Jennie Brooks, host of the Unstoppable Together podcast chats with Executive Vice President John Larson, who leads the strategy, business development, and delivery of Booz Allen's artificial intelligence and machine learning capabilities. Tune in as they discuss what AI is and is not, the imperative for a diverse AI workforce, and John's thoughts on bias mitigation techniques. Part 2 will explore AI education and how John's experience with dyslexia led him to a career in STEM.
Jennie Brooks:
Welcome to Booz Allen Hamilton's Unstoppable Together podcast, a series of stories that unite us and empower each of us to change the world. I'm Jennie Brooks with Booz Allen Hamilton, and I'm passionate about diversity, equity and inclusion. Please join me in conversation with the diverse group of thought leaders to explore what makes them and all of us unstoppable.
Hello everyone. Welcome to the Unstoppable Together Podcast. I'm your host, Jennie Brooks, and today I'm excited to be joined by Booz Allen, executive vice president John Larson. John leads strategy, business development and delivery of the firm's artificial intelligence and machine learning capabilities. John, welcome to Unstoppable Together.
John Larson:
Thank you, Jenny. It's really exciting to be here and looking forward to our conversation today.
Jennie Brooks:
I don't doubt it's going to be a pretty rich conversation. John, thanks for joining us. To get us all started, why don't you baseline us on a simple definition of what artificial intelligence is and maybe talk a little bit, demystify, what artificial intelligence?
John Larson:
It's a fascinating question because I think if you had asked that question even six months ago, the answer may have been slightly different. We have entered a whole new era with what is now called generative AI in ChatGPT. That is, I think, really transformed AI in a way that we haven't seen in the past. Let's start with maybe the basics. First of all, AI has been around since, literally the 1950s. The term came into vogue in the 1950s, and we've had periods of hype and expectations and then alternating periods of what I would call setbacks, troughs of disillusionment. And so it's been a science or a field for some time at its most basic level. It's really the theory of the development of computer systems that are able to perform tasks that normally require human intelligence. And the examples of these tasks can include things like visual perceptions, speech recognition, decision making under some degrees of uncertainty, learning, translation, things like that.
And so it's been around for a while. It tends to cover areas of computer vision, machine learning techniques, where the machines can learn and self-improve over time and natural language. And so that's the science, it's a broad moniker. Within that is this machine learning science where these machines get smarter and smarter. And then within that overall view, there's this deep learning where these algorithms, in probably the last seven to 10 years, have really developed a specialized collection of machine learning and approaches that allow you to build complex, convoluted recumbent neural networks that allow you to solve very, very difficult problems. And so they become very good. These models, these techniques become very good at giving the ability of machines to perceive, learn and understand. And they're able to do this in a way that I think, again, even in the last six months has changed dramatically with ChatGPT.
I think the ability for the algorithms to learn from these large language models and generate just remarkable content, that has really transformed our perspective of where AI is. So I think that's where we are. It's obviously been unlocked predominantly by three components. The rise of data. The more data you have to inform these algorithms, the better they are. The rise of compute, the complexity of the compute problem around this, it takes massive compute to generate these types of algorithms and create these insights that allow you to apply them to problem sets. And so the rise of both the CPU and Moore's Law as well as the creation of GPUs, which really have their foundation in gaming machines, allows you to solve these mathematical formulas more efficiently than the CPUs do. And so the rise of the GPUs and the compute power there has really transformed it. And then lastly, some of these novel algorithms around deep learning have changed a lot as well in the last seven to 10 years. So that's been the big transformation.
What is AI not? I think it's got to be said, it's not interchangeable with the term of data science. It is broader than data science. It's not able to perform generalized tasks that humans do. So there's this notion of the big demarcation is what's called narrow AI and generalized AI. And generalized AI is a machine that can do almost anything. Narrow AI is applied to very narrow tasks and applications. And so even in the chat sheet, BT world, very good at some things around language not as good when you ask it math questions. And so I think it's also important to say it's not what you see in the movies. It's not how from 2001 Space Odyssey, it's not SkyNet and the Terminator. These are not sentient beings. They don't have a conscience, they don't have intent. And so there's a tendency to anthropomorphize things. And it is important to not do that with these algorithms. These are machines that have been trained with parameters that humans have created and the data that informs them, and it's that simple.
