Check out this episode for fresh insights on AI from Kathryn Hume, Director of Business Development at Borealis AI. Tune in for her take on the AI ecosystem, how the media talks about AI, her talk “My Daughter’s Life on Artificial Intelligence”, and much more.
CEO and Co-Founder
Director of Business Development
Rob May: Hi, everyone, and welcome to the latest episode of AI At Work. I’m Rob May, Co-founder and CEO at Talla, and I’m your host today. And with me is Kathryn Hume. I’ll let her give a little bit about her background, but Kathryn and I met when she was Fast Forward Labs and just instantly started bonding over AI and where it was going and just being new in that space.
She’s had a couple of different AI related roles since then but is one of my favorite thinkers on AI and where it’s going. So Kathryn, welcome to the podcast today, and why don’t you tell us a little bit about your AI path and how you got here.
Kathryn Hume: Thanks so much for having me, Rob. I totally concur, you were one of our first customers at Fast Forward Labs, and I remember we were early on in the journey of creating the company, and it felt like you were a partner in our getting our act together around what we really wanted Fast Forward Labs to be as well. So I was always grateful for the insights that you gave.
I started working in the field of artificial intelligence in 2015. And my background is pretty unorthodox for the field. So I was trained as a theoretical mathematician, which is a little bit of a different way of approaching problems from a statistician. And then I have my doctorate in comparative literature with a focus on 17th century science and philosophy.
You’d say, how the hell did this person end up working in AI? But I actually find that the combination of both really having a sensibility for and understanding of the math plus thinking about how cultures and societies adapt to new scientific and technological changes which was actually my– when I say comparative literature, you might say, that sounds like it’s completely unrelated. But my focus of study was on how cultures adapt to new science, and I’m finding that the combination of the two has actually led to a really powerful combination and background to help think about applying AI in the enterprise.
In my professional career, I focus pretty much on B2B startups, and I’ve been as you mentioned working at Fast Forward Lab, just as a research lab helping Fortune 500 companies understand and adopt artificial intelligence. Followed by that, I worked for a startup that was working with consumer enterprises and helping integrate AI into their consumer marketing practices. And then recently joined Borealis AI, which is the research lab for the Royal Bank of Canada.
We have a pretty large group of researchers and engineers, and we’re looking to build transformative products for the bank.
RM: Let’s start by talking a little bit about some of the things that you’ve learned as somebody who didn’t come up a traditional AI path. There’s a lot about AI that’s very different. You’ve worked in business development, you’ve worked in product management. Is there anything that you’ve seen that you think, like advice that you would give for people for like, hey, when you’re dealing with AI, here’s how it’s different than just dealing with, say, software or maybe from a process perspective or a team perspective or anything like that?
KH: I think by this point, the software and tech community has generally moved from the traditional waterfall approach to managing products to the Agile Scrum oriented approach. In my mind, the difference boils down to the way in which we think about managing risk, and the way in which we constrain team oriented activities to deliver some particular feature and how we go about adapting those features along the way. And if you think about it, when you’re developing a standard software system, most of the time we know how long it takes to build a particular feature.
A great software architect will sort of sit down and be like, all right, this thing’s going to take this long, this thing’s going to take this long, and then they can manage the various tasks on a team according to the complexity of the subsystems that they’re building. When we shift into building AI products, there’s a lot more uncertainty in timelines and even the risk of being able to build a feature in the first place.
When we say AI, we’re basically in my mind this can be reduced down to products that are built using some sort of mathematical model on data. When we start off and as a product manager or business development person, the business will come in and they’ll say, all right, we’ve identified a business problem where building some sort of system using data could have impact. That will be stated as not necessarily a technical problem statement but more like a conceptual possible statement.
The first task needs to be resolving that down to a technically scoped specified problem. That exists in software, all software development protocol. In machine learning in particular, you have to sort of define a target variable or something you’re going to optimize for to pass that then over to the modeling team so they can dig into the data and see which algorithm might be able to meet some performance requirements.
