Jon Hyde: Hello, and welcome back to The Next Horizon, a Dell Technologies podcast. I’m Jon Hyde, and together we’ll explore the implications of several major emerging technologies for business, society, and most importantly, for you.
Jon Hyde: Welcome to Next Horizon. I’m Jon Hyde. Today I’m joined by Brons Larson, who’s a strategy lead for artificial intelligence and machine learning for John Roese’s Office of the CTO here at Dell Technologies. Brons, welcome. Thanks so much for joining us today.
Brons Larson: Yeah, thank you. I appreciate it.
Brons Larson: So I’m helping set the kind of longer term goal for AI, thinking a little further out, being a little more forward-leaning, and making sure that we can take all the learnings from the last four years on our AI journey, and kind of aggregate it and put some guardrails on it to unify it towards some universal North Stars and a universal vision that can help us leapfrog the market.
Jon Hyde: Yeah, you know, AI seems to be a hot topic today, but it has been around for a long time, as I understand it. It didn’t start with Skynet in the Terminator movies, it started way before that.
Jon Hyde: It seems like there’s this interesting confluence of technologies and economic drivers that have come together to get us to where we are. What can you tell us about that?
Brons Larson: Yeah, you’re exactly right. AI’s been around for decades. It’s going on close to 70, 80 years. It’s kind of had an interesting past where there’s been a lot of interest in it, and then that interest tends to kind of wax and wane over time as new innovation and new realizations occur. It could be hardware innovation, it could be new algorithms, new approaches, new applications.
Brons Larson: Neural nets are kind of the foundation of AI. And they have been around a long time as well, but they really need a lot of data in order to make them work well. So what really caused them to be very, very important today, is really the data revolution.
Brons Larson: Because they needed data in order to perform well, they weren’t terribly interesting 50 years ago, because there just wasn’t the amount of data to make them work well. Nowadays that’s changed, and that’s really what’s led to the current revolution and reinvestment and re-interest in AI, is the fact that we have these neural net applications, which really are combined with the data explosion, which is occurring around the world.
Brons Larson: And that leads us actually to another topic, which is a shift in AI as a whole. It’s kind of evolved over these three different waves in the past. The first one being where you just took the data scientists’ knowledge, and you tried to capture it in an algorithm. That was the first wave of AI. It’s handcrafted knowledge.
Brons Larson: The second wave was the stochastic learning aspect, which is what we’re in right now. The second wave of AI is neural nets. It’s all based on statistics. Needs lots of data to build up those statistics.
Brons Larson: But as we move forward… There’s some limitations of the current state of AI, and one of them is the fact that these systems cannot actually identify classes and things which they haven’t been exposed to before. So this idea of abstracting knowledge, what we humans can do very easily. Abstract this knowledge that has currently been trained up and apply it to something new.
Brons Larson: That is something that doesn’t work very well today, and there’s a big push towards a new algorithmic shift away from neural networks into one that is model-based. It’s not necessarily neural net models.
Brons Larson: There’s a whole other shift going on. Building up the core elements, the core models, for everything in the world that would allow us to then hypothesize stuff that maybe wasn’t in a training set. And it moves us from a data-driven world again, to this model driven world.
Brons Larson: So I see that as pretty much the future direction of whether you have a neural net based second wave system, or you have these non-neural based third wave systems. And an interesting aspect of it is, it’s not necessarily new. Just like neural nets weren’t new when they were brought under the umbrella of AI, they’ve been around a long time. It’s just that with this data explosion, they actually were a classification scheme that now had viability.
Jon Hyde: Yeah. If we start to look at some of the knowledge-sharing, when we look at training, and we look at models. You could have, in this institutional sharing of different types of models to implement across multiple different businesses, you could ultimately end up with a better trained set when you do that.
Jon Hyde: Because you’re looking at a much broader subset of data. You’re removing some of the institutional or individual biases that may have been in the individual models, right? So there’s this group-think value that may be very interesting.
Brons Larson: Yeah. And the thing about, I think a lot of people, as we talk a little more about the third wave of AI and the shift away from neural nets. The thing that’s kind of neat about building these fundamental models where you actually parameterize them and have knobs you can tune to make the model do what you want it to do, or look a certain way. That’s what allows us really to start to hypothesize scenarios which aren’t in a data set.
Brons Larson: And I’ll give you a concrete example. I’ve built a lot of different sensors over the years, smart sensors, and they tend to be more focused on third wave type kind of solutions. The reason for that, like an example that I built years ago to detect biothreats, really bacteria, viruses, toxins. It’s deployed around the world in subway systems, and post offices, and embassies, and it’s constantly doing air sampling to really try and detect if there is anything new coming out there.
Brons Larson: We didn’t want to just detect the known threats. For starters, it would be very difficult to actually, if you want to use a neural net approach, to actually gather enough data, genetic data, to actually even train something up to have the fidelity you need to make sure that you don’t false alarm. Because you sure as heck don’t want to raise a red flag that there’s a smallpox threat in New York City and shut down the city and it’s not true.
Brons Larson: What we wanted to also do is make sure that we could detect things that have never emerged before. And so by putting together the elements of what a particular threat would look like in terms of its family, genus, species, strain. Ways that we describe a threat.
Brons Larson: We started with models for each of those elements and were able to then tune and twist the knobs to hypothesize new threats which had never existed on Earth before, and then test those against the air samples that we had out there. That gave us the ability to now really detect anything that you could ever conceive of.
