Dr. Eric Daimler is an authority in Artificial Intelligence with over 20 years of experience in the field as an entrepreneur, executive, investor, technologist, and policy advisor. He is also the founder of data integration firm Conexus, and we kick our conversation off with the work he is doing to integrate large heterogeneous data infrastructures. This leads us into an exploration of the concept of compositionality, a structural feature that enables systems to scale, which Dr. Daimler argues is the future of IT infrastructure.
Dr. Daimler is an authority in Artificial Intelligence with over 20 years of experience in the field as an entrepreneur, executive, investor, technologist, and policy advisor. He is also the founder of data integration firm Conexus, and we kick our conversation off with the work he is doing to integrate large heterogeneous data infrastructures. This leads us into an exploration of the concept of compositionality, a structural feature that enables systems to scale, which Dr. Daimler argues is the future of IT infrastructure. We discuss how the way we apply AI to data is constantly changing, with data sources growing quadratically, and how this necessitates an understanding of newer forms of math such as category theory by AI specialists. Towards the end of our discussion, we move on to the subject of the adoption of AI in technologies that lives depend on, and Dr. Daimler gives his recommendation for how to engender trust amongst the larger population.
Key Points From This Episode:
Tweetables:
“You can create data that doesn’t add more fidelity to the knowledge you’re looking to gain for better business decisions and that is one of the limitations that I saw expressed in the government and other large organizations.” — @ead [0:01:32]
“That’s the world, is compositionality. That is where we are going and the math that supports that, type theory, categorical theory, categorical logic, that’s going to sweep away everything underneath IT infrastructure.” — @ead [0:10:23]
“At the trillions of data, a trillion data sources, each growing quadratically, what we need is category theory.” — @ead [0:13:51]
“People die and the way to solve that problem when you are talking about these life and death contexts for commercial airplane manufacturers or in energy exploration where the consequences of failure can be disastrous is to bring together the sensibilities of probabilistic AI and deterministic AI.” — @ead [0:24:07]
“Circuit breakers, oversight, and data lineage, those are three ways that I would institute a regulatory regime around AI and algorithms that will engender trust amongst the larger population.” — @ead [0:35:12]
Links Mentioned in Today’s Episode:
Eric Daimler 0:00
To the degree to which this technology will be closer to the utopia that we imagined, versus that the dystopia of Hollywood narratives is really up to all of us.
Rob Stevenson 0:12
Welcome to how AI happens. A podcast where experts explain their work at the cutting edge of artificial intelligence.
Rob Stevenson 0:20
You'll hear from AI researchers, data scientists, and machine learning engineers as they get technical about the most exciting developments in their field and the challenges they're facing along the way.
Rob Stevenson 0:33
I'm your host, Rob Stevenson, and we're about to learn how AI happens.
Rob Stevenson 0:45
To work in AI is to be surrounded by possibility. You'd be hard pressed to find an AI application without revolutionary and disruptive potential. So now that you're an AI expert, how do you even decide which world changing future you want to spend your time bringing about? Today's guest is a true AI powerhouse. Dr. Eric Daimler has spent 20 years in the field as an entrepreneur, investor, technologist and policy adviser. He served as the Assistant Dean and assistant professor of software engineering at his alma mater, Carnegie Mellon University, where he earned his PhD in computer science. Dr. Daimler has co founded six technology companies, including his current one connects us. He's the author of the upcoming book, The coming composability, the roadmap for using technology to solve society's biggest problems. And he served as the Presidential Innovation Fellow for the Obama administration. He assured me that Air Force One is every bit as cool as you think it is. Given Dr. Daimler storied background, I wanted to learn where he sees massive opportunity in the AI space, and how he decides where to point his expertise next. In short, his answer is the principle of compositionality. But I'll let him tell you why.
