Latest / Faster, Please! — The Podcast / ✨ AI and the future of R&D: My chat (+transcript) with McKinsey's Michael Chui
Transcript
- 0:06SPEAKER_00: Welcome to Faster Please, the podcast.
- 0:09SPEAKER_00: I'm your host, Jim Pethikoukis.
- 0:12SPEAKER_00: Several times a month,
- 0:13I'll feature a lively conversation with a fascinating and
- 0:16provocative guest about
- 0:18SPEAKER_00: how to make the world a better place through scientific discovery,
- 0:21SPEAKER_00: technological innovation,
- 0:23SPEAKER_00: and economic growth.
- 0:24SPEAKER_00: You're also going to want to check out my Faster Please newsletter.
- 0:27SPEAKER_00: You're on Substack throughout the week for fresh essays,
- 0:30SPEAKER_00: Q&As,
- 0:31SPEAKER_00: and stories from around the internet and around the world.
- 0:37SPEAKER_00: The innovation landscape is facing a difficult paradox.
- 0:41Even as R&D investment has increased,
- 0:44productivity per dollar invested is in decline.
- 0:48SPEAKER_00: In his recent co-authored paper,
- 0:50SPEAKER_00: The Next Innovation Revolution,
- 0:52SPEAKER_00: Powered by AI,
- 0:54SPEAKER_00: Michael Chewy explores AI as a possible solution to this dilemma.
- 0:59SPEAKER_00: Today on Faster Please,
- 1:00SPEAKER_00: the podcast,
- 1:02Chewy and I explore the vast potential for AI augmented
- 1:05research and the challenges
- 1:07SPEAKER_00: and opportunities that come with applying it to the real world.
- 1:10SPEAKER_00: Chewy is a senior fellow at Quantum Black,
- 1:13SPEAKER_00: McKinsey's AI unit,
- 1:15where he leads McKinsey research and AI automation and the
- 1:18future of work.
- 1:20Do we understand what explains this phenomenon where we seem to
- 1:27be doing lots of
- 1:28SPEAKER_00: science and we're spending lots of money on R&D,
- 1:31SPEAKER_00: but the actual productivity of that R&D is declining?
- 1:35SPEAKER_00: Do we have a good explanation for that?
- 1:37SPEAKER_01: I don't know if we have just one good explanation.
- 1:39SPEAKER_01: I mean,
- 1:40I think that the folks that we both know have been both working
- 1:44on what are the
- 1:44SPEAKER_01: causes of this,
- 1:45SPEAKER_01: as well as what are some of the potential solutions.
- 1:48SPEAKER_01: But I think it's a bit of a hidden problem.
- 1:50SPEAKER_01: I don't think everyone understands it.
- 1:51SPEAKER_01: There are a set of people who have looked at this, right?
- 1:53SPEAKER_01: And quite notably, Nick Bloom at Stanford.
- 1:57who published this somewhat famous paper that some people are
- 2:00familiar with.
- 2:01SPEAKER_01: But it is surprising in some sense.
- 2:03At one level,
- 2:04it's amazing what science and engineering has been able to do.
- 2:07We continue to see these incredible advances,
- 2:10whether it's in AI or biotechnology or whatever.
- 2:13SPEAKER_01: But also,
- 2:14SPEAKER_01: you know,
- 2:14what Nick and other researchers have discovered is that we are
- 2:17producing less for
- 2:19SPEAKER_01: every dollar we spend in R&D.
- 2:21SPEAKER_01: And so that's this little bit of a paradox or this challenge that we see.
- 2:25SPEAKER_01: And so,
- 2:26SPEAKER_01: you know,
- 2:26SPEAKER_01: what some of the research we've been trying to do is understand,
- 2:29SPEAKER_01: you know,
- 2:30SPEAKER_01: can AI try to contribute to bending those curves?
- 2:32SPEAKER_01: But it is remarkable.
- 2:34SPEAKER_01: You know,
- 2:34SPEAKER_01: in our research,
- 2:35SPEAKER_01: again,
- 2:35there are folks in the pharmaceutical industry who are well
- 2:39aware of this.
- 2:39SPEAKER_01: But, you know, I'm a computer scientist by training.
- 2:42SPEAKER_01: I love this idea of Moore's Law.
- 2:43SPEAKER_01: Every couple of years,
- 2:44SPEAKER_01: you know,
- 2:45you can double the number of transistors you can put on a chip
- 2:48or whatever for the
- 2:49SPEAKER_01: same amount of money.
- 2:50SPEAKER_01: There's something called E-Room's Law, which is Moore spelled backwards.
- 2:54SPEAKER_01: And basically it said...
