- il y a 2 jours
AI Agents and the Future of Programming
Catégorie
🤖
TechnologieTranscription
00:00Good morning everyone, welcome to day 2 of VivaTech.
00:06For those of you who were here just moments ago during the LVMH awards,
00:12well those were quite some awards, right?
00:14Congrats to the winners.
00:17For those who missed my intro yesterday, my name is Ariane Alcorta.
00:22I'm an international journalist and event moderator.
00:25And it is my absolute pleasure to be guiding you through the rest of the program.
00:32It is my third consecutive year with all of you here on stage 1,
00:37plunging into the world of AI.
00:41Speaking of which, I do hope you've got your dose of caffeine this morning
00:45because we have a packed day.
00:49Yesterday we welcomed some top-tier speakers on this stage
00:53and AI definitely took center stage.
00:57This morning we will be diving into the research and development side of AI.
01:04So if you're a nerd, an AI nerd, like we proudly are,
01:10well then you're in the right place.
01:12Especially if you want to get to know a bit more
01:15about the programming and development side of it all.
01:19So we're starting strong with the next evolution of software development.
01:26AI-powered software engineering agents, or SWE agents for short.
01:33In simpler terms, it's the future of programming,
01:36where intelligent agents can write, test, and deploy code on their own.
01:43So to guide us through this exciting terrain,
01:46we will be hearing from none other than the CEO of GitHub.
01:52If the name doesn't ring a bell, then thank Copilot.
01:59He is, or the company is, the creator of Copilot.
02:03And it is actively reshaping how code is written.
02:08So without further ado, I want you to give a really strong
02:13and energetic welcome to the following speakers.
02:17Thomas Duncan, CEO of GitHub,
02:19in conversation with Stephen Levy, Editor-at-Large at Wired.
02:26Good morning.
02:34Thank you.
02:35You all having a good time here?
02:38Okay, great.
02:39Well, I'm really delighted to introduce Thomas Dompke,
02:43the CEO of GitHub.
02:46He joined Microsoft in 2014
02:48and joined GitHub a few years later,
02:51became its CEO in 2021.
02:54I think it's fair to say that since you got there,
02:57probably the key challenge has been managing the AI tools,
03:01which, you know, revolutionize coding.
03:04You'll tell me if that's not fair, but you're nodding.
03:06It is, and that's what we're going to focus on in this chat.
03:10Welcome, Thomas.
03:11Thank you so much for having me.
03:13Really excited to be here again.
03:15Okay.
03:15So first, tell me a little bit in terms of numbers,
03:18what's happening in GitHub in terms of coding with Copilot,
03:24which is your AI system there?
03:28How many people are doing that?
03:30What's the percentages?
03:31You know, how has it transformed that platform?
03:35The last time we announced was that more than 15 million developers
03:39are using GitHub Copilot,
03:40and we have been on this journey for now a little bit under five years.
03:46Orchie, five years, June 2020 is when we started this journey.
03:50And if you look at where we are today,
03:52it's that most developers are using AI,
03:54are using Copilot to write code,
03:57to write programming language,
03:58to convert their ideas into reality,
04:01into, you know, the language of the machine.
04:06And depending on how you prompt the agent,
04:10I think most people can write 90,
04:12maybe even 99% of all code with the help of AI.
04:16Wait, what was that number again?
04:18That seems a little high.
04:20Well, you know, the 1% left is effectively the prompt
04:23that you're writing
04:24and the instruction that you're giving to the agent.
04:27And I'm sure, you know,
04:28some of you have seen the demos
04:30when you prompt agent mode or the coding agent
04:32and you tell it,
04:33I want to build a snake game
04:35or I want to build, you know,
04:37a web page that shows me all the apartments
04:39that are available in Paris.
04:41And then you write another prompt to refine that.
04:44And it's very similar to an image model.
04:46And we have all seen, you know, these image models.
04:48And the first image is almost never exactly what you want, right?
