- il y a 1 jour
Founder’s Story Scaling Against the Odds
Catégorie
🤖
TechnologieTranscription
00:00Sous-titrage Société Radio-Canada
00:30Sous-titrage Société Radio-Canada
01:12Sous-titrage Société Radio-Canada
01:14Sous-titrage Société Radio-Canada
01:28Sous-titrage Société Radio-Canada
01:32Sous-titrage Société Radio-Canada
01:39Life Sciences, Healthcare, Frontier
01:41and Climate in supporting innovative founders in building what we hope to be the next iconic
01:47companies. We have about $8 billion in management with Alphabet as our sole LP. So we don't invest
01:54strategically for Google or anyone else, but we have Alphabet as our sole limited partner.
02:00We have about $8 billion in management offices in the US and we've been in Europe since 2014 and in
02:08that time we've invested over half a billion dollars into over 40 companies in Europe and we're extremely excited to
02:14be doubling down and sort of increasing that pace as we go forward.
02:18And Milos who I've got with me today is one of our incredible founders. He's the founder and CEO of
02:25DeepSet, which is an enterprise at LM platform that he will be able to tell you about much better than
02:30I can.
02:31And Milos and I have been working very close together for the last few years and we're here to talk
02:36to you about scaling a startup against the odds in AI in very exciting, uncertain and sort of very kind
02:45of high change times as we can call it.
02:47And I think one of the really interesting things about DeepSet is that DeepSet started working on AI and large
02:55language models long before it was cool and it was the it thing.
02:59Actually, at the time, Milos, when you were doing it, it was a pretty contrarian thing to do.
03:04And I remember in our early conversations, I was really struck by your conviction and how deeply you believed in
03:10the technology and how sort of clearly you saw the future.
03:14But it really wasn't obvious to most people. And so I'd love to understand, just to start off with, you
03:19know, what was the founding hypothesis?
03:22What was the early sort of journey in starting DeepSet?
03:26Yeah, sure. Happy to share a bit more about that and also high from my side.
03:32So as you said correctly, before there was the term generative AI, there was actually it was called back then
03:38still NLP, natural language processing.
03:40And I think it's the year 2016, 2017. We simply spent a lot of time with NLP technologies, models, you
03:50know, what was around back then, all of the stuff of an earlier generation.
03:55And if I'm very honest, we were we were excited about the technology and we didn't really think or question
04:03the applications of this technology.
04:05It was quite obvious for us that search engines where we type in natural questions, virtual assistants, process automation, all
04:17of that will be empowered with natural language processing.
04:20So we actually didn't bother too much about what to use this technology for.
04:26What wasn't so clear to us at that point in time was rather how will all these businesses and products
04:36and business processes actually utilize the power of this technology?
04:41So how does that work? How does a NLP model become part of a product or a process?
04:48And this was really not obvious to us back then. So in 2018, when we found a deep set, you
04:55know, being so deep in this technological space, we felt the urge to actually answer this question.
05:02You have to think about it. 2017 was for us because we're so deep into it a little bit like
05:092023 is too many people right now, you know, so you can imagine what 2023 feels like to us.
05:17And in 2018, we started deep set and we started to simply apply natural language processing to very practical problems
05:24and practical applications in enterprises.
05:26So we were building custom NLP solutions to solve the problems of companies like, for example, Airbus or Siemens back
05:34then.
05:36So in startups, usually you say they should dog food or eat their own dog food, right?
05:43What we were doing is we were technically dog fooding before we actually even had a product that, you know,
05:49we could use.
05:50So this was a bit like how we wanted to expose ourselves to the building process.
05:53Now, in parallel to that, the technology, the technological foundation actually, you know, made a big, tremendous jump.
06:04Now, when we talk today about GPT, then we talk about the G, which means generative.
06:11We talk about the P, which means pre-trained.
06:14And then there is the T, and the T stands for transformer.
06:20Transformer is a model architecture that actually changed the whole state of the art and the capabilities, right, of natural
06:29language processing and what we can do with these models.
06:32But the first transformer was already born end of 2018.
06:37So this was when BERT, the first transformer, was released.
06:41And when that happened, we felt, hey, wow, this is now, this technology is now really accessible or at least
06:49from a performance standpoint, ready for mainstream adoption.
06:54These transformers, they will be in one way or the other part of the product.
06:58So this vision that we were having somehow felt very haptic.
07:02So what we did is, besides, you know, working on practical problems and solving them with NLP, we were really
07:09exposing ourselves to the technology.
07:11We were probably among the first teams outside of big tech to really, you know, apply these transformers in products
07:20and in production scenarios.
07:21And this is, in the end, how we came up with our product that should simply help organizations on a
07:28larger scale to adopt these transformer models or LLMs as we refer to them today.
07:37And, look, this was a journey where we actually, you know, took some time and were self-financed for a
07:43long time.
07:44And in 2021, when the market was still quite early, we already saw great adoption of our products.
07:50We saw many users.
07:52We had a growing customer base.