Jennie Brooks:
I want to go back to something you said right out the gate. The technology's been around since the 1950s. You mentioned troughs of disillusionment along the way and then you fast-forwarded us to today and these three drivers that are really accelerating the technology. How do you think about the troughs of disillusionment today, given where the capability is?
John Larson:
I think it's interesting because it is very informative. If you look back in time, there have been periods where I think the constraints of the compute power, the availability of data, have limited the ability to achieve the perception of where you wanted to go as a scientist in this field. I think that there are going to be periods of disillusionment that will still follow. Right now, there's a perception that ChatGPT is the panacea for everything. Almost any conversation you have, there's like, "How am I going to use ChatGPT to solve the problem?" And so I do think there's a hype cycle right now that we are going to come off of. We are on an endorphin high. We've seen the promised land and it is really amazing. But as you start to scratch the surface and look under the hood, you're seeing shortcomings and deficiencies. And so this is going to create, I think, a second trough of disillusionment where people will understand the limitations.
Even though it's been an amazing transformative event, you're going to start to see some of the shortcomings of it. And that's going to, I think, be important because it's going to rationalize everything. I think there is probably right now a period of irrational exuberance about what we're going to be able to do. I think there's unbounded opportunity, but I do think there is some need for perspective on how that opportunity will be perceived and realized as we go forward.
Jennie Brooks:
Speaking of perspective, on this podcast, we talk a lot about topics that span diversity, equity and inclusion. Talk us a little bit about AI in terms of DE&I work.
John Larson:
This is one of the areas where I have been really passionate about. This is something I think is very important, and there's several reasons why I think the application of AI and the Nexus with DEI is so important. The first one is, I think about as from the inputs to this AI opportunity that we see ahead of us. The NSCAI report, which was published, Jenny, three years ago this month, had a call to action for us as a country. It recognized the important role that AI was going to play in the future of this country, in the global economy and in the great power competition. And it said, as a nation, we needed to rise to the challenge. And it made a call for an all government approach to that.And I think that's very important because what it means is we need to bring talent to this problem, and we have a talent shortage. And why this impacts the DEI issue, is because there are swaths of our society that we are going to risk leaving behind if we don't proactively engage them and bring them along this AI journey. And so I am very, very passionate about this desire to reach into underserved communities and underrepresented communities and ensure that they have the skills and the opportunity to enjoy this amazing transformation in technology. This is the single most transformative thing, probably of a generation, if not since electricity. And if we don't ensure that everyone in this country at least understands it, knows how to be a consumer and user of it, but even better is a practitioner of it and can help us grow and develop it, we risk leaving portions of our society behind.
And that is going to create incredible hardship. It's going to put us as a nation at a disadvantage, and it's not going to allow us to, I think, bring about the AI we want that represents our democratic values. So that's one piece of the puzzle. The other piece of the puzzle, is that to bring these systems to bear and to be able to have the representation in these systems of all different views, we want to bring diverse teams. So my background as an economist, that's how I grew up. And if I ask a room full of economists to solve a problem, I can tell you exactly how that problem's going to be solved. It's going to be a lot richer if I get a lot of diverse backgrounds in there and solve the problem. I like to think of a problem as a globe, and if you don't have perspectives around it, you only see one side of it.
And so I really love this idea of enriching the talent base by creating a diverse opportunity for all so that you get a different perspective as you train and build these models that represents all of our society and not just the subset of our society. So I think that's really important. And there's some great examples of this, like facial recognition. Really important technique in computer vision. Nobody who was developing these algorithms back in 2012, '13, '14 thought, "Hey, if I trained this on public data, looking at internet images, I may end up training this model to be really good at white males, but no one else." And that is exactly what happened. And I think that if you had different people in that process training the model, you would've had people go, "Wait a second, there's a problem here. We only have a subset of the population represented, and we need to make sure that we're representing everyone."
And so that's a really good example of how in the past the developers of this, maybe didn't understand or perceive the risks. And by creating diverse and inclusive populations that are part of this solutioning, we're going to ensure we mitigate those types of things in the future.
Jennie Brooks:
So we're going to bring diverse representation into development of the use cases, the capability. I think we think, "Well, technology is inherently unbiased," but that's not accurate, right?
