- il y a 2 semaines
The Rise of Sovereign Scientific AI
Catégorie
🤖
TechnologieTranscription
00:00Bonjour à tous, merci beaucoup d'être avec nous à VivaTech aujourd'hui.
00:10Je suis très heureux de vous présenter à Patrick Grady,
00:14CEO de Tretra Science, Rory Care,
00:16Global Head of Business Development Healthcare at NVIDIA,
00:19et Emmanuel Frenard, Chief Digital Officer at Sanofi.
00:25Je vais commencer par vous, NVIDIA.
00:27C'est un peu comme le superstar de VivaTech cette année.
00:31La plupart des gens connaissent NVIDIA pour l'industrie du jeu,
00:35et pour l'AI, bien sûr.
00:37Donc, vous pouvez nous dire plus sur ce que vous faites dans l'espace ?
00:42Absolument, oui.
00:43Donc, 30 ans, NVIDIA a créé un type de compétition
00:47qui s'appelle Accelerated Computing,
00:48qui s'est réinventé l'architecture de l'architecture pour la première fois en 60 ans.
00:52Et évidemment, c'est le réseau de l'architecture de l'architecture de l'architecture à l'architecture.
00:58Mais Accelerated Computing n'est pas seulement sur les GPUs.
01:02Vous vous dites que NVIDIA est synonyme avec les GPUs,
01:05mais les GPUs font leur chemin à des systèmes,
01:08les systèmes font leur chemin à des centres de données,
01:11et vraiment, les campesnes de données des centres de données
01:14qui sont maintenant le mode de l'instrument du réseau de computing.
01:17Ce qui est le mode de l'infrastructure de l'architecture,
01:19qui est usheré dans cette révolution industrie.
01:21Et donc, à NVIDIA,
01:25nous avons aussi construit beaucoup d'informations
01:26sur l'architecture de l'architecture.
01:28Accelerated libraries,
01:30inference microservices,
01:32models,
01:33frameworks,
01:33blueprints
01:34qui permettent d'être facile pour les développeres
01:36de construire des applications
01:38à travers un nombre de différents verticales.
01:41Et donc, je spendais beaucoup de mes temps
01:42en l'architecture et en sciences de l'architecture.
01:44Et le façon dont nous avons l'architecture
01:45de l'architecture dans l'architecture,
01:48nous nous concentrons sur trois aspects.
01:51de l'architecture dans l'architecture.
01:54C'est l'architecture de l'architecture.
02:19La deuxième étape,
02:20nous sommes très, très focussés sur
02:22est l'architecture.
02:23Nous pensons à l'architecture
02:24d'agents
02:25pour transformer la expérience patient,
02:28avec un médecin ou un système de santé,
02:32avec des appareils,
02:33à faire des appointments
02:34à la visite de la routine,
02:37plus humaine.
02:37C'est l'architecture de l'architecture,
02:38où un docteur n'est pas seulement
02:39face à un computer,
02:40mais face à vous.
02:41Et la troisième étape,
02:43que je pense que nous allons parler
02:43beaucoup de aujourd'hui,
02:44c'est ce que nous referions
02:46à la biologie.
02:47Et comme nous le savons,
02:48il reste encore 10 ans
02:50et environ 2 milliards
02:51de dollars pour l'architecture
02:52de l'architecture.
02:53Et donc,
02:53il y a des opportunités
02:55de conduire l'architecture
02:57à travers l'architecture
02:58du phénomène de phénomène,
02:59mais surtout dans l'architecture
03:00et je crois que
03:01le développement de l'architecture
03:02est devenu industrie standard.
03:04chaque pharmaceutique
03:05est en train d'utiliser
03:06aujourd'hui.
03:08Donc, vous travaillez tous ensemble?
03:09Oui.
03:10Nous essayons de travailler
03:12très loin avec lui.
03:13Il a tout le monde,
03:13mais oui,
03:15nous travaillons ensemble.
03:15Ok.
03:17Donc,
03:18il y a beaucoup de choses
03:19qui parlent de l'architecture
03:20dans l'architecture
03:21dans l'architecture
03:21de l'architecture.
03:23Il aide à découvrir
03:24les drones.
