- il y a 2 jours
How Will GenAI Transform the Enterprise
Catégorie
🤖
TechnologieTranscription
00:00Sous-titrage Société Radio-Canada
00:37Hey Philippe, you're not the only French AI avatar. I also have mine. Unfortunately, I couldn't make it to Paris
00:43today, so you'll have to do with my avatar. Have fun on the panel.
00:50Wow. Hi, everybody. Thank you very much for being here for this panel about, we're going to say, the hottest
00:59topping for the past month.
01:01And you just saw a video. And I think, maybe, can you tell us, Victor, what it is? What was
01:09it?
01:09Sure, yeah. So the video you just saw here was generated entirely using AI. Of course, there's an avatar, Philippe,
01:17of myself and one of our other investors.
01:19The way you make these videos is simply by typing out the script in a web app, wait a few
01:23minutes, and you get videos like this, which looks almost real.
01:28So that's sort of the simple way of explaining it, but obviously there's a lot of very deep AI that
01:33underpins that kind of workflow.
01:35Yeah. So now I can introduce you. You're Victor, Ripa Belé, the co-founder and CEO of Synthesia, which is,
01:44of course, an AI video generation platform.
01:47And next to you have Florent Dueteau. You're the CEO and co-founder of Dataiku, which is a data science
01:54and machine learning platform.
01:56And Philippe Bothery, you are a partner at Accel.
02:01And we have a big news. You did a great announcement yesterday. Can you tell us a little bit more
02:07about it?
02:08Yeah, so yesterday we announced our Series C funding round led by Accel with participation from NVIDIA, valuing Synthesia at
02:16$1 billion, $90 million primary raise,
02:19which is a huge milestone for us and I think the AI industry in general.
02:23And we're super excited for this partnership to help us accelerate the fundamental AI research we do and deliver an
02:30even better product for our customers.
02:32Yeah. And Philippe, you're the one that invested in Synthesia. Can you tell us why this one?
02:39Yeah, we're super happy and super pumped to be partnering with Victor Steven, his co-founder, and having led that
02:49investment in Synthesia.
02:50I'm going to say that. It hasn't been easy to get that meeting. I had to use a lot of
02:55Synthesia video to convince Victor to take that meeting.
02:58But I can tell you these videos are very convincing because it worked.
03:02But so why did we invest? I think the first reason why we invest is just like the first time
03:07we sat in the room with Victor and Steven,
03:10we were like, wow, you know, they are very clear thinking about what this can become.
03:15And when we invest in a company like this, it is what you have seen today is great. It's actually
03:19more than great. It's amazing.
03:21But you think about what it can be in four years for now.
03:23And Victor and Steven were really carrying that vision with also, you know, a very, very strong focus towards execution.
03:31So that was the first reason.
03:33The second reason is the technology that they have built we think is pretty unique.
03:38I think a lot of people are saying, well, you know, AI is about an algorithm.
03:42You know, in three years from now, the algorithm they're using is going to be open source.
03:46Everybody's going to be able to do it, et cetera.
03:49But that's not true.
03:50I don't think that's true because it is, every algorithm needs a set of data to be trained on.
03:56And that set of data that they're using is actually, you know, very hard to get to and very proprietary.
04:02And within that research team that they have put together is actually amazing and will, you know, keep that differentiation.
04:08And if we look at the progress that the technology has made just in one year, I mean, you can't
04:14imagine what it's going to be, you know, in two, three, or five years from now, which is really amazing.
04:20And the last point, which has also hinted to, is that they not only have great technology, but they have
04:25combined it with kind of a great workflow layer, which produces a very strong ROI for enterprise.
04:32I mean, like, 35% of the 1,400 companies are using the technology and kind of reduces the cost
04:38of video production by 95%, which is amazing.
04:44Wow. So you have a lot of customers now.
04:47But how are, you know, corporates, big corporates companies feeling today about generative AI?
04:54Are they maybe worried? Are they doubtful? Are they excited?
04:59And maybe can you tell us, Filio?
05:01I think they are all of that. A bit frightened, but like with also lots of opportunities riding, because it's
05:09changing a lot of things.
