- il y a 1 jour
Building the Future of Coding with AI
Catégorie
🤖
TechnologieTranscription
00:00Hi, everybody. Can you hear me? Yes, you can.
00:03Hi, I'm Ingrid, a journalist based out of London.
00:07Great to see you guys.
00:10Thank you very much, you three, for coming.
00:12Yeah, I think I want to turn a tiny bit to face you guys.
00:15So we're talking about AI and coding and the bigger space,
00:20which is very, very hot right now, as a lot of you probably know.
00:24I'm just going to probably kick off with a little bit of a question
00:28around that, and I don't mean to start
00:32with a bit of a doomsday question, but I'm going to.
00:36So with AI and coding, obviously the discussion
00:42constantly seems to be walking around the idea
00:46of what happens to people when AI gets so powerful
00:50that people are no longer necessary.
00:53So I kind of wanted to open up with just a basic question
00:56of what do you see as the end game in that space
00:59when we're talking about AI in the space of coding.
01:03And that can really go to any of you.
01:05And maybe you might want to start with just a quick,
01:07your name before you start talking,
01:09since we didn't really have an introduction
01:11when we came on stage.
01:13So, yeah, maybe I can start.
01:15So, Stéphane Heidinger, I'm from AWS in France.
01:18I'm lucky enough, I've been 13 years in AWS,
01:21working with many companies from startups to big ones.
01:24Maybe let's do the round and then we can answer.
01:27Margarita, COO at Poolside.
01:29Been in deep tech for now 15 years,
01:32running everything from ops, finance, partnerships,
01:35you name it, anything that puts the heartbeat
01:37of a company in the right place.
01:39Hi, everyone.
01:40Guillaume Princeton.
01:41Very glad to be here.
01:43I'm leading EMEA for Anthropic.
01:46I've been there for a few months now.
01:48Before that, I've been in the technology space
01:50for roughly 15 years and excited for this panel.
01:54Okay.
01:55End game.
01:56Stéphane.
01:57It's hard to predict because,
01:58and I would say it's just like if you went back to,
02:02let's say, 1997, beginning,
02:04or let's say almost the beginning of the internet
02:06and trying to predict what will happen
02:08in the next 10 to 20 years.
02:10So, I mean, basically, I think nobody knows.
02:12But the one thing maybe we do see is today AI is helping
02:17and augmenting lots of people.
02:19And in three big areas,
02:21the first one is individual productivity.
02:23And we talk really about coding,
02:25but you can see this in many other areas.
02:27The second one is business process improving,
02:31so doing better, faster, cheaper.
02:34I think we'll talk about agentic on that one.
02:37But the last one,
02:38and I think where I have the biggest expectations
02:41from all of you in the room is think about
02:44what can AI and gen AI allows you to do
02:48as a product, as a service,
02:50that you were not able to do before.
02:52So the end game for me is just more innovation
02:55and new services that you have no idea today.
02:58So you don't think that AI will take over human roles?
03:02Maybe Margarita wants to say something on that.
03:04I just wanted to comment on Stéphane's.
03:07Augmentation, right?
03:09Because I think that's exactly what we believe in.
03:11And to your point on end game,
03:13end game is a very far-fetched point in time right now.
03:16At the pace that our industry is moving,
03:18I think medium term is probably the way to solve it.
03:21And augmentation is really it, right?
03:23I think if all of us in a room start thinking
03:26about the things that we get to not do
03:28because we don't have time,
03:29because we don't have the tools,
03:31because there's just not enough time and resources
03:34to do anything.
03:35And I often go back to the analogy,
03:38and AWS is probably a great one
03:40because of the size of the org,
03:41but who at AWS is a badged employee
03:44and who's a contractor, right?
03:46And I think if we think about all of the resources
03:49that we today bring in the form of contractors,
03:52in the form of different tools
03:53that we can build to our own liking,
03:55all of a sudden with everything that's happening
03:57in all of the AI,
03:59we will first augment,
04:01meaning that we need to learn how to use the tools.
04:03I think this was back to Stefan's first point
04:05in terms of individual productivity and usage of tools.
04:08And only then will we see any transformation
04:10in terms of actual workforce size.
04:13Guillaume, I would love to hear your answer to this question
04:16because, you know, obviously,
04:18you haven't been an ad-therapic for that long,
04:20but your CEO was very outspoken,
04:23made a comment only a couple of weeks ago,
04:27if not a week ago,
04:28about how he's predicting
04:30it's something like 10 to 20% of jobs will disappear.