- 2:56SPEAKER_01: For decades in the pharmaceutical industry,
- 2:58the number of compounds or drugs you would produce for every
- 3:01billion dollars of R&D
- 3:04SPEAKER_01: would get cut in half every nine years.
- 3:07SPEAKER_01: That's obviously moving in the wrong direction.
- 3:09SPEAKER_01: And so I think that challenge,
- 3:11SPEAKER_01: I don't think everyone is aware of,
- 3:13SPEAKER_01: but one that we need to address.
- 3:15SPEAKER_00: I suppose in a way,
- 3:17it does kind of make sense that as we tackle harder problems
- 3:21and we climb sort of
- 3:22SPEAKER_00: the tree of knowledge,
- 3:24SPEAKER_00: that it's just going to take more time,
- 3:28SPEAKER_00: maybe more researchers.
- 3:30SPEAKER_00: The researchers themselves may have to spend more time in school.
- 3:35SPEAKER_00: So it may be a bit of a hidden problem, but I don't know.
- 3:39SPEAKER_00: I guess it makes some intuitive sense to me.
- 3:43think there's a way to think about it that way right which is
- 3:45like all the easy
- 3:46stuff we already figured out right and so the low hanging fruit
- 3:49has been picked
- 3:50things are getting harder and harder it's amazing i mean you
- 3:52know you could look at
- 3:53some of the early papers in any field and it have a handful of
- 3:56authors right the
- 3:57dna paper three authors although it probably should have
- 3:59included rosalind franklin
- 4:00SPEAKER_01: but anyway
- 4:01SPEAKER_01: There's that whole thing.
- 4:02SPEAKER_01: And now you look at a physics paper or a computer science paper.
- 4:05SPEAKER_01: I mean, the author list just goes on sometimes for pages.
- 4:09SPEAKER_01: And so, you know, these problems are harder.
- 4:11SPEAKER_01: They require more and more effort,
- 4:15SPEAKER_01: whether it's,
- 4:16SPEAKER_01: you know,
- 4:16SPEAKER_01: people's talents or whether it's computing power or,
- 4:19SPEAKER_01: you know,
- 4:20SPEAKER_01: large scale experiments.
- 4:21SPEAKER_01: Things are getting harder to do with that.
- 4:24SPEAKER_01: And, you know, I think there's ways in which that makes sense.
- 4:27SPEAKER_01: Are there other ways in which we could improve the processes?
- 4:29SPEAKER_01: Probably too.
- 4:30SPEAKER_00: We could invest more in research,
- 4:32SPEAKER_00: make it more efficient,
- 4:33SPEAKER_00: and encourage more people to become researchers.
- 4:37SPEAKER_00: But to me,
- 4:37what's more exciting than automating different customer service
- 4:42processes is
- 4:43SPEAKER_00: accelerating scientific discovery.
- 4:46SPEAKER_00: I think that's what makes AI so compelling.
- 4:49SPEAKER_00: That is exactly right.
- 4:50SPEAKER_01: Now,
- 4:50SPEAKER_01: just as by the way,
- 4:52I think we need to continue to invest in basic research and in
- 4:57science and
- 4:57SPEAKER_01: engineering.
- 4:58SPEAKER_01: I think, yeah, I think that's absolutely important.
- 5:00SPEAKER_01: But that's worth noting because I'm not sure everybody thinks that.
- 5:04SPEAKER_01: So I'm glad you highlighted that.
- 5:06I don't think AI means that everything becomes cheaper and we
- 5:09don't need to invest
- 5:11SPEAKER_01: in both human talent as well as in research.
- 5:13SPEAKER_01: So that's number one.
- 5:13SPEAKER_01: But number two,
- 5:14SPEAKER_01: as you said,
- 5:15SPEAKER_01: we spend a lot of time,
- 5:16SPEAKER_01: and appropriately so,
- 5:17SPEAKER_01: talking about how AI can improve productivity,
- 5:20SPEAKER_01: make things more efficient,
- 5:21SPEAKER_01: do the things that we do already cheaper and faster.
- 5:25SPEAKER_01: I think that's absolutely true.
- 5:26SPEAKER_01: But one of the things, we had the opportunity to look over history
- 5:30and you know what has actually improved the human condition you
- 5:33know what has been
- 5:34one of the things that has been necessary to improve the human
- 5:37condition over
- 5:38decades and centuries and millennia is in fact discovering new
- 5:43ideas having
- 5:44scientific breakthroughs turning those scientific breakthroughs
- 5:46into engineering
- 5:47that turn into products and services that do everything from
- 5:50expand our lifespans
- 5:53SPEAKER_01: you know,
- 5:53SPEAKER_01: to be able to provide us with food,
- 5:56SPEAKER_01: all the sorts of more energy,
- 5:57SPEAKER_01: all those sorts of things require innovation,
- 6:00SPEAKER_01: require R&D.