It doesn’t always work, and it doesn’t often work. You have to sort of stage out room for iteration and failure in a way that I think goes beyond the traditional process mechanisms of Agile Management. It’s something that I can say certainly traditional enterprises just aren’t really accustomed to dealing with yet.
RM: What have you seen in terms of data science or machine learning teams and product teams and engineering teams trying to work together? Has that been a struggle at the companies you’ve been at, or have you found ways to bridge those divides?
KH: Yes, I’ve seen a standard evolution in just about every company that I’ve worked for where you’ll start off with maybe a heavy machine learning or data science team. In terms of the composition of people on the team, there’ll be a lot of data scientists. This is sort of companies that are AI first, so they’re built from the ground up to be building ML products.
Then you quickly realize that they’re building out often prototype models in Python or some other back end package that they’re using that are quite different from hardened modular scalable code that can be used in production to actually get these systems working once we use them in production. Then, we have to couple with the engineers, and there’ll be initially this sort of point of friction, where the engineers will come in and be like, oh my god, what are we supposed to do with this?
There’s a lot of work to take the sort of somewhat chaotic processes that often exist in the science community and then put them into a rigorous software development process. I’ve often seen that this mix and balance will evolve over time, and you’ll end up having a really different ratio. Say if it starts off with like five data scientists to one engineer, it will flip to maybe 10 engineers to one data scientist by the time the company reaches a little bit more maturity.
You’ll also see a push and an evolution in the kinds of skills that machine learning scientists need to actually be able to write code, functioning code that can be put into production. That doesn’t change the workflow, where you’d start off by training a model, testing it in simulation, and doing a lot of the statistical practices to make sure that the math is working like we’d like it to.
Then, that moves into the production system. There’s all sorts of questions that need to be asked regarding how frequently the model needs to be retrained, how frequently new data is coming in, where that data is going to be pulled from, how we’re going to manage the integration pipelines. Then even on the front end product side, how we transform the output of some probabilistic statistical model into something that makes sense for humans.
There’s all sorts of tricky different decisions that go into making a product really useful for the end user.
RM: Now, as you’ve looked at a lot of different companies over time and been at a couple of companies in the AI space, what have you learned or what opinions do you have now on the types of business models and techniques that are working? Because I feel like there was this broad explosion of stuff that people started trying like there is when any new technology really hits, sort of the mass entrepreneurial opportunity.
It seems like we’ve winnowed down now to a couple of things that seem to be working. What’s your take on that?
KH: It’s a good question. I think there’s a couple of ways to parse it. The one is, if we think about what is this whole AI revolution in the first place, considering that companies have been using statistical models and data science for a pretty long time? There’s one way of interpreting that and saying, what’s really happened is that deep learning is now moved from the research lab.
It’s existed, the algorithms have been around since the 1950s and ’60s. They sort of had a little bit of a technical and theoretical shift in the ’80s, and then suddenly in 2012, they’re working and people are paying attention to them. I think the biggest impact from that shift has been in our ability to build products that are automating processes that have images, video, speech, and text.
You’re not going to use a very large complex deep learning models, which I’d like to think about these things as universal function approximator. So as opposed to having humans come in and having to hand code which variables are tightly correlated, which outputs we want in some sort of system, you sort of throw a lot of really complex hierarchical data into the pot, like an image which has maybe 256 by 256 pixels, and then these models can make sense of it.
Computer vision really has been, I’d say the golden child of the last couple of years, where we’re seeing commodity scalable tools. You see the libraries coming out of Google, Amazon, Microsoft, et cetera to process images, and then that’s impacting the fashion industry, because there’s lots of information online in fashion that comes in the form of images. The marketing industries, and maybe some things that have to do with video.