04:35Half of all entry-level knowledge worker jobs
04:38are going to disappear.
04:40And let's put to one side
04:42all the non-badged employees
04:44who are the ones that can easily be let go.
04:48I think for this debate,
04:50it's really important.
04:51There's kind of the far future,
04:52but there's really what's going on today.
04:55And I think we need to root this debate
04:56in what's happening today,
04:58what are we seeing today,
04:59what's the usage of AI encoding today.
05:04You know, the technology today,
05:07as illustrated by, you know,
05:08the models we released three weeks ago,
05:11Cloud 4 with Sonnet 4 and Opus 4,
05:14has come to a point where you are able to do with,
05:18those models are able to run for hours autonomously
05:22to do very complex tasks, including coding.
05:26And I think that opens the way to several things.
05:30One, developers, you know,
05:32augmentation is the key word here.
05:34We're seeing that over half of the usage of these tools
05:37are now used for augmentation, not automation.
05:40That is today.
05:41Today, it's really about augmentation.
05:43The other element I think that's really interesting
05:45is that this enables non-coders, no-coders,
05:48to build applications and products.
05:51And I think that's a really interesting development as well,
05:54expanding the reach of what code can do.
05:56And the last element is even much larger companies,
06:00when you think about these very large legacy systems
06:03and legacy coding bases,
06:06are able now very relatively easy
06:08to rebase those legacy systems.
06:11and that's something we're seeing more and more as well.
06:14Now, back to the, I think, the initial question.
06:17I think the, nobody, I think that's where we are today.
06:21Now, where are we going tomorrow?
06:23I think that's still an open question.
06:25And I think what's important and what's important to us
06:27is to make sure that becomes a conversation,
06:32that becomes a conversation with academics,
06:34that becomes a conversation with economists,
06:36with policymakers.
06:38And what we're doing to help with that
06:39is we're making transparent,
06:41like, the data on what it is
06:44and how AI is used today
06:47as part of, you can all look it up,
06:49it's public information,
06:50the Anthropic Economic Index,
06:52to help feed those conversations.
06:54because that's not going to be a one-entity answer.
06:57It has to be a multi-entity,
06:59multi-thread, societal answer to that question.
07:02And the companies who are implementing these tools
07:05need to know, you know,
07:07what impact that's having inside their companies.
07:09So you don't see it replacing anytime soon,
07:12but it could be something.
07:13I mean, I feel like it's being held as a possibility
07:16when the leader of your company says,
07:18it's a possibility.
07:19That seems to be the most realistic idea.
07:22I think it's a possibility.
07:24I think it's a possibility.
07:25But I think it's,
07:26I think what, you know,
07:27what Dario was referring to
07:28is we need to have this conversation.
07:30We need to push this conversation.
07:32All the AI models and actors
07:34should be pushing to have this conversation
07:38in the same way we are.
07:40Okay.
07:41Listen, I just,
07:42I'm always very curious about
07:44how leaders in this space feel about it.
07:47but I'm going to switch over
07:49to talking about a bit of
07:50what you guys are doing in your businesses.
07:53Margarita, I just wanted to talk to you
07:54a little bit about Poolside.
07:56Very interesting company.
07:58Has been around for a while,
07:59but I don't believe has actually done
08:01a public shipping of a product yet.
08:06So I think for us, right,
08:09if we think about where others are right now,
08:11both in everything that's individual
08:14and, you know, everything on code
08:16that is Vibe Coding,
08:17we focus very much on the legacy systems,
08:21the hard parts to peers.
08:24We target the enterprise.
08:26We don't come out and make big announcements
08:29because we're working with very specific ones,
08:32very specific segments,
08:34a lot of which we can't even talk about
08:36in the first place.
08:37Everyone who's in the fence, gov,
08:39it's hard to come out in public
08:42and say what we're doing.
08:43I think for us,
08:44we talk about high consequence software.
08:47That's what matters.
08:48I think if we look at where
08:50we can make the biggest impact
08:53for humanity, for all of us here,
08:55outside of what everyone else is doing
08:57at an individual level already,
08:59and it's already a crowded space,
09:00is really in places where
09:04legacy software has eaten the code base,
09:08where onboarding an engineer takes six months,
09:11where people can't find things
09:13in our duplicating work, right?
