YT YouTube2Text

YouTube2Text → Transcripts → NVIDIA & Lilly: The AI Revolution in Drug Discovery | Jensen Huang & David Ricks

NVIDIA & Lilly: The AI Revolution in Drug Discovery | Jensen Huang & David Ricks — Transcript

by NVIDIA · 10,239 words · 1,700 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:03[applause]
0:04>> Hello. It's great to see all of you.
0:07You know, once a year I get the benefit
0:09of
0:10thanking all of you for the incredible
0:12work work that you do. As you know, we
0:14pioneered a new way of doing computing
0:16called accelerated computing.
0:19And we've been working on this for now
0:2033 years.
0:22The big idea
0:23The big idea
0:24is that if you could co-design, meaning
0:27if you understood the problems you're
0:29trying to compute for
0:32and you develop the entire stack,
0:34meaning the algorithms
0:36the computers and the processors within
0:38it.
0:39That if you co-designed
0:41you solve the problem from top down, you
0:43solve the problem from bottom up and
0:44inside out that you could accelerate
0:48these applications, which are incredibly
0:50computing intensive
0:52by a level that Moore's law would only
0:55dream of.
0:56Moore's law runs at about 10 times every
0:585 years, 100 times every 10 years.
1:01In the last 10 years we've accelerated
1:03AI about a million times.
1:05A million times versus 100 times.
1:07A million times
1:09compounded over 20 years
1:12gives us an opportunity to maybe address
1:14some of the
1:16most incredible, most impactful
1:18challenges of humanity. That's you.
1:21And so, we've been thinking about
1:23accelerated computing and this co-design
1:25idea for a very long time.
1:27Well
1:28after some 30 years or so
1:30um
1:31each different fields of of uh computing
1:34domains uh became within reach. One of
1:37the first ones was computer graphics.
1:39And somebody told me
1:40about
1:4210 years ago
1:43that it would take 30 years for us to
1:45solve this incredibly hard problem
1:48called ray tracing for computer
1:49graphics. Literally following every
1:50photon of light as it bounces around the
1:52room, eventually reaching your eyes.
1:55And using ray tracing you could solve
1:58for really interesting computer graphics
2:00problems and create photorealistic
2:02images like this.
2:03Well, that was
2:0410 years ago
2:06that it would take 30 years. It was the
2:08computer graphics problem that was
2:10always the future.
2:12And literally today we ray trace every
2:15single pixel in a computer graphics
2:16program, in our video games, and not
2:18only do we ray trace it incredibly
2:20beautifully, we do it at several hundred
2:23frames a second.
2:24As large of a resolution as you like.
2:26And we've taught an AI
2:28how to we've taught our computer how to
2:31simulate with first principle
2:33simulations
2:34very few pixels
2:36and then we use AI to predict the rest
2:38of it.
2:40We teach an AI what an image should look
2:42like and it generates the
2:45the image for us. It's temporal
2:47consistent, it's spatially consistent,
2:49it's incredibly beautiful. It looks even
2:51better than we compute.
2:54Well
2:55each one of these different domains are
2:57becoming within reach.
2:59About 10 years ago we started talking
3:01about self-driving cars. Today I think
3:03we are at the ChatGPT moment of
3:05self-driving cars. There's no question
3:07in my mind that within the next 10 years
3:10you know, almost every car that gets
3:11manufactured will have of course
3:14uh either robotaxi capability or if you
3:16like to drive yourself you still can,
3:18but it has the ability to be a robotaxi.
3:21That future is just around the corner.
3:23Each one of these domains are within
3:24reach.
3:26Now of course, we now have uh the
3:28ChatGPT moment. The last year we've seen
3:31incredible progress in generative AI and
3:35large language models.
3:36Uh
3:37these ideas
3:39made it possible for us to take an AI
3:41that we thought was magical
3:44but hallucinated a great deal. It was
3:46essentially cute and very interesting to
3:49something that is incredibly useful.
3:51This last year, we saw last couple of
3:53years, uh breakthroughs in multimodality
3:56AIs. Not only do you understand the
4:00word cat
4:02and the letter c a t you could look at
4:05an image and hear the word hear the
4:07sound meow and recognize all of those
4:10words, all of those information is
4:12aligned to the same place in your brain.
4:14And we can now do that with all kinds of
4:17different modalities, which is very
4:19important to you.
4:20Uh we made huge breakthroughs
4:23in reasoning capabilities.
4:25To be able to decompose through
4:27composition, break down a problem we've
4:30never seen before into routine things
4:33and through that composition solve
4:35problems step-by-step, reasoning through
4:38hard problems.
4:39The ability to do multimodality
4:41reasoning and to be grounded by
4:44different sources of information is
4:46going to be a revolution even for your
4:48industry. It is clearly a revolution in
4:51ours.
4:52The idea that we would use multimodality
4:55sensors
4:57very sensible.
4:58But the idea that a car, a self-driving
5:01car stack that can think
5:04because it can reason through
5:06and
5:07experience a situation it's never
5:09experienced before, reason through it
5:11step-by-step and be able to solve that
5:13problem, navigate through it incredible
5:16capabilities.
5:18It expanded the operating domain of
5:19self-driving cars tremendously. It no
5:23longer it's now the case that we are
5:26very certain we don't have to test every
5:29single car across every single scenario
5:31in order for that car to do well. Just
5:33like us. That we have we can now have a
5:36car that can think and be basically
5:38intelligent.
5:39These techniques
5:41has made uh and and of course agentic
5:44capabilities the ability for agents to
5:47reason, use tools, work with other
5:50agents completely revolutionary.
5:53And so these these breakthroughs um in
5:56in the day that I've been here uh is
5:58permeating your industry, permeating
6:01life sciences. Some of the most exciting
6:03work that I've seen in a very long time.
6:06And so today I we announced a very
6:09important uh very important partnership.
6:11Before I get there, let me say a couple
6:13of things.
6:14We announced of course
6:16uh some of the work that we've done uh
6:17with all of you.
6:19We've brought an accelerated computing
6:21platform uh but on top of that are a
6:24suite of libraries that we're very proud
6:26of. Parabricks for uh gene sequencing
6:29um
6:31MONAI, an open source platform for
6:33medical imaging
6:35uh
6:36and one of the most important ones,
6:38BioNeMo, a platform that understands
6:41uh geometric structures like uh
6:43molecules and proteins.
6:45And these platforms
6:48uh are used by many of you in the room
6:49and I'm very grateful for that.
6:51Uh the
6:53uh several announcements that we made
6:56uh before before I say that uh there
6:58there uh uh so many partners here,
7:00couple of couple of hundred different
7:02companies here. Uh there are pharmas,
7:03there are hospitals, there are startups.
7:05There are startups in robotics, surgical
7:07robotics, there are startups in robotics
7:10labs.
7:11There are agentic AI companies here
7:14um helping with health health care
7:16diagnostics.
7:17And also of uh some really cool work
7:19being done in AI scientists. You know,
7:21part of part of being at the JP Morgan
7:24uh event is uh large companies uh
7:27meeting
7:29uh small companies and so many of them
7:31are doing uh revolutionary work and and
7:34I'm so grateful that that all of you
7:35have chosen the NVIDIA platform to work
7:37on.
7:38And so so our platforms
7:40our platforms um
7:43uh ended up uh achieving some really
7:45great results.
7:46Uh MONAI has been downloaded some 6
7:48million times.
7:51We achieved a Guinness World Record of
7:53gene sequencing.
7:57You know, that's cool.
8:02Yeah.
8:03I've always wanted to be in the Guinness
8:05World Record.
