- 22 hours ago
The pharmaceutical industry is experiencing a seismic shift as AI transcends traditional human limitations. The scientific evidence base is exploding beyond human capacity as thousands of studies that could transform patient care are published daily. But we're falling further behind in our ability to synthesize this knowledge when decisions need to be made. AI agents are moving from pilots to production, fundamentally changing how we engage with scientific literature. Sanofi, Biolevate and NVIDIA share their journey of scaling AI agents to an enterprise level—now managing 100,000+ agents and millions of workloads—turning massive evidence bases into timely and actionable decisions. But can this productivity boost genuinely reshape how we work? And what does responsible scaling in pharma look like?
Category
📺
TVTranscript
00:01Hello, good morning everyone.
00:04First of all, thanks for making it to the black stage
00:08and it's our pleasure to open the VivaTech Day. I know there are many good sessions
00:12we're going to compete with and we'll make it very exciting and interesting for you.
00:16My name is Artemi and I work at Sanofi leading digital data
00:19and AI products for global trade and market access. And at Sanofi
00:23as we continue our journey to become the first biopharma company
00:27powered by scale, we're going to speak about an interesting topic
00:31which brings us here today, scaling. It becomes an increasingly
00:35important issue of the industry and as we are moving through this
00:39I'm very pleased to be here with two leaders, David
00:43the head of strategic partnerships from NVIDIA for
00:47EMEA region and Joel, the CEO of BioElevate company.
00:51So, I will pitch it into you, tell us a little bit more about you
00:55and what you do.
00:58You want to take this David?
01:00Okay, yeah. Hi everybody and
01:03Bonjour Paris.
01:04It's a pleasure to be here today.
01:06So, my role at NVIDIA is basically
01:09to detect the
01:11leading and growing
01:12basically successful stars
01:15in startups but also supporting
01:17biopharma such as Sanofi
01:19and all the others in
01:20the European, Middle East and Africa region.
01:24That's what I do.
01:25Yep.
01:27And as for me, I'm Joel Berafa.
01:29Very happy to be here.
01:30So, I'm CEO of a company called BioElevate.
01:33Our mission is to accelerate
01:35all the healthcare related processes
01:37starting with pharma,
01:39a challenge in putting drugs on the market
01:41and we do that obviously with AI.
01:44This is 2026.
01:45So, yeah.
01:47Excellent. Thanks for the intro.
01:49We're going to have a nice fireside chat,
01:51meaning that I would be mainly asking the questions
01:53and Joel and David would be answering them.
01:55And with no further ado, let me launch into the first one.
01:58So, just to warm up to both of you.
02:00In one sentence,
02:01what does pilot to production
02:03mean for you in regulated
02:06biopharma environment?
02:08Well, for me,
02:09it's basically delivering
02:11the value promised by your product
02:13or service
02:15at the scale of a company
02:17such as Sanofi or orders,
02:19but also following the constraints
02:23dictated by regulatory environment
02:25and IT best practices.
02:28Simple as that.
02:31Yeah.
02:31Delivering to scale is something
02:33quite interesting from the point of view
02:35of NVIDIA.
02:37For us,
02:38what we see between different
02:40basically solutions on the market
02:42is attention you pay to
02:44what we call tokenomics.
02:46So, you have basically AI
02:48factories, which are those
02:51systems that you find
02:52in various data centers.
02:54And on one side,
02:56you come a question.
02:57On the other side,
02:57come intelligence.
02:58And you need to pay attention
03:00to generate those tokens.
03:02So, those tokens are basically
03:04the money that you generate,
03:07the intelligence,
03:07and what Joel is basically
03:09making a business of.
03:11And what Sanofi is in fact
03:13using and making a business
03:16out of the intelligence itself.
03:18So, for us,
03:19we see the scaling process
03:21is critical to think about
03:24basically the return on investment,
03:25but also the margin,
03:27which is not a new field
03:29which is around tokenomics.
