because this is where we look at the specific domain as opposed to general intelligence. Because when you say general intelligence, we kind of conjure up these images of, you know, freaky robots jumping around from Boston Dynamics and so on. And, but I think people really understand intuitively when we talk about AI in the context of intelligence and the feedback loop in the learning
Because we all understand from a human experience what learning is and how that brings forward. So I know that's a non-technical explanation of that and I wouldn't get a PhD on that. But to me that's, you know, bringing it towards more of a consumer view of things is helpful in understanding the difference between machine learning and AI. Thank you. So I think we have a context, right? And so what I'm hearing is that, you know, we have big data as one of the causes of problems, right? So we have obviously lots of people moving around, right? So you guys have big data. And you guys have big data in images, right? But Vitaliy opened today with a kind of machine learning for the 99%. And he says, shallow learning is not solved yet. Right? So I wonder, and I know that Salesforce has big data in aggregate, but every individual silo is not necessarily big data
So it's kind of federated data, right? So I wonder if you guys can talk about what is the kind of amount of data needed for your customers, for you to make decisions. And, you know, does it have to be deep learning? Or if not deep learning, how do you deploy traditional machine learning? And how do you make a decision? How much data do you need for your business? Right? And how do you go about getting it? Well, if I may, because one image that I have that helps me understand it is a learning curve. And as you add more data, you have better performance. So as you have a lot of data, you can have pretty big performance. But it helps me to understand what the minimum amount of precision that is required for the uncanny valley to be crossed for the application to be useful. And you can do a lot with very little data in some domains. And it just turns out that when you deal with humans, they expect a fairly high bar for precision. But I always say that you should look at the learning curve to understand how much data you really need
Don't just assume that you're going to need billions of records. You have to really be aware of what your performance metric is so you know where you're going to get. And you don't need as much as you typically think you do. I think that's true. I think, though, a lot of it has to do with the amount of data. Or it's not the amount of data, but the quality of the data you get. And more data usually trumps very high-quality small data. But nevertheless, if you're at a company that has a small amount of data and the data isn't very good, your model's not going to work
And basically, there has to be a system around there to tell people, hey, this is why we can't solve this particular problem for you. This is what you need to do, whether it's get more data or whether it's clean up the data you already have. But then jumping back to what you were saying, there is an accuracy aspect of that showing at what point is it useful. You don't have to have 100% accuracy for a model for something to actually be useful for people. And to explain that to them so that if something has a score, that it's okay that if your accuracy isn't 100% all the time, as long as you're adding more information for the user to figure things out. Does that make sense? It does. And what my baseline is typically other people. So I hire other people, and they make mistakes, and now I level set
By the way, most people get it wrong 10% of the time. I don't expect my machine learning algorithm or my artificial intelligence system to do better than an expert. It's very common in medicine, especially in radiology with images, to have what you might consider big data. So like the algorithm that we had FDA cleared in January was trained with on the order of about 1,000 cases. And when you compress that into the format that you put into the TensorFlow model or whatever, that ends up being on the order of, I would say, tens of gigabytes. So it's not that big. It very easily fits into, on a computer's hard drive, like a web computer, like EC2, or maybe even into memory. So it's very common to have that much data, and really in medicine
And one sort of advantage of working in medicine is it's extremely imprecise, and this huge inter-rater variation. So in one sense, that's a disadvantage because it means that to get better models, you need quite a lot of data to trump that inter-rater variation, that noise in the data. But at the same time, the bar is low. So we were actually able to make a claim with our submission that we were as good as an expert annotator. And that's because expert annotators have a huge amount of variation in their results. So to some extent, it's helpful that there's this amount of variation. And to another extent, more data is always better. But often, we're operating in a small data regime where we can easily train on one computer with a few GPUs
I'd like to pick up on this because I think your case is actually very interesting. Because you're kind of one of the first companies whose AI was certified by humans. And so the medical decisions, I think the legal decisions will be ultimately decided by humans. So you have the government system where the court of law will be the final arbiter of a legal case. AI will never be a lawyer, at least in the current system of government. But in medicine, you still have doctors making decisions, and then you convince some high-level medical authorities that you are making similar decisions to a human doctor. So I wonder, what did you see from this? Can it be kind of a blueprint for certifying AI? Because for instance, we have Tesla Autopilot, which allegedly killed a person. So it will lead to government certifications