8:06So that's cool.
8:07Uh the the vast majority of the world's
8:10uh structure prediction models are built
8:11on NVIDIA today. And so I'm really proud
8:13of that.
8:14And and on top of the BioNeMo platform
8:18are a whole bunch of pre-trained models.
8:20You use it for data processing, you use
8:21it for training uh these uh biological
8:23models, you use it for fine-tuning. Um
8:26but we also have a whole bunch of uh
8:28pre-trained models and a bunch of data
8:30that that you can start with and you
8:32could uh apply your data to fine-tune uh
8:35your your molecules. And and some of the
8:36some of the models that that we've
8:38created and participated in jointly
8:41developed, uh really proud of Llama
8:42Protina for protein design uh RE SE NT
8:45for molecular synthesis
8:48uh Kermit for uh predicting toxicity
8:52uh Evo Two, a a uh DNA foundation model.
8:56Uh these these foundation models uh
8:58start of course where where I try to
9:01we're trying to learn to represent the
9:02representation of of proteins so that we
9:04can create a foundation model for
9:05protein. But the the moment that that
9:09time where where we're going to learn a
9:11representation and create a foundation
9:13model for cells is just around around
9:14the corner, the multi-protein systems.
9:17Uh
9:17Codon FM for RNA design.
9:20Uh just just really really terrific real
9:22really terrific progress.
9:24And so today
9:26uh we announced a uh
9:28uh a large number of partners that are
9:29working with us in BioNeMo and I'm very
9:31grateful for that. And we also announced
9:34a
9:35a really really a a a
9:37a land landmark partnership
9:40uh with one of the most important
9:42companies in the world and and a really
9:44good friend, Dave Ricks
9:46uh in Lilly.
9:47And so I without without I I won't tell
9:51you more about it. I'm going to I'm
9:52going to let the two of us describe it
9:53together.
9:54Um but uh Dave Dave
9:57he he's been the CEO of Lilly since
9:592017.
10:01And, um
10:02I In the context of the history of
10:05Lilly, 150-year-old company, he's been
10:08with the company for 30 years. That's
10:10So, he's basically a new college grad in
10:13the context of Lilly. And,
10:16and I'm anxious to to learn from him uh
10:19you know, how how he learned so much in
10:21just a short period of time. But, but
10:23he's been with Lilly now 30 years. And,
10:25as you know, uh he has helped guide and
10:28and help build Lilly into one of the
10:30most valuable companies in the world and
10:31surely one of the most important
10:32companies in the world. And so, if you
10:34could join me in welcoming my guest,
10:36David Ricks.
10:38>> [applause]
10:43[applause]
10:46>> Thank you. Great to be here.
10:48>> [applause]
10:49>> So, so it turns out the two of us have
10:51have um
10:54I can't imagine more different
10:56upbringing.
10:57As So, uh
10:59Dave was a was an Eagle Scout and uh
11:02went to a juvenile delinquent uh
11:05school.
11:06Just [laughter] to be clear, I I was
11:08placed there as an accident. It wasn't
11:12>> [laughter]
11:12>> It was No, I was It was a very good
11:14school. It was very affordable.
11:16Um
11:20And so,
11:21so at
11:222017 you became CEO. Yes.
11:24>> My question is this.
11:25What was Lilly
11:27like and what was its vision for
11:31uh the future of life sciences? And and
11:34then what is it like today and your
11:37vision for the future of life sciences
11:38in the context of Lilly?
11:40Yeah, well, so in 2017 we were
11:43almost 140 years old or just a year
11:45older than that. And a pretty
11:47well-established company.
11:49Uh we had been through a difficult
11:51period of time. Um In our industry, and
11:54we'll get get into this a little bit
11:55later, you know, we are plagued by, as
11:58many people in the room know, the cycles
12:00of innovation that are actually longer
12:02than the reward cycles Mhm. on average.
12:06So, you have this peak tro- you know,
12:08you're peaking and troughing and
12:10um in the troughs, of course, you either
12:13don't survive, you end up as a back end
12:15of a hyphen on another company.
12:17Um
12:18or you learn how to survive.
12:21And you do difficult things. Mhm. And I
12:24think it's true that Lilly is the only
12:28uh major pharma company that has not
12:29merged or combined ourselves to get
12:32through such a period. So, and we've
12:35been through a few of them. So, I think
12:36we've developed a bit of a DNA for
12:39scraping by in these tough times. And
12:42one of the things that really stuck with
12:44me, I was a leader on the executive
12:45committee at the time we went through
12:47our last cycle, which was 2010, '11,
12:48'12,
12:50was, you know, how you have to be so
12:52relentless
12:53on invention
12:55because that time cycle is brutal. When
12:58you have a patent cliff and you have
13:00nothing to replace it, you have to
13:02rescale your whole company. And imagine
13:04doing this every 10 or 15 years, it's
13:07impossible to get momentum. People
13:09leave, you're lose inherent knowledge,
13:11et cetera. So,
13:12we set about to do that. And I think
13:15that problem can be attacked two ways.
13:16One, you can say, "Okay, we'll just have
13:18better ideas." But, as you may know,
13:20that's difficult to sustain. Uh even if
13:23you have a good idea, Everyone thinks
13:25>> you're smart. They think about that
13:27about Lilly now, but um it's hard to
13:29sustain. So, we try it We, of course,
13:32worked on that um
13:33more invention, but the more important
13:36thing we did was speed up invention.
13:37Mhm. So, we systematically took apart
13:40the R&D process,
13:42literally like every step, and there's
13:43hundreds of steps from candidate
13:45selection to first approval,
13:48and we squished the time. Mhm. And so,
13:50still today probably Lilly is about 40%
13:53faster
13:55than the next scaled pharma company on
13:57average across all therapeutic areas.
13:59And so, we can invent things inside the
14:02patent cycle. Yeah. Then you can start
14:04to grow. Yeah. So, I think um that was a
14:07notion in 2017, which we've actualized.
14:10Along the way then, of course, we did a
14:12lot of things to drive more invention.
14:15We were lucky enough to be early on an
14:18idea called GLP-1 in 2006. We launched
14:20the first one in the world. Never heard
14:22of it. We fit Yeah.
14:23We fit We fit fiddled around with it Is
14:25it doing well? If it weren't for you, it
14:27would be the most important idea in the
14:29stock market. Yeah.
14:31>> [laughter]
14:34>> Never heard of it. Pretty good idea for
14:35mankind, and we fiddled around with it
14:37for like 12 years. And to their credit,
14:40our Sounds like a chip. GLP-1. Yeah, it
14:43does. Yeah.
14:43>> GLP-1. Is there Is there a two coming?
14:46Uh
14:47>> [laughter]
14:47>> there's a there's a three coming. And
14:50so, there is a GLP-2. That's actually a
14:52real uh Is it really? Okay.
14:54>> protein. It sort of does the opposite.
14:56Uh but,
14:58>> [laughter]
14:59>> Don't But, All right, don't want that.
15:01what we've done is combined uh two
15:04hormones
15:04>> Very small market. I got to tell you.
15:06Yeah, that's right. That's right.
15:08Extremely small market. Um I I don't
15:10know health care, but I'm guessing
15:11>> of on GLP-2 now cuz there's appetites
15:13two proteins fused together, and that's
15:15a technology we've developed, too. And
15:17there's a there's a a triple-acting one
15:19coming next year, '27. So, that'll be a
15:22pretty interesting. But anyway, we
15:24caught that wave. Our competitor
15:26innovated in obesity. We chased and
15:28executed, I think, quite well to become
15:31the market leader last year. So, we not
15:33only now have an engine that's faster,
15:35we have a tremendous success cycle with
15:37this one idea. The challenge before us
15:39now is like, how do we find another
15:42success cycle before that one runs out?