03:31tokenomics, okay?
03:32Tokenomics is implicating basically
03:34the right hardware
03:35for the right
03:37basically system.
03:38But it's also implicating
03:40the right architecture
03:41of your software.
03:42And you need to be careful
03:43attention of that.
03:44Otherwise,
03:45you will basically
03:46run into problems
03:47of competitiveness,
03:49but also return on investment.
03:52Excellent.
03:53Excellent.
03:53Thank you, Joel and David.
03:54That was a quick warm-up.
03:55So, we are hearing
03:56two essential truths here.
03:58The viable workflows
03:59and economics
04:00make it, break it.
04:02It gets a little bit difficult
04:03when the things
04:04get off the balance.
04:05So, let me bring it to you,
04:07let me bring you there
04:08with the next question is,
04:10can you recall a moment
04:11when you actually realized
04:13that a certain pilot
04:14would not really scale?
04:17Oh, yeah.
04:19And it's a bit surprising
04:22because, so I'm a tech nerd.
04:24You would call that.
04:26I love tech.
04:27And I would think
04:28that the first challenges
04:29would have been pure tech
04:31and related to optimizing
04:34the compute, optimizing
04:36the costs,
04:37and eventually transformation.
04:40But we were one
04:42of the first AI native
04:46agentic platform market
04:48more than two years ago.
04:50And we were lucky
04:51to realize
04:53that our biggest challenge
04:55wouldn't necessarily be
04:57having AI
04:58and 100,000 of agents
05:00working, delivering the value.
05:02We would eventually
05:03make the models work.
05:05But then,
05:07the people making the decision
05:09and being responsible
05:11for the whole system
05:12would remain humans.
05:14And, I mean,
05:15you don't want to see
05:18a human, like,
05:19chasing after
05:20what 100,000 agents
05:22or what AI has been doing
05:23through the nights.
05:24And we realized
05:25that we would need
05:27to reshape completely
05:29our system
05:29to optimize
05:30the ability for humans
05:32to oversee
05:36what's going through
05:37this AI system
05:38and connect the dots
05:39between something tangible,
05:41something real
05:41to avoid hallucination.
05:43because the more you do,
05:44the more you automate,
05:45the more there's a risk,
05:46the more there's a drift
05:47for AI
05:48to take a path
05:49you were not expecting
05:50and it still works.
05:52So, yeah,
05:53we decided to change
05:55our technology
05:56and also something
05:57very key
05:58in the way
05:58we build models
06:01to optimize,
06:02not the AI,
06:03but the way humans
06:05would oversee
06:06AI activity.
06:08Excellent.
06:09Thank you, Joel.
06:10And I think,
06:11for me,
06:12the takeaway
06:12what stands out
06:13and what you said
06:14is that the pilots fail
06:16not because the AI
06:16doesn't work,
06:17it's just,
06:18you said it many times,
06:19it's the systems around it
06:20which involves the humans,
06:22the processes, etc.
06:23So, David,
06:24probably you could get us
06:26a more concrete example
06:27from your experience
06:28what you've seen
06:30where the failure happens
06:31actually first.
06:32is it data,
06:33is it process,
06:34integration,
06:35governance, adoption,
06:36something you would recall?
06:39Yeah, I mean,
06:40I would be curious.
06:42So, we hope,
06:43I mean,
06:43with the panel
06:44to have five minutes
06:45at the end for questions,
06:46but I would be quite interested
06:47in your feedback
06:48on what failed
06:50in pilots
06:51to scale, basically.
06:53On my side,
06:54I mean,
06:54talking from personal experience,
06:56I was in pharma
06:57before joining NVIDIA.
07:00it's always an adoption process.
07:02And,
07:03it's an hybrid problem.
07:05There is many things in it.
07:06It's an eye drive.
07:08There is many heads.
07:09There is a UI.