- 6:01And what we've discovered is the potential for AI not only to
- 6:05make things more
- 6:06SPEAKER_01: efficient,
- 6:06SPEAKER_01: but for it to produce more innovation,
- 6:08SPEAKER_01: to more ideas that hopefully will lead to breakthroughs that help us all.
- 6:13SPEAKER_00: I've heard described as an IMI.
- 6:15SPEAKER_00: It's an invention that makes more invention.
- 6:18SPEAKER_00: It's invention of a method of invention.
- 6:21SPEAKER_00: That sounds great.
- 6:23SPEAKER_00: How's it going to do that?
- 6:25SPEAKER_01: Well, there are a couple of ways.
- 6:26We looked at three different channels through which AI could
- 6:29improve this process
- 6:31SPEAKER_01: of innovation and R&D.
- 6:33And the first one is just increasing the volume,
- 6:36velocity,
- 6:37and variety of different candidates.
- 6:40SPEAKER_01: I mean,
- 6:40one way you could think about innovation is you create a whole
- 6:42bunch of candidates
- 6:44SPEAKER_01: and then you filter them down to the ones that might be most effective.
- 6:47Number one,
- 6:47you can just fill that funnel faster,
- 6:49better,
- 6:50and with greater variety.
- 6:51SPEAKER_01: That's number one.
- 6:52SPEAKER_00: The candidates could be a molecule,
- 6:54SPEAKER_00: could be a drug,
- 6:55SPEAKER_00: it could be a new alloy,
- 6:57SPEAKER_00: it could be lots of things.
- 6:58absolutely or designed for a physical product i mean you know
- 7:01what one of the
- 7:02interesting things is right this this you know quote unquote
- 7:05modern ai ai's been
- 7:06around for 70 years right is based on foundation models these
- 7:09ideas these large you
- 7:11know artificial neural networks trained on huge amounts of data
- 7:16and they produce
- 7:16unstructured outputs like in many cases language we talk about
- 7:20llms but the
- 7:21interesting thing is you can train these foundation models not
- 7:24just to generate
- 7:25SPEAKER_01: language but you can generate
- 7:27SPEAKER_01: a protein or a drug candidate, as you were saying, right?
- 7:31SPEAKER_01: You imagine the prompt being,
- 7:33SPEAKER_01: please produce 10 drug candidates that address this condition,
- 7:38SPEAKER_01: but without the following side effects,
- 7:39SPEAKER_01: right?
- 7:39SPEAKER_01: It's not exactly how it works,
- 7:41SPEAKER_01: but roughly speaking,
- 7:42SPEAKER_01: that's the potential to generate these things,
- 7:44SPEAKER_01: or generate an electrical circuit,
- 7:47SPEAKER_01: or a design for an airfoil or an airframe that has these characteristics.
- 7:51SPEAKER_01: being able to just generate those.
- 7:53SPEAKER_01: And the interesting thing is not only can you generate them faster,
- 7:56but there's this idea of you can create more variety because
- 8:00we're usefully proud
- 8:02SPEAKER_01: as humans about our creativity,
- 8:04SPEAKER_01: but also that judgment or that training that we have,
- 8:08SPEAKER_01: that experience sometimes constrains it.
- 8:10SPEAKER_01: So the famous example was,
- 8:13SPEAKER_01: You know,
- 8:14SPEAKER_01: when some folks created this machine called AlphaGo,
- 8:18which was meant to compete against the world champion in this
- 8:21game called Go,
- 8:22SPEAKER_01: a very complex strategic game.
- 8:24SPEAKER_01: And famously, it beat the world champion.
- 8:26SPEAKER_01: But one of the things it did,
- 8:28SPEAKER_01: there's this famous Move 37,
- 8:29SPEAKER_01: this move that everyone who was an expert at Go said,
- 8:33SPEAKER_01: that is nuts.
- 8:34SPEAKER_01: Why would you possibly do that?
- 8:35SPEAKER_01: And because the machine was a little bit more unconstrained,
- 8:39SPEAKER_01: actually came up with, you might describe as a creative idea.
- 8:43SPEAKER_01: And so I think that's one of the other potentials of using AI,
- 8:45that it could both absorb some of the experience that people
- 8:50have as well as
- 8:52SPEAKER_01: stretch the bounds of what might be possible.
- 8:56SPEAKER_00: So you come up with a design and then a variety of options,
- 8:59SPEAKER_00: and then AI can help model and test them.
- 9:02SPEAKER_01: Exactly.
- 9:03So you generate a broader and more voluminous set of potential
- 9:09designs,
- 9:10candidates,
- 9:10right?
- 9:11SPEAKER_01: Whether it's molecules or chemicals or what have you.
- 9:14SPEAKER_01: Now you need to narrow that down.