05:09But first, I would like to confirm for the audience that actually, Philippe and Victor are human, especially for those
05:15watching us of our videos.
05:17They are actually 3D, so unless it's actually very, very well done, which is, I think, not within the realm
05:21of like Synthesia possibilities right now.
05:23Only 2D right now? Synthesia?
05:25Yeah.
05:26They are actual humans, which is good.
05:28Which is good. And so indeed, enterprise are, well, all enterprise I've been talking to are currently exploring generative AI,
05:37and most large organizations would build roadmap with potential use cases for it, but also understanding the risks and what's
05:45doable, what's not,
05:46and understanding what the business impact, and all over the spectrum, for sure.
05:50And what about your customer potential?
05:53Yeah, look, I think, I totally agree with Florian. I think everyone is feeling all of those things at once.
05:58AI is not completely new, right? Like, people have been talking about AI for many, many years, and the big
06:03shift the last six months have been generative AI.
06:06AI. And I think what is different about generative AI compared to predictive AI, which sort of is like what
06:12came before it,
06:14a big part of it is the accessibility.
06:16AI. Previously, you'd have to collect the dataset, train the models, take a very long time before you even know
06:22if whatever you're trying to do actually worked,
06:24and how well it worked. Today, with tools like ChatGPT, Synthesia, StableDiffusion, you can go in, and in literally five
06:32or ten seconds, you can discover for yourself how powerful these things are, right?
06:36And you can also start to build proof of concepts, you can play around with them in a very different
06:41way than what was possible before.
06:43And so, when I talk to our customers, I think a lot of them, of course, are also sort of
06:47fearful of, like, is this going to replace my whole business?
06:50But most of them, I think, really see opportunity. And that's all the way from, you know, can we optimize
06:56the way that we, you know, build our databases to, can we optimize the way we communicate by using Synthesia,
07:02for example?
07:02So, I actually feel that because it's so accessible, it's easy for people to grasp and understand. And I think
07:10that's also one of the reasons why we've seen so much interest in AI the last six months.
07:14And Philippe, do you think every company can jump into this new revolution? Or is it just for big corporates
07:22or just tech companies that have, you know, the resources?
07:26I think this is something that is going to have a radical impact on society as a whole, on any
07:33individual, any worker in any company in the world, just at different level.
07:40I think if you think about generative AI and AI in general, I think, to me, this is just a
07:45way to enhance human achievements.
07:48And if we take the example of the enterprise, suddenly people are able to automate things that they could not
07:54automate in the past.
07:55I mean, if you look at the first layer of automation in the enterprise, I mean, it really started with
08:01companies like UiPath, which developed kind of an RPA solution, basically interconnecting system.
08:06And now you look at what UiPath has become, it's a full AI automation platform, and the AI basically gives
08:13the ability to increase the number of processes that you can automate.
08:18And as we push the boundary of AI, we're going to push the level of processes that we can automate
08:24to achieve even more productivity.
08:28And what are the main use cases today that companies can, you know, try or implement in this generative AI
08:37space, Florian?
08:40I think there are like a variety of use cases, but we have to be cognizant of the fact that
08:44there may be a misconception on the fact that you just have to, all of those use cases will be
08:49sort of like a chatbot and only GPT-based.
08:52Meaning, real-life use cases will be a combination of different agents, multiple models, and different techniques, including traditional ML,
08:59and you'll have to stitch all of that together in order to build things at work.
09:04And the real-life use cases will be that.
09:06And in terms of initial use cases that most companies explore, I think they focus on topics where it's okay
09:13to hallucinate a little bit, which effectively is more towards marketing than, let's say, legal, for instance.
09:21And I also think that there is also currently, within large organizations, some fear or, let's say, questions pending related
09:29to regulations.
09:30So they focus mostly on use cases that are more defensible in that regard and potentially more focused on public
09:36data compared to very private data.
09:40Victor?
09:40Yeah, I would say, I mean, I think it has so many implications across every area of the business, but
09:45the one that we sort of mainly deal with is around communications, right?
09:49So if you take the online economy today, it's very, very obvious that people want to watch and they want
09:54to listen to that content.
09:55People don't want to read that much anymore.