04:51But you know that you write another prompt
04:53and then you get a better image
04:54and the same is happening with software development.
04:57So I want to kind of get into the implications of that.
05:00But first, I know that you've been talking a lot
05:02about the AI agents doing the programming
05:06as opposed to just using Copilot.
05:09What do you mean?
05:09What is the difference between introducing agents
05:12into the process and, you know,
05:16technically what are those
05:18as opposed to just using, you know,
05:21Copilot as it was introduced
05:23and, you know, and then developed
05:25over the last couple of years?
05:26Yeah.
05:27In the first few years,
05:28Copilot was simply auto-completion.
05:30It would predict the next word,
05:32the next few lines of code, you know,
05:35whatever comes next after where the cursor is.
05:37And then with ChatGPT,
05:39we went into the chat modality, right?
05:41And now you could ask questions,
05:42explain things, explore topics.
05:45And the agent adds now what we call tool calls to that.
05:48And so the agent can actually run commands
05:52on your comment line
05:53or can call into something called MCP servers,
05:56model context protocol.
05:57It can generate files on its own.
05:59So in contrast to Chat,
06:01it's not only a question and answer game.
06:03It's actually you give it a task
06:04and then it does that task for you.
06:06It might modify existing files,
06:08add new files, run those files,
06:10look at the output,
06:12you know, analyze the error message
06:13and solve that problem by itself.
06:16Well, you know, I think, you know,
06:18everyone says this is the year of the AI agent, right?
06:22But the thing preventing AI agents
06:25from coming into widespread use
06:28and the fear of that
06:30is that when you give some autonomy to AI,
06:34things can go wrong.
06:36You know, it can hallucinate or whatever.
06:40code, things have to be pretty well taken care of there.
06:45How have you dealt with the problem
06:46that AI may not always get things right?
06:50We are dealing with it
06:51as we deal with most problems in companies,
06:54which is we need humans.
06:56You know, there's the myth
06:59that AI is going to require less humans in our companies.
07:02I think that's only if you assume
07:04that we're building exactly as much
07:07as we're building today.
07:08But if I'm saying AI is going to write 90% of the code,
07:11that means we're writing so much more code.
07:13Who's going to review all that code?
07:15Who's going to make sure
07:16that what the AI has generated
07:18actually aligns, you know,
07:19with the business goals,
07:21with the security compliance posture,
07:23with the correctness, if you will,
07:25of what we're trying to build?
07:26And so I think we actually have the human
07:29as the operator, the conductor,
07:31the person that reviews all that work.
07:35And what was I going to say?
07:37And I think that's going to be crucial
07:39that we still understand code.
07:41In IT, we have this concept of zero trust.
07:44In many companies,
07:46you don't let any developer
07:47just deploy to production.
07:48You need a second developer
07:49or a third developer
07:50to review the work of the first developer
07:52to make sure that what goes into
07:54onto your production service
07:56actually doesn't introduce
07:57security vulnerabilities
07:58or an insider attack,
08:00these kind of things.
08:01With AI, we're going to take that
08:02to the next level
08:03where we constantly have to look at
08:04what the agent actually done
08:06and is that good for us to ship?
08:08So those are responsible coders
08:11who are going to oversee
08:12what the AI agents do.
08:15Isn't there, you know,
08:18an expectation
08:18that people get used to a tool
08:20just like a lot of people
08:22when they self-drive their Teslas,
08:25maybe they don't put their hands
08:26on the wheel,
08:27maybe they, like,
08:28look at the phone
08:29or go to sleep
08:30or something like that,
08:31that a lot of people
08:32will do their coding
08:33and they won't check,
08:35you know,
08:35because, you know,
08:36it could be so productive,
08:38you know,
08:38just prompting the thing
08:39to do it at once.
08:41I mean, I'm sure
08:42there's going to be people
08:43that are just YOLO-ing it,
08:44you know,
08:45you only live once approach
08:47in software development.