It’s similar for unstructured text. This might happen in the compliance departments of large enterprises that are processing all sorts of texts that has a lot of similarities to it and some structure to it. You’ll see it in applications like due diligence and contract analysis in law firms or in sort of compliance departments in large companies. That’s one way of thinking about it.
I think, it’s not to say that computer vision is now a solved problem. There’s still lots of lots of progress to be made there, but it really has changed over the last couple of years. The other way to think about it, and I think this is where it really goes from a technical capability to having business impact, is to sort of parse up these systems into tools that can predict things, classify things, and optimize things.
Those could be business processes, those could be a part of a task or process, and the product work goes into really defining exactly what aspect, what phenomenon at work we’re looking to parse in some way. But if we take the example of marketing automation, in the non machine learning world, a lot of the efficiency gains came from the ability to do things that scale very cheaply.
So, mass email marketing campaigns, where we could send a similar kind of message out to many, many, many customers basically for free, so super high margins vis a vis direct mail where there was a lot more cost associated with that outreach. When we move into the ML oriented approach to the same process, we can optimize the economics of outreach much more elegantly, where we can take a customer, probabilistically make a guess of that person’s likelihood to act, likelihood for their behavior to be changed given some incentive, likelihood to acquire however many products a company might have over the history of their lifetime at the business, and then a business can make optimization choices to say, if we only have so many resources, we can go beyond just the scale of high margin email, and we can think about really targeting high value, high cost resources like people’s time, like discounts, like offers to those people that our models probabilistically indicate are going to be likely good targets for that.
That shift in the economics of outreach in some way I think is something that really has settled in as a process impact in the enterprise of machine learning systems over the last couple of years as well.
RM: Cool. Do you feel like there are any major opportunities out there that people are ignoring or aren’t being addressed?
KH: That’s a great question. I think one of the opportunities is just the way in which people frame the AI, the AI opportunity in the first place. So I give a Ted talk back in October, and in preparing for it, I had to sort of– Ted Talks are small, 12 minute segments, so I had to whittle down all of my theses on AI into like one core message that I could deliver. And in doing that work, what I arrived at was that for a CEO, so somebody who’s not technical at all, they don’t need to go in, they don’t need to know the difference between a neural network and a standard regression algorithm.
If they could only know one thing, they should know that the opportunity in running an AI first organization is to shift the mindset around what a business process is from a vehicle to drive standardization and efficiency, which is sort of like the use of software in thinking about efficiency in a business process, to a vehicle to collect unique knowledge about the world that can only be known via the processes that you have and the customers that you have. Basically, all businesses are almost these knowledge receptacles, where based on years, based on what they’ve built, based on their unique expertise and also their customer base, every time they do something, they are creating a reflective data mark of that action that can be reused using algorithms to do something different in the future.
To sort of summarize this, I think the missing opportunity is a conceptual opportunity, and it’s one where business owners need to flip the value of a business process to being a knowledge and data collection mechanism from being a means to drive efficiency. I think if they don’t make that shift, then most of the AI applications in business are just going to be efficiency plays, where we then lead to fear of replacement of the mania around AI taking jobs, because it’s always around sort of automation and optimization as opposed to, how can we use this everything that we know and invert it to create new value in the future?
RM: That’s one of the more interesting things I think people have ever said on this podcast actually, and it reminds me of when James Cham from Bloomberg Beta was on, and he talked about this idea that people might move from doing things to training models. Like, the job that you do needs to become a model in an organization. How do you get there?
That ties back to my favorite Wall Street Journal op ed of the past year was one that was written by Steve Cohen, who basically said, models will run the world. I thought that was pretty interesting, because that’s not how your average business person is thinking about it yet.
KH: Yeah, that’s right. I think part of the reason why they’re not is there is this knowledge gap, and the field has yet to mature to the point where there’s really the distribution of tasks between the high level exec, like what does the CMO need to know, the CEO, the CXO, the sort of director level, and then down to the people that are sort of building the technology.
