- il y a 9 heures
What’s Next For GenAI and Where Europe Can Lead
Catégorie
🤖
TechnologieTranscription
00:00...to be with us this afternoon.
00:01I'm really excited about this panel.
00:04It has been one of the newsiest months in AI, in tech, really.
00:10It's been an incredible couple of weeks.
00:13Not least because of people like the panel I've got this afternoon.
00:16It gives me great pleasure to welcome Charles and Philippe to the stage.
00:21And let's just dive straight into the news.
00:23So, Charles, you raised just a tiddly amount the other week, a $220 million seed round.
00:32And the laundry list of investors include, of course, Excel, UiPath, BPI France, Eric Schmidt, Amazon, Xavier Neal, FirstMark.
00:42I could go on, I could go on.
00:45Tell me, what is H?
00:49That's an amazing question.
00:52So, thank you.
00:53H is, we would call it as a new AI venture.
00:59Holistic, humane at the same time.
01:02Heuristic, heterogeneous.
01:04H can be basically interpreted the way you want.
01:08The way we want to build this company is with a differentiated strategy
01:13from all the well-known AI ventures that you find in the world.
01:17Think of the open AI's and tropics of this world.
01:19But with a focus straight on a next generation of models that are not just like a monolithic approach, basically,
01:29to LLMs.
01:30We try to go beyond just the language aspect and really like think of H as a holistic manner, basically,
01:38to approach the market.
01:41And H, you can see it first as a kind of AI company, just with a strong research department and
01:49a research focus.
01:50But H is also like a new AI-like direction for enterprises, for the workers.
01:57And our mission, in a sense, is to kind of drive the productivity forward of billions and starting with the
02:05workers and enterprises.
02:06So that's kind of the...
02:09And the differentiating point, if I understand it, is that your models are more kind of actionable?
02:14Is that how you're thinking about them and less kind of text-based?
02:18Yes.
02:20Basically, like many points, the long-term vision is really around, like, building AGI.
02:27That's the long-term vision.
02:29Now, in the mean term, you need to be realistic.
02:31Yes.
02:32And you need to have, like, strong-term kind of approaches.
02:37On the short term, basically, what we want to achieve is building a next generation of models that are action
02:42-driven.
02:43For that, you need to have, like, a different stack of data.
02:46You need to build, like, new flywheels of data.
02:49You need to have, like, a very aggressive and strong, like, a go-to-market strategy to get that data.
02:53Because internet scale will not be enough enterprise scale as well.
02:58So you need to have, like, a whole bunch and mix of data that are action-driven.
03:02So that we can build, like, these large action models.
03:06Or, if I may say, like, a mix of LLMs and action models.
03:09And that's the kind of strategy that we have here.
03:11And just in case for people in the audience who aren't aware, could you talk a little bit about your
03:15background and also your co-founders?
03:17Because they came from a fairly famous AI company.
03:21Sure.
03:22So H is also, like, holistic and kind of heterogeneous in terms of approach to the team.
03:30Myself, I was doing some AI at Stanford prior to funding H.
03:36My co-founders are generally, like, all from a deep mind.
03:40Laurent, our CTO, was leading some of the LLM efforts.
03:47Carl, on the chief operations and research side, was leading the multi-agent.
03:51At Paris, we have also, like, Julien, who was working on game theory and multi-agency.
03:58And Dan is joining us very soon.
04:00And he was leading a lot of things related to representation learning.
04:03So we have, like, this approach, which is not only, like, on scaling and large models, as well as a
04:08team.
04:09We want to bring, like, different components to the market that are around, like, the next generation of models with
04:16multi-agency.
04:17With new representations, we go beyond just the single transformer.
04:21We try to combine, like, different approach.
04:23And I think we saw, like, on the market a lot of trials and efforts around, like, hybrid architectures with
04:30RNNs.
04:30So it's a kind of return of RNNs, as well.
04:34And that's the team that we've built.
04:36And I think it's also, like, a great illustration of the people that we will hire and that we've already,
04:41like, hired, in a sense.
04:43Not just, like, scaling-oriented, but also, like, across, like, a broad spectrum of AI competencies.
04:49And what can they do inside H that they couldn't do inside DeepMind, inside Google, inside Alphabet?