08:49There's this term
08:49Vibe coding,
08:50and I think, you know,
08:51when you do work
08:53on a hobby project
08:54on a Sunday,
08:54that's fine.
08:55You know,
08:55when I work on my own
08:56hobby projects,
08:57I don't write any test cases
08:58and I don't send a pull request
09:00against myself,
09:01not because I'm perfect,
09:02because I'm doing this
09:03out of fun
09:03and I don't have to ship
09:06something where millions
09:07of people are relying
09:09in my hobby projects,
09:10of course,
09:10but as the CEO of GitHub,
09:12I have a huge responsibility.
09:14If I lose any one
09:15of your private source code,
09:16you're not going
09:17to be happy with GitHub.
09:18lose in the sense
09:19of, you know,
09:20making it accessible
09:20to an attacker
09:21or leaking it
09:22to the internet.
09:23And I think that's going
09:24to be true
09:25for most companies.
09:26They're going to have
09:27a huge responsibility
09:28to verify
09:29and to understand
09:30what the agent
09:31is actually doing.
09:32And we're really far away
09:34where we just tell an agent,
09:35you know,
09:36book my summer vacation,
09:37and you don't verify the work
09:38and you don't check
09:39the dollar amount.
09:40You just show
09:41at the airport
09:42and you try,
09:44kind of figure out
09:45what did the agent
09:45do for you.
09:46So I think
09:46if you take
09:47these simple examples,
09:49like you realize
09:49that the agents
09:50can only help us
09:52in translating
09:53our own ideas
09:54into reality,
09:55they cannot generate
09:56those ideas for us
09:57and they cannot verify
09:58that what was created
10:00actually matches
10:01those ideas,
10:02those business goals,
10:03you know,
10:03the mission and vision
10:04that we have
10:05for our companies.
10:06So if I'm a coder
10:08and I'm not using AI,
10:11am I riding a horse
10:12in the age of the automobile?
10:14Well, there's still coders
10:15that are working
10:16on mainframes,
10:18building,
10:19working in a programming language
10:20called COBOL,
10:21which was invented
10:22in the 50s
10:23when Eisenhower
10:24was the president.
10:25You don't have to raise your hand
10:26if you're one of those people,
10:28I'm sorry.
10:29But that's the reality
10:30of the software,
10:32of the technology industry,
10:33that we have a lot of
10:35legacy code,
10:37technical debt,
10:38you know,
10:38things that are around forever.
10:40I saw some news article
10:42the other day
10:43that some federal governments
10:45are still looking
10:45for people
10:46that have experience
10:47with Windows 95.
10:49That's 30 years ago,
10:50you know,
10:51and so I think
10:52there's still people
10:53that will have to manage
10:53all this old technology
10:56in old school ways,
10:57but the top of the game
11:00today is using AI
11:01and we see a lot of startups
11:03that are able to
11:04ship so much faster
11:06with the help
11:07of a co-pilot.
11:09It's interesting
11:10because some people feel
11:11this is sort of
11:12an apocalypse
11:13in the coding world,
11:16but you,
11:17I read a blog post
11:19that you did
11:20where you kind of
11:20contextualize it
11:22in sort of like
11:23different waves
11:24in the history
11:26of programming.
11:27Can you talk about that?
11:28You know,
11:29Mark Andreessen
11:30once wrote this blog post
11:33saying software
11:34is eating the world
11:34and I think,
11:35you know,
11:36the world is drowning
11:37in software
11:38and most developers
11:39are drowning
11:39in their work.
11:40They're drowning
11:41in millions
11:42of lines of code.
11:43The job of a developer
11:44in almost any company
11:46unless it's day one
11:47is going to be
11:48to manage
11:49somebody else's code
11:50and fix all the bugs
11:52that somebody else
11:52has introduced.