09:15It's all of the hard deployments,
09:18on-prem,
09:19or a bunch of other deployment types
09:21that no one is tackling right now,
09:24and that's why we also work very closely with AWS,
09:27so that we can actually offer
09:28to all of our customers the possibility
09:30to own their own models,
09:32own their weights,
09:33securely deploy in the privacy of their homes,
09:37and ultimately leverage their code bases
09:39to the best, most efficient way.
09:41So you're not building anything
09:43that you're going to ever release
09:45on general release
09:46in the same way that others are?
09:48I didn't say that.
09:49Okay.
09:49What is the timeline for that?
09:50When can we expect something from you?
09:52I don't answer timeline questions.
09:54Okay.
09:56It's coming, Ingrid.
09:57It will be on Amazon Bedrock for sure.
09:59It will be on Bedrock
10:00whenever it happens,
10:03120%.
10:03And it will be in everywhere else
10:05that everyone is thinking of too.
10:08I think the thing is,
10:10there's so much...
10:11Is the pressure
10:11that you haven't done it though?
10:12I'm sorry to interrupt you,
10:13but are people coming to you
10:15going,
10:16why haven't you done it?
10:17Or is working behind the scenes...
10:19I think everyone who's worked with us...
10:21A viable way of building the business here.
10:23I think everyone who's working with us
10:25does not see any pressure
10:27because they see what we're building.
10:28And so it's a matter of
10:30either we give in to the crowd
10:32and things that we don't necessarily believe
10:35are the right way
10:35to go and tackle all of these problems today,
10:38or we go do it our own way
10:41and then prove everyone that...
10:42If I can turn it around
10:43before Stefan responds,
10:44do you think that companies
10:45like Anthropic and OpenAI
10:47who are, you know,
10:48racing to release
10:49an incremental model here and there,
10:51particularly OpenAI
10:52does that quite a lot,
10:54lots of updates.
10:55Do you think that's almost like
10:56a fool's,
10:57expensive fool's chase?
10:59Absolutely not.
11:00I think what OpenAI
11:00and Anthropic have done
11:01for the world
11:02is unquestionable
11:04and unmeasurable.
11:05It's just very different ways
11:07of going to market.
11:08It's also very different pockets too.
11:10Right.
11:11Okay.
11:11Stefan, you wanted to say something.
11:12This is just interesting.
11:14We have the two extremes
11:15of the spectrum.
11:16I mean,
11:16and it's both writing new code
11:19for new services,
11:19but also, as Margarita said,
11:21the legacy
11:22and a huge, huge potential
11:24and a huge pain
11:24for many of the customers.
11:26Interestingly,
11:27when we look at the data,
11:29any developers in the room,
11:31by the way?
11:32Okay.
11:33We have a bunch of them.
11:36The data we have
11:37shows that developers
11:38spend less than one hour per day
11:40actually writing code.
11:41All the rest is about designing,
11:44it's about maintaining,
11:45it's about documenting,
11:46testing,
11:47meetings, maybe, also.
11:49You name it.
11:50So, it's really about
11:51tackling the entire life cycle,
11:55software development life cycle.
11:57And at the same time,
11:59to come back to the evolution
12:00of the role,
12:03I had one of our architects
12:04at AWS,
12:05last week,
12:06he did a new demo
12:07on Amazon Connect,
12:08so it's a call center.
12:09He did it in two hours
12:11with QChat,
12:13so the long running process
12:15with many supervising the model.
12:18It would have taken him
12:19like two weeks
12:20to do that,
12:21and he was referring,
12:22it's just like having
12:24like three or four
12:25junior architects
12:26working for him,
12:27or somebody
12:28called them minions,
12:29but you take the word
12:31that you want.
12:32And what's interesting
12:33is for one...
12:34He was building it
12:35with AI,
12:35you're saying,
12:36is that what your...
12:37No, he was building
12:37a demo on cloud
12:39using AI.
12:41Which model did he use?
12:42What tools did he use?
12:44He was using
12:45Amazon QDeveloper,
12:46what we call QChat.
12:47I don't know exactly
12:48what the models are behind,
12:51probably a combination,
12:52and that's my point as well.
12:54Very often,
12:55I get the question
12:56from CIOs,
12:57what is the best model?
13:00And, I mean,
13:01vendors have their opinions
13:02on that,
13:02but I would argue
13:04this is not the right question.
13:06And let's shift a little bit.
13:07When you are going
13:08to an IT vendor,
13:09do you ask the question,
13:11what is the fastest database
13:13in the world?