Eric Daimler 2:05
I was influenced not only by the time I've spent in and around artificial intelligence, and its various expressions, but certainly the time I spent in federal government, and during the Obama White House, that was really a different perspective around AI, where I saw the limitations of these grand plans, both for organizations the size of the US federal government, and also very large commercial expression. So I came to the conclusion that the limitations of AI were manifesting themselves in ways that were not appreciated. You know, I think that people have gotten the memo about collecting data, the phrase, the data is the new oil or some such thing. People understand that. And they've been collecting a lot more data. But it really feels like there's a misinterpretation of what that even means. What I didn't intend is for people to give me timestamps on US equities, at the second level or millisecond level, where previously was at the minute level, just to use a analogy, you know, we have high frequency trading in equities, and we've solved that problem. But you can create data that doesn't really add more fidelity to the knowledge you're looking to gain for better business decisions. And that's one of the limitations of that I saw expressed to the government and other large organizations, you know, what I saw was that many, many organizations were collecting data and then storing it in strange ways that didn't comport themselves to what the business actually needed. Getting knowledge, getting wisdom. That's what I'm working on right now. I found a research project that happened to have gotten funded by the US Department of Defense and the US Department of Commerce simultaneously, to solve two very large problems that were deemed to be intractable at the level of computer science. I started working with this group, once I got out of the White House and began to put money into it personally, to explore what it could be as a business, then we have turned that into something that's commercial, and I think can help fulfill on people's expectations around AI, the bringing together of large, heterogeneous data infrastructure to allow for better business decisions.
Rob Stevenson 4:39
When you say the approach to AI was different or the perspective on AI was different. Was that a matter of scope of thoroughness? What was the difference between your previous experience in the White House experience?
Eric Daimler 4:50
I am grateful for spending time in and around AI from the perspective of an academic researcher at Stanford at Carnegie Mellon University of Washington, Seattle. And as a venture capitalist on Sand Hill Road as an entrepreneur, full time entrepreneur, but my time and policy gave what you just said, Rob is a bigger scope, it gave a broader visibility about these really, really large problems. You know, sometimes their organizational, sometimes they're technical, most often they're both. And I just got to see these NSA circumstances, terrible stories up front, where breathtaking amounts of money were wasted from doing ETL processes that are terrible, in all respects, they're terrible for the individuals having to execute them, because they can be vocational level it work for people, often with graduate degrees in computer science, they're often you know, terrible for the investors and Board of Directors of the client company, because they can most often produce disappointing results. And then they can be frustrating in the whole IT team, because the results are often frustratingly fragile. You know, it's part of my training as an engineer, that I like to solve technical problems, but I like to build it, and then have a result that I can be proud of, you know, ETL processes are, are just not that. They're just often these long slogs that cost a lot of money with large teams that often produce these results that need to then be redone. As soon as the world changes, you add in the idea that companies may not just acquire or divest of assets, but you add in the the idea that technology is evolving all the time, and that I might store some of my data on SAP and other an Oracle, and any number of other cloud infrastructure combinations, and it just becomes a mess. And it's no fun. The infrastructure that supports AI is often ignored. You call it data engineering, because these terms are going to change around a bit. But in that data engineering infrastructure, the jobs are bad, they're unfair, unpleasant, and they are limiting what the AI researchers can accomplish in the AI professionals in a commercial expression can accomplish, it's helpful for us to always remember, you know, what's the point of us using these learning algorithms? You know, what's the point of our deployment of these systems, the collection of data, the processing of data, the execution of a decision, and then learning from it? What's the point of these things, you know, the point of these things is to make everyone's lives easier, so that we really don't have to think about it much, you know, we kind of put that sort of stuff in the background a little bit, it should make our lives kind of less, less stressful, more enjoyable, let us do things that are more human that we really want to do. That all gets lost when the collection of the data itself comes into question, or the collection of data itself becomes a burden either being a pain or unpleasant, or become costly, untrap, fragile or unpredictable. So this
Rob Stevenson 7:59
is the chief problem you are addressing with your current company connects us. What is that outcome? You said you are outcome oriented? You like that payoff of solving the problem? When you think of like big picture, what problems that you're hoping to solve with Connexus is implementation. What does that look like?