11:53Software development
11:54inherently
11:54is the job
11:56where you work
11:56in the entropy,
11:58the chaos
11:59that somebody else
12:00has created
12:00and that entropy
12:02is ever growing,
12:03you know,
12:03because we're all
12:04adding more code
12:05and now with AI
12:06we're adding even
12:06more code to it
12:07and so we need AI
12:09to deal with the complexity
12:11that AI itself
12:12introduces
12:13into our systems.
12:15Well,
12:16you know,
12:17I mean,
12:17I think it is
12:19something different.
12:19I mean,
12:20the question is
12:20if people are dependent
12:22on AI
12:24to do the hard work
12:27of coding,
12:27you know,
12:28do they lose
12:28the facility
12:29for understanding
12:31really what's
12:32going on there?
12:35That shortcut
12:36means that,
12:38you know,
12:38like a generation
12:39of coders
12:40won't ever learn
12:41the hard stuff
12:43and, you know,
12:44and then
12:44not have a real grasp
12:46about how the machine
12:47really works.
12:49Some definitely
12:50will lose,
12:51you know,
12:51some knowledge
12:52and we have
12:53seen that,
12:54you know,
12:54ever since,
12:55you know,
12:56the 1990s.
12:57Who here,
12:57you know,
12:58still knows about,
12:59you know,
13:00the constraints
13:01of MS-DOS
13:02and that you can
13:03only use 640k of RAM
13:04or how to write
13:06assembly language
13:07to optimize a game
13:08for a pension processor,
13:10right?
13:10That knowledge
13:11hasn't gone away,
13:12but there's a really
13:13small group of developers
13:14that still know that,
13:15while the majority
13:16live, you know,
13:17in the web
13:18on clusters
13:19of Kubernetes
13:21and Docker containers,
13:22you know,
13:23running in the cloud
13:24where you don't even know
13:24where the server is,
13:25let alone what chip
13:26is in the server
13:27because the virtual machine
13:28you're running on
13:29has, you know,
13:30a simulated
13:32hardware environment.
13:33And so I think
13:33that's just the nature
13:34of what we do
13:35in software development,
13:36that we're always
13:37moving up the abstraction
13:38ladder.
13:39We have gone from,
13:40I have to teach it
13:41all myself
13:41with books and magazines
13:42to,
13:43I go to internet forums
13:44to,
13:45I use open source libraries
13:47to now AI
13:48and, you know,
13:49nobody here in this room
13:50is not building
13:51on top of 90%
13:53open source
13:53in their stack.
13:54They're not building
13:55an operating system,
13:56they're not building
13:56a programming language,
13:57they're not building
13:57an editor,
13:58they're not building
13:59the node package manager
14:00or the Ruby on Rails
14:03architecture.
14:04They react,
14:05you know,
14:05all these components
14:06that we're already using.
14:08So we already live
14:09in a world
14:09where we essentially
14:10gave commit access
14:11to millions of open source
14:13developers around the world
14:14that contribute
14:15to our systems
14:16and we're building
14:17the 10%
14:18of the top
14:18of the stack
14:19and any
14:20software development company,
14:22any company
14:22I'm meeting with,
14:23they're all telling me
14:24we have way too many
14:25things in our backlog,
14:26we can't get things
14:27done fast enough,
14:29we have all these
14:29compliance
14:30and regulatory requirement,
14:31you know,
14:32here in Europe
14:32with GDPR
14:33and DSA
14:34and DMA
14:35and some other acronym
14:37that you also have
14:37to comply with
14:38and developers
14:39are longing
14:40for something
14:40that ultimately
14:41gets them back
14:42to what they actually
14:43love doing
14:43which is taking
14:44their ideas,
14:45you know,
14:45all the ideas
14:46of the product managers
14:47and making them reality.
14:49Can,
14:51you know,
14:51maybe you could extend
14:52into the future
14:53a little bit,
14:54but are these...
14:55I'm living the future.
14:56Oh, well, okay.
14:58But, you know,
15:00can,
15:01will this create,
15:03will the AI agents
15:03create great code?