07:54And that was the time when we felt, hey, we're on a good way with what we're having to really
07:59build probably a market leader for these tools and platforms that enterprises will use to adopt NLP and LLMs.
08:09And that was then also the choice when we decided to actually raise venture capital for the first time and,
08:15yeah, and pursue the vision to build this major player.
08:18And that was also the time, of course, when we partnered back then and, yeah, one thing actually I think
08:28would be interesting for the audience is I remember in 2021 when we met, it was still early for the
08:35market, right?
08:36So there were only a few people who knew the term transformers or who actually thought so much about NLP
08:42back then.
08:42It was quite exciting because you were already spending so much time in it.
08:46I think it would be quite interesting to understand how you actually approach your investments and especially this particular one.
08:53How did you actually build the conviction?
08:56Because to us it felt there was a lot before you even met us.
09:00Yeah, no, it's a good point.
09:01And actually it's sort of, it was an unusual journey in for me to LLMs and NLP in that I
09:09started off as a lawyer.
09:11I went to law school.
09:12I worked as a lawyer.
09:13And because of that I got very interested in how technology can manipulate language and how that could influence business
09:19settings.
09:19And so it was around that time, a bit early, just before we spoke actually, that I saw my first
09:26demos with transformers.
09:27And was really kind of, I had my chat GPT moment back then and was blown away by the things
09:33that they could do.
09:34It felt really magical.
09:35But one of the things as investors we see a lot, we see a lot of very, very exciting technology
09:40and we're privileged to be able to see that.
09:43What gets us particularly excited and motivated is when you see that technology can be used in real practical settings.
09:51And actually that really wasn't obvious for transformers or larger models back then.
09:57They were amazing, but they were still very rough around the edges and people were really trying to think, okay,
10:03how can we actually use this?
10:04And so when we went out there and we were speaking to lots of founders and people and builders in
10:09the space, well, I say lots of, but it really wasn't that many at the time that were working on
10:13it.
10:13And one of the things that really excited me about our first conversations was that they started at the place
10:19of solving problems for users, solving problems for customers.
10:25And what was really interesting and I think would be, I think it'd be cool for the audience to learn
10:30about because it really was at the time, like it now feels very obvious to be for an investor and
10:36a founder of a large language model or an AI company to be sat on stage talking about scale and
10:42hyperscale.
10:43Because that's something that the market has developed around, but you were very, very early, you know, you were building
10:48this at a time when it really wasn't obvious.
10:50You did a lot of things that are fairly unconventional for startups to do in, for example, bootstrapping your first
10:57few years, doing lots of services.
10:59And what was very appealing to me as an investor was that you were clear in your vision of building
11:05a hyperscale startup and building a large company.
11:08But early on, you were using the services and, and, and, and your sort of deployment into your customers as
11:15a way to learn about their problems, learn about the solutions that would help them in practical real world settings.
11:22And then combining that with your deep understanding of the technology and seeing how you could build for those use
11:29cases in mind.
11:30So actually, I mean, I, I sort of, I'd love to understand, like, what were those early days like for
11:34the, for the founders in the room?
11:35What were the early days like of when you were building for those customers and, and sort of trying to
11:40solve their problems with this immersion technology?
11:46So, you know, if, if, if I look, if I look back to this time, I, I, I'd say it's
11:52a bit like of a twofold memory.
11:53Number one, it's always a bit nicer when you look back and you think like, oh, was that really that
11:59hard or, but then sometimes I remember some moments where I felt like, oh my God, this is incredibly hard.
12:05So, it's, um, in the early days, you know, no one, no one even knew what the abbreviation NLP meant.
12:12Uh, we were really, we really had to trailblaze our way somehow also into these enterprises.
12:19You know, we, we couldn't come from a technological pitch or value proposition.
12:23We had to come actually from use cases and how can we solve problems they have.
12:27Um, still, what I think, what was useful for us was to really, you know, um,
12:34take the time to, you know, to, to become, to become experts, not, not in the technology from a theoretical
12:44standpoint,
12:45but really in how to apply it, what does it mean to deploy it, what is an enterprise requirement, uh,
12:52um,
12:53and, you know, these are things you cannot leapfrog them.
12:55You cannot, you cannot come up with a product, uh, on a whiteboard validated in five customer meetings,
13:02and then, you know, scale this up, um, in a space that is so early, you need a lot of
13:08stamina
13:08and you need simply to take the time and the patience.
13:10So, when I, when I think back, this was probably really a good choice for us to take this time
13:15and to have the time and to have the space and probably even to not be venture capital financed,
13:20if I'm very honest, because of course, you know, there are expectations and it might, you know,
13:25resources also are very appealing and because, you know, they give you opportunity to probably
13:31do more things and maybe we would have done more marketing and, you know, I don't know,
13:36whatever, things that didn't matter because everything that mattered was
13:40what is the scalable product that we see or we want to build, yeah?
13:47And we simply took the time to figure, to figure that out and focus on that.
13:50So, that's definitely the positive side of it.
13:54The bad side of it is being bootstrapped is really, really, really hard.