07:11You need to nail down the UI
07:12because you should not forget
07:13that, basically,
07:14it's a world made for humans.
07:16So, we need to interact
07:17with such a software.
07:18I think it's a job
07:20that Biolevet did fantastically.
07:22They really nailed down their UI.
07:26And,
07:26you should not forget
07:27that it needs to be integrated.
07:29So, you need the adoption.
07:31You need to have the right tool.
07:33You need to be gluing
07:34to your users.
07:36Very, basically,
07:37Y Combinator style.
07:39And, from there,
07:41evolve with them.
07:42Okay?
07:44And, further than that,
07:46nowadays,
07:47moving into,
07:48we move into a meta space,
07:50not meta as a company,
07:52but a meta space where
07:54adoption and AI
07:56AI is, in fact,
07:57also happening
07:59within the system.
08:00So, agents
08:01need to have the right harness
08:03around them.
08:05So, they need to have tools.
08:07They need to have coach.
08:08They need to have, basically,
08:09a grumpy reviver, number three.
08:11They need to have, basically,
08:13a set of judge.
08:14And those systems need to evolve
08:16by themselves
08:17and improve over time.
08:19It's interesting that we bring
08:21the same, basically,
08:23logic of,
08:24how do you use a tool?
08:25How do you train those systems
08:27within those systems,
08:28which are quite intelligent?
08:29And the same thing
08:30applied to us.
08:32So, failing happened,
08:34for me, mostly at the adoption layer.
08:37The tech looked good,
08:38but the people don't use it
08:40because it doesn't solve
08:41the real problem.
08:43Very good.
08:44So, we got it loud and clear.
08:45The technology itself is not the challenge.
08:48The enterprise realities
08:49comes integration,
08:51adoption,
08:52governance.
08:53And let's be optimistic.
08:54For sure,
08:55we start to learn
08:56how to get it right.
08:57And if we get it right,
08:59and when it comes to the scale,
09:00it's not anymore
09:01about running
09:02an agent
09:03or a pilot
09:04or an application.
09:05It's running a multiple,
09:07hundreds of thousands
09:07workloads
09:08in production
09:09at scale.
09:10So, Joel, back to you.
09:12Can you share us a little bit more
09:14what does it really look like
09:15to do that?
09:16And how do you control that?
09:19It's difficult.
09:20I won't lie.
09:21Especially,
09:22we were
09:24very dedicated
09:25to the mission
09:26of accelerating
09:26the healthcare ecosystem.
09:28And we realized that
09:29if we needed
09:30to have an impact,
09:32we would have to scale AI
09:34much faster
09:34than any other industry.
09:37And
09:39we,
09:40two years ago,
09:41two years in AI
09:41is like two centuries ago,
09:44once upon a time.
09:46People were doing
09:47mostly chatbot,
09:48interacting
09:49with simple answers,
09:52running simple tasks.
09:53And we were already
09:54focused on,
09:56okay,
09:56we're going to do long-lived
09:57processing
09:58with difficult use of tools
10:01because these tools
10:02would decide
10:03whether or not
10:04a therapeutic product
10:07is safe
10:08or whether or not
10:09we are compliant
10:11to deploying
10:11for certain subpopulations,
10:14something really serious.
10:15So,
10:15we had to build something
10:16that would run
10:18100,000 of agents
10:19concurrently.
10:20And something that is
10:21even more difficult
10:22than running
10:24a lot of things
10:25on a computer
10:26is running things
10:27concurrently
10:28on this computer.
10:31knowing that,
10:32like I said before,
10:34one human babysitting
10:36100,000 agents
10:37is a challenge
10:38of itself.
10:38And,
10:40David mentioned it
10:41as well,
10:42you need a lot
10:44of things around
10:44these agents
10:45to complete
10:46complex tasks.
10:47It was really about
10:49making these agents
10:50self-evil
10:51and make certain
10:52that they learn
10:53from the environment
10:54and they don't expect
10:56humans to tell them
10:57the very single detail
10:59they need to complete
11:00their task.