- 9:16And traditionally,
- 9:17you would narrow it down either one through physical testing,
- 9:20right?
- 9:20So put something into a wind tunnel or run it through the water
- 9:24if you're looking
- 9:25SPEAKER_01: at a boat design or something like that.
- 9:28or you know put it in an electromagnetic you know chamber and
- 9:31see how the antenna
- 9:32operates you either test it physically and then of course you
- 9:36know lots of people
- 9:37figure it out how to use physics mathematical equations in
- 9:41order to create you know
- 9:42quote unquote digital twins right and so you have these you
- 9:45know these these long
- 9:46SPEAKER_01: acronyms like cfd for
- 9:49SPEAKER_01: Computational fluid dynamics, right?
- 9:51SPEAKER_01: Basically a virtual wind tunnel or what have you.
- 9:54SPEAKER_01: Or you have finite element analysis,
- 9:56SPEAKER_01: another way to model how a structure might perform,
- 9:59SPEAKER_01: or computational electromagnetic modeling,
- 10:03SPEAKER_01: right?
- 10:03SPEAKER_01: All these ways that you can use physics to simulate things.
- 10:07SPEAKER_01: And that's been terrific, right?
- 10:08SPEAKER_01: That's one of the things.
- 10:09SPEAKER_01: But some of those models...
- 10:11SPEAKER_01: actually take a huge like hours, sometimes days to run these models.
- 10:16And it might be faster than building the physical prototype and
- 10:19then,
- 10:20you know,
- 10:20modeling it.
- 10:21SPEAKER_01: You know, again, sometimes you just wait till something breaks, right?
- 10:24SPEAKER_01: You're doing failure testing.
- 10:26SPEAKER_01: then you could do that,
- 10:27SPEAKER_01: you know,
- 10:27SPEAKER_01: in a computer using these models,
- 10:29SPEAKER_01: but sometimes they take a really long time.
- 10:31And one of the really interesting discoveries in quote-unquote
- 10:34AI is you can use
- 10:35SPEAKER_01: that same sort of neural network that we've used,
- 10:40SPEAKER_01: you know,
- 10:40SPEAKER_01: to simulate cognition or intelligence,
- 10:43SPEAKER_01: but now you use it to simulate physical systems.
- 10:46So in some ways it's not AI because you're not creating
- 10:49artificial intelligence,
- 10:50SPEAKER_01: you're creating an artificial...
- 10:52you know wind tunnel right it's just a different way to model
- 10:55physics and sometimes
- 10:56these problems get even more complicated when you have to both
- 11:00model the physics
- 11:01right if you're trying to put an antenna on an airplane you
- 11:03need to know you know
- 11:04how the airflow is going to go over it but you need to know
- 11:06whether or not it you
- 11:07SPEAKER_01: know the radio frequency stuff works out too right all that rf stuff
- 11:11And so these multi-physics models,
- 11:13you know,
- 11:13the complexity is even higher.
- 11:14SPEAKER_01: And you can train these neural nets,
- 11:16SPEAKER_01: which is interesting to do this,
- 11:18SPEAKER_01: but again,
- 11:19SPEAKER_01: even faster than these physics-based models.
- 11:21SPEAKER_01: And so we have these things called,
- 11:23SPEAKER_01: you know,
- 11:23SPEAKER_01: AI surrogate models,
- 11:25SPEAKER_01: you know,
- 11:25SPEAKER_01: they're sort of surrogates.
- 11:26So it's,
- 11:27you know,
- 11:27two steps removed in some ways from,
- 11:29you know,
- 11:30actual physical testing.
- 11:31SPEAKER_01: And when that can happen faster,
- 11:33SPEAKER_01: I mean,
- 11:33SPEAKER_01: literally we've seen models that can run in minutes rather than hours or,
- 11:38SPEAKER_01: you know,
- 11:38SPEAKER_01: an hour rather than a few days.
- 11:40SPEAKER_01: that can accelerate things.
- 11:41We see this in weather forecasting and a number of different
- 11:44ways in which this can happen.
- 11:45SPEAKER_01: And so again,
- 11:46SPEAKER_01: if you can generate more candidates and then test them faster,
- 11:50SPEAKER_01: you can imagine the whole R&D process really accelerating.
- 11:55Does achieving your estimated productivity increases depend
- 11:59more on further
- 12:00SPEAKER_00: technological advances?
- 12:02or does it depend more on how companies adopt and implement the
- 12:05technology i guess
- 12:07is is the bottleneck still in the tech itself or is it more
- 12:10about business
- 12:11SPEAKER_00: adaptation
- 12:13SPEAKER_01: Mostly number two.
- 12:14SPEAKER_01: All right.
- 12:14SPEAKER_01: So like, yeah, the technology is going to continue to advance.