15:44Uh hopefully a lot sooner than it runs
15:46out. And
15:47you know, sort of get exit velocity,
15:49which really no major company in our
15:51industry has had. There's a They've just
15:53had these cycles. So, Yeah, you know,
15:54you You said something You said
15:56something earlier. And and
15:58uh I I I I have several of my employees
16:00here, and they they hear me say this all
16:02the time.
16:03You know, the fundamental purpose of the
16:05CEO is to create the conditions by which
16:08you know, not our job isn't to create
16:10ideas. Our job is to create the
16:11conditions by which great ideas can
16:14systematically,
16:15you know, arise.
16:17And and uh
16:19what what you just described is
16:20essentially, you know, Lilly with wisdom
16:22that that um you have to create the
16:24conditions by which these great
16:26inventions can happen. One of One of the
16:28you know, we we've spoken before about
16:30the early days of Nvidia getting into
16:32AI. And
16:33and I I would say I would say that that
16:36without exception,
16:37um probably the single uh most important
16:41thing that we did
16:43was investing in
16:45the infrastructure,
16:47the creating the conditions by which
16:49amazing AI researchers can come to our
16:51company
16:52and uh great great uh inventions could
16:56arise. Right. You know, today
16:59we're we're we're quite silent about
17:00this. Um but, people are in in-
17:03increasingly discovering that that um
17:05over the over the course of last 10
17:07years,
17:08one of the best decisions that I made in
17:09the beginning was was uh and at the time
17:12we could we couldn't afford it, um but
17:14nonetheless, I built one of the largest
17:15supercomputers in the in the world for
17:17Nvidia to use. Mhm. And at the time, I
17:20had nobody to use it.
17:22But,
17:23because Nvidia had very few AI
17:25researchers at the time. But, the
17:27problem is
17:28if if you don't have AI researchers, and
17:31therefore you don't have the instrument,
17:32if you don't have the instrument, you'll
17:33never have the AI researchers because
17:35they need Scientists need the instrument
17:38you know, necessary for for their
17:39science. If you If you want to attract
17:42the world's greatest uh particle
17:44physicists, if you don't have a uh a a
17:47particle a proton, you know, a collider,
17:50a large hadron collider, you're you're
17:52going to have a very hard time doing
17:53that work. And so, I came to the con-
17:56conclusion that the right answer is to
17:59to create the conditions for great uh AI
18:02work to be done. And I built that
18:03supercomputer. And we attracted some
18:05amazing people. And today, we do
18:07frontier work in physical AI and
18:09biological AI and robotics AI. And um uh
18:12the work that we do in self-driving cars
18:14is revolutionary. And And so, so which
18:17led led led our conversation Yep. to uh
18:23that that AI is is starting to to make
18:26enough progress that we might be able to
18:28apply it to tackle some of the most
18:31extraordinary challenges of humanity,
18:33which is biology. Yes.
18:35And And the two of us thought it'd be
18:37amazing if we have the the largest
18:40computer company in the world and
18:43computer science company in the world
18:44partner with the world's largest life
18:46science company.
18:47And it's as if if we created this uh
18:51lab where we are essentially one company
18:55combining all of the best amazing things
18:57together, which led us to our
18:59announcement today. Yeah. And so, So
19:01exciting. Yeah. Today we announced Yeah.
19:04>> [applause]
19:06>> Yeah.
19:07And so, and as you know, so there's like
19:09four parts to this. We we bought a bunch
19:12of chips, and we're building, I think
19:13it'll be finished this week, actually,
19:15the largest dedicated on-prem
19:18uh biology supercomputer in the world.
19:20Um that'll be in Indianapolis, Mhm.
19:22which is the world's center for
19:25biological discovery, by the way.
19:27You know, I was going to say I was going
19:29to say earlier [laughter] because of the
19:31selection of the headquarters of Lilly,
19:33it is exactly the reason why it has
19:35never merged with anyone. It maybe so.
19:37That could be true. That could be true.
19:40There's a lot of artifacts
19:41with that.
19:43I don't Yeah. You know, I think
19:46seriously, there's something to like if
19:48you manage to make it all the way to
19:50>> Everybody knows what I mean. Yeah.
19:52If you manage to make it all the way up
19:54that list
19:55and you're a discovery chemist or
19:57biologist, you really aren't going to
19:59leave, right? I mean, you made it all
20:00that way. Yeah, you're going to stay.
20:02So, we have great retention. Um anyway,
20:05we're building the this huge So, we have
20:06massive compute. Dave's invited me to
20:09visit him many times.
20:10>> Yeah, he's never
20:11>> And I'm Well, Actually, you did come in
20:122018. I don't know if you remember that.
20:14>> Yes, I did.
20:15>> Yeah.
20:15Actually, that just a sidebar, like that
20:18was the first time we met and that you
20:20gave a talk to our management team room
20:22about this size about AI in 2018. Yeah.
20:25And uh I guess I was lucky enough to
20:28100% of the audience I could see that I
20:30could see it in their eyes, nonsense.
20:32Yeah.
20:33Honestly, it blew my mind. It What you
20:35said blew my mind to the point where I
20:37spent several months like reading about
20:39some of these things which sounded like
20:41science fiction.
20:42>> Yeah. And now it's fact. Yeah. It's
20:44incredible what you've built and um it's
20:46great to partner with Nvidia on this.
20:48So, supercompute, we're going to put
20:50together a research team here in the Bay
20:52Area. So, if there's any smart um
20:55uh scientists or computer scientists who
20:56want to work at the center, it'll be a
20:58joint Lilly Nvidia
21:00lab AI lab. We're going to develop new
21:02data cuz anybody who knows about this
21:05problem we have is you can train models,
21:08but if the data is an inch thin you
21:10don't get very far. That's right. And
21:12biology is largely unknown. So, we need
21:14to create massive experimentation for
21:16the purpose of training machines and
21:19then we need to build on the work that
21:20Nvidia's done uh with the models you've
21:22created to ladder up to even more
21:24sophisticated predictions. So, um super
21:27exciting announcement.
21:28>> so, you know, one of the the important
21:30things to understand about AI is it
21:32it's easy to think about it as a
21:33multi-layer cake. I mean, at the lowest
21:35level you need some energy
21:37uh for these computers. And then, of
21:39course, you have the chips in the
21:41computers.
21:42You have the infrastructure layer. That
21:44infrastructure layer is is about, you
21:46know, how do you how do you cause this
21:49computer? It's kind of like the
21:50operating system of AI. How do you cause
21:52this computer to learn, process the
21:54data, learn? How do you guardrail it?
21:57How do you fine-tune it? How do you
21:58teach it? And so, that layer is is kind
22:01of like a the HR department of AI, you
22:04know? It's like it's like, you know,
22:06Only they're they tell better jokes.
22:08Exactly. [laughter]
22:10And so, that
22:11that's it. Yeah. And and so, they
22:14That that's the infrastructure layer and
22:15then the layer above that is the models.
22:17Yeah. And the layer above that, the
22:19domain the domain uh specific data and
22:23flywheel. And one of the areas that
22:25that's really exciting
22:27is it is a
22:30uh the idea that we would we would be
22:32able to uh train models
22:34uh synthesize proteins or chemicals go
22:37put it into an a robotics lab, collect
22:40more data, take that data back into our
22:42model and this this flywheel this
22:45flywheel
22:46of uh of um
22:48uh
22:49of data generation or pro- or model
22:52uh protein generation uh using AI, uh
22:55collecting ground truth, experimentation
22:58and then putting that scientific
23:00flywheel
23:02on steroids.