11:00So,
11:01it was a technological challenge.
11:04We had to throw away
11:05a lot of open source
11:07models
11:08or tech
11:09that we were trusting
11:10and we loved.
11:11And,
11:12we had to build
11:13a lot of things
11:13that back then
11:14were not existing.
11:15But more importantly,
11:19the supervising process
11:21and the best practices
11:24to let our customers
11:26and partners
11:27know what's going on
11:28regardless of the scale.
11:30Excellent.
11:31I was listening
11:33to your answer
11:34and I just wanted to highlight
11:35a couple of points
11:36which really struck me.
11:37Number one,
11:39scaling AI
11:40does not just amplify
11:41the value,
11:41it also amplifies the risk
11:43associated with this.
11:45And the second point
11:46is this tiny balance
11:48between what is really human,
11:50what is self-learning,
11:51self-adjustable.
11:52So, probably the next one
11:54would be for you, David.
11:55If you could share
11:56your point of view,
11:59what does responsible
12:03scaling look like
12:04in the industry?
12:05And what do you think
12:07at this point of time
12:08because we don't know
12:09the future,
12:10as Joel said,
12:10it's evolving fast.
12:12It's better still
12:13to stay with the humans
12:14and what could be already
12:15given to the machine.
12:18Well, I think we cannot live
12:20without AI,
12:21but I'm a bit
12:21for my own shop.
12:24But it's,
12:27I think you need to use
12:29those tools
12:30and consider them
12:31as tools
12:32which are enabling us.
12:34And I think
12:34we can all agree
12:35they truly enable us
12:37to do things faster
12:38than we used to.
12:39to do.
12:41It's still
12:42in its end fancy.
12:43We are just
12:44scratching the surface.
12:45And I think
12:46we need to pause
12:47and reflect on that.
12:48It's still
12:49just the surface
12:50of what they can do.
12:52We have new models
12:54which begin
12:54to be so powerful.
12:56We certainly heard
12:56in the news
12:57like in the last two weeks
12:58basically on Topic
12:59and OpenAI
13:00released new models
13:01that they are like,
13:02in fact,
13:03they need to test it.
13:05there is some restriction
13:06around it.
13:06Not everybody has access to it.
13:08And that tells you
13:09about where is it going.
13:11Okay?
13:12Those things
13:13will enable us
13:14to make sense
13:15of the hardest problem
13:16we have in front of us
13:17which is,
13:18in my opinion,
13:19biology,
13:20understanding us,
13:21understanding disease,
13:25and making basically
13:26the world a better place.
13:29Solving climate,
13:31being able to reach consensus
13:33with large groups,
13:34complicated environments,
13:35I think that will be
13:36a complicated environment
13:38in medicine,
13:39for example,
13:41with conflicting proof,
13:42conflicting data,
13:44trying to resonate
13:45about it.
13:47And so coming back
13:50to the topic
13:51of responsible AI,
13:53I think responsible AI
13:54is about moving along
13:58and forward along
14:01the basic,
14:02trying to solve
14:02our biggest problem
14:04because that's our responsibility.
14:05Okay?
14:06Responsible AI is that
14:07you're in the loop,
14:08you guide,
14:09and you still work
14:10on the problem using
14:11an augmented version
14:12of yourself.
14:13Okay?
14:14Digital twin of yourself,
14:16which is, in fact,
14:17the crowd of different agents.
14:20and I think that's what is
14:21responsible AI.
14:23Using those tools,
14:24analyzing the result,
14:25moving forward,
14:26and going faster
14:27towards solution.
14:29Excellent.
14:31So, basically,
14:32we are speaking here
14:33about leveraging
14:35the accelerated capability
14:36of technology,
14:37still being bounded
14:39by the human judgment
14:41and decision making
14:42where it matters most.