03:25Vous avez dit
03:26qu'il y a des patients
03:28qui peuvent détecter
03:29les cancers
03:30plus vite
03:31mais
03:33vous avez le hype.
03:34et qu'est-ce qui existe ?
03:36Qu'est-ce qui fonctionne aujourd'hui ?
03:38Patrick ?
03:39Donc, le hype
03:39versus la réalité.
03:41C'est ça ?
03:42Oui ?
03:43Donc,
03:45TETRA Science
03:46operates
03:47across
03:48the biopharma
03:49landscape.
03:50Donc,
03:51nous avons une perspective
03:51vraiment intéressante.
03:52Aujourd'hui,
03:52nous avons environ 40
03:53leading biopharma customers
03:55including 12
03:56top 25.
03:57Et nous operate
03:58across the entire value chain
03:59from research
04:00to development
04:01and manufacturing.
04:02and what I would say
04:03from our vantage point
04:04is that we are
04:06really inspired
04:06and there's a lot
04:08to be hopeful about,
04:09a lot to be excited about
04:10and we are seeing
04:11what I would characterize
04:12as green shoots,
04:14emerging green shoots
04:15of innovation
04:15and the application
04:17of AI
04:17across that value chain.
04:19Whether it's
04:20in silico
04:21and digital twin
04:22technologies
04:23and drug discovery
04:24and through
04:25the development phase
04:26all the way through
04:27some early innovation
04:29manufacturing.
04:30but it's a bit
04:31of a mixed bag
04:32and I think we are
04:33going to dig into this
04:33which is
04:35where there's a lot
04:36of hope
04:36and a lot of hype,
04:37the reality is
04:38there are structural barriers
04:41to fulfilling
04:42the potential
04:43of scientific AI.
04:44And the prince,
04:45from our perspective,
04:46the principal obstacle
04:48to scientific AI
04:49is data.
04:51It's the preparation
04:53of the data,
04:54the availability
04:55of the data,
04:56the organization
04:57of the data.
04:58It's not whether
04:59we have enough data.
05:01Companies like Sanofi
05:02produce ungodly
05:04amounts of data
05:05and we'll touch on this
05:06I think in a minute
05:07but it's a data
05:08preparation issue.
05:09So McKinsey
05:10earlier this year
05:12published a survey
05:13with the top
05:13pharmas
05:14and 80%
05:16of the respondents
05:17cited data readiness
05:18and preparation
05:19as the number one
05:19obstacle
05:20to achieving scientific AI.
05:22So we're really excited
05:24by what we've seen.
05:25Sort of the art
05:26of possible
05:27is keeping us
05:28going every day
05:29but until and unless
05:31we solve these
05:32structural data problems
05:33we're going to come back
05:34every year and talk
05:35about the hope
05:35and promise
05:37of scientific AI.
05:38That's sort of
05:39our vantage point.
05:40And that,
05:41sorry to interrupt,
05:41that data specificities.
05:43When we think of data
05:44we often think of
05:45very structured data
05:46so that you can
05:47recognize
05:48first name,
05:49last name, etc.
05:49In our industry
05:51we've got very unusual
05:53types of data.
05:54It could be a scan,
05:55it could be a blood sample,
05:56it could be a biopsy,
05:57it could be...
05:57So how do you store that
05:59and how do you
05:59exploit that?
06:00That's where I think
06:01the magic of what
06:03technology can do
06:04help us exploit that data.
06:08You wanted to add something, Patrick?
06:10What's that?
06:10No, I thought you wanted
06:11to add something.
06:12No, no, no.
06:14But what's really,
06:15what we're missing today
06:17so you know
06:18you can go to the next
06:19level if I may say that.
06:22Because, of course,
06:22we have the compute today
06:23with NVIDIA,
06:25of course.
06:25We have the big pharma,
06:27we have the skills.
06:29So what's missing?
06:30It's really the data.
06:32Patrick?
06:32Yeah, so I think
06:34you know
06:35we have a pretty passionate
06:36and we think
06:36a pretty informed view
06:37of this.
06:38So to your point,
06:40Charlie,
06:41the Sanofis of the world
06:44have assembled
06:45some of the greatest minds
06:46in the world
06:46in science.
06:47they have vast amounts
06:49of data
06:50and I should probably
06:50dimension that
06:51for the audience.