13:14No,
13:15because it doesn't make sense.
13:16Because, yeah,
13:17a vendor can provide it,
13:19but it will be
13:19not cost effective whatsoever.
13:21It's exactly the same
13:22for models.
13:23I mean,
13:23some models are very good
13:24on one side
13:25to supervise,
13:26and cloud, for instance,
13:28is super good
13:29for long-running process
13:30and supervising,
13:31good for coding as well.
13:33But we see also
13:34lots of other smaller models
13:36for very specific tasks.
13:37And actually,
13:38when you start
13:39to do agentic AI,
13:41you need to have
13:42this combination
13:43of all those models,
13:44very, very narrow models
13:46for one task
13:47and a big one
13:48to supervise.
13:50Yeah,
13:50I can maybe just
13:51follow up on that
13:52with two things.
13:53One is kind of
13:54a fun fact,
13:57kind of bringing
13:57a few things together here,
13:59product, coding,
14:01the role of AI.
14:04I talk to customers
14:05every day,
14:05but, you know,
14:06our best kind of user
14:07is actually ourselves
14:08at Anthropic
14:09and how we use
14:10these tools.
14:11The fun fact is,
14:13so we released
14:14Claude Code,
14:15which is our coding agent,
14:18a few weeks ago,
14:19publicly,
14:20openly,
14:21when we released
14:22our new models.
14:23And that Claude Code product,
14:2790%,
14:2790%
14:28of the code
14:29that we wrote
14:30for Claude Code,
14:32the product,
14:33was written
14:34by Claude.
14:36There's a lot of Claude
14:37in there.
14:38but basically
14:39that new product
14:40release
14:40was written,
14:4290%
14:43of the code
14:43was written
14:44by an AI,
14:46which was Claude,
14:46obviously.
14:47That's an interesting fact.
14:49I think back
14:49to what you were saying.
14:50Just, Guillaume,
14:51actually,
14:52we have the same data.
14:53When we released
14:54QChat,
14:55this tool
14:55for long running,
14:56same story.
14:57It was written
14:58by a non-developer
14:59at AWS
15:00using AI.
15:01And I think
15:02the interesting part
15:03related to the augmentation
15:04is,
15:05you know,
15:05did we get rid
15:06of 90%
15:07of the engineers
15:08working on,
15:08no,
15:08of course not.
15:09What we did
15:10is they were
15:12redeployed
15:12to work on other stuff
15:13and release
15:14other products
15:15faster,
15:15right?
15:16So that's how
15:16this is happening
15:17right now.
15:18The other element
15:19I wanted to bring,
15:20you know,
15:21vis-à-vis
15:21what you were saying,
15:23is we're entering
15:25this stage
15:25of kind of
15:26the new AI
15:28virtual collaborator.
15:29When you think
15:30about a real
15:31human collaborator,
15:33what you have,
15:34what you need
15:34is two things.
15:35You need the ability
15:36to work autonomously
15:37and then you need
15:39the power
15:40to take action.
15:42With today's models
15:44and Claude IV's
15:46release is,
15:47you know,
15:47improving that
15:48over,
15:49we have basically
15:51gotten to that point
15:52where autonomous work,
15:55again,
15:55Opus IV can do
15:56up to seven hours
15:58of work autonomously.
16:00and on the
16:01taking action side,
16:03you can now,
16:04especially through,
16:05you know,
16:06various connections
16:07and our integrations
16:08launch,
16:09you can connect
16:10to all sorts
16:11of digital tools
16:12that you have
16:13and so you now
16:14have the possibility,
16:16we're talking about
16:17multi-agents,
16:18you now have the
16:18possibility
16:19to take action
16:21and work autonomously
16:22at the same time,
16:24by the way,
16:25with these agents.
16:26so we're now
16:27really entering
16:27a new era
16:28of AI collaboration,
16:31like human AI collaboration
16:33that we hadn't entered yet
16:34and I'm very excited
16:36to see what we're
16:36going to build
16:37around that.
16:38Okay, Margarita.