09:57And I think all businesses know this, and all businesses have for many years been struggling with producing video content,
10:03producing audio content, because it's just not scalable.
10:06It's so hard to deal with a camera.
10:08Once you've shot something, you can't go back and edit it.
10:10What we're seeing really is a general up-leveling of all the text in a company that can now become
10:16video.
10:16And it's a very, very fundamental thing as to why we are successful and I think why people are so
10:22excited about this technology.
10:23It isn't really so much about replacing traditional video production.
10:27It's actually about taking all the text you have inside the company and making it into video.
10:32So let's say that you're one of the world's biggest fast food companies, right?
10:35And you have to onboard millions of people in your restaurants every single year.
10:39That used to be done with a 40-page handbook that you'd have to sit down and read.
10:43That's a pretty terrible experience for the employee who don't remember anything, and it's also terrible for the company who
10:48gets an employee that isn't trained the best that they can, right?
10:51Now they can make video instead in 120 different languages.
10:54The information retention is much higher.
10:56But the unlock here is that this is something that everyone can do.
11:01With Synthesia, what that means is video.
11:03But with ChatDBT means you can become a much better writer.
11:06With Stable Diffusion, you can generate a much better image.
11:08And what we're seeing in Synthesia really is that our market is less video production and more PowerPoint.
11:15And how many people in an organization makes PowerPoints?
11:18Almost everyone at some point, right?
11:20So we've built the platform.
11:22It's collaborative.
11:23It's easy to work with your colleagues.
11:24And what we're seeing is that it's everyone but the video production department that uses the technology to essentially enhance
11:31what would otherwise have been a text article or PowerPoint slides.
11:35In sales enablement, training, customer support, customer success, and a wealth of other use cases.
11:40And I think that's actually the exciting thing here.
11:42It's really about less text, more video, more than it's about replacing sort of traditional video production.
11:48And I'd say that's very much what we're seeing.
11:51I think that's the competitive advantage that our customers have today when they use us.
11:56And just to echo on what you're saying, because we just need the due diligence to have a couple of
12:00interesting practical use cases.
12:03I mean, one company was actually using the avatar of their CEOs to send a personalized message to every customer.
12:12So suddenly you're a customer.
12:13You buy the product.
12:14And you have the message of the CEO saying, hey, Mr. Jones, thank you very much for buying our product,
12:19et cetera.
12:19And even though it is a synthetic video, it really has an impact on the customer.
12:25Another use case which I thought was pretty interesting was a large automotive manufacturer.
12:30And they are sending every month a lot of information to their dealership globally.
12:37And unfortunately, the dealerships are not very good at reading the tens of pages of PDF to have all the
12:44information.
12:44So that translates into a lot of calls from the dealership to their call center.
12:48And now they've basically shifted.
12:50Instead of sending PDF, they send videos with the content of this PDF.
12:54And suddenly people are watching it.
12:56And the number of calls going into their call center has been going down drastically.
13:01So to me, it truly shows the kind of application where you're basically not trying to replace something,
13:07but just you're enhancing the things that you can do.
13:10And to me, that's the real power of AI.
13:12And I would add to that that I think when you look at these videos, there's so much more to
13:17go on how we push this technology to build much more rich, complex, expressive, emotive avatars.
13:23There's a lot to build still, right?
13:25But even in its current phase, it's very important when you look at these videos that most people, the alternative
13:30generally isn't video.
13:32The alternative is text.
13:33And I think in the case of the dealerships, for example, right, that is just such a clear value proposition.
13:38Sit down and watch this two-minute video instead of read five pages of PDF documents, which goes in and
13:42out, right, for most people at least.
13:45So we can see all the benefits.
13:47So it saves time, of course.
13:48But does it save money, you know, for the companies that you don't have to hire more developers, Florian?
13:58Yeah, I think that there is, well, for the second time in my life, I was able to witness a
14:05technology with whom I think, like, everyone would perceive some value.
14:10Like, first time was, like, a search back 99 or 20, something like that.
14:16And now it's already the case, like, I think everyone in this room can imagine things they could do with
14:21generative AI from a practical standpoint.