06:53We have our friends
06:54here at NVIDIA.
06:55They bring all the world's
06:57compute to bear.
06:58We have companies
06:59like Databricks
07:00and Snowflake
07:01that are really good
07:01at data infrastructure.
07:03We have the cloud vendors
07:04that bring a lot
07:05of value to the table.
07:07What's essential
07:09to capitalizing
07:10on this opportunity
07:11is we,
07:12as an industry,
07:13hold hands
07:13and we treat this
07:15much more
07:16as a collaboration
07:17on behalf of humanity
07:19and downstream patients
07:20versus a very standard
07:22customer,
07:23vendor,
07:24partner relationship
07:25because all of us
07:27bring to bear
07:28key ingredients.
07:29None of us
07:30can deliver
07:31a final product.
07:32I do think
07:33I need to baseline
07:34everybody on the data.
07:35So in the R&D space
07:38that we're talking about,
07:39so preclinical research
07:40and development,
07:41there are 10,000-ish
07:44biopharma companies
07:45globally
07:45and there's a fathead
07:47long-tailed
07:47or giant global
07:49entities like Sanofi
07:49and then smaller biotechs.
07:51But within that cohort
07:52of 10,000 or so companies,
07:55you have 20 exabytes
07:57of data-ish
07:59and growing rapidly.
08:01So it's one of the
08:01largest data sets
08:02in the world.
08:02It's possibly the fastest
08:04growing data set
08:05in the world.
08:05And to Emmanuel's point,
08:07it is probably
08:08the most complex data
08:10on Earth.
08:12None of us
08:13can solve this alone.
08:14So bringing
08:15the deep expertise
08:16of a pharma
08:18to bear with
08:19the expertise
08:20of a Tetra
08:20that understands both science
08:22and the modern data stack
08:23and AI
08:24with massive
08:25horizontal partners,
08:26we can crack the code
08:27on this
08:28much faster
08:29than most people believe.
08:31But we actually
08:31have to move forward
08:32collaboratively.
08:33And that's what
08:34we're all trying
08:34to work on now.
08:35I mean, one of the reasons
08:37why we've seen such
08:38incredible progress
08:39in the natural language space,
08:41the LLM space,
08:42is because data
08:43was free for a lot
08:44of these companies.
08:45The internet exists,
08:46right?
08:46And in biology,
08:48there is no internet
08:48of biology.
08:49Yes, there's a lot of data
08:50that exists within the walls
08:52of pharmaceutical companies
08:53and research institutions,
08:54but they are within the walls
08:56of the pharmaceutical companies
08:57and research institutions,
08:58right?
08:59And so there is no
09:00large corpus of data
09:01for everyone
09:02to be able to train on.
09:03And so I actually think
09:04if you ask,
09:05what is needed?
09:06What do we need to do
09:07to realize scientific AI?
09:08I think about two things.
09:09One is we need countries,
09:12companies to invest
09:14in an industrialized approach
09:16of generating large-scale
09:17experimental data.
09:19That is the source code
09:21to artificial intelligence
09:23that can give rise
09:24to discovering new medicines,
09:26understanding disease,
09:28driving efficiencies
09:29and knowledge work,
09:30right?
09:30From a scientific standpoint.
09:32I think that is one thing
09:33that is essential
09:35for every organization,
09:37every country
09:37to be thinking about
09:38because it's expensive
09:39and it is a commitment,
09:41if you will.
09:41The other thing
09:42that I think is missing,
09:43and Emmanuel,
09:44I'm sure you experience this
09:45within your organization
09:47as well.
09:47From a technology standpoint,
09:49you can build the tools.
09:50The tools can be wildly capable.
09:53You still got to get people
09:54to use them, right?
09:55And that change
09:57from this is how I used
09:59to do things
09:59and this is how you want
10:01me to do things now,
10:02that's hard, right?
10:02And so you either need
10:03to bring people along
10:06through education
10:07and development
10:09or you need some old trees
10:11to burn down
10:11and some new trees
10:12to grow up, right?
10:13And so I think those
10:15are just a few things
10:15that, you know,
10:16maybe are a barrier
10:17for us to recognize
10:18the potential here.
10:19I would love for Emmanuel
10:22to weigh in on this.