- 12:17SPEAKER_01: And right, as a technologist, I love all that stuff.
- 12:21But really,
- 12:22as usual,
- 12:23a lot of the challenge here are organizational challenges.
- 12:27We know that companies are using AI surrogates,
- 12:31deep learning surrogates already.
- 12:33SPEAKER_01: But, you know, is it being applied as many places as possible?
- 12:37SPEAKER_01: No, it isn't.
- 12:38SPEAKER_01: And a lot of these things are organizational, right?
- 12:40SPEAKER_01: Does it match your strategy, for instance, right?
- 12:42SPEAKER_01: Do you have the right talent and organization in place?
- 12:45SPEAKER_01: Let me just give one very specific example.
- 12:47SPEAKER_01: In a lot of R&D organizations,
- 12:49we know there is a separate organization for physical testing
- 12:52and a separate
- 12:53SPEAKER_01: organization for simulations.
- 12:56SPEAKER_01: You know,
- 12:56SPEAKER_01: in simulation,
- 12:56SPEAKER_01: in many cases,
- 12:57SPEAKER_01: physics based,
- 12:57SPEAKER_01: but you add these deep learning surrogates as well.
- 13:01SPEAKER_01: That doesn't make sense at some level.
- 13:03SPEAKER_01: You know,
- 13:03SPEAKER_01: I'm not saying physical testing goes away,
- 13:05SPEAKER_01: but you need to figure out when you should physically test,
- 13:08SPEAKER_01: when you should use which simulation methods,
- 13:10SPEAKER_01: when you should use deep learning surrogates or AI techniques,
- 13:12SPEAKER_01: etc.
- 13:13SPEAKER_01: And that's just, you know, one organizational problem.
- 13:16difference that you could make if you were in an organization
- 13:20that was actually
- 13:21SPEAKER_01: taking this whole testing regime seriously,
- 13:24where you're actually parsing out when the optimal amount of
- 13:29physical testing is
- 13:30SPEAKER_01: versus simulation,
- 13:30SPEAKER_01: etc.
- 13:31SPEAKER_01: So there's a number of things where that's true.
- 13:34SPEAKER_00: Even before AI,
- 13:35SPEAKER_00: historically,
- 13:36SPEAKER_00: there's a gap between novel new technologies,
- 13:40SPEAKER_00: what they can do in lab settings,
- 13:42and then how they're applied in real-world research or in
- 13:45business environment.
- 13:47SPEAKER_00: That gap,
- 13:48SPEAKER_00: I would guess,
- 13:48SPEAKER_00: probably requires companies to rewire how they operate,
- 13:52SPEAKER_00: which takes time.
- 13:54SPEAKER_01: It is indeed.
- 13:55SPEAKER_01: And it's funny that you use the word rewiring, right?
- 13:57SPEAKER_01: My colleagues wrote a book entitled Rewired,
- 13:59SPEAKER_01: which literally is about,
- 14:00SPEAKER_01: you know,
- 14:01SPEAKER_01: the different ways together that you need to,
- 14:04SPEAKER_01: as you say,
- 14:05SPEAKER_01: rewire or change the way an organization operates.
- 14:08SPEAKER_01: And only one of those six chapters is around the tech stack, right?
- 14:12SPEAKER_01: It's still absolutely important, right?
- 14:13SPEAKER_01: You've got to get all that stuff right.
- 14:15But it is mostly all of the other things surrounding how you
- 14:18change an organization operates and
- 14:21in order to bring the full value of this together to reach
- 14:24scale right i mean we
- 14:25also talk about pilot purgatory right we did this cool
- 14:28experiment seems to but you
- 14:29know when is it good enough that the cfos talks about it the
- 14:32quarterly earnings
- 14:33report that requires the organization to change the way it
- 14:35operates and that's the
- 14:37learning we've seen all the time we've been serving thousands
- 14:39of executives on
- 14:40SPEAKER_01: their use of ai you know for seven years now and then you know
- 14:45nearly 80% of organizations say they're regularly using AI
- 14:48someplace in the business.
- 14:50But in a separate survey,
- 14:51only 1% say they're mature in that usage,
- 14:54right?
- 14:54And so there's this giant gap between just using AI and then
- 14:59actually having the
- 15:00SPEAKER_01: value be created.
- 15:01SPEAKER_01: And by the way,
- 15:02SPEAKER_01: You know,
- 15:03SPEAKER_01: organizations that are creating that value are,
- 15:04SPEAKER_01: you know,
- 15:05SPEAKER_01: accelerating their performance difference.
- 15:07SPEAKER_01: Again,
- 15:07SPEAKER_01: if you have a much more productive R&D organization that churns out,
- 15:13SPEAKER_01: you know,
- 15:14SPEAKER_01: products that are successful in the market,
- 15:16SPEAKER_01: you're going to be ahead of your competitors.