Now, this is all patchett. This is working with the assumption that these kinds of hierarchies still should and will exist in the future. Given the current situation, they’re certainly still in practice. So then you say, all right, as somebody who knows a lot about their business but can also think through basic common sense first principles work like what aspects of my business change frequently, and what aspects stay the same frequently?
We can all think that way. If we think that way, then that’s actually a really relevant question to ask as it relates to drift in your machine learning models. So when we train our first model, we do it at a moment in time. That moment in time is going to capture a snapshot of relationships between let’s say a process in your customers in the world on the day that you build your model, train your model, and put it into production. And then there’s questions like really tactical, technical questions around, as I mentioned earlier on, how frequently should this thing be retrained?
If it’s the kind of process that doesn’t change all that frequently, you don’t need a real time optimized streaming data model. That’s just not required. You could use something that is re-trained once a month, maybe even once a year. And then there’s other kinds of things where the phenomenon is changing constantly. Think about something like capital markets. There, you really do need a different kind of engineering architecture.
Anybody in the company, anybody in the business and certainly on the executive team can ask those first principles fundamental questions around the cadences of activity in their business, the future projections that they’re trying to think towards. They can also do sort of a design oriented approach and think about where their customers are going to be. Who they are, what they look like, how they’re behaving in the world in five to 10 years, and then working backwards from that to think about what the right toolkit would be to sort of serve those needs.
It’s not always AI, but I think it’s a helpful framing so that it can start to catalyze a conversation between people who are approaching things from a different perspective.
RM: Now on your personal blog, you wrote a blog post recently called My Daughter’s Life on Artificial Intelligence about a talk that you gave which from the notes that you left about the talk seemed to be that some people sort of really enjoyed it, and some people really sort of didn’t get it. Tell us a little bit about what it was about, and why do you think you had the mixed reaction like that?
KH: Great question. Glad you asked it. OK, so the origin of this talk, I actually gave a different version of it back in 2017, when somebody asked me to talk about the future of AI in five years and 10 years. I thought to myself, I can’t stand these, just because you’re an expert in a field, that somehow makes you an oracle for what’s going to happen 10 years out in the future.
A lot of my initial resistance to questions like that come from the fact that I believe predictions on where technology is going to go are often linear in nature. They’re coming from the perspective of somebody within one local slice of the world and their purview of what they can see and what they think about. The truth is, the social reception of new tools that we build is very infrequently predictable and often involves a complex emergent combination of many technologies and not just sort of the one linear stream of, let’s say, AI or machine learning or blockchain or whatever it may be.
Given that irritation, I said, I think all talks about the future are fictional anyway, so why don’t I just embrace that? And instead of giving my pseudo comforting oracles on where things are going to go, I’ll just pretend there’s somebody born on the day I’m giving my talk and think about how that child’s life might be impacted by some of the new technologies that we do know exist today. With a little bit of a hint towards where they might go.
In this talk, I gave it at the MIT Tech Conference in late April. I imagined my future daughter. I don’t have a daughter yet but hopefully will someday soon, and basically what it would be like, what her life would be like and the aspects of childhood that are going to remain the same. So somebody born when I was born in the 1980s, somebody born in the 1950s. There’s certain human essential rites of passage that we all go through no matter where we’re born or when we’re born.
Then, there’s sort of variations on those core rites of passage that might be inflected using algorithms and technology. And I thought it was interesting to think about that. As an example, Clara who’s the heroine of the story, when she is in kindergarten, she thanks to these advances in computer vision goes on a field trip, and she’s able to bring her iPad and focus it on any old tree that she walks by. And immediately, there’s all this augmented reality information about its genus and its species and much more information than Rob, you and I had when we were kids.
I grew up around Boston area, and I had to watch those stupid videos of the Voyage of the Mimi with the young Ben Affleck. I wasn’t using cool technology. What people liked– I think there were a few people in the audience who found this extremely stimulating. One, because it forced them to question some of the standard narratives around social reception of technology that they’re accustomed to hearing.