23:03Um that that future is just incredibly
23:06exciting. And so, so we're we're uh
23:08we're systematically uh bringing
23:10together some of the brightest minds in
23:13in the field in the domain expertise of
23:15drug discovery
23:17and some of the brightest minds in
23:18computer science.
23:20And uh together, you know, between the
23:22two of us in our co-innovation lab,
23:24we'll have essentially every layer of
23:26that cake. Yeah. And put together uh you
23:29know, hopefully hopefully a blueprint
23:32for uh what is possible in the future of
23:34drug drug discovery. I was going to tell
23:36you this I was going to tell you earlier
23:37I was going to tell you the story. 40
23:38years ago
23:40So, um uh
23:42I I guess 43 years ago
23:44I I was the first generation of
23:47engineers where the concept of
23:50computer-aided design Mhm. was invented.
23:54Before before my generation
23:57uh
23:58my manager
23:59uh
24:01would
24:02that generation of engineers, they would
24:05they would design a chip
24:06and
24:07and uh it would be a hit.
24:10But, they would go several years without
24:12another hit. Mhm.
24:14And they would design another chip and
24:15then it would be a hit. And they would
24:17go several years without a hit.
24:18>> works now. Exactly. Yeah.
24:20Exactly. And and our generation came
24:23along
24:24and we were able to represent the
24:27transistor, the logical gates, the
24:29functionality. Mhm. We learned the
24:31representation and described it in
24:33software. Yeah. We did 43 years ago
24:37um finally a computer could could
24:39represent your domain and accurately
24:42simulate its its behavior.
24:43>> That's right. To do as much in situ in
24:46silicon as possible. To do as much as in
24:49silico as possible. And so, today we
24:52design
24:53massive chips. 15,000 people come
24:55together to design a computer system and
24:57we just we just taped out Vera Rubin.
25:00Billions of dollars of R&D goes into it.
25:02Uh the computer when it's done is 3
25:04tons, you know, million and a half parts
25:07and and uh uh six brand new chips, 220
25:11trillion transistors. I mean, it's
25:13gigantic systems.
25:1615,000 people coming together and that
25:18computer is going to work first time.
25:20It does exactly what we expected. And
25:23every single generation we build upon
25:25the last generation. Every single
25:26generation build upon the last
25:27generation. I think your time
25:30has come for that. Yeah. Finally, the
25:32computer, you know,
25:35a million times a million times a
25:37quadrillion times of computing power
25:39later I think finally we might be able
25:42to represent one of the hardest problems
25:43in the world, which is the human
25:44biology. And I'm really hoping that your
25:47industry moves from drug discovery,
25:50which is kind of like wandering around
25:52the forest looking for truffles
25:55Which literally sidebar we used to have
25:58a department of soil discovery.
26:01>> [laughter]
26:01>> Seriously, where we would send people
26:03into tropical forests and scoop up soils
26:05and bring them back to Indianapolis.
26:07Uh-huh. And refine out what became
26:09antibiotics. In fact, the most important
26:11antibiotic in the world still is
26:12vancomycin. It's a last resort
26:14antibiotic in the hospital. We couldn't
26:16run hospitals without it. Was discovered
26:18in Borneo in a soil sample by Lilly.
26:20Now, how did they know to go there? They
26:22would they would just go to exact
26:23because fungus produce
26:25uh as
26:26tools to fight bacteria and so, they'd
26:28dig in the soil and discover these
26:30naturally occurring defense mechanisms,
26:33refine them into drugs. Huh. But, that
26:35is true drug discovery when you think
26:37about it, right? That's that's like you
26:40could get malaria doing that job.
26:42>> [laughter]
26:43>> That was that's and that was in the
26:45'50s. That was, you know,
26:4770 years ago.
26:47>> Well, see, now your job gets to be just
26:49like my job. You never have to leave
26:50your office.
26:51>> There we go. Yeah, we do it all in
26:52silico. Yeah. Yeah, so I I think that
26:54that that's what really excites me, the
26:56concept that we could have
26:57computer-aided drug design. Yeah. And
27:00that that this is going to be of course
27:03the the science behind it is going to be
27:04incredible. It's incredibly complex.
27:07The human biology is incredibly complex.
27:09It's going to be still a decade yet, but
27:12I I I really think that this
27:13announcement between us you know, could
27:15be a blueprint
27:16>> start of something.
27:17>> of something. I hope you're right. I
27:19mean, we need it, right? I think if we
27:20think of all the unmet needs left at a
27:22conference at JP Morgan, you're reminded
27:24of like the incredible pipeline of the
27:26industry but most of that is sort of
27:28like
27:29chipped out of stone, you know, it's
27:31like hard work. And and especially um
27:34occasionally we we get we we can harness
27:36systems like when monoclonal antibodies
27:39came around, we we basically hijacked
27:41like biology's way of making solutions
27:44to infections and other things. But,
27:46mostly what we do is like really
27:48empirical and like chipping away and
27:50very difficult. Each small molecule
27:53discovery is like a work of art. Mhm.
27:55There's only one of them and
27:57you know, one brushstroke and it's
27:58wrong. Um if we can make that an
28:01engineering problem
28:03versus a you know, the sort of discovery
28:05and artisanal drug making problem uh
28:08think of the impact on human life.
28:11The first the first chips that that I
28:13worked on, my my first generation of
28:15computers were quite empirical. There's
28:17some rules that we used and we apply
28:20some intuition. We have no ex- we have
28:23no clue how it's going to perform. We
28:25know that it's it's going to perform
28:26better than the last one we built
28:28because we you know, we made mistakes
28:29and and so, we could we could at least
28:32correct those mistakes and we'll have
28:33more transistors and we'll more
28:34everything. So, it's going to do better.
28:36Yeah. However, the functionality of it
28:39unknown, it's full functionality
28:41unknown, its performance completely
28:43estimated. Today we know the
28:45functionality of the systems we created
28:47down to every single bit, the
28:48performance we could measure, we
28:50simulate everything. Yeah. And so, we
28:52know exactly what we're going to
28:53produce. I I really believe that that
28:55that could revolutionize the future
28:57>> It would if we can make that come true
28:59even for a part of human biology. Yeah,
29:01that's right.
29:02And so, tell me
29:04um
29:05GLP-1
29:08I
29:09you were saying that it's really in its
29:10third phase. Yeah, already. Yeah. And
29:13and where does it go from here? What
29:15happens? What other things can you use
29:16it for? How could it change everything?
29:19Yeah. Well, I think as it turns out, you
29:21know, obesity is like a key lynchpin
29:24driver of chronic disease in in adults.
29:27And um
29:29like a lot of things in evolution
29:32um most the elements of today are
29:34weren't weren't concerns in our
29:36evolutionary past, but we evolve much
29:39slower than our environment is and so,
29:41we're maladapted to the
29:43to the modern world and obesity is one
29:46of those things. We grew up in a world
29:47of scarcity. Sounds like you grew up in
29:50a world of scarcity, but most people
29:51here didn't. Uh, ourselves, our
29:53ancestors did. And so, um, that's behind
29:56me.
29:56>> There was no
29:59>> [laughter]
30:00>> I just I don't want I don't want you
30:02guys to get the wrong impression.
30:04There was no need for an off switch on
30:06hunger.
30:07>> The word is abundance now. Abundance,
30:09yes.