16:39Yeah, I just want
16:41to comment
16:41like on both
16:42the usability
16:42which I think
16:43is the point
16:43you're trying
16:44to make
16:44around model performance
16:46being a number,
16:47right,
16:48and usability
16:49and actually
16:49what you get
16:50on the ground
16:51from using,
16:52excuse me,
16:53from using these tools
16:54on the day-to-day basis
16:55with agents
16:56being able
16:57to take action,
16:58right,
16:58and I think
16:58agents being able
16:59to take action
17:00is something
17:00that when the model
17:03leaves outside
17:04of your premises,
17:05right,
17:05it becomes
17:06an IP conversation,
17:07it becomes a do I trust
17:09all of this information
17:11that is so proprietary
17:12to us,
17:13that is so important
17:14in everything
17:14that we've been working on,
17:16how do I give it access
17:17in a safe way,
17:18right,
17:19and so to your question
17:19of does the world care,
17:22I am not sure
17:23there is an option,
17:24right,
17:24I think in the long run
17:26the world has to care
17:27about where these models
17:28are being deployed
17:29in order to bring
17:30and merge
17:31the usability
17:32with the ability
17:33that the models
17:34have today
17:35and the agents
17:35have today
17:36to actually
17:37real work
17:38on a day-to-day basis
17:39to increase productivity.
17:41Well this actually
17:41leads to something
17:42I wanted to ask you
17:43about later
17:43so maybe we could
17:44just move into that,
17:45I mean I think
17:45that I can kind of
17:47see that there are
17:49challenges
17:49on the side
17:51of what do you
17:53believe the models
17:54have yet to do
17:55or what are the problems
17:57with how the models work
17:58and then the other side
17:59of the challenges
17:59I think when you're working
18:01especially in any kind
18:02of an enterprise environment
18:03is what are the enterprises
18:05asking for,
18:06do they even know
18:07what to ask for?
18:09Maybe you guys
18:10can just talk
18:11a little bit about,
18:11I mean you mentioned
18:12what you're effectively
18:14talking about
18:14is model sovereignty
18:15and this idea
18:16of everyone owning
18:17their own models
18:17or at least
18:18their own data
18:19that powers the models.
18:21I guess I'm wondering
18:22how far along
18:24are the companies
18:25you're talking to?
18:26Obviously if you're
18:27talking about AI coding
18:29it's you're speaking
18:30with a very specific audience
18:32that understands
18:33what they're doing
18:34but the minute
18:34you leave that
18:35it's a very non-technical
18:37group of people
18:38I'm guessing.
18:39So yeah maybe
18:39I don't know
18:40Stéphane
18:41I can start
18:42maybe on that one
18:43first and Marguerite
18:44you touched upon
18:45you need data
18:46of course
18:47and there's a saying
18:48from one of our experts
18:49or I was saying
18:49in French
18:50is
18:50no data
18:53no chocolate
18:53it doesn't translate
18:54very well anyway
18:56but it puts back
18:57on the top
18:58of the pile
18:59all the data projects
19:00because if you
19:01just use models
19:02on public data
19:03it doesn't bring
19:05much value
19:06and same thing
19:06for code as well
19:08you need to leverage
19:10your code base
19:10of course.
19:12The second thing
19:13is
19:13when you say data
19:15you need security
19:15of course
19:16because it's
19:17all the knowledge
19:18of your company
19:19we have some
19:21technology
19:21and very high
19:23standards of security
19:24especially one
19:25called Nitro
19:25if you want to know
19:26about Nitro
19:27go to the AWS booth
19:28we'll explain to you
19:29the idea
19:30out of placement
19:30please
19:30keep going
19:31the question I get
19:34is how to protect
19:34data from US laws
19:36or extra European laws
19:37the message is
19:38we have answers
19:40for that
19:40and the last thing
19:41is really to control
19:43the outputs
19:43of the models
19:45prevent hallucinations
19:47prevent harmful content
19:49think of that
19:49so it's always
19:50also working
19:51on what you call
19:52the guardrails
19:53making sure
19:54you don't have
19:55offensive content
19:56and one thing
19:57we're exploring
19:58as well
19:58is using
19:59automated reasoning
20:00for guardrails
20:01ok
20:02what does it mean
20:03it means that
20:04we have a team
20:05working on
20:06automated reasoning
20:07which is math
20:08so how to prove
20:10math theorems