10:23I would just add to that
10:26by saying,
10:26and this is a bit provocative,
10:29I understand,
10:29but the way you can think,
10:32I know many of you
10:32are not from the industry.
10:34You're tech folks
10:35but maybe not scientific natives.
10:38The way I want you
10:38to think about this
10:39is the operating model
10:41for the biopharma industry
10:43since time immemorial
10:45has been that of an artist colony.
10:49Companies like Sanofi
10:50have assembled
10:51the greatest scientific minds
10:54in the world
10:55and they've given them
10:56a wide berth.
10:58they've given them tools
10:59to paint and create art, science,
11:03as they deem appropriate
11:04given how brilliant they are.
11:06And that transformed humanity.
11:09It has improved
11:10and extended all of our lives.
11:12But we have reached
11:13a state of diminishing returns.
11:15And to Rory's point,
11:16we now need to augment
11:17those brilliant scientific minds
11:19with industrialization
11:21of scientific data
11:23and the use cases
11:25and come to these scientists
11:27and I know that you have
11:29to deal with this every day.
11:30Come to these scientists
11:31and say,
11:31look, we come in peace.
11:33We come with tools
11:34to wield in all new ways.
11:36It's going to augment.
11:37It's not to diminish your brilliance.
11:39It's to augment it.
11:40So I think change management
11:42for the industry writ large
11:44and for the scientific community
11:46specifically is a major obstacle.
11:48I don't know if you agree or not.
11:50Yeah, I think it's fair
11:51that adoption is always a challenge.
11:54Irrespective of which industry
11:55the technology adoption is,
11:5670% of the difficulty.
11:59And we're all very proud
12:00when we make the tech
12:01because we think it's very difficult.
12:04But really,
12:04tech can be orphaned
12:06of a user rapidly.
12:09So that's number one.
12:11Number two,
12:11I want to go back on the data side
12:13because I have a bit of a plea, Charlie.
12:15Since you've got us on stage,
12:16I might as well use the form.
12:17You know, national data is important.
12:20Human data.
12:21Our intention in the future is
12:23we'd love to be able to say
12:24we can develop a drug
12:25and we're discussing it backstage
12:28and simulate as much as possible.
12:30So we talk about in silico
12:32or digital or virtual if you want.
12:35Could you really start to create
12:36the molecule, test it,
12:39simulate it against a virtual patient population
12:41before you even go to a human?
12:44Nothing is more heartbreaking
12:46than a drug that fails
12:48because it's a failed promise
12:50to people who have a need.
12:51So could you test all of this
12:53with the technology,
12:54with the logic,
12:55with the likes of Tetra Science
12:56and NVIDIA,
12:58but could you do everything
12:59as much as possible
13:00in a virtual world?
13:02And then you get to the real world,
13:04your world,
13:05just at the last,
13:06really the last time
13:07and just to validate it.
13:08And to do that,
13:09we need more,
13:11what we call longitudinal data,
13:13is individual anonymized patient data
13:16over history of patients.
13:18To create a,
13:19if you want a twin or a replica of a human,
13:22you need to have a human history.
13:24And the more human histories we have.
13:25So my plea is,
13:28health agencies all over the world
13:30or yourselves as individuals,
13:32submit,
13:33share your data
13:34because that data will help others,
13:37maybe not you,
13:38but your next generation.
13:40Maybe your data will help
13:41your grandchildren's health.
13:44And so that's my plea today.
13:46Let's make sure that we are,
13:48we are generous
13:50with our individual data,
13:52knowing that it will serve the purpose
13:54of future generations.
13:59AI across the value chain and pharma.
14:02A lot of our work
14:03in the healthcare space,
14:04healthcare team in NVIDIA
14:05over the past five years
14:07had been almost exclusively focused
14:09in early discovery.
14:10Right?
14:10And that's where these models
14:11were really starting to show value first
14:13and being able to generate molecules
14:15that have the desired characteristics
14:16that are likely to be successful
14:18within clinical trials
14:20when you have a development candidate.
14:21but the value is still ten years out.
14:25Right?
14:25It still takes a really long time
14:27to go from molecule to market,
14:29inpatient.
14:30Right?