30:09>> Yes, yes, yes. And we live in a world of
30:11caloric
30:11>> Abundance of intelligence, abundance of
30:13productivity, abundance of those kind of
30:15things. I don't know what you were
30:16thinking about, but my life is
30:18surrounded by abundance. But
30:20unfortunately, we're all in caloric
30:21abundance now. And there was no need for
30:24us to wire in an off switch
30:26evolutionarily. So, um,
30:28that's why we have obesity, and that
30:30leads to many, many chronic diseases
30:33that ail people and probably shorten
30:35life. Um, so, um, I think by having a
30:39pretty effective set of tools, going to
30:41get more effective, and I'll get back to
30:42that in a second. Um, to really help
30:44people manage that, to create an off
30:46switch synthetically in a way. Yeah. Um,
30:48we can down regulate a lot of diseases,
30:51over 200 chronic diseases. No kidding.
30:53So, the the two big evolutions that are
30:55going to happen in the next 24 months
30:58are a real progress on more choices. So,
31:01GLP-1 is one peptide, but it's in a
31:04family. And by combining two in in
31:07tirzepatide, GIP plus GLP, we see we can
31:09get better performance and better
31:11tolerability. I think you can expect
31:13more of that kind of thing in the
31:14future. So, within this super family of
31:16proteins, we can harness their combined
31:19strengths to really have much more
31:21tolerable drugs that are even more
31:22effective. Mhm. Um,
31:24effective in the maintenance phase,
31:26effective in inducing weight loss,
31:27different, uh, patient characteristics.
31:30That's coming, and I as I said, our
31:31triple
31:31>> reducing weight loss has so many So many
31:34downstream
31:35>> And that's the second thing, which is
31:36use cases. So, we know you lose weight
31:39if you take our Zepbound today, you'll
31:41lose 23% of your body weight on average.
31:43But what we're starting to uncover is a
31:45series of obvious things, like reduced
31:48heart attack and other metabolic systems
31:50that improve. Diabetes, prediabetes,
31:53um, conversion to diabetes drops by 93%.
31:56Imagine if we treated the country where
31:59we have 70 million people with
32:00prediabetes with Zepbound, we would we
32:03literally would have 93% less diabetes
32:05in our country. That that's a profound
32:07finding, and hopefully with the
32:09government access deal we signed with
32:10the Trump administration, we'll be able
32:11to do something like that.
32:13But there's many, uh, now non-obvious, I
32:16think, use cases, or less obvious.
32:17Inflammation being a primary one. So,
32:20um, being obese causes excess
32:23inflammation, not acutely, like in
32:26response to an event, but chronically,
32:27and that's not good for you. Um, it's
32:29not good for
32:31your, uh, cardiovascular system. It's
32:33not good for your joints. And so,
32:34spontaneously in our trials, early on,
32:37people would report, "Gosh, I've lost
32:39weight, but actually, like I can
32:41I can stand up from a sitting position
32:44for the first time in 10 years.
32:46My knees don't ache as much. My back
32:48feels better."
32:49Or some people with, uh, Crohn's and
32:51colitis reported kind of resolution of
32:55their symptoms. So, why is this
32:56happening? It's because we're sort of
32:57taking down that chronic inflammation at
32:59a system level. So, we're doing a bunch
33:02of studies, the first one read out last
33:03month, on top of, uh, medicine for
33:06psoriatic arthritis, a kind of
33:08arthritis, and it boosted the effect of
33:10that medicine by about 50% to co- uh,
33:13prescribe Zepbound with it. So, we're
33:15going to take that on medically, because
33:16that's how the medical system adopts
33:18innovations. They need the use cases.
33:20Maybe similar to adopting AI technology,
33:22you need to prove the value in a partic-
33:25a particular setting.
33:27And then, we're also starting a number
33:29of brain health studies right now, and
33:30we've have a uh, special purposed uh,
33:33GLP-1 GIP for brain conditions,
33:36particularly addiction,
33:37which, you know, is really a major
33:39modern problem.
33:41And, uh, it turns out that for similar
33:44reasons that these these medicines down
33:47regulate your appetite,
33:50which is
33:51it turns out not really driven by our
33:53need for more calories, but driven by
33:54other self-learned behavior loops.
33:58Sounds like AI more and more, doesn't
34:00it? Um,
34:01that, uh, it can do the same thing or a
34:03similar thing for other
34:06harmful self-learned behavior loops,
34:08like
34:10gambling, like alcohol consumption,
34:13smoking cessation. So, we're exploring
34:15all of these as well, which are, you
34:17know, including all we're going to take
34:18on.
34:18>> doesn't I hope it doesn't affect working
34:20too hard.
34:21>> No. Okay.
34:22>> It's only the adverse ones we're trying
34:24to edit out, yeah. It's only the adverse
34:26ones. Um, but it's a common pattern with
34:29over food consumption, actually. There's
34:31a parallel for that.
34:31>> That's just a comment for all the
34:32startup CEOs. You got a lot of startup
34:34CEOs in the in the room here. Be
34:36careful. If you find your your your
34:38willpower starting to diminish,
34:41stop microdosing. Yeah, stop
34:42>> [laughter]
34:43>> microdosing.
34:45Yeah. So, um, anyway, that's very
34:47exciting.
34:47>> interesting. Yeah. And then, you know, I
34:49think on the horizon, uh, people are
34:51working on much longer acting, so you
34:53can take these less frequently. Of
34:55course, we'll launch our oral GLP-1 this
34:57spring, which I think is a major
34:59breakthrough, not just for convenience,
35:01but actually global reach. Um, it looks
35:04simple, our medicine in a auto-injector,
35:07you know, but it's actually the most one
35:08of the most complicated drugs marketed
35:10in the world, uh, from manufacturing
35:12perspective. And we have to do it times
35:14hundreds of millions based on the
35:16current demand.
35:17We just launched in China and India and
35:20in Southeast Asia. There's a lot of
35:22obesity there. We cannot possibly meet
35:24all that demand with just injectable
35:25systems. So, we need to get
35:27>> Southeast Asia
35:28really serious diabetes. Huge problem.
35:30Yeah. Yeah. And Southeast Asians sort of
35:33lost the genetic lottery of obesity
35:35because they're much at lower body
35:38weights,
35:39small gains in obe- small increases in
35:42the amount of obesity trigger more
35:43diseases.
35:45So, um, they're
35:47they need to to use, uh, something to
35:49mitigate obesity. Hopefully dietetics,
35:51if not, they can use our medication. But
35:53we have to make enough, and having an
35:54oral form that's a chemistry, it's not
35:57relying on protein synthesis and then
35:59putting it in these sterile injectable
36:01systems, just an oral pill, we can scale
36:03that massively. Man. Um, so, that will
36:06launch this year as well. So, huge, uh,
36:08This one idea, you're going to ride this
36:10for another 50 years.
36:11>> Right, right. We're hoping to. We're
36:13learning from you.
36:14>> My gosh.
36:15>> [laughter]
36:16>> Hang on a second.
36:18I'm convinced I'm convinced
36:20GLP-1, it's the second best technology.
36:23That's right, [laughter] exactly, yeah.
36:25Ours is not universal
36:27>> Blackwell, Blackwell's number one. GLP-2
36:29GLP-1 number Ours is not universal to
36:32the entire economy, just to the human
36:34economy.
36:34>> [laughter]
36:36>> That's incredible.