04:55Like, why is it that they have decided to strike out on their own and with you?
05:00One big point is probably, like, the first point is that a lot of the GAFAMs, so when you think
05:08about Google, Amazon, Microsoft, and so on,
05:11are converging towards that kind of monolithic approach where you scale the models.
05:16And it takes a lot of efforts.
05:17It takes a lot of people.
05:20It takes, basically, like, massive amount of capital, as well.
05:23And if you want to be focused, laser-focused on that kind of approach, you need to have all your
05:30people, basically, centered around that direction.
05:34And the thing is, like, it constrains the capabilities that you can have in small groups, basically the sandboxing capabilities.
05:42Right.
05:42And so at H, right now, what we're trying to achieve is to be really fast at execution, really, like,
05:49oriented by a market.
05:52And so the two points, to answer your question, are fast execution, flexibility, basically the ability to iterate super fast
06:00by sandboxing some ideas, jetting some ideas or taking some ideas and scaling.
06:06And, of course, like, the opportunity to really have an impact on the market and partnerships, which is less the
06:13focus, sometimes, of these big companies, more driven by the tech and the scaling, but less by the revenues that
06:20you can bring and the impact that you can have on the productivity, for example.
06:23Great, that makes sense.
06:25And, Philly, I mean, H is an example, yet again, of Excel betting on the European tech market.
06:33And you recently just raised your eighth fund, $650 million, to invest in European and Israeli startups.
06:40What's your investment thesis?
06:43What has Europe got to offer that other markets don't?
06:46Thank you.
06:47Yeah, so Excel was founded in Silicon Valley 40 years ago, so we celebrated our 40th anniversary last year, and
06:55we opened our European office 20 years ago.
06:59So we were one of the first venture firms from the U.S. to really believe in the potential of
07:06Europe and to understand that Silicon Valley did not have the monopoly of innovation.
07:12So we think that great companies can come from anywhere in the world, and that's why we took a really
07:18global approach.
07:19We also have an office in India to cover India and Southeast Asia.
07:22So getting to Europe, if you look at the evolution of the ecosystem in the past 20 years, 25 years
07:28now, it has been incredible.
07:31I mean, I remember, like, in 2008, 2010, I mean, journalists were asking the question, well, can Europe create a
07:40billion-dollar company?
07:41I think that was the question.
07:42I think that question has been answered.
07:45I mean, we saw in 2021, UiPath was a larger software IPO globally.
07:50It was a company out of Romania where we had the chance to lead the Series 8 with a $35
07:55billion IPO.
07:57So I think Europe has really proven its capacity to create global winners.
08:03They have done that in different sectors, and I think now they really have a real shot at proving that
08:09they can do it in AI as well.
08:10That's why we're super happy to be a core investor in AI.
08:15And when we think about AI and the opportunity in AI, I mean, we think that the potential here is
08:21tremendous.
08:22I think it's something that's potentially bigger than the shift to cloud or the shift to mobile.
08:28And, you know, it's very likely that new AI models will be part of every piece of software that's going
08:35to be written in the future.
08:37And so when we think about this, yes, I mean, I think a large part of our new fund that
08:42we announced, you know, a couple weeks ago
08:43is going to go into companies that are either inventing AI, like H is doing, or leveraging heavily AI to
08:52basically bring the next level of automation and performance.
08:56And how did the relationship between Accel and H come about?
09:00How did you first kind of get to know the company?
09:02Yeah, so the thing which is interesting about AI is that if you think about, you know, before Genitive AI
09:11really came along,
09:14we're, I mean, investors were building relationships with investors and with engineers,
09:19but not really researchers, because researchers, they were really way ahead.
09:23And I think what Genitive AI has shown is that now basically researchers are part of the company.
09:28So the R in R&D really means something for a Genitive AI company.
09:35And so we started, we understood that, you know, a couple of years ago,
09:39and we started building relationships with some of the key researchers, key talents.
09:45I mean, I'm focusing on the European side, of course.
09:47That's how we met Laurent Sive, who then introduced us to the team.
09:51And now that's history.
09:53And where are some of the interesting kind of hotbeds of that talent across Europe?
09:58Where are some of the interesting universities and research houses?