20:11automatically
20:11and the thing is
20:13we take documents
20:15like policy insurance
20:16for instance
20:17and the models
20:18recreate
20:19a kind of
20:20mathematical models
20:21of what is allowed
20:22what is not allowed
20:23in the policy insurance
20:24which is then reused
20:25when we ask
20:27for an inference
20:27for instance
20:28is a claim
20:29covered by this
20:30policy insurance
20:31the model will do
20:32its work
20:33but then we apply
20:34the mathematical models
20:36to make sure
20:37we are within
20:38what's allowed
20:39by the model
20:40so it reassures
20:42our customers
20:43into the fact
20:44that the model
20:44is not going
20:45completely south
20:46and not hallucinating
20:48are you seeing
20:49though generally
20:50I mean it's quite
20:50interesting
20:51because I would say
20:51poolside
20:52because as you said
20:55you're only working
20:56within enterprises
20:57you're not having
20:59this kind of
21:00wide release
21:01facing approach
21:02are the conversations
21:05typically
21:05getting
21:08are what companies
21:10asking for
21:11or organizations
21:11asking for
21:12close to what
21:13you can deliver
21:14or is there
21:14still a huge gap
21:16so I think
21:17to Stefan's point
21:18data is all that matters
21:19and there's a reason
21:20I mean
21:20that's all they care about
21:22or are you saying
21:23data is all that matters
21:24using the data
21:25effectively
21:26and leveraging it
21:27to the extent
21:28the tools
21:29can leverage it
21:29today
21:30is what matters
21:31right
21:31at the end of the day
21:33there's a reason
21:34why we went
21:34first to
21:36FSI
21:36I mean
21:36financial services
21:38where the
21:39leveraging the internal data
21:41makes a world
21:42of a difference
21:43right
21:43gov defense
21:44to that point
21:44in terms of
21:45proprietary information
21:48to your question
21:50do they know
21:50what they want
21:53they think
21:53they know
21:54what they want
21:54I think
21:55what they want
21:56is an end goal
21:57I think
21:58the hard thing
21:59right now
22:00is a step-by-step
22:01process
22:01to lead
22:02all of us
22:03in the right direction
22:04what are the
22:05critical steps
22:06today
22:07tomorrow
22:07this year
22:08that need to be
22:09taken
22:10and how to start
22:12how to start
22:13doing change management
22:14because at the end
22:14of the day
22:15I think
22:15to my point
22:16on usability
22:17and my biggest
22:18pet peeve
22:18being model benchmarks
22:20versus actual
22:21usability
22:21in a day-to-day
22:22people are not
22:24using it
22:25to the best
22:25of the extent
22:26the tools
22:26are there
22:26right
22:27I mean
22:27Claude
22:28I mean
22:29generally speaking
22:30from what I understand
22:31there aren't
22:32any universal
22:33deployments
22:33even in the
22:34largest enterprises
22:35they're still
22:35working in
22:36pretty small
22:36project
22:37based
22:37absolutely
22:39beta testing
22:41it's a change management
22:42it's not in live
22:43action
22:44it's a huge transformation
22:45and I think
22:46it's a transformation
22:47both at a systems level
22:48and all of the tools
22:49that any enterprise
22:51can leverage
22:51and deploy today
22:52but as well
22:53on a processes basis
22:54right
22:55because if the processes
22:56are not changing
22:57when the tools
22:58are changing
22:59there's a
23:00big disconnect
23:01like there's
23:02it's really hard
23:03to implement
23:04you think the long-term
23:05goal is going to be
23:06or long-term
23:07end game
23:08is going to be
23:09around businesses
23:11holding their own
23:13data
23:14for every
23:15AI process
23:16or do you think
23:17there will be a mix
23:17of some
23:18who are
23:18doing that
23:19and some that are
23:20happy to give
23:21everything to
23:22you know
23:23the big
23:23foundation model
23:24companies
23:25to put into the mix
23:26I think
23:28sovereignty
23:29will
23:29definitely
23:31will reign
23:31reign
23:32yeah
23:32I do think
23:33everyone will still
23:34to a certain extent
23:35have to make peace
23:37with some of it
23:38not living inside
23:39and most of it all
23:40because I think
23:41we've all been
23:42living with
23:42data that we're
23:44gathering from
23:44multiple places
23:45right
23:45if we take
23:46an insurance company
23:47for that matter
23:48right
23:48you're leveraging
23:49a bunch of data
23:49from elsewhere
23:50so there are