So in another context, humans will have to say, this AI is good. It will not kill a person. This AI is bad. It might kill a person. So you kind of passed through something similar to that. Can you talk a little bit about that? Sure. And I think that the latest that I've heard about that Tesla incident is that Tesla was found not at fault. Because they had pretty clearly messaged the capabilities of their system, and the operator of the vehicle was not using it in compliance with what they said
And then that's a huge part of FDA clearance. What you get cleared to do, essentially, is to message. You say, this is what the software can be used for. This is what it's capable of. And the same piece of software that does something can achieve different levels of clearance, depending on how you message around it. And so our software is, from the FDA standpoint, a semi-automated procedure. It's a fully automated system, but the radiologist has the opportunity to make modifications to the results and change them. And if our system makes mistakes, and on occasion it does, they can always change them and fix them
So I think that's kind of a critical component of, especially these early stage AI systems, that allow human in the loop to make modifications and to see and make changes to the results. And I think that's a critical component to how we were able to get cleared. Thank you. So maybe I'll ask folks to start lining up next to the mic if you have questions, and I'll ask another question, and we can switch to that. So human in the loop, let's maybe talk a little bit more about that. So I think it's recently become very common, more common than before. So can you talk about human in the loop in your companies? When does a human have to look at the decisions made by AI, and how do you integrate the human in this process? So actually when Daniel mentioned having a radiologist look at the results, that reminded me of something we went through in GREV as well. Not for certification, but to give you an example, it was when we used, kind of automated our fraud detection systems
So we have drivers, a very small percentage of drivers who set out to essentially defraud the system because of incentives, of promotions, and so on. When we first started, we would sort of catch them by the ground teams, would tell us, these are the ways in which we think drivers are frauding us, and then we would go in and look for those instances. So when we automated it, and then the system then spit out kind of a list of names of drivers that were doing bad things, and all the operations teams were up in arms because they said, these guys aren't fraudsters, there's no way. We don't believe it. We don't believe the machine. So how then did we kind of bridge that gap? So what we have now is essentially something that's, I guess, we can call it semi-supervised. So someone essentially validates a lot of what comes out of the machine. And the reason why that's necessary is also because fraud is an interesting problem
It's not like predicting the weather. It's more like you do something and then they react. And then you do something and they react again. So if you're always kind of one step behind, then you're at a risk of always being ineffective. So we placed kind of a human in that loop to, I guess, validate and to refine how the machine was identifying these fraudsters. And over time, the humans had to get less and less involved. But I think that was a necessary step, both for validation as well as to earn the trust of people who had to kind of live with the results of what came out of that machine. The phrase that comes to mind is trust but verify
So you have to have these systems that you need to be able to trust. But you do verify. And so for us, it was annotators. So every time that every day we would have several people look at some of their predictions. And all of our predictions had some confidence. And so we looked at the low confidence predictions and then remedy those. And so continually learn that way. So always having a human in the loop because you have to build training data to know what you want to evaluate
And if you don't have those people looking at your prediction, then maybe you don't know what you're predicting. So that investment in people who go verify what the predictions are seems like a good one for any artificial intelligence system so that you know what you're trying to accomplish. So I had an interesting case. You know, we talked about medical, Tesla, self-driving cars. And it kind of makes sense that you want a human in the process in there. But I was working with a financial company. And when we talk about the feedback loop and automatic learning and how we can improve the model in every duration, they immediately said, we like everything you say except that one thing. Okay? Nothing goes in automatically
So it goes back to the issue of trust. And I think the trust is being underestimated in how important it is. So I think, you know, we are going to trust the humans, even if the humans defraud us more often than the machines. We are going to trust the humans every time. We're not yet ready to trust the machine. So I don't see anybody next to the mic asking a question. Like, raise your hand if you want to ask a question. Okay
Folks are queuing up. Awesome. So I'll just ask one more question while you guys are coming there. Explainability. It just struck me recently. We demand that machines explain themselves. How often do you demand your colleagues explain themselves? How often do you demand humans explain how did you arrive at a decision? Are you even capable of explaining why? Except maybe I don't like it, right? Maybe my gut feeling is I don't want it, right? So, you know, why are we demanding that machines explain themselves if humans cannot explain themselves? Should we do this? You know, any thoughts? So I'll come back to that very same case I was talking about in the financial company. Actually, they're quite sophisticated in their data science and AI