Eric Daimler 8:16
What connects us is working to do is solve the problem of compositionality as an abstraction compositionality is is a wave that's going to take over the next 10 to 20 years of IT infrastructure, that the term compositionality that connects us addresses is often confused with modularity. So people can often think of modularity as Lego blocks. This gets confused in in the discourse quite often. You know, Gartner actually messes this up. Gartner is not not a good source of truth in this regard, because they did did a report on compositionality. And really, they could have just put it in the headline. compositionality is modularity oops, it just made that mistake. compositionality is fundamentally different. In that modularity, the Lego blocks plug and play is analogous to a train where you can swap in and out boxcars as often as you'd like. They're generally interchangeable. But you can only make a train so large couple of miles, say, but compositionality says that I can do that, regardless of the state, essentially meaning scale. So whether the train is 100 meters long, or a kilometer long, it doesn't matter. I can keep adding. In that analogy. The train would be modular, the train system is compositional. It's that sort of infinite building block expression. That is the future of our world and you see some beginnings of this. You just don't read about them because they're not terribly sexy to talk about. One example is smart contract. Next, you know Aetherium and all Ethereum is competitors on the blockchain, those actually would not be possible without the level of mathematics that expresses compositionality type theory, categorical logic categorical theory, those branches of mathematics are demanded to even enable smart contracts to exist. You know, another example is quantum computers, it's certainly sexy to talk about qubits. And everything else around quantum computers that I'm just fascinated about that stuff is as anyone, those wouldn't be able to exist without that same level of math that specifically, you can't prove the efficacy of a quantum compiler. That's how it would happen. You can't prove the efficacy of a component, quantum compiler without type theory without category theory. So that's a way in which quantum computers wouldn't exist. Without this level of math, that's compositionality, you actually need this level of math. To prove whether or not these compositional systems are doing what they say are as effective, are as stable as intended, you have to have them because we can't interpret quantum compilers, we can't program in any other way. smart contracts. You can even say Minecraft is compositional, because it's infinitely self referential. It's this whole other universe that expands in infinite ways, in every dimension. So in some ways, Minecraft is compositional. And in that sense, you'll begin to see compositional systems beginning to emerge all over there, they are not themselves infinitely compositional. But there they are beginning to be compositional. So the media ecosystem is an example of this where it's completely self referential. And you can begin to add in little parts to it distinguished from modular is that it can be infinitely growing in multiple dimensions, those little parts will become automated over time. And then it'll become truly compositional, like a smart contract, like quantum computer compiler, or perhaps, Minecraft, if that's either the result or you know, some intermediate step. That's the world is is compositionality. That's where we're going. And the math that supports that type theory, categorical theory, categorical logic, that's really going to sweep away everything underneath it infrastructure over the next decade or so that's where the world is going. That's what connects us addresses. We are not the only ones doing it. As I'm, as I'm describing, we just happen to be connected as happens to be the leaders in the enterprise software expression of this discovery and math, specifically, that in bringing together databases, so connects us brings in databases in a way that makes them useful for colleagues doing executing on AI projects.
Rob Stevenson 12:52
I'm really interested in the mathematician lair of the AI practitioner, it's kind of a handful of times and if compositionality is the future of the space in a lot of ways, as you say it is at least in the IT infrastructure area. It sounds like having a grasp on some of these more advanced even like, frankly, compared to geometry, for example, newer math is an important stack in the AI practitioners skill set right and important toolkit in the tool belt. Would you agree? Do you think having a deep understanding of some of these new and more advanced branches of math is crucial to achieving AI mastery?
Eric Daimler 13:31
is something I can talk about in my career around AI is that I balanced? Trying to find out what was interesting to me with how I'm growing as a professional, what skills I need to acquire with the changes in the world, you know, what is going on in the world that is now demanding maybe a new composition of skills. So to your question, generally about what is the future of, of AI skill set, it's difficult to predict in some respects, I want to be careful about how how much prediction I give for that I couldn't have predicted 10 years ago, 20 years ago, maybe 20 years ago, the the increasing importance of statistics and probability that we might we might have today, I couldn't really have predicted the decline of geometry, trigonometry, even calculus, and its relevance. So if you might say the more math the better. But you frame this out pretty well Rob, which is categorical logic, it type theory categorical algebra. If I was to choose I would I would replace calculus, trigonometry and geometry with category theory and statistics and probability. They're more relevant math to the 21st century the more relevant math to a digital age. The place where you see the math of continuity calculus is still to be found in applications such as mechanical engineering. aerospace engineering, but really just less appropriate to the applications we see today. But a relational databases were built on relational algebra coming out in 1872. That's just not great for us in 2022, when we're talking about trillions of attributes on trillions of data sources, each of which are expanding quadratically, year over year, it's, it's, it's a really different world, you have to be thinking at a level of abstraction, that a couple of million and maybe even a billion data points didn't quite demand at the trillions of data, trillions data sources, each growing quadratically, what we need is category theory.