36:37>> Very exciting. And you know, what So,
36:39you know, one challenge we have is like,
36:41can we keep inventing things that are
36:42better and better enough that people
36:44will pay for them to improve beyond the
36:46patent cycle. But has there ever been a
36:49a drug has there ever been a life
36:51sciences technology that has benefited
36:54so broadly humanity? The The ones
36:58>> The only other I I'd say it's a class of
37:00drugs, and the only class that did
37:02something like this, I alluded to
37:04earlier, which is antibiotics. The
37:05discovery of penicillin invented modern
37:07medicine. If we subtracted antibiotics
37:09from our lives, we couldn't have modern
37:11dentistry. Yeah, right.
37:13>> We couldn't have surgeries. We couldn't
37:15do a lot. So, now we take that all for
37:17granted.
37:18But in, you know, when Fleming just, you
37:21know, self-dosed this weird mold he grew
37:23in his lab, that was the start of like a
37:25major wave, and that lasted into the
37:27'90s. So, that was a 60-year
37:30cycle with 100 plus new medicines.
37:34I think this could be like that. Yeah.
37:36Um, which would be profound, but this is
37:38not for an acute infection, it's for
37:39chronic disease of the modern homo
37:42sapien. That's That's what we're
37:44affecting.
37:44>> did you ever realize that you were going
37:46to have such an easy job?
37:48You know,
37:50my wife asked me that a few times, like,
37:52I mean, just Why aren't you working on
37:54>> one incredible discovery, one incredible
37:56I mean, you were just running around the
37:58forest and you discovered this
38:00this thing called GLP-1. Yeah. That's
38:02one of Imagine, you're digging up stuff
38:04and you found this incredible truffle,
38:06and it's so it's solid for 60 years.
38:08Yeah.
38:09Actually, you know, so the What's the
38:10history of this? So, um,
38:12>> It goes to the '70s, and, um, a German
38:15scientist discovered this thing called
38:16the incretin effect, and we call these
38:18GLP-1s part of this incretin family.
38:20These are hormones excreted by our gut,
38:23which signal the rest of the body that
38:24you're fed.
38:26And, um, we overboost those, and that's
38:28what is happening. But he observed that
38:30if you
38:31give sugar
38:33orally
38:35versus intravenously,
38:37you get a very different insulin
38:39response. Insulin's like the opposite,
38:40it stimulates absorption of sugar. It
38:43was slower, um, if you give it orally.
38:46And so, that's what's called the
38:48incretin effect, which was the first
38:50insight that somehow your body's doing
38:53something different
38:54when the gut is involved than when it's
38:57not.
38:58And that then led in the '80s to the
39:00discovery of actually GIP first, then
39:02GLP. Those are the two parts of
39:04Zepbound. And later, a whole 'nother set
39:06of, um,
39:08hormones, but, uh, one of the things I
39:10hope we can crack with our supercomputer
39:12and our team
39:13is here we had targets,
39:16but really difficult to drug. The
39:19half-life of the native human form of
39:20GLP is like 7 minutes.
39:23So, to give that as a medicine, the
39:25native peptide sequence, which was
39:27identified in the '90s, you'd have to
39:29like infu- walk around in with an
39:30infusion all day and all night. So, that
39:33doesn't make for a product. Um, it
39:35wasn't until we figured out how to
39:36extend that life, first to twice a day,
39:39and then to once a week, and now people
39:41are working on monthly and even beyond
39:43that we could make that a effective
39:46product. So,
39:47one
39:48thing in AI drug discovery that is
39:50really working is this ability to
39:52uh create massive combination
39:55possibilities and then test those in
39:56silico and and filter out good ideas.
39:59We're doing that at scale now. We can do
40:00that at an even bigger scale with our
40:02collaboration.
40:03The other side is discovering so that's
40:05the key discovering the lock.
40:07Can we find more biology targets
40:10using uh AI? And that is the really the
40:13holy grail cuz if we put those two
40:15things together then we can model the
40:17whole system at once. Could be could be
40:19super exciting.
40:20>> Yeah, one of the things that that that
40:23that Dave really wanted to do was to put
40:25the AI lab the co-innovation lab here
40:28Yeah. in Silicon Valley and even close
40:30right here in AI City, you know, San
40:32Francisco is really epicenter of a lot
40:34of AI researchers and we have an office
40:36here and and so this is going to be
40:38either either in South San San San
40:40Francisco Which is really the biotech
40:42hub of the Bay Area or we could do it up
40:44here. We We want to go where people want
40:46to live. Yeah. Yeah.
40:48Yeah.
40:48Well,
40:50San Francisco's pretty cool. It's cool.
40:52Yeah. Yeah, yeah. It's cool.
40:53Indianapolis is cool as well. Yeah.
40:56Yeah. Yeah, I've heard from from from
40:59from just one person.
41:01From
41:02one other person. We're very adaptive,
41:04you know, we we used to be good at
41:05basketball but now we're good at
41:06football. Uh-huh. And so
41:09the uh
41:11uh
41:12so we used to be good at
41:14chemistry but now we can be good at AI.
41:15Or you're just good at something. Yeah,
41:17we're good at something. Not sure.
41:18Better than the alternative.
41:19>> Yeah, yeah, it's it's incredible. And so
41:22so um
41:24I
41:27GLP-1 beyond
41:30you know, one of the things that that uh
41:32a fair amount of Silicon Valley's
41:34talking about
41:36longevity and Right. Aging. Yeah.
41:40And and the absence of Yeah. Yeah.
41:42Subtracting aging. Subtracting aging,
41:43yeah.
41:44>> Yeah. And and so so tell me is there
41:46what is the technology behind that? Is
41:49it
41:50Yeah. Well, I mean it's an interesting
41:52topic. Obviously it's it's attractive to
41:54a lot of people because who wants to get
41:57old?
41:58Um
41:59the Sometimes I just want it to be over.
42:01There are a number of yeah. Yeah. Yeah,
42:04I get that too. Yeah. [laughter]
42:06Um the
42:10there's a lot of biology on what why
42:11individual cells expire and you can
42:15extrapolate that to cell systems and
42:17then organs and then potentially the
42:18whole organism.
42:20But as we were talking off stage, that
42:23is not really a problem we work on that
42:26way. Mhm. So that idea is essentially
42:28can we extend the maximum shelf life of
42:32a human
42:33to something beyond what that is under
42:36perfect conditions. So let's just say
42:37under perfect conditions we're meant we
42:39can live up to 100 years.
42:41And then for maybe good reasons
42:44the system shut down, right? I think
42:46there it would be a weird world if
42:48people live forever, you know, it would
42:50why would we reproduce? How do we
42:52evolve? Like the there's a lot of
42:54problems with that kind of
42:56even philosophically.
42:58>> nature fix that original design? Yeah.
43:00Even philosophically though like then
43:02yeah. It yeah.
43:05We would have a a very weird world I
43:07think. But I understand why people
43:09pursue it. We're more focused on a more
43:11practical problem which is if the if the
43:13maximum shelf life is 100 years, how do
43:15we get more people a better shot to get
43:17that full life? Mhm. And the way we do
43:20that is by knocking out disease. So the
43:22reason people perished, if you look at
43:24life expectancy in our country in 1900
43:26it was like 46 years old for a male.
43:29And the biggest factor there was child
43:31death.
43:32So you lost life because you died when
43:34you were young at a pretty high rate
43:36from infection mostly and by having
43:38antibiotics and vaccination and other
43:41tools to manage childhood disease that
43:43quickly shot up.
43:45From 1960 till now
43:47we've added about 10 years of life
43:49expectancy mostly through chronic
43:51disease extension. Cancer care has
43:53improved dramatically. Most companies in
43:56this room have had a big role in
43:58those chronic disease management. HIV
44:01AIDS and other new pathogens have been
44:02conquered and that's you know, now we're
44:04we're up to almost 80 years in our
44:06country.