But this is one thing they absolutely demanded is the governance. And it was not because they wanted the governance. It's because they're working in a regulatory environment. And any time a decision is made, you absolutely have to be able to explain how the decision was arrived, whether it's by machine or human. Okay. What was the data that was used to train and score the model? Okay. What were the circumstances? And it all has to be kept. So that's, to me, that's one good example of where compliance is when demands that we ask machines to explain themselves
I always also go back to some of the simpler models or methods to do learning. And that's a decision tree. So you can easily walk through a decision tree, explain why you made the prediction, and there's the path. So that often helps the business people to understand. And that's as good as most explanations by experts as well, who are at post hoc explanations that just have reverse engineered some answers that they think the other person will like. And this one just has data to support it. So I think it's almost as good. Well, I think it's funny because people, like, depending on how the person acts and talks depends on whether you trust them or not
So it's actually, it's, if someone has, we refer to my family as voice of authority, if they say something strongly, yes, this is a good lead. This is a person you should talk to. You're just going to believe them. And you don't actually even care what the, you know, what, what their reasons are. And as a result, you're more likely to be led astray. I think it's actually better in a sense to have the machine say, hey, this is, this is why this model is important. And that that should be built into part of every model that you create should be the insights that go along with the model explaining why the decision was made the way it was. It helps user adoption
It helps for regulation. And it's just a, it's a, it's a better system than just believing someone on the sound of their voice. I think I'll just say quickly that we have the advantage, for example, for our model that does this automated segmentation that the radiologists used to do manually, where they know what the answer is. They just don't want to do it from scratch themselves. It's a tedious process. So to that extent, or to some extent we are obviated from needing to explain ourselves. If occasionally we're wrong, the radiologists will see that occasionally it's wrong and that's kind of lame, but I can fix it. I don't need to know why it was wrong because I can fix it and make it right
And so that's sort of an advantageous position. Whereas if we were doing classification and just say, um, this patient has heart failure, this patient doesn't with no other information, that wouldn't be acceptable. Because a lot more explanation would be needed because it's not obvious. Thank you. So now we'll open the floor to the audience. Please, uh, uh, ask, uh, so introduce yourself and ask a question and you can direct it at anybody on the panel or everybody. Hi. Hello
Hi. Uh, Nikki Rota here. Thank you again for the panel. This is wonderful. I'm an interaction designer actually. So I was very excited for a conversation today, speaking about how users are interacting or what are the use cases that we're thinking about for AI. And it's been spoken of here, but I want to sort of broaden it out to what has been spoken of by previous, um, presenters as well as how are, how are we thinking about how people consume the information or the results of the various models that we're using? I mean, we've talked about trust and we've talked about building trust and maybe a business context, but there are end users too, who these are con there are consequences to the choices that our models are making. You know, whether it's, you know, uh, uh, cab comes five minutes versus three minutes or, or, you know, a radiology result
Um, and thinking about, um, how we as the purveyors of the data, the experts, if you will, of the data sets that we're working with, but we might not necessarily have the, all the mechanics necessary, right? Like a data scientist is an expert maybe in, um, data engineering or, or statistics, but maybe not specifically the domain that their models are being applied to. And how do we build trust if data is also has social implications? Anyway, thank you so much. Well, I have, uh, sort of the scientist answer if that's okay for me to start that way. Um, so I look at intrinsic measures of performance. So these are measures that my team can sort of geek out on and, but maybe the impact on the user is lost. So that's not good. And so that's where you're getting at a little bit, but there are extrinsic measures of how is it really behaving out in the field. And ideally those measures connect to the experience of that, that user, which, you know, a lot of what we service now, a lot of websites and apps that we're creating are fairly impersonal to begin with