It also gave them food for thought. It was surprising. They don’t normally hear things like this or hear anecdotes mentioned like this in Ted talks. For other people, I think it didn’t feel all that useful, because it was too imaginative. It wasn’t based on fact. And I think it maybe lacked for them the practical utility they were looking for in helping them make decisions on where they might guide and push their career.
I think that’s for me, the difference in response was interesting. I guess for me, it raises and imposes a large question about the overall reception of certain design practices in companies. Because I think a lot of great product design is actually built using the imagination and the imagination applied to very particular situations and very particular user needs as opposed to some sort of general concept.
Even maybe the value of what the right place for the imagination, the humanities could be in the realm of technology today.
RM: It’s interesting, because I think it highlights the way that AI is right now. I divide the market from a conceptual perspective in terms of just how people think about it into super research-y techie people, and the general public that’s sort of like we don’t know what to make of this. It sounds scary but also sounds cool.
I still feel like even though we’re a handful years into this wave of AI, I still feel like that practical applied wave of understanding this stuff and working with it is still way, way too thin and not what I would expect it to be. There’s all of these articles about who are self-driving cars going to kill given the choice? I feel like there’s all these other different day to day concerns that you should have about AI and the more that it runs your life.
Death by autonomous vehicle is not really going to come up that much. You compound that by the fact that almost all the AI news that you read is actually about people who– it’s research based. it’s StarCraft that’s coming out of Deep Mind and Google Brain and Open AI and all that kind of stuff. And I think there’s a lot of confusion about where is this actually applying to our lives day to day?
So as somebody who writes a lot about this and reads a lot about it and everything else, what do you think about how the media communicates around AI?
KH: Back to the sort of fascination with the big stories, I really like the AlphaGo movie. When DeepMind, which now is acquired by Google, large research lab working on some really cutting edge research in machine learning, they I think very intelligently decided to pick a challenge for their technical problem that could be readily cognizable by the general public as an achievement in some sort of form of intelligence, because it was it was a game.
There’s this great article written by John McCarthy, who was one of the early pioneers in AI about Arthur Samuel. Arthur Samuel in the 1950s built a checkers playing system.
He was actually interested in studying the phenomenon of machines that learn. An early pioneer in the field where learning was specifically defined as the ability for a computer to improve performance on a task after having been presented with new data and experiences in completing that task. That’s sort of this– it’s kind of a boring– it certainly wouldn’t make public attention if we were to be talking about that.
But, he astutely realized that in order to make this extremely abstract mathematical computational thing meaningful and excitement around it, meaningful to the general public, he had to find a metaphor, sort of a convenient means of designing a system that could express these ideas. He chose checkers. Then they staged competitions against the world’s fourth best checker player, I believe.
You see this foundational moment in the history. McCarthy talks about it, and he compares the use of games in machine learning to the use of fruit flies in genetics, saying that fruit flies are cheap and easy to keep. He says that it’s kind of bizarre that geneticists study fruit flies, but there were a great testing ground for revealing larger principles about some abstract science.
There’s something kind of similar in the public perception of AI in that it’s easy for us all of us to think about what qualifies as intelligence, perceive internally as we’re going through our day the acts that we think are more emotional, the acts that we think are more prefrontal cortex and rational, and then the kinds of feats of human activity that are deemed super smart in intelligence by different cultures. So it’s a little bit different than the click bait news cycle around self-driving cars that crash that sort of taps into a different primordial fear impulse.
The narratives around what AI is are compared and contrast to what we think the feats of human intelligence are. In my experience when you’re getting into the real applied side in business, the biggest questions are much more mundane but actually much more sort of tricky and hard to work with. Things like with this particular business process, how comfortable are we if 10% of the time, the system outputs the wrong answer?