10:02Yeah, so, I mean, we're very fortunate that there are a lot of, you know, great talent in Europe.
10:07And we think that the, I mean, for the two main hubs, we'd say are, you know, Paris and London.
10:14And Paris, because we have great, great school here.
10:18So, you know, obviously I'm waving the French flag.
10:20But you have Ecole Polytechnique, Ecole Normale Supérieure.
10:24We have the MVA master, which is, you know, great master.
10:28I'm sure Charles can tell you more about that.
10:32And, you know, large companies really understood that potential.
10:36That's why Yanlo Kuhn created FAIR in 2015 in Paris.
10:40That's why in 2018 Google created first AI research center in Paris.
10:44Now they're doubling down with, like, a new center with 300 people.
10:49And in London, it was very driven by DeepMind and great university in Cambridge.
10:54But it's not only, I mean, you have the, you know, outside of these two hubs, you have great hubs
10:58in Germany.
10:59In Munich, for example, you have the LMU who actually invented the stable diffusion algorithm.
11:06You know, you have great researchers in Israel as well.
11:10So I think the region is very rich.
11:12But, Charles, maybe you want to comment on that as well.
11:14Yeah, and it's, I'd say, like, thanks.
11:17It's definitely, like, a, I'd say, like, an ecosystem which is thriving right now.
11:23I think in France we had huge expertise.
11:27And we're lucky to have, like, that kind of huge expertise in mathematics.
11:31And when it comes to applying mathematics to artificial intelligence, it works generally well.
11:37And you have, and you see, like, in diverse companies, like, that kind of split between research engineers, research scientists,
11:44software engineers.
11:45And we have a big pool of talent around science stemming from France, but also, like, in the UK.
11:52And I think it's from the rigor of the, basically, like, the French system.
11:57So there's definitely, like, a pool of talent that we need to leverage.
12:01And if you bring, also, like, in addition to science, that kind of entrepreneurship vibe and really, like, the goal
12:10to achieve missions that are market-driven or impact-driven,
12:16based on the non-profit side but also on the for-profit side.
12:19I mean, on the non-profit, you have also, like, an amazing, like, Qtai, for example, being set up in
12:23Paris.
12:25On the for-profit, you have, like, Mistral, you have now H and other companies.
12:29And so, basically, we are happy to, basically, like, have the chance to show the strengths of this ecosystem
12:38and to potentially, like, build the next wave of startups, the next generation.
12:43When it comes to building large language models, though, you know, raising $220 million is not anything to be sniffed
12:52at.
12:52Neither is a $650 million fund.
12:55But when you have reports that Sam Maltman is talking about raising, you know, trillions of dollars, potentially, for open
13:01AI,
13:02how can European startups compete when there is just so much money flowing to the US?
13:09So, basically, you have, like, different funds for different stages.
13:15Right now, we have the chance to work with Accel, but also, like, other funds.
13:20And generally, like, the funds that were involved here were funds regarding, like, seed stages or growth or high net
13:27worth investors.
13:29But, basically, you have, like, many stages possible.
13:32And we think that we will find the money both in Europe.
13:36You'll have your trillion dollar day yet, basically.
13:39We'll find it.
13:39Talking to you too early.
13:40We'll find a way.
13:42And it's just a question, also, of, basically, like, making people aware of the opportunity that we have in Europe
13:49with that pool of talent.
13:51And it needs to be with, first, potentially, like, the investors that we have, like, American investors and a mix
13:58of European investors as well,
14:00just to show the way and try to think about, like, delivering these opportunities, showing that we're going to have
14:06an impact,
14:06showing that we can have, like, real traction or so with revenues from our strategics, and then opening the doors.
14:12It's more a question of awareness, I would say.
14:15But we have the potential here.
14:17But I would say that, I mean, capital is global.
14:21So, and capital is going to go towards the interesting opportunities.
14:25And I don't think we're talking about trillions of dollars.
14:27I mean, I think that's kind of looking very, very far into the future.
14:31I think right now what you're talking about is hundreds of millions of dollars.
14:35And when you do a first seed at $220 million, it really shows that, you know, you can have great
14:40talent, great team,
14:41who can attract that kind of money coming out of Europe.
14:44So, I'm not worried that the next stage is going to come.