So it's whether an AI is creating an impersonal experience or some designer like yourself does an impersonal experience, they're both are bad outcomes. And so if we have a way to track, let's say the engagement by the user through this change, is engagement there or did they walk away? So those extrinsic measures I've found to be really important because it keeps both yourself and my team honest in terms of what changes we make. I think if we even take a step before that, I mean, one of the first things we tell anyone who joins Grab during the onboarding is, you're not here to solve a technology problem, you're here to solve a market problem. Uh, and what that really means is that you're taking what we call your superpowers and you're applying them to problems that you know, exist for your consumers and you're trying to make life better for them. Uh, and so what we try and do is to have a very close working relationship between our data science teams as well as the business teams. I think, uh, we fully kind of see that both sides have something to bring to that conversation. Uh, to give you an example, um, you know, there's obviously an optimal point at which demand needs supply when you set a price at a certain level and that price can go dynamically up and down. But the public acceptance is of that pricing is something that, that isn't, you know, it's not something that we can necessarily predict, uh, through our existing data
And it's also something that will evolve over time. Right. So that, that kind of local knowledge is something that the business teams will bring and they are the ones who suffer the consequences in a way. Uh, so they can be kind of a feedback loop in, into the machine to say like, Hey, you know, how far should we take this? Um, at what point should we start to kind of, kind of optimize it for, for kind of how people are feeling, not just like what they're paying. You guys do search pricing too? Uh, we do some dynamic figure pricing. Is there any difference between different countries to people responding to search pricing? Uh, very much so. Which is like where people don't like it the most. No one likes search pricing
I would say if you talk about like price sensitivity, like, so the elasticities, um, Vietnam is probably the downside. Uh, most prices, price sensitive, uh, both upwards and downwards. Is it correlated with income? Uh, I don't think it's a hundred percent related with income. I think that's definitely like, uh, it's probably related to like presence of substitutes. That's just, uh, a bit logical. But I think it's also like a cultural aspect to people like not wanting to feel like they're getting cheated by a machine. So, there's a bit of that. So, so it's interesting that you mentioned the cultural aspect of this
I mean, of, uh, had it recently had an experience of age aspect. So we, you know, as data scientists, we would all talk about the accuracy of our models. And it turns out the acceptance of the solution has very, very little to do with the accuracy of the model. Uh, so I hate to draw it again on financial institution, but they put a chat bot in place. And, uh, they found out very quickly that it was the presentation of the chat bot that drew the adoption and the like of it more than the accuracy of that. Specifically, uh, what they noticed is when they introduced, uh, an avatar, uh, the acceptance of the suggestion by the chat bot became so much higher, especially with the senior citizen population. So I had absolutely nothing to do with the suggestion by the chat bot. I had a chat bot in place and, uh, they found out very quickly that it was the presentation of the chat bot that drew the adoption and the like of it more than the accuracy of that
Uh, so I had absolutely nothing to do with accuracy. They could have just light through their teeth, uh, with a chat bot, but the people just love the advice that they were getting from an avatar versus an impersonal text, uh, type of suggestion. Do you think that's more of a short term, uh, thing and then like over time the usage will drop off in the accuracy of the data? I don't know if I, that's a good question, right? It's a good question, but they actually saw a, a, a market increase in the, in the MPS, uh, score for that. Thank you. I think we have our next question. Hey, uh, my name is Daniel. I want to thank you all for talking about AI in the industry and in the wild. Um, so my question sort of has two sides that naturally fold into each other
How do you find and hire good AI talent? And if you want to be that good AI talent for industry, what do you think the preparation is? Well, uh, so my innate, what I look for is always innate talent or giving them problems that relate to artificial intelligence and fairly basic ones. And seeing how naturally it comes to them and answering the question rather than taking a forced route. So how do you prepare for that is to continually practice that, that aspect of, of not necessarily programming, but working with information and working with those problems so that it becomes second nature to you so you can answer it really clearly and well. And then also then you can determine whether you actually like doing that kind of work. For, for us, I think knowledge of machine learning is, is very important, but sort of an underappreciated aspect of, uh, at least the work that we do is general software engineering skills. And that means unit tests working collaboratively and get, for example, um, writing code that has like a very high probability of being correct and is understandable and so on. And I think that's not emphasized very much maybe because general data science, uh, is more of a, a solo endeavor. But when we're making, um, uh, uh, models that we intend to use for production and whenever, when, uh, we have all of our team working on a common code base that is used to create and train those models, the correctness and understandability of the software is very important
And I think that's something that, um, a lot of people don't have as much skill in that I think would, would be really beneficial. Yeah. Um, we always, I always think of it as there are actually two different jobs, right? There's the data scientist and the machine learning engineer. Mm-hmm. And the machine learning engineer has to be able to understand the data science and productionalize it and actually put it to scale. And as you were saying, have the unit tests, make sure that the system is strong and solid. And I guess the question for you is sort of which direction are you sort of more interested in? Uh, depends on what your answer would be. I'm, I'm with you that it's, it's really good to have good software skills in order to be a really good machine learning engineer