14:23So even if you have lots of questions unanswered in terms of, like, the implication and the cost or, like,
14:28the changes for society,
14:29it's like something where we can perceive a day-to-day user.
14:32That's unique.
14:33Like, empirically, it's happened every 10 years.
14:35So we are lucky.
14:36Yeah!
14:37But then, from a practical standpoint, I think that in the enterprise, it will take a few years to percolate
14:43in order to understand who you turn that into
14:46very practical use cases, changes in the business, understand how you can change an email from a PDF to a
14:53video and actually win some time or money, then, and so forth.
14:56And potentially, you have to unlock hundreds of use cases in the enterprise in order to actually unveil the real
15:03value of AI.
15:04And that's an economic change.
15:05It takes 10 years.
15:06And I think that's what we are all working for in this room, or almost.
15:12Philippe?
15:14Yeah, I mean, I would say that at the end of the day, it is all about the cost saving
15:19for a company.
15:20I mean, the current environment, everybody is trying to see, well, how do you become more efficient?
15:26How can you, you know, increase your bottom line?
15:31And generally, AI, I think, and AI in general, that's, I think, the best way to do that.
15:35I mean, if you look at the example that I've gave, the first example when you're sending video to a
15:40customer,
15:40I mean, I think this is going to play in the long term.
15:43So how much do you save right now, I think it's probably harder to assess.
15:47But in the second case, you look at their call center, the number of calls in your call center going
15:51down.
15:51I mean, that is pretty obvious.
15:54And if you look at the way AI is applying the enterprise today around being able to automate processes
16:01with, you know, deeper document understanding capabilities.
16:04So today, you can take any unstructured document, and you can analyze, extract the right information,
16:10send it to the right system, and put it in the right workflow to automate.
16:14Like, obviously, like, all these technologies are really improving the efficiency and translates into cost savings.
16:21And Victor?
16:23Yeah, I would echo very much, I think, what's already been said.
16:26Ultimately, it's about moving business metrics.
16:29One of the core philosophies at Synthesia is what we call utility over novelty.
16:34There's a lot of cool demos right now.
16:35There's a lot of things making the rounds on Twitter.
16:38And I think a lot of those things will become very, you know, real technologies that deliver us value.
16:44But I think if you are trying to build a business, and a long-lasting, iconic, durable business, like we're
16:49trying to do,
16:50you have to focus on utility.
16:51The metrics that we see moving are the one, the obvious one, is sort of on cost savings, right?
16:56I think that's been covered.
16:58The other one is on productivity gains.
17:00So let's say you have 4,000, a sales force of 4,000 people, one of our customers.
17:05How much would you pay to make that team, you know, 95% up-to-date with everything that's happening
17:14in your business,
17:15versus 80% if you send text?
17:17That delta is quite important if you have 4,000 people taking calls with prospects every single day.
17:22You want to make sure that sales team is educated well, right?
17:24And that's something that's very measurable.
17:26It's also very measurable how many calls to the call center, how many calls to customer support.
17:30How quickly can your customers onboard and adapt your product in a customer success setting?
17:35And I think that is very much what our customers care about.
17:39There's, of course, it is also cool, but ultimately that's what they care about.
17:44That's why they expand their licenses.
17:46That's why we see more and more teams in these big enterprise companies using the technology.
17:51And I think it's kind of obvious, but I think sometimes when you are in these hype cycles,
17:57people can forget that, you know, over a longer timeline, it is about utility.
18:02It's not just about novelty.
18:03Okay, so we are in Europe, so we have to talk about regulation, of course.
18:10Do, you know, if a company goes to, is trying to, yeah, to implement, to test generative AI,
18:18can it be compliant today?
18:22What are the risks?
18:25I think, first of all, I think these are very powerful technologies.
18:28We should have the right regulatory framework.
18:30We should make sure that we put up the right guardrails.
18:33That's something we, just as a company, care deeply about and spend a lot of resource on.
18:37In terms of regulation, I think there's a lot of different discussions going around,
18:42around many different aspects of it.
18:44One of the ones that is very prominent right now is about the use of training data.
18:48So if you take a lot of very large foundational models,
18:51they kind of have to be trained on internet-scale data,
18:54which means that almost per definition,
18:55you're not going to have the license to train on that content.