14:47And it's going to probably go from $200 to, you know, $300, $400, $500.
14:51And, you know, there is a natural progression.
14:54I think what's going to be interesting is that at the same time, I think the compute power, the price
14:59of compute is going to decrease as well.
15:02So, there is going to be more powerful chips of potentially faster time to train, so lower costs.
15:08So, there are many aspects where it's hard to predict, you know, in three years from now, what is going
15:12to be the cost of compute that's really going to be required.
15:15So, what we're looking at now is, well, is a couple hundred million dollars enough to achieve what we want
15:21to achieve in the first milestone?
15:22I think the answer is yes.
15:24So, I think for now, the company is well-funded.
15:28I'm aware we've only got a few minutes left, but I did want to get your assessment on how we
15:33should think about who the winners and losers are going to be in AI.
15:38You know, who's going to have the best footing?
15:40Is it the ones with the most money, the most processing power, the most talent?
15:44Is it those that go vertically integrated?
15:46Is it those who, you know, I'm kind of keen to get your opinion on how we should assess that
15:52at this early stage.
15:56My take, and again, it's just my take, is that you need to combine, like, different pillars.
16:01The first pillar is, of course, like capital.
16:03So, you need to have some money.
16:05The second pillar is basically around a compute strategy.
16:09You need to either, like, rely on a very strong partnership with various pillars, or you need also, like, to
16:15think about your own cluster strategy.
16:17That's the second thing.
16:19The third element is you need, like, a talent strategy.
16:24In the end, it's always about, like, the talent that you manage to attract as well.
16:28The fourth pillar that we'll add is also, like, around, like, the strategic partnerships that you can have around data.
16:34And so, one thing that was missing in some companies today is probably, like, this attractiveness when it comes to
16:42very differentiated data.
16:44And so, one of the goals that we have here is to make sure that from the get-go, and
16:49I think we've shown it with a strategic partnership with UiPath, for example, to start with, but we will also,
16:55like, create, like, more relationship here.
16:57We want to have some data flywheels that are pretty unique.
17:02And that's the key, basically, to not only, like, build strong technologies, strong models, but also, like, we'd say, like,
17:09differentiation.
17:10Do they have to be exclusive for that to happen?
17:13You don't necessarily, like, need exclusivity.
17:16When it comes to accessing some private data, you can then use some techniques of anonymization or even, like, synthetic
17:22generation and whatsoever, where you can build your own IP internally and then just generalize from there.
17:27So, the goal is more for getting access first, making sure that you build that network, but never let aside,
17:34like, one of these three pillars in addition to the data.
17:37So, I would say, like, always relying on the four pillars, positions yourself as a potential winner.
17:44And would you agree, Philippe?
17:46Yeah, no, I think I agree with what Charles said.
17:49I think, to put it in perspective, I'm not sure there's going to be one winner in AI.
17:54I think there's going to be multiple winners.
17:56I mean, if you look at the field of AI, I think it's a very diverse field.
18:00I mean, it's basically making software more intelligent.
18:04Like, software is part of everything in the world, right?
18:08From, you know, from your car to your phone to your, you know, your computer, you know, even your refrigerator.
18:15Like, there is software everywhere.
18:17So, I think there's going to be multiple winners in different categories.
18:20I mean, if I look at, you know, last week I was in San Francisco.
18:24I took a Waymo first drive in a fully driverless car.
18:28I mean, that's one type of AI that you can think about.
18:32This morning I was at a meeting with a company called Wondercraft in Paris who is building exoskeleton to help
18:40people walk.
18:41I mean, now leveraging AI instead of more traditional algorithm is just going to be like a step function in
18:47terms of the ability of these exoskeleton to kind of move around.
18:52So, I think the AI revolution is, in the next few years, like super exciting and I don't think there's
18:58going to be only one winner.
18:59I think there's going to be a different field, different winners.
19:02Can't wait to see it.
19:04Yeah, well, I'm afraid it's time to wrap up.
19:07I'm really excited to see what you're going to build and come back next year with some really interesting products
19:13and examples.
19:14We will.
19:14Give a great round of applause to Philippe and Charles.
19:18Thank you.
19:20Thank you.
19:21Thank you.
19:21Thank you now.
Commentaires