Rob Stevenson 15:45
I like how you framed your approach, as you know, the center of the Venn diagram between what I'm really interested in, and what is sort of in demand are coming about in the marketplace currently. For you, then the middle of that Venn diagram, is that the disruption to the IT infrastructure that's happening with compositionality is that sort of the sweet spot between what you're really interested in and what you think are the problems of today and tomorrow,
Eric Daimler 16:10
I'm working with my two fantastic co founders, on the promise of compositionality is what we think is in the future for us. And for a greater IT infrastructure, we think that part is inevitable, there's really no other pathway we can envision, that somehow limits the impact of compositionality, and the math that supports compositionality. That's why I'm here. That's why my co founders here, that's why our team really gets excited about the future that we are creating. We think this is really foundational, the timing at which this is going to impact people in the that aren't experts in the art, who knows it or whether it ever enters common nomenclature, who knows. But this will be foundational for all aspects of, of IT infrastructure, calculus is going to become in relational algebra, all that it's going to become a little bit like Latin, it'll still be interesting. It'll be useful that you know, people don't really talk about Latin much, except when they talk about their bonus. I think that some of that math, it's just going to become less and less relevant, we need to prove compositionality. And the reason we need to prove compositional systems is because the world is becoming more and more complex, how it's becoming more complex, is that people are trying to add in their own knowledge that you know, the future I am super excited about for the average person in an organization, because very soon, they will be able to every one of every one of the people that find themselves in a large organization will begin will have new sets of tools to begin be able to express business rules will be able to express their own knowledge will be MP able to express their own subject matter expertise inside of a framework that can then be applied or removed as the needs require to be part of better business decisions. This is impossible that it you know, many jobs are currently just a constant series of rather low level decisions, you know, they're not actually taking advantage of, of people's expertise, because they're often just too, too nuanced. And the training, for most of us hasn't required a level of thinking about exactly what we're doing or exactly what we want to have happen. That's exciting, and a little scary, because that'll be the retooling of skills over the next decade, is that all of us will benefit from thinking more carefully about what we want to have happen. And exactly how you know the example in AI is is that classic problem about the the automated car coming across the crosswalk, they notice a stop, slow down, keep going when they when it's a shadow, is it a person? Is it a tumbleweed? You know, what percentage of the probability is the is the system comfortable with that decision? Ultimately, as a society, we will need to make a decision around the liability for that decision to just sit with a programmer to sit with a manufacturer doesn't sit with the driver. That level of specificity has yet to enter public discourse broadly, that type of clarity of decision is going to show up more and more throughout large enterprises. So when my company Connexus was looking during the pandemic for a client, they asked where is my personal protective equipment, you know, where's my PPE? And it was funny to think that you know, these companies with 100,000 employees and that and 1000 chips each one of our Just 10,000 containers on it, that they couldn't immediately find out, Oh, hey, it's in the Baja ocean, and it needs to go to Houston or Hanoi or whatever, you know, these sort of decisions took a couple of days, there was some friction involved in these decisions. And that's because there hasn't been the systems to capture all of that knowledge in a way that's then easy to integrate with other systems, that then could engender a business decision upon which the organization could execute and allocate resources, you know, often those are just done in Excel, you know, the, the, the list of, of items that a particular container on a particular ship and a particular fleet, that that's evolving, but I'm describing the way in which it's evolving in the dynamic that supports that evolution, which is not merely the digitization of data, but also the reliable guaranteed integrity of the semantics, integration of that knowledge that's acquired across a large organization. That's the future. And that's, that's going to be an exciting place for people comfortable with clear thinking and clear decision making to be working over the next decade.