18:58And there's a lot of questions around, is that okay or not, both morally and legally?
19:03I don't think I have the answer to that.
19:04For a company like us, we only train on data that we actually have the license for.
19:08I do think that if we fast forward a few years,
19:12this will be something enterprises care about.
19:14We're already seeing companies like Adobe, who released Firefly,
19:17which is their generative AI.
19:19It can generate images, for example.
19:21That has been trained only on images that they have licensed.
19:25And part of the motivation for that is because they know
19:28that their big corporate customers will not feel comfortable
19:31using generated content that potentially infringes on copyright.
19:35So I think that's maybe more on the front end of what will businesses care about.
19:39But obviously a lot of this will ultimately stem from what happens with regulation.
19:42In the UK, where we have our HQ, it's a very pro-innovation approach,
19:46looking much more at the outcomes of the technology.
19:49I do think that's the right approach.
19:51Whereas in Europe, there's a lot of focus around trying to regulate
19:55the specific technologies, the specific algorithms,
19:57with very specific governance requirements.
20:01I think the idea is good.
20:02I do worry a little bit.
20:03That will become overly restrictive.
20:04And that, unfortunately, a lot of AI companies will decide
20:08to not kind of headquarter in Europe.
20:11So, I mean, it's exciting to see how it pans out.
20:13But I think regulation is important.
20:15And it's important we get it fast.
20:17Because clarity is probably the most important thing,
20:21I think, for an AI company today.
20:22Some companies abide by one set of rules.
20:24Other companies abide by a different set of rules.
20:26We need to quickly kind of make the right playbook
20:30for how you build these companies.
20:32And Florian, because DataQ was born in Paris,
20:36but now the HQ is in the U.S.?
20:39Yeah, well, yeah, we've got the luck of seeing both sides of the topic,
20:43the U.S. side and the European one.
20:45Well, from my vantage point,
20:47I think we are at the beginning of like a three, five-year timeline
20:49in terms of changing a significant aspect of data,
20:53intellectual property, and AI regulations,
20:56where in data you have to answer to the question of like
20:59whether a large and large model is a database, including private data.
21:02and so is it GDPR compliant?
21:04From a copyright standpoint, we have to redefine fair use.
21:07And from an AI standpoint,
21:09I think EU regulation will actually drive the discussion,
21:13but like with an implementation that will actually be in 2025.
21:17And so from the perspective of organization,
21:19it's like a three, five-year window of effectively uncertainty,
21:22which will require actually some work in advance
21:25in order to understand how you navigate
21:27the moving regulatory environment.
21:30That will be fun.
21:32Thank you.
21:34Philippe, you want to add something?
21:36Yeah, I think, you know, I echo what has been said.
21:39I think having a clear set of rules,
21:41I think would really help the ecosystem know
21:44in what direction it needs to be moving.
21:47I think I agree that right now,
21:49I think a lot of the questions are
21:50what are the data used to train model?
21:53I mean, if you take the example of coding, for example,
21:57like Copilot can really drive a lot of productivity
22:00for any, you know, engineering team,
22:03like up to 50%, and everybody's using it.
22:06But if you talk to a lawyer today,
22:07they're going to say, wow, you're investing in this company.
22:09They use Copilot.
22:10So we cannot guarantee the integrity of the code
22:14that they are producing.
22:15And at the end of the day,
22:16it boils down to what data has been used
22:19to train the model that is, you know,
22:22that helping the engineers develop its code.
22:25And if it is only trained on open source license,
22:28which is very permissive,
22:30then everything is fine.
22:31If it trains on more restrictive open source license,
22:35then suddenly it is not fine.
22:36And so there needs to be a clear boundary
22:39and a regulation that basically set the rule
22:41and say, here's the type of data you can use,
22:43here's the type of data that you cannot use,
22:45so that as an industry, we can use the model
22:49that are basically compliant
22:51and move forward with that.
22:54So I guess we'll talk about that next year.
22:58Thank you very much for being with us
23:00and for sharing your experience.
23:02Thank you.
23:02Thank you.
23:03Thank you.
Commentaires