14:43I think the complex thing
14:45here is actually
14:46to take time
14:47to reflect
14:48and stand back
14:48because we cannot
14:49have all this context
14:50sometimes just in the mind
14:51of the humans.
14:52And that's where
14:53responsible AI comes
14:55with the element
14:55of risks,
14:56governance,
14:57what should
14:58and what we potentially
14:59should not yet
15:00do with the technology
15:02accelerating so fast.
15:03David,
15:04I would stay on you
15:04if you allow me.
15:05and we speak a lot
15:09about acceleration,
15:10so let's step into
15:11the technology.
15:12How do you see the role
15:13of accelerated compute?
15:15I think I'm asking a very
15:16obvious question,
15:17but give us an interesting
15:18answer to it.
15:20Accelerated compute.
15:21It's going in the right
15:22direction.
15:23We are bringing,
15:24I think there is an evolution,
15:26there is a faster demand
15:28for accelerated compute
15:29than what we are able
15:30to come up with.
15:32So there is a bit of a delay
15:33between what we are able
15:34to build as hardware
15:37and what is the demand
15:39for such a compute.
15:41So on one side,
15:42I mean,
15:42working with Bioelevate
15:44and you, Sanofi,
15:46I mean,
15:46it's definitely something
15:48where we perceive,
15:50I mean,
15:50if we give you more compute,
15:53you would consume it directly.
15:55It's not a problem.
15:57You want something
15:58which is running faster.
15:59And so the next generation
16:00of NVIDIA solution,
16:03the Vera Rubin platform,
16:05will be basically bringing
16:07this type of innovation
16:08to the market.
16:10And finally,
16:11allowing us to explore
16:12this leading edge
16:13of reasoning
16:15within AI models.
16:16But not only thinking
16:18about things which are
16:19analyzing
16:21literature,
16:22making sense of
16:23very complex data,
16:23but also moving into
16:25drug discovery.
16:25I mean,
16:26you have some very hard problems
16:27there.
16:27Protein folding,
16:29pushing the limit
16:30of co-folding,
16:31binding,
16:31prediction,
16:33physic informed models.
16:35All of that basically
16:36need to be basically
16:37empowered.
16:38There are still some
16:38applications which have
16:40not enough
16:42basically compute.
16:43The thing don't work
16:44because we don't have
16:46a solution for it.
16:47It does not work
16:47in the current hardware.
16:49Excellent.
16:51So we are approaching
16:52the last third
16:54of our session.
16:54So let me connect
16:55a few dots here.
16:56So from what we've heard,
16:58I think infrastructure,
17:00governance,
17:00operating model,
17:01responsible AI,
17:02they have come all
17:04close together
17:04as of the day one.
17:06So there is no space
17:07to design architecture
17:08and any of these elements
17:09later after you deploy.
17:11I think it would just
17:12make the things worse.
17:13That's what I'm hearing
17:13from the answers.
17:15Thank you very much,
17:16David, for actually
17:18I understand also the other
17:19element of responsible AI
17:20is not to release a new
17:21thing every three months.
17:22So really,
17:23thanks a lot.
17:24It's nice too.
17:25I mean, yeah.
17:25It's nice too, but I think
17:26it is really a perspective
17:28which is just,
17:29I'm just opening it for myself
17:31as we speak,
17:32that sometimes it's really
17:33good not to have the new
17:34thing every three months
17:35so that we can make
17:36the best use out of it.
17:37Otherwise,
17:38we will consume and enjoy it,
17:39but we need time to process.
17:42And Joel,
17:43I'm getting into the last one
17:45before the last question.
17:46If you were to start
17:48a new pilot now,
17:49what would you do
17:50in the first 90 days,
17:52personally?
17:54Yeah.
17:55I've learned something,
17:58first of all for myself,
18:00through the recent months.
18:05First of all,
18:06this technology has been going
18:09so fast and transforming things
18:11that are turning a lot of people
18:13super optimistic.