15:05You know,
15:06there's the myth
15:07of the 100X programmer,
15:08right,
15:09who's so much
15:10more valuable,
15:10you're going to pay
15:11that person
15:12a million dollars.
15:14Yeah.
15:14Will those agents
15:16be 100X coders?
15:20They will be 100X coders
15:22because they can
15:23output so much
15:24more code
15:25in so much
15:26short amount
15:27of time
15:28and while,
15:29you know,
15:29developers ultimately
15:30have a physical limit
15:32of what they can do.
15:33Look,
15:33you know,
15:34we all have
15:34infinite number
15:35of ideas,
15:36but we only have
15:37a very limited window
15:38during our day
15:39where we have
15:41high energy,
15:42high creativity.
15:43You know,
15:43for me,
15:43it's often
15:44after two coffees
15:45in the morning
15:46when I'm at my best
15:47and the longer
15:47the day goes,
15:48the more I'm
15:49still able to do,
15:50you know,
15:50the chores
15:51and doing the dishes
15:52but I'm certainly
15:53not able to be
15:54the, you know,
15:55a good developer
15:56and a creative developer
15:57and it all ultimately
15:59comes down to trust
16:00and in my head,
16:02you know,
16:02as a parent
16:02of two boys,
16:03the question is,
16:05do I trust
16:06the Uber driver
16:07in San Francisco
16:08or do I trust
16:09Waymo in San Francisco
16:11to take my 13-year-old
16:12to soccer practice
16:13and I actually think
16:14the answer
16:14for many parents
16:15is already
16:16I trust the Waymo
16:17more than I trust
16:19the human.
16:19Now,
16:19in software,
16:20it's still the other way around.
16:21We trust humans more
16:22and that's a good thing
16:23because of all the
16:24security issues
16:26and incidents
16:27that we had
16:27in recent years
16:28but we got to build
16:30the trust
16:30with those agents.
16:31We got to trust them
16:32to a certain degree
16:32to do some of the work
16:34because those that do,
16:35and those that can use
16:36these agents
16:37to the most efficient grade
16:40are going to leapfrog
16:41everybody else
16:42in the industry.
16:43So you've just compared
16:44coders
16:45to, like,
16:47taxi drivers.
16:48No,
16:49I compared coding agent
16:50to a self-driving car
16:52and, look,
16:53a self-driving car,
16:54you know,
16:54has a huge
16:57trust question,
16:58right,
16:58because we have seen
16:59the news of self-driving cars
17:00causing accidents
17:01or even just traffic jams
17:03and some of them
17:04are funny
17:04and others are not,
17:05but it's the same problem.
17:07How do you build
17:09the trust for this agent
17:10and we have seen
17:10how Waymo did that
17:11in San Francisco
17:12and I think we're going
17:12to see that
17:13in many other industries
17:14as well.
17:15Well,
17:15I guess,
17:16you know,
17:16if you talk
17:17to young people
17:18who,
17:19you know,
17:20have just been trained
17:21in coding,
17:22they're graduating
17:23from college,
17:24they're angry.
17:25They're saying,
17:26wait a minute,
17:26I prepared
17:27for this job
17:28and they think,
17:30maybe you have
17:31a different opinion,
17:32there's going
17:32to be fewer
17:33of those jobs
17:34because the people
17:35who have the jobs
17:36will be so much
17:36more productive
17:37because they're using
17:38AI agents
17:39and, you know,
17:40co-pilot for GitHub.
17:42What do you say
17:43to them?
17:44Do you say,
17:45you know,
17:46only, you know,
17:46most of you
17:47aren't going
17:47to be employed
17:48or is this a world
17:50where everyone's
17:52a coder
17:53and, you know,
17:54the opportunities
17:55expand?
17:56I would love
17:57a world
17:58where everybody
17:58knows how to code
18:00in the same way
18:00that I love a world
18:02where everybody
18:02knows how to read
18:03and write
18:03and everybody
18:04has a basic
18:05understanding
18:05of math
18:06and art
18:06and science.