44:07Actually the the the number would be
44:08more like 84 if it weren't for opioids
44:10and traffic deaths
44:12which is exceptionally high in the
44:14United States both of those. So
44:16self-driving cars can help with one of
44:17those.
44:18Maybe GLP-1s will help with the other
44:20because we're studying opioid addiction
44:21actually
44:22as a as a solution there. But we chip
44:25away at it this way. And of course as
44:28people live longer the incidence of new
44:30maladies of aging rise.
44:33So the next frontier for us beyond GLP
44:35that I'm really
44:37excited about and interested in is is
44:40dementia and diseases of the aging
44:42brain. Mhm. Because as we as our the
44:45rest of our bodies
44:46are able to carry us into the 80s and
44:4890s this is something we haven't really
44:51figured out
44:52how to address and for quality of life
44:54but also quantity of life. Um that's
44:57that's going to be important frontier.
44:59We're spending a lot of effort in that
45:01area. We've got some exciting programs
45:03running now in studies but I think you
45:05know, through drug discovery that's um
45:07simulating brain tissue and how to look
45:10for new targets and the interaction
45:12between targets we don't understand.
45:14There's inflammation plus uh protein
45:16misfolding happening in the brain the
45:18aging brain. Perhaps we can discover
45:20something really exciting and new. And
45:22so if you take a step back and look at
45:24our
45:25now looking at our co-innovation lab the
45:27the first thing that we have to do is we
45:29have to set up the computing platform
45:31for the researchers to do work. And so
45:34we have we have the computers, we have
45:36we have the the infrastructure layer and
45:38we have BioNeMo and uh we have a whole
45:41bunch of of
45:42of models that model models and model
45:44architectures we've already invented and
45:47and we'll get our scientists together
45:48and they they can they can start
45:50collaborating in in that particular area
45:51just building the foundations. But but
45:54if you had a wish list for the
45:56researchers you know, I know that that
45:59and certainly it is the case with with
46:00me. I try to influence the research
46:03>> the direction of it.
46:04>> But I try not to be directive.
46:06>> Yeah.
46:06And and obviously the reason for that is
46:08because we don't know what we don't know
46:09and we don't know what can be discovered
46:12um and there are many things that are
46:13discovered that quite frankly we didn't
46:15even know was possible. Yeah. But once
46:17it's it's created uh you know, our our
46:20our innovation brain could kick kick in
46:23and and together try to figure out how
46:24to bring that to the market.
46:26Uh but but if you had a direction that
46:28you would like to set them on uh in in
46:30what direction would you would you hope
46:32that they do some research? Yeah, I
46:34think there's like two
46:36basic problems and I think within these
46:37we'd have to narrow the scope to make
46:39progress and feel achievement. Um the
46:41first one there's a lot of momentum um
46:43in our space already which is what I
46:44talked about the the the drug design uh
46:48engineer drug design which I mean you
46:50could argue a lot of drugs these days
46:52you fall into that category. You know
46:53the target cuz it's been around and well
46:54profiled and you're better customizing
46:57kind of the the key to the lock.
47:00Um but there are new modalities that
47:02have emerged you know, RNA
47:06uh gene therapies etc. They really
47:08haven't been optimized because they
47:09haven't been around as long. So I think
47:11that whole uh suite of problems around
47:14uh drug engineering drug optimization is
47:16pretty tractable and I think the NVIDIA
47:18scientists coupled with ours could make
47:20a lot of progress pretty quickly there.
47:23Um that solution by the way would be um
47:25highly amenable to our Tune Lab um kind
47:28of interface we've built Yeah. um for
47:30all of our biotech partners as well.
47:32Better drug making through uh in silico
47:35design engineered And and just so you
47:36guys know there's anybody who doesn't
47:38know about Tune Lab this this is this is
47:40a a platform
47:42where where uh Lilly and startups or
47:46third parties uh could work together to
47:50uh con both contributing data Yeah. uh
47:53to advance a particular space and it's
47:55built on top of a platform
47:58that is called federated learning. It's
48:00called MV Flare that we contributed to
48:03uh so that so that um
48:05developers and researchers
48:07uh don't have to aggregate all of your
48:09data commingle your your data but you
48:10could still train a train a model
48:12together so that you can benefit
48:14together. The the thing that's really
48:16since our announcement all day long the
48:19number of startups have been really
48:21excited about Tune Labs and and
48:23co-innovation and how they could be
48:24involved are just coming out of the
48:26woodwork and
48:27and so I think the the the vision you
48:29have for Tune Labs and the work that you
48:30did in building up Tune Labs is
48:32completely visionary and and it makes
48:34possible for collaboration of AI
48:37researchers that are that are not in the
48:38same company and it's very Exactly,
48:40yeah. Protecting the intellectual
48:42property but the process of improving
48:46their idea improves the training of the
48:48model itself. So we all gain by
48:50participation. There's no fees. We do
48:53curate it so there's like legitimate
48:54biotech projects but um we will add the
48:58product of our work together to the Tune
49:00Lab interface and so not just Lilly
49:02scientists can benefit many biotech
49:03scientists. The other side of the
49:05equation of course is target. And here I
49:08think we need our robotics uh effort to
49:11really kick in.
49:12Why do we have more targets and why do
49:14we understand the targets that we have
49:15better which are probably um
49:18structurally more complicated than we
49:19understand and that's one of the reasons
49:21drug drug development fails is we think
49:23we have the target but we don't really
49:24understand it. And that's because we
49:26need massively more amounts of um
49:29experimental data on specific good
49:32targets.
49:33Um so I think here we'll work with you
49:36on flushing out We already have one
49:38robotic lab um running, wet lab, but
49:41we'll build more. And we can have them
49:44running 24/7 experimenting around a
49:48target space to really fully profile
49:50that and then couple that with those
49:52engineered drug designs. So, that's a
49:54little bit harder problem. It's more
49:56empirical, but I think machines are made
49:59to work day and night to solve this
50:02problem. And that flywheel is going to
50:03is going to fly. I think you know, Dave,
50:05if you if you look at some of the the
50:07advances that we've had in other
50:09domains.
50:11Um the first the first thing is you need
50:13the flywheel to eventually land a ground
50:15truth. So, you have to have physical
50:17world experimentation, collect that
50:19data.
50:20Um and and you close the loop that way.
50:22Uh eventually that model becomes
50:24sufficiently good. That foundation model
50:26becomes sufficiently good.
50:28And you could teach it how to generate
50:30new be generative.
50:32And eventually
50:34there's a there's a model
50:36that is being tested against another
50:38model.
50:39And this other model is the world model
50:42or the target model.
50:44And and uh the target model along with
50:47the model that you you're training uh
50:49runs in a closed loop. So, essentially
50:52the the lab is another model.
50:55And so, you eventually have to to to
50:57ground it again by producing and
50:59synthesizing the proteins and and
51:00collecting real data. But we we get
51:02ourselves in a synthetic data flywheel.
51:05And that synthetic data flywheel is
51:07incredibly effective.
51:08>> I don't think that's been attempted in
51:09our industry and at scale and I think
51:11we'll do that here. But I think really
51:13uh taking a handful of targets which we
51:16have highly suspect will be useful in
51:19key diseases.
51:20And really doing the full sort of
51:23profiling of those in
51:26um robotic wet labs to train the data
51:28set and then use the machines to fill in
51:31the
51:31fill in the space, make better drugs cuz
51:33we understand
51:34>> so exciting. It's very it's very it's I
51:37think it's singularly true that finally
51:40we're going to have we're going to have
51:41a lab where the expertise and the scale
51:45of that lab is going to be sufficient to
51:47attract people who really want to do
51:49their lives work in that intersection.