- 15:17SPEAKER_01: And that's what we're seeing, too.
- 15:19Is there a specific problem that comes up over and over again
- 15:23with companies,
- 15:24SPEAKER_00: either in their implementation of AI?
- 15:27SPEAKER_00: Maybe they don't trust it.
- 15:29SPEAKER_00: One, they may know how to use it.
- 15:30SPEAKER_00: Or what do you think is the problem?
- 15:32SPEAKER_01: Unfortunately, I don't think there's just one thing.
- 15:35SPEAKER_01: I mean,
- 15:35SPEAKER_01: you know,
- 15:36SPEAKER_01: again,
- 15:36SPEAKER_01: my colleagues who do this work on Rewired,
- 15:39SPEAKER_01: for instance,
- 15:41SPEAKER_01: you kind of have to do all those things,
- 15:43SPEAKER_01: right?
- 15:44And so you do have to have the right talent and organization in
- 15:47place,
- 15:48right?
- 15:48SPEAKER_01: You have to figure out scaling, for instance, right?
- 15:50SPEAKER_01: You have to figure out change management.
- 15:52All of those things together are what underpins,
- 15:56you know,
- 15:56outsized performance.
- 15:57SPEAKER_01: So all those things have to be done.
- 15:59So if companies are successful,
- 16:02what is the sort of productivity impact you see?
- 16:05SPEAKER_00: We're talking about basically the current technology level, give or take.
- 16:11SPEAKER_00: We're not talking about human level AI, super intelligence.
- 16:14SPEAKER_00: We're talking sort of about AI more or less as it exists today.
- 16:18SPEAKER_00: Everybody wants to accelerate productivity.
- 16:19SPEAKER_00: It's like governments around the world, companies.
- 16:22SPEAKER_00: So give me a feel for that.
- 16:23SPEAKER_01: There are different measures of productivity,
- 16:25SPEAKER_01: but here what we're talking about is basically how many new products,
- 16:29SPEAKER_01: successful products,
- 16:30SPEAKER_01: can you put out in the market,
- 16:31SPEAKER_01: right?
- 16:32SPEAKER_01: And our modeling says,
- 16:33SPEAKER_01: you know,
- 16:33SPEAKER_01: depending on your industry,
- 16:35SPEAKER_01: you could double your productivity,
- 16:37SPEAKER_01: in other words,
- 16:38SPEAKER_01: of R&D,
- 16:39SPEAKER_01: right?
- 16:39SPEAKER_01: In other words, you could put out...
- 16:41SPEAKER_01: double the amount of products and services,
- 16:44SPEAKER_01: new products and services that you have been previously.
- 16:46SPEAKER_01: Now, that's not true for every industry.
- 16:47SPEAKER_01: And by the way,
- 16:48SPEAKER_01: the impact of that is different for different industries because,
- 16:51SPEAKER_01: you know,
- 16:52SPEAKER_01: for some industries,
- 16:54SPEAKER_01: your dependent,
- 16:54SPEAKER_01: your pharmaceuticals,
- 16:56SPEAKER_01: you know,
- 16:57SPEAKER_01: a lot,
- 16:57like the majority of your value comes from producing new
- 17:00products and services over
- 17:01SPEAKER_01: time because eventually the patent runs out or whatever.
- 17:04SPEAKER_01: There are other industries,
- 17:05SPEAKER_01: you know,
- 17:07SPEAKER_01: we talk about science-based industries like chemicals,
- 17:09SPEAKER_01: for instance,
- 17:10SPEAKER_01: right?
- 17:10SPEAKER_01: I mean,
- 17:12SPEAKER_01: the new product development process in chemicals is very,
- 17:14SPEAKER_01: very close to the science of chemistry.
- 17:16SPEAKER_01: And so these levers that I just talked about,
- 17:19SPEAKER_01: right,
- 17:20SPEAKER_01: producing more candidates,
- 17:21SPEAKER_01: you know,
- 17:22being able to evaluate them more quickly and all the other
- 17:24things that LLMs can do
- 17:25SPEAKER_01: just in general,
- 17:27you know,
- 17:27we could see,
- 17:28again,
- 17:29potential doubling in the pace of which innovation happens.
- 17:32SPEAKER_01: On the other hand,
- 17:33SPEAKER_01: the chemicals industry,
- 17:35SPEAKER_01: let's leave out specialty chemicals,
- 17:36SPEAKER_01: but,
- 17:36SPEAKER_01: you know,
- 17:37SPEAKER_01: the commodity chemicals,
- 17:38SPEAKER_01: they'll still need to produce ethylene,
- 17:40SPEAKER_01: right?