When we move from a deterministic software system where we’ve deliberately simplified some complex things in the world down to a clear set of rules that do exactly what we tell them to do every time unless there’s a bug or unless it’s hacked to this more fuzzy, probabilistic world of machine learning where most of the time, it does the thing that we would like it to do but because there is a threshold for error, we try to bound that error. Every once in a while, it might do something we don’t want it to do.
That’s where people just don’t know how to manage that. I think if we were to– while I think there’s a place for the big, grand scale human narratives on man versus machine in the domain of games and even in the domain of scientific research, on the applied side, we really have to have almost a lot more openness and curiosity around the mundane, banal and yet fundamentally important questions that are the make or break questions for getting these systems into production and actually having a business adapt them and transform.
RM: All very excellent points. My last question for you before we wrap up is, as somebody who spends a lot of time in this space and thinks both pragmatically and day to day about building AI products but also more philosophically and long term just sort of by the nature of who you are, where do you fall on the spectrum of– there’s a lot of people that are– there’s sort of three schools of thought on AGI, right?
There’s people who are like, oh my god, we should be super worried because this is happening? And there are people who say, well, maybe we should be worried someday but not now. And there’s people that sort of feel like, eh, it’s all overrated. And I’m interested in your general perspective on that topic.
KH: I was listening recently to a podcast with Alison Gopnik. She’s on Sam Harris, I believe, or he interviewed three of the authors in a new book Josh Brockman put out about AI. I like what she said where she felt like we would be in a better position, discourse around the topic would be better today if we were to just call it large scale statistical inference on big data sets. Like that’s kind of what just happened over the last couple of years versus we’ve actually cracked causal reasoning.
Judea Pearl, he wrote a book recently. He’s sort of been in the causal side. I think I fall a little bit on the side of folks where without being a Chomskian rational essentialist who believes that we all are born hardwired with the ability to have a universal grammar and reason about things, I do think that there just to me seems to be a big gap between the strict inductive inferential power of statistical systems as they are today, and this will come up in sort of deep learning explainability, where we realize that the patterns and correlations that the systems are identifying that seem to enable human like cognition to call a cat a cat or a dog a dog are actually sometimes quite mysterious, and they’re just actually picking up on sort of shadows or something else in these pixels.
For me, the difference between what would be a thinking, reasoning, sentient, emotion experiencing system are quite different from the capabilities that we’re seeing today. It’s hard for me to get worried about the long term existential risk around AGI or even think it’s sort of the most interesting problem to solve today. I think as it relates to the risks, we’d be better positioned to be analyzing and addressing some of the risks that already occur in society today based on less the smarts of AI and its limitations.
As an example, we human software designers still have to choose what target variable we’re going to optimize for when we deploy and implement an AI system. I think, and this is coming from my philosophical background, there’s a lot of epistemology that goes into this, so there’s a lot of finesse around thinking about, all right, the worlds out there.
The mechanisms we put in place to capture data, sensors, measurement mechanisms, only capture a summed subset of the actual sort of reality. Then, we take an algorithm and we transform in some way this raw subset of data, transform it into an output that is then re consumed, and then there might be a feedback loop. There’s a lot of risk inherent at each point in that process.
I think it’s a risk that we haven’t thought enough about, and it would behoove the philosophical community to think more about. My call to action would be as opposed to our sort of speculating on the future of AGI, we should do some digging in and thinking a lot about the nature of time, space, agency, judgment, what we can know and not know, what systems can know and not know, ontologies. Like these for me are the meaty, philosophical questions in AI, and they’re present and observable today.
That’s my take.
RM: That is a great answer to end on. So Kathryn, thanks for being on the program. Thank you all for listening, and if you have questions you’d like us to ask or guests you’d like us to invite onto the podcast, please send those to email@example.com, and we’ll see you next week.
Subscribe to AI at Work on iTunes, Google Play, Stitcher, SoundCloud or Spotify and share with your network! If you have feedback or questions, we’d love to hear from you at firstname.lastname@example.org or tweet at us @talllainc.