And I think that probably in a larger company like Salesforce, there are more verticals where there's the data scientist and there's machine learning engineer. But a lot of times they're the same. We have, we have, we're a pretty small company and yeah, it's the same. Everybody does, does everything for us. So yeah. no, we, we, we definitely have teams that are combined where it's the data science and the machine learning. Uh, and frankly, it's actually easier if they're the same person. Uh, uh, back in a consulting company, I did a little bit
It's, it's hard when you have the data scientists come up with this beautiful model and it's gorgeous, but you can't actually use it. Uh, because they haven't thought about how to scale it and how to actually use it in a production environment. And it's much better if you can to actually create it the first time around with those considerations in mind. Yeah. So in my experience, I, I don't know the answer to your question, uh, uh, very well, but we deal with a lot of people who are trying to put together a data science team. Uh, it's published out there by our Institute for the business value and I recommend people pull that down. Uh, and the, the kind of, uh, the result of the study has been that you always want to start by selecting your kind of achieved data scientist and building the team around them as opposed to start with a particular expert in let's say data engineering or, or, or, uh, you know, machine learning and so on or domain expert. And, and that, that that's been kind of been proven to be a, a very valid technique for building data science teams
And, and I agree that, you know, you really, when you're entering into an existing data science team and you're being interviewed, you probably want to make a mental note in your mind that says, which way do I want to go? Uh, as opposed to just looking for a team to join. Thank you. Uh, next question. Hello, my name is Diego and, um, thanks again for the panel. I, the other question askers in the same, my question actually goes along the same lines as Daniel's, but it focuses on management. So managers, be that product or engineering, what skills, traits, you know, natural talents do you think are required or beneficial for somebody who will be managing a AI driven or numerical driven team as opposed to a regular software engineering team? Well, so my, uh, my focus is on having a, first of all, as a manager, I have a career track for, uh, data scientists. So having, uh, them be able to see forward. It's the track for a data scientist or an AI specialist quite different than software engineers
So coming up with that track is, is sort of important for your team to know what the direction is. And that's, I think it has multifaceted areas. So how much, how well are you with managing data? How well are you with statistical concepts? How well are you with algorithms? So going through that process is, uh, and tailoring it for the individuals is, is quite important. But then also having metrics for everything that you do so that you can, since most of these people are fairly quantitative, um, being able to express to them what it is that they're achieving and making them visible. Like that project you just did, had an ROI of four months, uh, would be, I found to be very valuable to get them, keep them excited. I think one, one critical thing is an interest in management because a lot of maybe the, the top level developers or, um, data scientists are promoted into management. Really all they want to do is get into the data science and, or, or write code. And if you sort of wishing you were doing that and reluctantly being a manager, you probably won't be as successful
So I think there are probably a lot of managers like that, that sort of wish they were still writing the code and probably those managers are less successful. So someone who sort of is very interested in the elements of that role is, I think, a critical quality that some managers lack. I think it depends also on the stage of the company. I, I do remember when we first built out a big data science team in Grab, uh, there was a lot of misunderstanding as to what this team did and like how it would benefit the business. I mean, I gave some examples about, you know, when we did the fraud machine and so on. So in a way, you know, if you're at that stage, then the manager needs to be someone who, you know, as Daniel said, maybe is okay with not doing so much of the actual technical work, but defending his team and kind of helping to sell that to the rest of the organization. And when you reach a greater level of maturity, like where we are now, that, that person then becomes a business partner. He's no longer, you know, he doesn't have to like be a shield anymore, but he's more like, he's kind of the, the thought partner to the business side
And, and what they do is really collaborate on like how to solve business problems. So I would say that's, that's how I would look at it based on like where your company is at right now. So then let me ask you guys, do you think your data science manager needs to be a data scientist? I would say yes. I think there's, there's so many elements of the actual data science or machine learning that are important. And for a manager, I mean, a manager has a lot of roles. And one of those roles I think is in sort of coaching the team and helping them with the technical aspects. And if, if there is no sort of senior leadership that can help them technically, I think there can be a lot of, you know, going in the wrong direction. So I think that is, do they have to, I can't say have with a hundred percent certainty