18:14But trust me,
18:15and maybe this is one of the key
18:17message you want to go back
18:19with today.
18:21You want to be optimistic about
18:22what AI will impact
18:24in your daily life.
18:26We've been seeing things like
18:29treatment completely
18:31discovered by AI system
18:32that we built.
18:33And when you hear that in the press,
18:35and you're happy about that.
18:38when you see that happening
18:40before your eyes,
18:42and then you meet someone
18:43with the gene specifics
18:45that give the hints
18:47that you might be able
18:48to cure that person
18:49a few years later
18:51with certainty,
18:52is a totally different thing.
18:54But what I would do differently,
18:57however,
18:59is sit with the professional
19:01in the field.
19:01because tech moves forward
19:04and will always do.
19:06But in order to have
19:09this technology having the impact
19:10at the right place,
19:11you want to understand
19:13how people act, behave,
19:17and interact on a daily basis.
19:20And what we did was just
19:23asking questions.
19:24So you remember our collaboration
19:25with Sanofi.
19:26You've been part of this.
19:28We had a process.
19:28We've been extending information,
19:31fruitful information, sure.
19:33But it was only when we realized
19:35what was your daily life,
19:38what were your actual critical process
19:41that we knew where to take the technology
19:45from a strategic perspective.
19:48So yeah.
19:49Not touching the keyboard,
19:50sitting and being an intern
19:52in every place
19:53I want to have an impact on.
19:56So...
19:57Excellent.
19:57Well, thanks for bringing
19:58this perspective, Joel.
20:01Which leaves me...
20:02If there is one thing
20:03I would personally take away
20:04from this room
20:05and from the session,
20:06I would say scaling
20:07from pilots to production
20:08is not a technical step.
20:10It's the whole transformation
20:12of the system.
20:13It's who are the people
20:15you put behind the task.
20:16It's how do you identify
20:17what is the actual problem
20:19to be solved.
20:20It's how do you make
20:20the value in economics work well.
20:23Because, as David said,
20:24we can consume as much as we can,
20:26but it does not always justify
20:28the value of the task
20:30we are trying to solve
20:31with this compute.
20:32And I want to thank you both for this
20:34because that is exactly
20:35the type of thinking
20:36we really need in this room
20:37and the audience
20:37as we are really progressing
20:38and the industry evolves
20:40and we'll see more
20:41and more exciting
20:42and fascinating things
20:43to come ahead of us.
20:45and I think it's a good moment
20:47actually to introduce
20:47a couple of questions,
20:48to invite a couple of questions
20:49from the audience
20:51if there are some.
20:52We are happy to answer.
20:55So there should be
20:56some microphones
20:57which are around.
20:59We do have a question there.
21:04We've got a couple of,
21:06we've got hand raised here
21:07in the front line.
21:08It's coming.
21:09We have microphones.
21:11Yeah.
21:12It's coming.
21:18Just there.
21:21And after that,
21:22we'll have one more
21:23on the right side.
21:24Okay.
21:26I would like to know
21:27from all of you actually,
21:30a board prediction.
21:31where do you see AI agents
21:34in five years
21:35in health and longevity?
21:41Wow.
21:42I could say
21:43I would not see myself
21:45on the stage 18 months ago
21:47when we just started
21:48to work on a few things
21:49with Joel.
21:50So it's going to be
21:51a bit of a moonshot.
21:53Yet,
21:53I think it's going beyond
21:55just the agent,
21:56first of all.
21:57In five years,
21:58we would likely not be
21:59speaking agents anymore.
22:00That is the first thing.
22:03Second of all,
22:04where I would see
22:05the technology evolves
22:06and how it helps biopharma,
22:08I think we would be able
22:09to crack
22:10some of the problems
22:11of today
22:14where David and Joel
22:15have been anticipating them.
22:17It's the drug discovery.
22:19It is solving
22:20complex mathematical
22:21model tasks.