Rob Stevenson 21:16
When you say the retooling of skill sets. Do you mean allowing workers to focus on high leverage activities? Things that are more judgment calls, things that only a human can do and automating sort of like low leverage decision making? Is that what what the retooling is,
Eric Daimler 21:31
there's a often a discussion about humans being able to work on their own empathy. And that's true, the degree to which people are effective in collaboration with others, or in discussion with others. That is definitely a skill set that other people have talked about. And I would encourage, the type of human attributes that are just difficult, if not impossible to replicate in a computer are absolutely the future, I am describing a different part of the work equation, where we are all part of large organizations. And we need to think about how we want work to be expressed inside of these large organizations, we're working Connexus is working with a large energy company, they have 10s of 1000s of employees, these employees have particular functions, to maintain the safety of the discovery and the distribution of energy. Broadly, there is certainly something to be said about each one of the people inside of Connexus his clients expressing empathy with each other in order to be more effective. But there is a day to day responsibility over the next decades to distinguish exactly what it is in their jobs that they need to formalize or you'd say automate it or just to distinguish, in order to be able to work with more reliability to a broader system. It what's often missing in in artificial intelligence deployments, is this sort of appreciation that the current state of the technology exploring the subset of ml or just deep ml and, or deep learning and, and neural networks, is has become a sort of religious argument that, you know, people are going down that that hole, because it's sexy and interesting to be sure, but but also just because of the momentum over the past decade from these often not terribly practical applications, and that there is a lot to be gained from going back up the stack towards more basic machine learning techniques, and even deterministic AI and the other, the other often neglected subset of artificial intelligence. I am personally of the belief that many of the most sophisticated applications of AI over the next decade will use a combination of probabilistic and deterministic AI. I think at the board level, it is very important to at least have that awareness that this is a possibility. And the reason is manifold. But I'll take just one is that when you're talking about probabilistic outcomes in ML in neural networks, you're talking about systems that are not really by their nature, 100% fault tolerant if you have probabilities, that means, by definition, there is some probability of failure built in there, which is fine if what you're talking about is applications in digital advertising. But if you have contexts that are life and death, or other similarly catastrophic events for an organization, relying on probabilistic outcomes is often just not where or you're going to need to end up the testing and repeated testing of automated vehicles is a is a good use case here that you see outside my office, you see UK's cars driving by every couple of minutes with way mo cars and in pursuit, I guess that those will not there's an argument to be said, will not be satisfying to safety authorities or to a driver at even a trillion miles of testing. And synthetic data creation, for which there's some great companies being built and commercializing results is not the final solution to that problem. These are all possible solutions, they could be pathways to solutions. But you think that commercial airline manufacturers have done meant much of this before. And yet, airplanes fall from the sky, you know, people die. And the way to solve that problem, when you're talking these life and death contexts for commercial airplane manufacturers, or in energy exploration, where the consequences of failure can be disastrous, is to bring together the sensibilities of probabilistic AI and deterministic AI. It's in deterministic AI, that we can guarantee integrity of the semantics, the integrity of meaning across a heterogeneous data infrastructure. That's what I've seen, what I've seen at Connexus is that we find value our clients find value, us being able to prove that this harder knowledge across a large organization is preserved. As it moves across an organization, you know, what's worse than to have a lot of effort be put into capturing the knowledge and the wisdom on top of the data that's collected, but then not be able to trust that that's going to be not misinterpreted, as it moves to another context with the organization. That's just a terrible waste of will say organizational energy.
Rob Stevenson 27:03
When you say that the public, the market wouldn't be satisfied with, for example, a trillion miles of of data, synthetic or otherwise. Is this like a marketing problem? Is this a messaging problem? Or do you think in the case where AI is being used to make decisions that are life or death? It needs to be perfect? Or does it just need to be marginally better than a human in the same scenario?
Eric Daimler 27:28
This is a great question, Robert, this gets to my motivations for even being in this business. The issue about whether something's a marketing problem, I think we might characterize differently, because us humans are not really rational in general, often, and we're we are especially bad around probabilities. So if we present to the general public, today, I got a automated vehicle here that is 95% accurate, plus or minus 7%. You know, they'll say, killer robot. But then you think, Well, hey, do you know what your own efficacy is, as a driver, you know, people will completely miss that point. And even if you tell them their own numbers of probabilities of getting injured in a car, you know, per per mile or per drive, you know, they don't care,
Rob Stevenson 28:21
they don't care, or they're like, I'm a safe driver, it wouldn't happen to me. Yeah, exactly.