14:31And so, some of the work
14:32in the clinical development space
14:33where you're taking these
14:34multi-modal foundation models,
14:36you're combining imaging data
14:38with electronic health record data,
14:39with omics data,
14:41and getting that representation
14:42of the patient,
14:43and then allowing you to predict
14:45which patients are likely
14:47to respond to which medicines.
14:49That is huge value
14:51for pharmaceutical companies.
14:52And if you can accelerate
14:53clinical trials,
14:54which is where 70% of the dollars
14:57in time goes into
14:58for bringing a drug to market,
15:00the impact, both to patients
15:02and to enterprises,
15:04is massive.
15:05Absolutely massive.
15:06Yeah, if I could, Charlie.
15:08So I want to put a value on this.
15:10I want to, I'm going to tie together
15:12a few things you've just heard.
15:13I talked about reaching
15:14a state of diminishing returns.
15:16Manuel talked about economics.
15:18You're touching on economics.
15:20This is important to understand.
15:22If I asked for a show of hands,
15:24I suspect all of you
15:26have heard of Moore's Law.
15:28Yeah?
15:29Which is the observation
15:30that the number of transistors
15:31doubles every 18 months,
15:33so price performance improves.
15:35Yeah?
15:35And now we just call it
15:37Jensen's will.
15:38As Jensen wills it
15:39to happen seemingly daily.
15:40But the point is,
15:42the point is,
15:43many of you
15:43have leveraged Moore's Law
15:46every day.
15:47Your iPhone is a function
15:48of Moore's Law,
15:48as an example.
15:49Many of you have gotten wealthy
15:51off of Moore's Law.
15:52In biopharma,
15:54there is another law.
15:55It's called Eroom's Law.
15:57And Eroom's Law
15:59is,
16:00Moore's Law spelled backwards,
16:02because it's the inverse.
16:04It takes longer,
16:05every year,
16:06every year,
16:07takes longer,
16:08and costs more,
16:09to go from magic molecule
16:11to market.
16:13We are now at a point,
16:15now at a point,
16:16where it is 12 or more years,
16:18and between 2.5 and 6 billion U.S. dollars,
16:23to get to market.
16:25The economic model of this industry
16:27is collapsing.
16:28It is the most important industry on Earth.
16:32and the economic model is collapsing.
16:34So what we're trying to do collectively
16:36is arrest that Eroom's Law
16:39and get that curve going the other way.
16:42And to Rory's point,
16:43this is really important,
16:44the really cool stuff we love to talk about,
16:47we all do,
16:48is new drug discovery,
16:49the molecule discovery.
16:51That's awesome.
16:51It's very cool.
16:53Finding that silver bullet,
16:54that needle in a haystack.
16:55But that value chain,
16:57that comprises that 12 years,
17:00all of that can be optimized
17:02through the application of AI.
17:04All of it.
17:04And there are hundreds,
17:06and hundreds of use cases,
17:07in a company like Sanofi,
17:09with many thousands of scientists,
17:11that,
17:12no matter how talented they are,
17:14are currently sub-optimized.
17:15So we all are committed to both
17:17shrinking the time,
17:19and I can't give you a forecast yet,
17:21but we can do better than 12 years,
17:22right?
17:23And cutting billions off that number.
17:26That's the objective.
17:27Maybe a word about sovereignty,
17:32in your sector.
17:34Of course, Sanofi,
17:35I'm sure,
17:36is deeply concerned about sovereignty AI.
17:40But what about you,
17:42Nvidia,
17:42and about Twitter saying,
17:44maybe you just want to work everybody,
17:46that's it.
17:48Yeah, I mean,
17:48obviously we're an American company,
17:50but we are a global platform,
17:53right?
17:53And we work with every cloud provider,
17:57every OEM system,
17:58and frankly,
17:59because this is such fundamental technology,
18:03that can impact economic development,
18:05public health,
18:08you know, boosting startups in your country,
18:10we believe it is really important,
18:12for nations to have a sovereign AI strategy.
18:15And that's frankly,
18:16why we've co-located GTC,
18:17the world's number one AI conference,
18:20with VivaTech,
18:21Europe's rendezvous,
18:22is because we want to enable Europe,
18:25to build sovereign AI healthcare services,
18:28in Europe,
18:29for Europe,
18:30so that Europe is not dependent on anyone else.
18:32You don't have to be, right?