23:50things that you
23:51will allow
23:52to leave outside
23:53of your premises
23:53the majority
23:55of it
23:55won't
23:56right
23:56we had a
23:57customer last
23:57week that
23:58you know
23:59would open
24:00AI's acquisition
24:01and obviously
24:02you guys
24:03wait
24:03acquisition
24:04of Windsurf
24:05of Windsurf
24:06and you guys
24:07turning off
24:07well it hasn't been
24:07confirmed yet though
24:08I mean
24:08do you know
24:10that it's definitely
24:10closed
24:11no no no
24:11I'm not saying
24:12that Ingrid
24:12especially because
24:13you're a journalist
24:13I know where
24:14this is going
24:14I'm definitely
24:15not confirming
24:16everyone talks
24:16about it
24:17as if it's happened
24:18but there's been
24:18no confirmation
24:19to make
24:19is that obviously
24:20Intravic
24:21turned off the tap
24:22right
24:22yes they did
24:23yeah
24:23I am going to
24:24ask you about that
24:25and I think
24:26to a certain extent
24:27this is giving them
24:28all the
24:28oh my god
24:29okay this can
24:30actually happen
24:30I can actually
24:31all of a sudden
24:32be turned off
24:32from the model
24:33that I've been using
24:34and that I love
24:35and that makes
24:36people think
24:37yes
24:38exactly
24:38I mean
24:39when an AI company
24:41turns off the taps
24:42you know data
24:43is a big deal
24:44if an AI company
24:46doesn't want
24:47its data being shared
24:48then why does
24:49any insurance company
24:51want it
24:51or anyone else
24:52for that matter
24:53any individual
24:54please talk
24:55so a few
24:57important points here
24:59one
25:01Anthropic
25:01is an AI safety
25:02first company
25:03that builds
25:04safe
25:05responsible
25:06ethical
25:07models
25:08that are at the frontier
25:09that is what
25:10we're focused on
25:11and that is
25:12back to the topic
25:13that is what
25:14enterprise companies
25:15look at first
25:16and are really
25:17interested in first
25:17I'd say in particular
25:19in Armenia
25:19in particular
25:20in Europe
25:21because that element
25:22of trust there
25:23that you won't get
25:24hallucinations
25:25that you won't get
25:26all those
25:26side bad effects
25:28are key to
25:30enterprise customers
25:30and the way they think
25:32about their business
25:33I talk to these
25:34customers every day
25:35and they're saying
25:37essentially two things
25:39one is
25:42I don't want
25:43my data to be used
25:44for training purposes
25:46that's a really
25:47important one
25:48and that's something
25:49we never do
25:50two is
25:52my big problem
25:53is
25:54my company
25:55has data
25:56all over the place
25:57how do we solve that
25:59and that's where
25:59MCP comes along
26:00that's what we call
26:01we call
26:02the context crisis
26:03which is
26:05what these large
26:06companies are trying
26:06to solve for
26:07is complexity
26:08data complexity
26:09internal complexity
26:11and with the appearance
26:12of MCP
26:13model context protocol
26:15I think we're going
26:17to see
26:17I mean we are
26:18starting to see
26:19usage
26:21of AI
26:21that we've never
26:22seen before
26:23in enterprise companies
26:24because we're able
26:25to bring together
26:26things
26:26that are not only
26:28related to one function
26:29but across the company
26:30and there's huge value
26:32there
26:32now just
26:33I mean
26:33just you know
26:35we're fans
26:35I think it's
26:36you know
26:36we haven't mentioned
26:37a customer yet
26:37I think it's worth
26:38mentioning a few
26:39customers
26:39and what they do
26:41make it quick
26:42because I have
26:42another question
26:43I want to ask
26:43before we have
26:44to leave the stage
26:45so what I was
26:46relating to
26:47is you know
26:48we're working
26:49with SNCF
26:49with Bujoo
26:50with Buik Didikum
26:52on specifically
26:53those topics
26:54and others
26:55and that's
26:56really what
26:57they're looking for
26:58and what I'm hearing
26:58day in day out
26:59from talking
27:00with these customers
27:00okay
27:01I want to talk
27:02about vibe coding
27:03please
27:03before we have
27:04to get off the stage
27:05is it a myth
27:06or a reality
27:07I mean you mentioned
27:08some queue
27:09was built
27:09by a non
27:10coder
27:11yeah more
27:12and more
27:12and especially
27:14so it's a reality
27:15of course
27:16of course
27:16and I
27:17so I mean
27:18the term vibe coding
27:19there are some
27:19very cheesy
27:20it's very cheesy