John Larson:
It's really interesting. I think the way I would characterize it, Jenny, is that technology represents the world around it in many ways, and these AI algorithms essentially reflect the data that you inform them with, and the parameters with which you prescribe to them on how they're going to learn. And so when you look at a system that has had inherent biases in the past, really good example is the mortgage lending. There have been discriminatory practices and mortgage lending in the past, and what that has done, is that meant that minorities were not entitled to mortgages at a rate that was commiserate with non-minority populations. If I take that data and put it into a model and train a model on it, it will start to entitle or approve mortgages for non-minorities at a higher rate than minorities because the bias is already baked in the data.
And so that's a really good example where understanding your underlying bias in the data, you will now understand how that bias will be perpetuated and amplified in an algorithm as it goes out to performance activity. That algorithm isn't bias. It just learned from the existing data, and so it's important that we train and hone our talent to understand those types of systemic biases and data that exist and think about how we mitigate them. The one thing I like to talk a lot about is we cannot eliminate bias. Bias exists in everything, but if we understand it, we can perceive it, we can measure it, we can do our best to mitigate it. And there's mitigation techniques from response bias sampling where we can over sample certain populations and correct things in the data.
Or there's other techniques that we can use where we might bring synthetic data to the problem set and say, "Look, the data doesn't exist, but I can make a population that's more representative by generating synthetic data, inserting it in and now I can create a model on this data because the data represents more of what the world should look like than what the world perhaps looks like because of those biases that exist."
Jennie Brooks:
What does broad governance look like today with where we are right now? You read in the media that various leaders are calling for a pause. There's a robust discussion around ethical AI. What does a snapshot of today look like?
John Larson:
I would characterize it as diffuse, somewhat cloudy and unclear. I think there is a desire to figure out how to regulate this space. I think we need to be very cautious about how we go about doing that. And what I mean by that is I think there's a tendency to rush in and apply regulations, where perhaps there's sufficient regulations in place. Let's go back to the mortgage example that I gave. It's a great example where there's a lot of regulations that have been put in place to mitigate these types of biases that have historically perpetuated themselves through the mortgage lending process. And so those regulations exist. There's a lot of different agencies that have prescribed those regulations, and financial institutions are required to follow those. What we need to do with AI is say, "Okay, we want to engineer models that reflect those regulatory requirements in place today.
And so it becomes, as you said, a question of governance and guide rails. And so now we can take those regulations, and this is something that we're doing right now as we look at some of our venture capital investment partners like Credo AI, who we announced an investment in about a month ago. We're looking at how do we take those regulatory frameworks already in place and how do we create policy packs around those that prescribe the conditions that you must meet from a regulatory perspective and govern the AI learning? And this is the fundamental difference. Data science and modeling has been around for a long time. The fundamental difference now is that these machines can learn without any guidance from humans. And so what it means is when you set these machines up, in the old world, you built a static model and you prescribe the features of that model, you would say, "I'm going to use these features to define the problem I'm trying to model, and I generate the model and then I apply the model."
Today, the power of the machine, of the AI, is that it can learn. It can look at the data, it can figure out the features it wants to use to build and prescribe the solution to the problem. And then as it applies itself, it's looking at the residuals, the errors that comes back and it's reinforcing itself and saying, "Ooh, I want to fix that error. I'm going to go find a feature that helps me reduce that error," and it's going to bring it into its algorithm and it's going to self-learn. And what we need to do is we need to ensure that that learning doesn't allow that model to figure out that a feature like race or zip code, which is highly correlated with race, may be relevant to reducing error in some of its predictions. And so we have to put the governance and guide rails around the learning to say, "We are going to throttle the way it learns to ensure that it learns in a way that comports with the regulatory guidelines and requirements."
And so I think the preference right now is to allow those guidelines and governance to reflect existing regulatory frameworks, because fundamentally, the algorithm has not changed the representation of bias that's out there. It just needs to reflect those regulations. And then if we learn things through the application that maybe introduce new problems, we can ask ourselves, "How might the regulation need to evolve?" But I don't think there's a need to go out and immediately regulate differently. It's simply how do we bridge the gap between what's already been established and the new techniques of autonomous learning to ensure that we don't end up violating what already has been defined as not permitted within regulations.
Jennie Brooks:
This wraps up part one of our conversation with John Larson. Be sure to join us again for part two where he addresses AI education and how his experience with dyslexia led him to a career in STEM.
Thanks for listening. Visit careers.boozallen.com to learn how you can be unstoppable with Booz Allen. Be the future. Work with us. The world can't wait.