18:33And so,
18:34that was part of the reason why we've come here,
18:37and we've been working with an incredible ecosystem,
18:39of startups and enterprises alike,
18:42who are well on this journey.
18:43Well on this journey.
18:44and so we,
18:44we do believe that sovereign scientific AI,
18:47sovereign AI,
18:48generally speaking,
18:49is wildly important.
18:50Sure.
18:51Yeah.
18:52So I,
18:53so I agree with all of that,
18:54so I would just put this to the audience,
18:56right?
18:57I don't have all the answers to all the things.
18:59If,
18:59if you believe,
19:01that AI,
19:03is the most powerful tool,
19:05ever wielded by mankind,
19:07if you do,
19:07you may not,
19:08and that's okay too,
19:10but if you do,
19:12then you must agree,
19:13that the moral imperative,
19:15is to deliver on scientific AI for all of us.
19:19Number one,
19:19that's first.
19:20Second,
19:21to save the most important industry in the world,
19:24biopharma,
19:25by fixing the economic model.
19:27And third,
19:28going to sovereignty.
19:30Every nation state in the world,
19:32every developed nation state in the world,
19:35if you,
19:36they believe,
19:37if Macron believes,
19:38AI is the most powerful tool,
19:40then he has an obligation to the people of France,
19:44a moral and fiduciary responsibility,
19:47to ensure that AI,
19:49is harnessed properly for national security,
19:53and economic security,
19:55and health security.
19:56That's not an option.
19:57I don't think that's even debatable.
19:58It's a moral and fiduciary obligation.
20:01So,
20:01I agree with Rory.
20:02Sovereignty isn't about American companies selling things.
20:06It's actually quite distinct from that.
20:09It is ensuring that every developed nation can safeguard
20:12its most valuable data,
20:14in furtherance of its most consequential use cases and applications,
20:19and harness the power of AI in a safe and secure manner.
20:23And then, by the way,
20:24if you have allies,
20:26you can federate that,
20:28and federate those capabilities to your allies.
20:30But the idea that adversaries,
20:33or peer adversaries,
20:34or frenemies,
20:36should be sharing AI capabilities,
20:38is kind of nutty,
20:40to be honest.
20:40So that's what we talk about around sovereign scientific AI,
20:44is enabling nation states to take control of the most important data they have.
20:48Hopefully that makes sense.
20:49I'm going to add one topic on this.
20:51I'm going to disagree a little bit with you.
20:52It's okay, it's okay.
20:53The two of you.
20:53This is good.
20:54And we can resolve it.
20:55We don't want to be boring.
20:56We need to mix it up.
20:58Sovereignty cannot be a false excuse to protectionism and over-regulation.
21:03I think sovereignty has to be the result of a very thoughtful process of sovereign data.
21:09So what are the reasons for sovereignty?
21:11I fear that we're entering in a phase where we used to live in a big planet called the globe,
21:17and we're starting to live in a very fragmented globe.
21:20And as we all see geopolitically every day of our lives,
21:23it's increasing, it's becoming more complicated.
21:26We're not trying to solve issues that have borders.
21:31Patients don't have borders.
21:33Patients are not French or Chinese or Russian or American.
21:37Patients are not humans that belong to the same cohort that we have,
21:40which is called Planet Earth.
21:41So I do struggle with the fact that sovereignty is being misused as a principle
21:48to introduce unnecessary regulation and protectionism, which is commercial,
21:54versus what it needs to be.
21:56Protect your own data as individuals.
21:58That's a must.
21:59As Patrick was saying, that's an absolute must.
22:02There's elements, of course, of science and of defense that have to be protected.
22:08But let's not make the false shortcut that protectionism equals sovereignty.
22:14Those are two very different principles that are being mixed up at the moment.
22:19I think we all agree.
22:20I do agree with you.
22:21Look, I think we don't need to get into geopolitics, but I think we can agree.
22:26The world kicked off an unprecedented experiment about 25 years ago in hyperglobalism.
22:32The pendulum is swinging maybe too far, too far to the other side.
22:37I think most people would agree it needs to maybe come back to the middle in a measured way.
22:42And at its center, science should not be politicized.
22:45We can all agree on that.
22:46For sure.
22:48For sure.
22:49For sure.
22:49Maybe we can finish with maybe the borders.