27:21and lots of fun
27:22being made about that
27:23but it really changes
27:24the job of a developer
27:25in a way
27:26as we mentioned
27:27is removing
27:28some of the
27:30boring stuff
27:31in some ways
27:31but as you mentioned
27:32Guillaume
27:33MCP servers
27:34so it's really
27:35to connect
27:35your models
27:37to any
27:38application
27:38or any data source
27:40I mean
27:41we see
27:41every day
27:42tons of new
27:42MCP servers
27:43coming out
27:44I would urge you
27:45don't write code
27:46just ask Claude
27:48to do it
27:48I mean
27:48it is probably
27:50the first time
27:50to build
27:51those new components
27:52just don't write
27:53anything
27:53it's all there
27:56yeah
27:57Margarita
27:57I probably have a very different opinion
27:58so maybe
27:59Guillaume
27:59do you want to go first
28:02vibe coding
28:02myth or reality
28:03yeah I mean
28:03vibe coding
28:05as I mentioned
28:05at the beginning
28:06I think
28:07the level
28:08the level
28:09at which
28:09we are on the technology
28:11you know
28:11Claude 4
28:12Opus 4
28:12Sonnet 4
28:13basically enables
28:14non-coders
28:15to be able to code
28:16you know
28:17all of a sudden
28:18is it really working though
28:19I mean
28:19everything I've heard
28:20this is happening
28:21I have no data
28:22like I have no data
28:23to you know
28:24to kind of discuss
28:25today on that
28:25but we're seeing it
28:26yeah
28:27let's you know
28:28let's see if it's
28:29long-term trend
28:29but I'm seeing it happening
28:32yeah Margarita
28:33one way to look at it
28:35I mean
28:35IT started with
28:37assembly language
28:38and more and more
28:39evolved
28:40programming language
28:41just think about it
28:43it's just
28:43your new programming language
28:44in a way
28:44is French
28:45or English
28:46or Arabic
28:47perhaps
28:49Margarita
28:49coding is the key word
28:51I think software engineering
28:52and software development
28:53is very different
28:53from just the pure
28:54active coding
28:55and I think there's
28:56a lot more to it
28:57so whether or not
28:58we can automate
28:59a few things
28:59and I would definitely
29:00call it automation
29:01almost more
29:02than coding
29:04is one thing
29:06whether we can deploy
29:06and we can build
29:07and we can keep
29:08maintain
29:08you know
29:09and actually scale
29:10those systems
29:10is a completely
29:11different thing
29:12do you think
29:13there'll be vibe
29:13maintaining
29:15vibe compiling
29:16I think we'll all
29:17be able to do
29:18much more
29:19right
29:19to the initial point
29:20on augmentation
29:21we'll all be given
29:22the tools
29:22to do anything
29:23we want
29:23the level of creativity
29:25will be individual
29:27right
29:27what we choose
29:29to do
29:29with those tools
29:29will be completely
29:30up to us
29:32and I think
29:33when it comes
29:34to coding
29:35I seriously hope
29:36when we all
29:37code
29:38or when we all
29:39engineer a system
29:40it won't be
29:41through an IDE
29:42it won't be
29:42exactly as we're
29:43seeing it today
29:44right
29:44there's
29:46no way to scale
29:47vibe coding
29:48if we keep it
29:50in the tools
29:51that we've used
29:51up until now
29:52and that the developers
29:53only use
29:54right
29:54I think
29:54we're going
29:56to see
29:56a lot more
29:57and we're going
29:57to see an evolution
29:58that will be
29:58magical for all of us
29:59and maybe just
30:01to kind of
30:01bring the two
30:02sides together
30:03I think
30:03there's something
30:04about the complexity
30:05of the coding
30:05work as well
30:06right
30:06the most complex
30:08coding tasks
30:09will always have
30:11a piece
30:13or an important
30:14piece
30:14of human feedback
30:15in there
30:16of needing
30:17to have the human
30:18feedback
30:18to kind of
30:19so is it
30:2110
30:2220
30:22is it
30:2380%
30:23of that
30:24kind of
30:25coding activity
30:26that's going
30:26to be run
30:27by AI
30:28is a question
30:29and depends
30:29on the task
30:31but again
30:32there is some
30:32vibe coding
30:33where most
30:33of that
30:34it can be
30:34maybe not
30:35for the most
30:35complex pieces
30:36okay
30:36we're now
30:37over time
30:38I wish
30:39all of you
30:40who are coders
30:41out there
30:42that you'll
30:42stay and work
30:44but thank you
30:45so much
30:45for your time
30:46you three
30:47it's been very
30:48interesting
30:48thank you
30:49thank you
30:49thank you
Commentaires