18:07That doesn't mean,
18:08you know,
18:08just because you learn
18:09all these things
18:10in schools
18:10and I think
18:11actually computer science,
18:12we should teach
18:12that from first grade
18:13here in France
18:14and Germany
18:15and in the US
18:15everywhere around
18:16the world
18:16as a fundamental
18:17skills
18:17because you all
18:19have a computer
18:19in your pocket
18:20that runs software
18:21yet not all of you
18:22are able to actually
18:23build software
18:24for that computer.
18:24It's a fundamental
18:25human skill
18:26that everybody
18:27has.
18:27That doesn't mean
18:28everybody has to
18:29become a professional
18:29software developer
18:30just as I didn't
18:31become a physicist
18:32even though I wanted
18:33to become a physicist
18:34and I tell students,
18:36you know,
18:37whether they're in high
18:37school or my kids
18:39are middle
18:39in high school
18:40that this is the most
18:41exciting time
18:42to become a software
18:43developer
18:43because you see
18:44people leaving
18:44college and university
18:46and they're building
18:47startups at fantastic
18:49valuations
18:50with a much smaller
18:52team
18:52that was ever
18:53done before
18:54and we have seen
18:55actually
18:55we're seeing a repetition
18:56of what happened
18:57with the internet
18:58in the 1990s
18:59where all of a sudden
19:00you could build
19:01a worldwide
19:02business
19:02with worldwide
19:03payment processing
19:04and worldwide
19:05shipping
19:05out of your garage
19:07and that garage
19:08no longer has to be
19:09in Silicon Valley
19:09or Seattle
19:10it can now be here
19:11in Paris
19:12it can be in Berlin
19:13it can be in Hyderabad
19:14and I think that's
19:15what AI really
19:16enables for all of us
19:17so GitHub
19:18besides being
19:20a repository
19:21of code
19:21it is a community
19:23that has been
19:24just a key
19:25component of it
19:26through its history
19:27I'm wondering
19:28as AI agents
19:31proliferate more
19:32you gave me
19:33those like
19:33unbelievable numbers
19:34the percentage
19:35began with a 9
19:37whether
19:37this will just
19:38be a community
19:39for robots
19:40who shares the library
19:41the agents will
19:43you know
19:43will the aspect
19:45of GitHub
19:46which is so beloved
19:47the community thing
19:48be turned over
19:49to the agents
19:50GitHub today
19:51is the largest
19:52community
19:53of humans
19:55around the world
19:56working together
19:57think about this
19:58what else can you find
20:00where people
20:00here in France
20:02people in China
20:03people in Singapore
20:04people in Australia
20:05are all
20:06working together
20:07collaborating
20:08on a thing
20:09that they are
20:10passionate about
20:11which is open source
20:12which is software
20:12and if you go
20:14into open source
20:15projects
20:15they are also
20:16drowning in work
20:17they are like
20:17we have way too
20:18many issues
20:18we are fighting
20:19all these bots
20:20we have security
20:21vulnerabilities
20:21you have to fix
20:22and we have a long
20:23long backlog
20:24because all of a sudden
20:24your project is successful
20:26you know what happens
20:26when your project
20:27is successful
20:28people come and
20:29say Stephen
20:30I need this feature
20:31why are you not
20:31doing this thing
20:32why is this bug
20:33still not fixed
20:35even though
20:35five years ago
20:36I reported it
20:37and so open source
20:38maintainers will also
20:39be using AI
20:40to get through
20:41your backlog
20:41faster
20:42and then you know
20:42hopefully they
20:43come here to
20:44Viva Tech
20:44and meet other
20:45developers
20:45and enjoy the
20:46lovely weather
20:47in Paris more
20:47well that's all
20:48we have time for
20:49I want to thank
20:49a very optimistic
20:50Thomas Domfke
20:51thank you so much
20:52everybody
20:53thank you
20:54I want to thank you
Commentaires