Brons Larson: One of the first it was able to detect back in 2003, the coronavirus, SARS, correctly identified as a coronavirus with some of the attributes associated with its virulence, its transmission factor. And it had never actually existed before. The fact that we were able to actually hypothesize it by tuning these knobs with this different way of doing AI, it allowed us to abstract two things that we had never actually seen before.
Brons Larson: And to build the models, we actually didn’t need any data to do it. So we were able to design these systems without any data, which really captures the essence of what the ultimate goals of AI are. The ability to sense something, learn from it, abstract to something you’ve never seen before, and then provide some reasoning logic into what it is and what you should do about it.
Jon Hyde: That is really, really powerful. It’s an amazing thing to think about.
Jon Hyde: When you consider how you built that type of a system, where do you see the applications of that going for a lot of our customers today? How do you see that going out into the market, and what are the benefits for them?
Brons Larson: One of the big benefits is the reduction in data training. You don’t need to aggregate data. You don’t need to manage data. If you’ve ever worked with the current neural net approaches, there’s this 80/20 split between the amount of time you spend curating the data, and classifying the data, and tagging it, and actually 20% of the time, doing the inference scene and actually making decisions. So it reduces that 80%. You don’t have to have all that data managed.
Brons Larson: It impacts privacy and security, because you may not need to secure that data, or the privacy issues associated with it. Because you’re building a model that is not based on a large amount of data being analyzed. So it does reduce some of those requirements and some of those issues.
Brons Larson: And then really, it’s just anything where you have a situation that you’ve never seen before, that ability to abstract. That’s the real benefit here is, it can actually think ahead.
Brons Larson: So imagine if you’re dealing with connected cars or autonomous driving, you can imagine a current neural net approach where if it seen, let’s say, a ball cross the road, a car is going to stop. But if it’s never seen the idea of a person running after the ball, a kid running after the ball, it may start to go again and hit that little kid. Because it’s never been exposed to that.
Brons Larson: Whereas a third wave approach is going to be able to combine models of, let’s say, the game of baseball with driving. And so it says, “Well, I’ve never seen this scenario right in front of me right now, but I know from baseball people chase balls. And I see a ball in front of me. I’m going to combine that in some new logic and abstract it to something I’ve never seen before.”
Brons Larson: And I’m going to be able to say, “I’m going to hold off on continue driving a few seconds longer to see if there’s a person chasing that ball.” And that’s the real power of third wave AI. In addition to the, I’ll say simplification of the data management.
Jon Hyde: That’s some pretty interesting and heady stuff to think about it. The implications are, from my perspective, fairly limitless in how that could be applied to a lot of different capabilities.
Jon Hyde: It’s equal parts interesting and equal parts kind of terrifying. Because I think there’s always this perception in the public mind that AI will eventually rule the world if we’re not careful.
Jon Hyde: And it could be as simple as what the movies have done to exemplify that through Terminator and others, or it could be as grandiose as the things like the paperclip theory, which in and of themselves are not all that farfetched in today’s society. They’re possible. Those things could happen. There are limitations.
Jon Hyde: What’s your thought on the future of AI? This third wave is amazing. What’s it going to lead us to?
Brons Larson: That’s a big question to ask. I mean, the good thing is we’re still far away from really mimicking human cognition and having the ability of having a Skynet out there that we all kind of fear as the evil future. We’re far away from that.
Brons Larson: I mean, current neural nets in many ways, if you think about it, they’re almost like a big Excel spreadsheet. There’s a bunch of numbers in cells and they’re all linked together and they get updated based on information content.
Brons Larson: So, when you think about it in that perspective where we are right now, it’s a great classification scheme. It’s neat in the way that it’s able to uncover nuances, but it still is kind of a static, almost large Excel spreadsheet. I use that as a analogy, which isn’t quite as scary as I think most people think.
Brons Larson: When you move to the third wave concept, with the idea of these fundamental features that we kind of build upon, and we use these underlying features to build higher level models and higher level models in much the way humans do. Well, it’s interesting because basically six-year-olds are able to aggregate all of the thousands of different geometries out there really by age six.
Brons Larson: And so, the level of cognition that we’re trying to even get to in third wave AI, is below a six-year-old’s level of capability at this stage. So we’re still far away off, but again, it’s technology can change in an instant and hopefully the time it takes to build those systems will allow us to have lots of intelligent discussion on the ethics and the usability and the safeguards.
Jon Hyde: Yeah, I think the promises of the first two waves was very, very interesting. And it was insightful for us to learn what might be possible. But this is a new chapter, a very new idea that companies need to be able to grab onto.
Jon Hyde: And I think it’s going to, from my perspective, it opens up the accessibility of AI much more than the first or second wave does, because you’re not constrained by a lot of the challenges of the first two waves.
Brons Larson: Absolutely. Yup.
Jon Hyde: Well, Brons, this is super insightful. I appreciate the time and thank you for sharing it with us. I’m sure that our listeners will get a lot out of this.
Brons Larson: I really appreciated speaking with you today.
Jon Hyde: I enjoyed the view. So thank you for that.
Brons Larson: Take care.
Jon Hyde: You too.
Jon Hyde: For those of you who enjoyed this podcast, you can find it at www.delltechnologies.com/nexthorizon, along with feature podcasts and other great content focused on emerging technologies. Thank you so much for listening and be sure to subscribe. Until next time, I’m Jon Hyde, and this is The Next Horizon.