51:51Yeah, we'll be so excited. Yeah, I'm
51:52super excited about that. And so, one of
51:54the things that that that is really cool
51:56about this conference
51:58is large companies and small companies
52:00getting together. Um this year uh
52:04AI scientists
52:06Uh there's a company I met today and
52:08we've been working with them for for
52:09some time. I hope they don't mind that I
52:10mention their name, Edison, is really
52:12terrific.
52:14Uh
52:14um a
52:16a foundation model
52:18uh company who who is going to build
52:21um both
52:22the life sciences lab as well as the AI
52:25lab
52:26uh where the the the factory
52:29a a real robotics factory and and a a
52:32large scale system uh Lyra uh
52:35uh incredible companies. They're they're
52:37probably in the audience. Uh what are
52:39what are some of the companies that
52:40you've seen? So, so robotics labs
52:42companies, um AI scientists companies,
52:45of course uh agentic health care um
52:49assistants
52:50uh the work that Abridge is doing and
52:53Open Evidence is doing. Really really
52:55cool work that's that's being done
52:56there. Uh what are some of the some of
52:58the areas that that you've seen that are
53:01really cool?
53:02Yeah, I mean it well, I think at the for
53:04AI [clears throat] the most tractable
53:05problem now that you're seeing more and
53:07more on is health care service. So, this
53:09is not as complicated as what we're
53:11talking about here today, but rather how
53:13do you replace human level service with
53:16agentic service, AI service. And it
53:18seems so obvious for what is like the
53:20least productive part of the entire
53:22economy
53:23to adopt AI, but there are a lot of
53:24frictions for that. So, a lot of our
53:26discussions are well, how can we
53:28collaborate, you know, we run I think
53:30it's probably the largest pharmacy
53:32direct pharmacy platform in the country,
53:35Lilly Direct. Um it's only a year old,
53:38but um it's annualizing close to a
53:41billion dollars uh a quarter right now,
53:434 billion a year, just selling
53:47pharmacy. So, people want to say, "Well,
53:48can Lilly work with us
53:51to kind of use your market weight to
53:53force adoption and sort of work against,
53:55you know, let's be honest, insurance
53:57companies that like they don't they
53:58don't into change.
54:01And physician groups are frightened by
54:03some of these technologies, but we
54:04really do need to make the US health
54:06care system more efficient and in doing
54:08so, we can we can actually have more
54:10room for innovation, more room for
54:13adoption of cutting edge medicines and
54:16maybe more accurate and comprehensive
54:18workup of people.
54:20Um so, that's that seems inevitable, but
54:23what can we do to speed that up and
54:25really harness that? I'm there's a lot
54:27of new modalities as I was mentioning
54:29earlier to drug targets. And really in
54:33the last 10 years without AI in the run
54:36rate, I think the productivity of the
54:37industry on new modalities is
54:39accelerating. And that's exciting.
54:42And yet it's early days. So, there a lot
54:44of companies are
54:45combining new and old modalities into
54:47bifunctional medicines. That's a whole
54:50hot field where you have a targeting
54:53part and then a warhead part um say to
54:55kill cancer cells.
54:57And even just refining how we get these
54:59medicines to the right tissue
55:01is a important problem and there's a lot
55:04of interesting new approaches, different
55:05approaches happening that I took
55:07meetings on today. That was pretty
55:09interesting to me. So, um that's the
55:11drug making side and I think here again
55:14AI can help us accelerate that, but a
55:16lot of smart people working on on those
55:17problems. Um
55:19yeah, and of course a lot of people want
55:20to talk about GLP-1s. That's that's my
55:22life. Yeah.
55:24Yeah, but it was good you said like
55:25small companies and big companies
55:27getting together. So, here like we're
55:29the small company and you're the big
55:30company. Yeah, yeah, cut it out.
55:32>> And and we're getting together. Yeah.
55:33Cut it out. Cut it out. But I was I was
55:35going to say like, you know, the
55:36probably Lilly has a reputation 10 years
55:39ago when I started, you were asking
55:40about 2017, of like a little bit of a
55:42sleepy uh Midwestern company. But here's
55:46a like fun facts for the computer
55:48people. I think the first
55:51you may may not know, but that the first
55:52approved biotech product made by living
55:55organism, a medicine made sort of in the
55:56most modern way, was a Lilly product,
55:58Humulin, which we collaborated with
56:00Genentech on.
56:01And then 12 years later
56:03we bought the first supercomputer in the
56:06pharmaceutical industry and put it in
56:08Indianapolis.
56:09It was a Cray supercomputer.
56:11It was called Big Red.
56:12Our our company logo is red.
56:15Very clever. Yeah.
56:16And um we designed the first approved
56:19medicine designed on a computer
56:22on that thing. It's called insulin
56:24lispro, which is Humalog. It's still uh
56:27used by type type 1 diabetics and type 2
56:29diabetics globally today.
56:31By playing around with amino acid
56:33sequences empirically and solving the
56:34>> trying to convince me that Lilly is
56:37cool. We are cool.
56:39We are cool.
56:40You know. We are cool. And uh we've been
56:42working with computers and the cutting
56:44edge of science and computer technology.
56:47>> Yeah. I I was convinced at hello.
56:50>> just makes tons of sense. This is
56:51[laughter] a natural evolution of our
56:53company.
56:55So awesome. It is it is so awesome to
56:57work with you and your team. And I am so
56:59excited about the work that we're about
57:01to do together. I I think that that um
57:04I
57:05I can't imagine a more worthy, a more
57:08complex uh more exciting field to go
57:12apply computer science to.
57:14And and uh
57:16for someone who's dedicated a a whole
57:18career in basically one job uh building
57:21computers uh the opportunity to work
57:23with with you and and Lilly and the
57:26scientists there uh in service of of the
57:29next breakthrough can't be more
57:31exciting. And and so, I want to thank
57:33you for your partnership and uh for
57:35personally driving this and and uh
57:37the two of us uh uh you know
57:40teaming up together, pulling together
57:42the the mights of our two company, you
57:44know, hopefully we could go bend bend
57:46the arc
57:47of of history and
57:49see we can make a difference. I'm super
57:51excited.
57:51>> exciting to hear you say that. Thank you
57:53for helping make this come together.
57:56We're just getting started, but this
57:58could be you know, a new day. Yeah. And
58:01if we can accelerate progress even even
58:04if we set modest goals, I think we can
58:06accelerate progress in, you know, a use
58:09case for computer technology that has to
58:11be the most personal and impactful um
58:15for our country and for for the world.
58:17And for an industry that really needs to
58:19find more ways to be more productive
58:22with R&D dollars. So, it's it's a worthy
58:25tribe, but I we're going to do more
58:27and win.
58:27>> We're going to make it happen. Ladies
58:29and gentlemen, David Ricks and the
58:31second most I'm just kidding.
58:33>> [laughter]
58:34>> Thank you. [applause] Thank you. Thank
58:35you. Great. Yeah. Thank you. Thank you.
58:39>> [applause]
58:40>> That's fun. That was fun to do. Yeah.
58:41Thank you. Yeah.

About this transcript

This page contains the full transcript of NVIDIA & Lilly: The AI Revolution in Drug Discovery | Jensen Huang & David Ricks by NVIDIA, generated from the public captions YouTube serves with the video. The transcript has 10,239 words across 1,700 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.

What you can do with it

Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to choose a caption track available on the video.

Free YouTube transcript tool

YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, caption language selection, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.