- 17:41SPEAKER_01: And so,
- 17:41SPEAKER_01: you know,
- 17:42SPEAKER_01: to a certain extent,
- 17:43SPEAKER_01: while the R&D process can be,
- 17:45SPEAKER_01: you know,
- 17:45SPEAKER_01: accelerated a great deal,
- 17:47The EBIT impact on the industry might be lower than it is for
- 17:51pharmaceuticals,
- 17:51SPEAKER_01: for instance,
- 17:52SPEAKER_01: but still it's valuable.
- 17:53SPEAKER_01: And then again, if you're in specialty chem, it means a lot to you.
- 17:56SPEAKER_01: And so again,
- 17:57SPEAKER_01: depending on where you sit in your position in the market,
- 18:02SPEAKER_01: it can vary,
- 18:02SPEAKER_01: but the potential is really high.
- 18:05Will future R&D look more like researchers augmented by AI or
- 18:12AI systems assisted
- 18:13SPEAKER_00: by researchers?
- 18:14SPEAKER_00: Who's the assistant in this equation?
- 18:17SPEAKER_00: Who's working for who?
- 18:18SPEAKER_01: Yeah, yeah.
- 18:19SPEAKER_01: No, it's all of the above.
- 18:20And it depends on,
- 18:21you know,
- 18:22how you decide to use these technologies,
- 18:23right?
- 18:24SPEAKER_01: But, you know, I mean, we even write in our paper, you know, that...
- 18:27SPEAKER_01: We need to be thoughtful about where you put the human in the loop.
- 18:30SPEAKER_01: Every study,
- 18:31SPEAKER_01: you know,
- 18:32SPEAKER_01: the conditions matter,
- 18:33SPEAKER_01: but there are lots of studies where you say,
- 18:35SPEAKER_01: look,
- 18:35SPEAKER_01: the combination of machines and humans,
- 18:37SPEAKER_01: right?
- 18:38SPEAKER_01: So AI and researchers is the most powerful combination, right?
- 18:44SPEAKER_01: And so each brings their respective strengths to it.
- 18:48But the funny thing is that sometimes the human biases actually
- 18:52decrease the
- 18:52SPEAKER_01: performance of the overall system,
- 18:53SPEAKER_01: right?
- 18:54SPEAKER_01: And so, oh, maybe we should just go with machines.
- 18:57At least for the medium term,
- 18:59we're not going to be able to get rid of all the people,
- 19:01right?
- 19:01SPEAKER_01: The people are going to be absolutely important to the process.
- 19:05And so I think when is it that people either are necessary to
- 19:10the process or can be
- 19:12SPEAKER_01: helpful,
- 19:12SPEAKER_01: right?
- 19:12SPEAKER_01: In many cases, it is around things like
- 19:15SPEAKER_01: you know,
- 19:15SPEAKER_01: when is it that you need to make a decision that's,
- 19:17SPEAKER_01: you know,
- 19:18SPEAKER_01: a safety critical decision,
- 19:19SPEAKER_01: a regulatory decision where you just have to have a person look at it,
- 19:22SPEAKER_01: right?
- 19:22SPEAKER_01: I mean, that's the sort of necessity argument for people in the loop.
- 19:26But also there are things that machines just don't do well
- 19:30enough yet,
- 19:30right?
- 19:31SPEAKER_01: And there's a little bit of that, yeah.
- 19:33SPEAKER_00: If we can get more productive R&D and then businesses...
- 19:37get better at incorporating this into their processes and they
- 19:41could potentially
- 19:42SPEAKER_00: generate more products and services.
- 19:45SPEAKER_00: Do we have a government ready for that world of accelerated R&D?
- 19:50SPEAKER_00: Can we handle that flow?
- 19:52SPEAKER_00: My bias says probably not, but please correct me if I'm wrong.
- 19:57I mean,
- 19:57I think one of the interesting things is,
- 19:59you know,
- 19:59people talk about AI regulation.
- 20:01SPEAKER_01: In many of these industries, the regulations already exist, right?
- 20:06SPEAKER_01: I mean,
- 20:06SPEAKER_01: we have regulations for,
- 20:09SPEAKER_01: you know,
- 20:09SPEAKER_01: what goes out as,
- 20:11SPEAKER_01: you know,
- 20:12SPEAKER_01: in pharmaceuticals,
- 20:12SPEAKER_01: for instance,
- 20:13SPEAKER_01: right?
- 20:13SPEAKER_01: We have regulations in the aviation industry.
- 20:16SPEAKER_01: We have regulations in the automobile industry.
- 20:18And in many ways,
- 20:20you know,
- 20:20AI in the R&D process doesn't change that because maybe it
- 20:25should.