22:23And I think
22:24where the mind
22:25of the industry
22:25would be,
22:26there would be
22:27on the next question is
22:29how do we deploy it?
22:30Because discovering
22:31the drug
22:31is just one step
22:32in the entire value chain.
22:34The question is
22:34how can you bring it
22:35faster to the patient
22:36across the entire value chain?
22:38So there would be
22:38a lot more problems
22:39to solve.
22:40And that's why I'm saying
22:41it's not just a single agent.
22:43It would be really
22:44a harness,
22:45as David was explaining
22:46to us,
22:47around the entire solution
22:48which would allow us
22:49to accelerate
22:50the whole cycle.
22:51that's my bold prediction
22:53for the next five years.
22:54Let's record it.
22:54Let's see.
22:55Maybe one thing
22:57I would mention
22:58is that
23:00the
23:01agentic AI
23:03stack is already
23:04quite powerful
23:05here and there.
23:07However,
23:07where I see a lot of potential
23:09is the ability
23:09for even more agents
23:10to uncover
23:12the mechanics
23:12of a lot of things
23:14that we know
23:14statistically
23:15are working.
23:17And
23:17these mechanics
23:18that
23:19all these agents
23:20would uncover
23:21would help
23:23discovery
23:24at a crossroads
23:24of multiple
23:27therapeutic areas.
23:28I want to mention one.
23:31We know that
23:32research in
23:34women's health
23:34for instance
23:35has been
23:36quite of a slowdown
23:37because you don't want to
23:38for instance
23:39you don't want to take risks
23:40with pregnant women
23:42and everything you have
23:44in theory
23:44there's always
23:45this
23:46molecular
23:47versus
23:47super molecular
23:49when you combine
23:49two things
23:50and then you combine
23:51two billion of these things
23:52on this left
23:53and on the right.
23:54For instance,
23:56I think that
23:56understanding
23:58this type of mechanics
23:59using a lot more agents
24:01is going to help
24:02a lot
24:03longevity
24:03for example
24:05because
24:06we have the
24:07answers
24:07hidden
24:08and human intelligence
24:10will never have
24:12enough resources.
24:13So maybe scaling
24:14agent
24:15one or two
24:16order of magnitude
24:17more
24:18is likely to
24:19crack that.
24:21I would answer
24:22the first part
24:23of your question.
24:24So where agent will be
24:26quickly
24:27is that
24:27I think
24:28we generate hypothesis
24:29by the ton.
24:31So with new agentic system
24:33and
24:34the bottleneck
24:35will be
24:36our ability
24:37to test
24:37those hypothesis
24:38at scale.
24:39So you will need to have
24:40automated labs
24:41that come in the loop.
24:43And those
24:43automated labs
24:44are driven
24:44and connected
24:45with agents
24:46to AI system
24:48which are generating
24:49hypothesis
24:50to be tested.
24:51Like for example,
24:52the AI system says,
24:53yes, there is an ambiguity
24:54into the data.
24:55I reviewed the entire
24:56literature.
24:57I come up with a set
24:58of tests
24:59to be performed.
25:00You need to do
25:01basically cell assays,
25:02you need to generate that,
25:03you need to generate that,
25:04and then come back to me.
25:06So agents
25:07will be interacting
25:08with us
25:09as basically effectors
25:10and become
25:11an integral part
25:12of our daily work.
25:14But they need to be enabled
25:16with another system
25:17which is automated testing
25:19and automated labs.
25:21Longevity
25:21is a bigger topic.
25:24It's just like
25:24everything is longevity.
25:26Like cure disease,
25:28everybody will have
25:29a better life
25:29and live longer
25:30in a way.
25:33That's actually
25:34a very good perspective
25:35coming from
25:36three different places
25:37of the industry.
25:38We have one more question
25:38over there.
25:40Guillaume Cabon
25:41from Bio101.