Eric Daimler 28:25
Exactly. So for a variety of reasons, I don't think this is so easily dismissed as merely a marketing problem, we have to and my motivation is to communicate with the public in a way that gets them involved in the conversation. So it's not just selling them, it's having them be engaged, you know, the next decade is going to involve a lot more people doing traditional AI ish jobs. You know, something that happened when I was in the federal government is that people were held holding some technical leader leadership positions, did not have technical degrees. And, you know, some people might be disappointed to hear this, but it's really the nature of the world. We're going to be living in his people, smart people, well, meaning people are going to are going to be making judgments across technical teams across technical infrastructure. That implies that suggests that we need to have a bigger conversation across society about what we want these technologies to look like, you know, for Connexus. What we do and how we express it is connects us make sure that we don't even see metadata, we don't see data. We don't see metadata, we let organizations bring their databases together in a way that guarantees semantics in a way that guarantees meaning. But for many, many organizations, we will need to have a larger ethical conversation, a larger moral conversation about how we want data to be collected, redistributed or merged or discarded. So those are non technical people. Little bit having a conversation in a technical context of AI of FAI systems, if I will, what I want to do is make sure that we have a society that is comfortable with the technologies that are being introduced, such that they get embraced, I want these technologies to be welcomed, not resistant, because the technology often can can save lives, the technologies behind automated cars are being introduced into cars, you and I buy, am I my car has this, these fancy technology is keeping me in the lane, you know, making sure that I don't crash into the car in front of me in traffic, those are fantastic, you know, makes my my drive much, much more pleasant. I want to see those embraced by a large part of the population with enthusiasm. Because they save lives, they also have a little another side benefit, which is they will they will encourage the further development of these life saving technologies, which then has a fantastic virtuous cycle for the maintaining of we'll say Western civilization, you know, relative to Western civilization's competitors, globally, you know, we companies go where the customers are. And if we have Cust customers in western civilization, that that encourages the adoption of this technology, then we will have companies that generate those technologies, and we will have the concomitant benefits from that adoption.
Rob Stevenson 31:34
Interesting. So you view the gap between AI right now and AI being ubiquitous and accepted not as a factor of where the technology is, or because there hasn't been enough of a ad campaign, public service announcement, you know, trying to whip up support for this, you think it's about bringing people into the process and making this a slate of technologies developed, beside the people it serves to build products for,
Eric Daimler 32:03
I want to speak about this technology at the right level to the right audience at the right time, you know, there is a time to be talking about doing queries on Postgres that can import or data stored in Oracle, that's fine. But the general public is going to get tired quite quickly, of our own fascination with our own brilliance. If we've talked too much in the weeds, about how this data gets implemented. It what the general public cares about is, what this technology means to them, what it can provide, what the benefits are, and how they can be sure, they can be sure that the benefits are actually being expressed. There. One of the frustrations I have about AI regulation and proposed legislation is that it is not addressing some really low hanging fruit that can address some concerns for the general public. And this issue can be relevant for technologists developing this infrastructure, you know, what is the engagement of circuit breakers in this technology, you know, the development of some of these automated technologies can be linked to other automated technologies, it doesn't mean that it should be linked just because it can be. And I appreciate that everybody is, is working on their own quarterly incentives or their own project manager, product managers. So really, there is no there is no matter there is no Twitter, there's no Apple, there's just a whole bunch of teams that have their own quarterly objectives, right. But all those teams might just think about how these automated systems can be broken up in individual compartments, you know, so that we can get some human interaction. The good example here is my car, you know, my car is it's driving down the road, it will occasionally asked me to intervene, or says, driver intervention necessary, right? That's a type of circuit breaker. It's essentially saying, Hey, I'm comfortable. But now I'm not comfortable, I was comfortable that I'm not comfortable. So you take over, you know, those sorts of circuit breakers, were just humans are going to touch an automated system is really helpful. Another example about that is that, you know, I like many people during COVID was reassessing our own housing situation. So I was just looking at looking at houses, you know, if I just inputted, oh, I work at Connexus. And then a list of houses appears, I would say, wait, wait, what happened in the middle there? You know, you took my income and my credit score, and then determine what I could afford, you know, I want to I want to break