22:55So what borders do you think remain unsolved regarding AI?
23:00What's going to happen in the next 10, 20 years?
23:04Yeah, I think, you know, we've over the past, call it 10, 12 years, we've gone from perception AI to
23:13generative AI,
23:14where we can kind of translate from any domain to any domain, text to image, image to video, sequence to
23:22protein structure.
23:23We're now entering agentic AI, where agents can use tools, they can take action, they can reason, they can interact
23:32with other agents or with humans, transforming knowledge work as we know it today.
23:37And we're starting to enter into physical AI.
23:40And so your question around what is missing, I think connecting the digital world and the physical world, digital agents
23:49and physical robots is going to become really interesting, especially as it relates to science.
23:55And I believe that this is going to reinvent the scientific method as we know it, where digital agents will
24:04be able to generate hypotheses, they will be able to design experiments,
24:08they will be able to call a fully automated robotic laboratory that is physically enabled with AI.
24:16It's going to be able to run and generate experimental data that then can be fed back into the AI
24:22systems that goes through that entire cycle again, at rapid pace and at massive scale.
24:27You've heard maybe a little bit about the concept of AI factories while we're here, like turns raw data into
24:35intelligence and tokens.
24:36Science factories are going to start to emerge as well, where you're filling up giant warehouses with automated robotic laboratories
24:46to generate experimental data at scale.
24:49And to me, that connection between digital agents and physical robotics is going to be really, really exciting.
24:55Yeah, and I would just say, because we're working with some of the same pioneers in this area that Rory
25:01just described, this is, with execution, and if we solve the data challenge, which we will, this is on the
25:09near-term horizon.
25:10This is not, it sounds futuristic, but we are that close to these kinds of breakthroughs that will fundamentally transform
25:17science.
25:17I could not agree more.
25:20We all have to remind ourselves what an incredible time we live in now.
25:25It's an unbelievable period of our lives.
25:28And AI is here accelerating what we could not do before.
25:31So I love what we could do now with technology, with AI, identifying a new target that you could potentially
25:38treat a disease, simulate it, et cetera, et cetera, et cetera.
25:41I think the most difficult thing to solve yet is human biology.
25:46We have to be, we have to remain very humble in front of human biology.
25:51There's so much that we can emulate, simulate, and so much that we can fantasize about.
25:57Yet, you are very complex machines to understand.
26:00Each one of you in this room are very unique.
26:03And so there is so much more that we have to learn about biology to be able to get to
26:10a point where we'll be able to do personalized medicine.
26:13We'll be able to really solve individual diseases at a person level.
26:20But that's, I think, the holy grail that we're all working towards.
26:23It is the most humbling industry of all.
26:25Yeah.
26:26It is.
26:26No doubt.
26:26It is for sure.
26:27Just to maybe comment on understanding human biology, you guys are probably familiar with the Human Genome Project.
26:34It was, call it, biology's space mission for, I don't know how many decades it actually went on.
26:40We're amidst what is probably the next Human Genome Project, this concept of virtual cell.
26:46Every bio-institution around the world is generating large-scale, perturbed, single-cell data, spatial imaging data.
26:55Trying to understand the causal biology of disease so that we can start to map out a virtual cell where
27:03you can almost think about, like, prompting a diseased cell.
27:07Which perturbation, which medicine would return you to a healthy state?
27:11Like, that is the vision of a virtual cell that many of these, I can't even say companies yet because
27:17it's still, it's early, it's research, are working on.
27:20And it truly requires this combination of industrial scale, digital biology generation, scaling models to be able to understand the
27:31complexities of biology.
27:32And then, of course, you need incredibly talented both researchers and data scientists to be able to make this happen.
27:39But it's a wonderful, wonderful vision and an exciting future.
27:43Amen.
27:44Well, thank you very much. We are almost out of time. It was very, yeah, very interesting.
27:50I'm sure we hear new things about AI in the earth industry for the months to come.
27:57So thank you very much again for coming to VivaTech to talk about it.
28:00Thank you, Charlie.
28:01Thank you, Charlie.
28:02Thank you, Charlie.
28:02Thank you.
28:02Thank you.
28:02Nice job, guys. Thank you.
28:06Right on time.
28:08Always.
Commentaires