25:42I want to understand
25:44better what BioElevate
25:46together with Sanofi
25:48and NVIDIA.
25:49Give us some more examples
25:51of where you have improved
25:53in terms of either
25:55drug discovery
25:55or clinical trials.
25:57Is it like drug design?
26:01I'm trying to understand
26:02where you have improved
26:03the proposition
26:05because as you know
26:06when you put a drug
26:07in a human
26:08then you see
26:09if the drug is going
26:10to be effective or not.
26:12So is it more
26:13you're designing
26:13the targeted population?
26:15What is it exactly
26:16that you have already achieved?
26:18I will...
26:19It's an excellent question.
26:20First of all,
26:21thanks for the opportunity
26:22to do a little bit
26:22more marketing
26:23inviting you all
26:24to the Sanofi booth
26:25at the ground floor.
26:26You will see it
26:27in 1D01.
26:28That's where we exactly
26:29show what we do
26:30with BioElevate
26:32and of course NVIDIA
26:33makes it all happen
26:34powering it by the GPUs.
26:36Giving a little bit
26:37more concrete answer.
26:39There is a lot of knowledge
26:40which is already existing
26:42in the scientific literature
26:43and the problem is
26:44that the amount
26:45of this knowledge
26:46increases day by day
26:47with the speed
26:48which is just enormous.
26:49Human reviewer,
26:50human researcher
26:51cannot cope
26:52and manage this
26:53at the pace
26:54the knowledge grows.
26:55So what we do
26:56concretely with BioElevate
26:57today is
26:58we started to crack
26:59the nuts of
27:00AI assisted literature
27:01reviews which enable
27:02different parts
27:03of the value chain.
27:04It's in
27:05health economics,
27:06it's in
27:07health economics
27:08outcome research,
27:10it's in R&D,
27:11it's in medical,
27:11it's in clinical.
27:12So that's where we
27:13actually start to get
27:14some knowledge faster.
27:15you will see also
27:16the number and concrete
27:17proof points,
27:18how it works in production
27:20beyond the pilot.
27:22And probably,
27:24Joel,
27:24if you would like to share
27:25a little bit more of your
27:26perspective, what you do
27:27with us.
27:28So yeah,
27:30without breaking any
27:31confidential information,
27:33we've managed to cut
27:34months and months of
27:35processes,
27:36various processes.
27:38What you want us
27:38understand in the nature
27:39of our technology is that
27:41it is meant to take
27:42the knowledge and then
27:42take a lot of different
27:44tasks.
27:44So it can be automating
27:46things like clinical
27:47protocol or writing
27:50or clinical study
27:51or writing.
27:53But we also did a drug
27:55discovery or automated
27:57with AI.
27:57So right now,
27:58we have patented five
28:01treatment candidates
28:02with major discovery in
28:03oncology.
28:04I can already talk about
28:05leukemia because this is
28:07now known.
28:09we discovered something
28:11that is going to change
28:12life of 25% of our
28:13patients in leukemia.
28:15But we also work and
28:16have breakthrough in
28:17neurology and dermatology
28:18and lots of different
28:22therapeutic areas.
28:24And we've also worked
28:25with NVIDIA on some key
28:28tools that is going to
28:29help other researchers.
28:32So for instance,
28:32we had to understand the
28:34impact of non-coding
28:38genes and we had to create
28:42dedicated tools in
28:44transcriptomics to crack
28:45that and optimize the way
28:48it would run on GPUs so
28:50that other researchers would
28:52take advantage of this
28:54progress for even more
28:57programs.
28:58So probably we draw the
28:59line here just to be
29:00respectful to the next
29:01speakers.
29:01Thanks a lot to both of
29:02you.
29:03Thanks to every one of you.
29:05We wish you to enjoy the
29:06VivaTech.
29:06Come see us at Sanofi booth.
29:08We are open for networking
29:09and probably a few more
29:10questions to follow up in
29:11the Q&A.
Comments