Yeah. But I would say that is a big, big bonus. So I, I think that's, it's a really good question, but I don't, I'm not sure if I agree with the answer that Daniel has provided. I'd say that you have to have enough understanding and enough practical experience to be able to provide the mentoring. But manager, first of all, I, I, sorry to put a cliche out there. Manager is not a manager, manager is a leader. If you're, if you're managing and not leading, you're not going to do well by your team. So to be a leader, you need to enable others to do things and not to do it for them or not to show them and kind of take them by the hand
So, you know, sometimes it takes, you take a roundabout way of getting there, but you, you need to enable your people to kind of come to their own. And yeah, it is, it is mentorship and you do want to show them. But I do believe that you really probably want to have a senior person on the team, if you, if you have ability to do that, to provide the technical mentorship. And the leadership has to come from a manager. And from that point of view, I'd rather have my manager be more leader than other scientist. So I think we have one last question. Great. So my question is for Daniel, but if anyone else has an opinion from the perspective of a heavily regulated industry, I'd be curious to hear that as well
My impression of the FDA is they can be pretty traditional and rigid with their expectations. But on the other hand, AI models are constantly growing and learning. What is your experience been with working with the FDA in the sense that, you know, do they expect an entire process change when you update a couple of weights? Or is it a matter, they give you a kind of a longer leash for the development of your model between needing to communicate with them when changes are made? Yeah, I think that's a great question. I think the, the FDA gets a bed wrap because a lot of people say it stands between them and progress, for example. But I think its role in ensuring that any products that you make are safe and effective, and those are the two sort of magic words for the FDA, is really important. Because I think, I mean, I think we've all been there and made a bad model that it's over fit and it doesn't actually work. And there's always motivation to get that out there. And the FDA is the gatekeeper making sure that whatever models you put out there are good and have been tested and have been proven that they're safe and effective
I think that they could do better work in terms of guidelines because you have to, they don't tell you what safe and effective means. They say, make it safe and effective, and with some exceptions, essentially how you decide on what that is, is up to you. And then it's like a paper review where they look at it and they say, we don't agree with that definition, try again. Or they say, that's correct, or whatever. So guidelines where you don't have to guess as to what they want would be really helpful. In terms of updating a model, I think people like to talk about active learning, where you create a model with a corpus of data, you deploy it, and then users use that application with the model. And they maybe make some tweaks, and that gets fed back, and it gets updated and done in a continuous way. The update is not like that
Because you have gone through a lot of trouble to show that your model is safe and effective. And if now you have a new model, what makes you think that's safe and effective? What if someone dumped a whole bunch of bad data in there and it turned into garbage? You would need to know that. So active learning is out for medical devices, for sure. But an occasional update, let's say maybe you could do something every three months or something, there is a process called a letter to file, and there is more detail that you don't want to know right now or that everybody doesn't want to know, wherein you can say, we're making a minor update. Here is how we still believe it's safe and effective. It's extremely similar to what we already submitted. And in that case, that's actually, that does not get reviewed by the FDA. If you have internal processes that you tell the FDA to convince you that it is safe and effective, and they will take you at your word until proven otherwise in that case
So first shot, you do an application that gets reviewed. Updates can be this simple process that you just sort of send them a memo and then just do it. But you can't do it continuously. Does that answer your question? Yeah, that was great. Thank you. So I think that basically concludes our panel. So I want to thank everybody on this panel. I want to mention a couple more things
So actually, I want to congratulate you on Gabor because he started his Sony role today. So we started this conference. And Gabor was a frequent speaker and teacher at this conference before. Another thing, coming back to the careers, we have the job board, which is upstairs. It has two sides. One says, I want to get hired. The other says, we want to hire. And guess which one is bigger? Any guesses? We want to hire or we want to get hired? We want to hire
So actually, at our previous conference, it was so disproportionate we tweeted this. But we'll see how this goes here. We have it upstairs. Please, you know, affix yourself on a post-it there in either column or in both. And, you know, we'll share it. And so now we'll basically proceed to the happy hour. And I think the drinks will be served here at the bar. The food will be at the mezzanine
And you guys are welcome to roam the whole place. So we'll have it until 8. So thank you very much. That concludes our first day. And, you know, enjoy yourself. And we'll see you tomorrow. Thank you.