- 20:25SPEAKER_01: I mean,
- 20:25SPEAKER_01: people talk about can you actually accelerate the process of,
- 20:30SPEAKER_01: you know,
- 20:30SPEAKER_01: approving a drug,
- 20:31SPEAKER_01: for instance,
- 20:31SPEAKER_01: right?
- 20:31SPEAKER_01: But that wasn't the, you know, the thing that we studied.
- 20:35SPEAKER_01: And so in some ways, those processes apply now already, right?
- 20:39SPEAKER_01: And so that's something that doesn't necessarily have to change.
- 20:42SPEAKER_01: Now,
- 20:42SPEAKER_01: that said,
- 20:43are some of these potential innovations gated by approval
- 20:48processes or clinical
- 20:50SPEAKER_01: trials processes?
- 20:51SPEAKER_01: Absolutely.
- 20:52SPEAKER_01: In some of those cases,
- 20:53SPEAKER_01: the clinical trials process gate is not necessarily,
- 20:58SPEAKER_01: you know,
- 20:58SPEAKER_01: not necessarily a regulation,
- 20:59SPEAKER_01: but we know there's a big problem just finding enough potential subjects
- 21:04SPEAKER_01: in order to do clinical trials, right?
- 21:06SPEAKER_01: And that's not a regulatory problem.
- 21:07SPEAKER_01: That's a problem of finding people who were, you know,
- 21:12SPEAKER_01: are good candidates for actually testing these drugs.
- 21:15SPEAKER_01: So,
- 21:15SPEAKER_01: yes,
- 21:16SPEAKER_01: in some cases,
- 21:17SPEAKER_01: even if we were able to double the amount of,
- 21:19SPEAKER_01: you know,
- 21:19SPEAKER_01: candidates that can go through the funnel on a number of these things,
- 21:22SPEAKER_01: there will be,
- 21:24SPEAKER_01: you know,
- 21:24SPEAKER_01: these exogenous,
- 21:25SPEAKER_01: you know,
- 21:26SPEAKER_01: issues that would constrain,
- 21:27SPEAKER_01: you know,
- 21:27SPEAKER_01: society's ability to bring these to market.
- 21:31SPEAKER_01: And so that just says,
- 21:32SPEAKER_01: you know,
- 21:32SPEAKER_01: you squeeze the balloon here and it opens up there,
- 21:35SPEAKER_01: but,
- 21:36SPEAKER_01: you know,
- 21:36SPEAKER_01: Let's go solve each of these problems.
- 21:38SPEAKER_01: And one of the problems that we said that,
- 21:39SPEAKER_01: you know,
- 21:40AI can help solve is increasing the number of things that you
- 21:44could potentially put
- 21:45SPEAKER_01: in the market if it can get past the other necessities.
- 21:48For a general public where so much of what they're hearing
- 21:51about AI tends to be
- 21:53SPEAKER_00: about,
- 21:53SPEAKER_00: you know,
- 21:54SPEAKER_00: job loss or are they stealing copyrighted material or,
- 22:01SPEAKER_00: yeah,
- 22:01SPEAKER_00: people talk about these huge advances,
- 22:03SPEAKER_00: but they're not seeing them yet.
- 22:04SPEAKER_00: What is sort of your elevator?
- 22:08SPEAKER_00: optimistic pitch why you may have heard,
- 22:10SPEAKER_00: you may be worried about the impact of AI,
- 22:11SPEAKER_00: but here's why I'm excited about it.
- 22:13SPEAKER_00: Why are you excited by it?
- 22:15SPEAKER_01: By the way, I think all those things are really important, right?
- 22:17SPEAKER_01: I mean,
- 22:17SPEAKER_01: I think all those concerns and,
- 22:19SPEAKER_01: you know,
- 22:19SPEAKER_01: how do we reskill the workforce,
- 22:21SPEAKER_01: all those things,
- 22:21SPEAKER_01: and we've done work on that as well.
- 22:22SPEAKER_01: But here, the thing that I'm excited about is
- 22:25you know we need innovation we need new ideas we need
- 22:29scientific advancements and
- 22:31engineering that turns them into products in order for us to
- 22:34improve their human
- 22:35condition whether it's living longer lives or living you know
- 22:38higher quality life
- 22:40you know whether it's having the energy whether it's to be able
- 22:43to support that in
- 22:44SPEAKER_01: a way that doesn't cause you know other problems all of those things
- 22:50SPEAKER_01: We need to have them.
- 22:52And what we've discovered is AI is one of those things that can
- 22:55produce potentially
- 22:57SPEAKER_01: more of those ideas that can,
- 22:59SPEAKER_01: you know,
- 22:59SPEAKER_01: underpin the,
- 23:00SPEAKER_01: you know,
- 23:01SPEAKER_01: hopefully an improved quality of life for us and our children.