that up a little bit, you know, or, or I go to the doctor with a pain and the and the physician says, Oh, hey, here's a drug, it may have some side effects. Like I want to say, hey, well, what was the chain here? What was the decision point in here? I want to have some agency in that sequence. So the first part is circuit breakers. You know, the second part is an audit, it'll people are gonna need to, we need to find a better word for this because audit oversight can give people a bad feeling sometimes correctly. But the big efficiency that we gained from automated systems might benefit by being managed slightly by experts in the art, still taking a look at the data models, and the degree to which they are doing what's intended. So separate the data from the data model, and then get let experts outside outside individuals that would be appropriate for such things, evaluate the degree to which a all algorithm is doing what it claimed to have done, use a zero knowledge proof if you want, or actually just investigate the actual code of the algorithm. But that's the second part is do a sort of oversight or audit separating the data from the data model. And then the third part that we need to all implement for the regulation of algorithms is lineage and provenance that's guaranteed. So often, when data is normalized, when it's integrated across, we used Postgres or Oracle database, the early earlier example, or any sort of kind of heterogeneous data infrastructure, that the fidelity of the underlying meaning often gets lost? They don't we had one example where we were working with a Connexus was working with a hospital group where one group talked about diabetes, and the outcome is yes, no, another talked about diabetes in the same hospital group diabetes, you know, how are we treating it, you know, I don't just want those to be normalized, and have that fidelity be lost. This happens for us with clients and drug trials, it happens in the logistics companies, I want that fidelity to be guaranteed to be maintained. So that's lineage and provenance. I want to know how that data actually came about. So circuit breakers oversight. And data lineage shows are three ways that I would institute a regulatory regime around AI and right around algorithms that I think will engender trust among the larger population,
Rob Stevenson 36:58
engender trust, and also build tools worthy of this technology, build tools that kind of bring about a, you know, our fair, that are understandable that don't perpetuate biases, and some of the nastier things about human nature from our past. Right. Super, super crucial. Before I let you go, Dr. Daimler, I want to put it on you to tie a nice little bow on this episode. What can AI practitioners do in their daily work in the way they structured their careers to ensure that the way they're using their skill set and the tools they're working on, bring about a better future,
Eric Daimler 37:33
I really admire those that can be self aware enough to look around about how their technologies are implemented. In the real world. I think there are some real opportunities for technologists to engage in the conversation with non technologists about what's possible and what's not possible. You know, there, there are people that may not be experts in the art that with good intentions, have a difficult time looking at a problem as hard. And they might think it's easy, or looking at an easy problem. And they're thinking it's hard. You know, I think it's the job of technologists and those that are studying this field every day to engage so that those that are looking to help deploy this technology in real world environments, can understand what's hard, and what's easy, so that we make the appropriate trade offs. I'm gonna give an example here, where the City Council of New York City voted to implement this requirement to expose the automated algorithms behind resume screening. You know, it's just the most ridiculous legislation because the first of all, the idea that somehow if it was computer automated is offensive. But if it was done by 2000, humans, it's not going to fundamentally misses the point about what this technology does, you know, so we need to engage with others that are scared about these automated tools to bring out exactly what are the concerns exactly what do you what have happened? Exactly? What are you concerned about? You know, bias can show up in a lot of different ways. For example, it can show up in all the ways we know about with datasets, you can also show up in the absence of datasets. So Connexus had a customer that was working on ESG reporting and environmental social good reporting, and we found from our client connects us did that we had a lot more data available in developed countries than developing countries. That's a type of bias that would skew further investment in these technologies. As we report our clients had they had the good intention to report so these are conversations we all need to be part of. We can't just allow for non technologists to accept or reject our technology, the degree to which this technology will be closer to the utopia that we imagined versus that day. dystopia of Hollywood narratives is really up to all of us.
Rob Stevenson 40:03
Dr. Daimler. This has been a fantastic conversation. Thank you for being here and for your time and for sharing your expertise. I've loved chatting with you.
Eric Daimler 40:09
Thanks, Rob. It's been great.
Rob Stevenson 40:15
How AI happens is brought to you by sama. Sama provides accurate data for ambitious AI specializing in image, video and sensor data and rotation and validation for machine learning algorithms in industries such as transportation, retail, ecommerce, media, med tech, robotics and agriculture. More information, head to summit.com