13:59And, uh, you realize that, you know, doing taxes and everything, you, you understand that
14:05there's a lot, a lot that runs in the back of a company that you're not even aware of.
14:09And even for a small company, administrative overhead is a mess.
14:13And, you know, taking care about your finances is a lot of effort and having no resources
14:17for all of that is often very frustrating, but I wouldn't trade the experience.
14:22And I would recommend everyone who wants to enter an early space.
14:25And I would say, to be honest, LLMs are still, this is still an early market, right?
14:30We're maybe at 1%, if at all, of the total potential adoption LLMs will see, right?
14:37So, it's very, very early.
14:38So, I would say, um, try to find a setup where you can afford to take a bit of time
14:44to figure
14:45out how your product in this space should look like.
14:48Yeah, it's a really interesting point.
14:50And we, you know, by, by definition as investors, we see a lot of companies that are coming to
14:56us to, to raise money.
14:59And it's a really interesting point you mentioned in terms of this idea of scaling at the right
15:03time.
15:04And, and there is definitely such a thing as premature scale.
15:08And, um, in, in times, in prior times, we've seen, you know, between the sort of 2020, 2021
15:14funding environment, there were companies getting funded with, with large amounts of capital much
15:20earlier in their life cycle than they previously would have been funded.
15:23And, and you really took the opposite approach, even though you were raising a, a venture round
15:28in, in 2021 as your first rounds, again, it struck me in our conversations about how focused
15:35you were on staying lean, staying nimble, not prematurely scale, scaling.
15:41And I think a lot of that came from that experience of the early days of being bootstrapped.
15:46And I, I think the audience would like to hear a little bit about, you know, you were small
15:50for a long period of time.
15:51And when you were sort of working with your customers trying to solve their problems, a
15:57lean organization, not a large scale organization, not a large amount of operational overhead.
16:03What were some of the benefits of staying small for a period of time before getting to a point
16:09when you realize, okay, now is the time to press the button, for example, around decision
16:14speed, operational speed, execution, et cetera.
16:19So, I mean, the big benefit of, you know, one of the biggest benefits in the end of being
16:24a small organization is that alignment happens naturally in your day to day work.
16:28If you get up in the morning, your team, let's say your team of 10 people, you get up in
16:32the
16:33morning, you have your stand up, even though you're working on different things, every idea,
16:38every change of direction, every iteration you run, you know, that's how I try to pitch the
16:44product today, or this is, this is, I don't know, this is a feature I want to work on today.
16:50All of this is, you know, is in everyone's mind.
16:53Everyone is aligned constantly.
16:55Yeah.
16:56And that, that is, you know, from an alignment and communication perspective, super efficient.
17:02So you don't waste any time on caring about, okay, how do we align everyone?
17:06How do we make sure we all go into the right direction?
17:09How do we ensure that teams get information they need from each other?
17:13How do you ensure that sales talks to product?
17:15How do you ensure that product talks to engineering?
17:17How do we ensure that backend talks to frontend?
17:20You know, all of these problems, you simply don't have them.
17:23And everything is, the moment you have an idea, the moment you decide something, everyone's
17:28aware of it and everyone knows about it.
17:29And that's great because as you grow, it's very important to maintain this alignment, but
17:34it becomes really a lot of work and a lot of effort.
17:39So that's, that's definitely the benefit.
17:41When is a good time to, to really grow?
17:44I would say when you really are sure that you can achieve more with more people because
17:49that's not always the case.
17:50There are setups when you can achieve more with less people, right?
17:53Because if you don't have that friction of how to manage communication, you know, how to
17:59align everyone, this is simply time.
18:01This is mind space that you can invest, for example, into building product.
18:05So I would say when you really think, hey, the reason why we're not achieving more is
18:12because we're too few people, then, then you should, then you have to accept that you will
18:17have to hire more people and you will have to introduce friction and you will have to work
18:20on how to align them, how should they communicate, you will need processes also as much as, as
18:26much as, uh, people in startups are opposed to it.
18:29Um, it's really, it's really essential, you know, at least in a lightweight format to have
18:35some meaningful processes.
18:37Yeah, that makes a, that makes a ton of sense.
18:39And it's interesting.
18:40So we've, we've talked about the early days.
18:41We talked about staying small, the lack of scale and the benefits of that.
18:45Now, when we use the word scale and growth, I don't think we've ever seen scale and growth
18:51like we're seeing the AI market right now.
18:53Like it's, it's, it's absolutely incredible what's happened over the last six to 12 months.
18:59Uh, everybody has had their chat GPT moment.
19:02Now it is the, it's almost, it's almost a little bit trite to say that we're interested in AI,
19:07but as, as someone who's been deploying the technology, working with the technology in real world
19:14settings for a number of years now, where do you see the state of the AI market, the large language
19:21wallet market, and actually, how do you see that panning out for, for deep set as well?
19:28So in general, I think what we see is, you know, um, today we have a lot of awareness about
19:34the capabilities of the technology, we have some early adoption, um, it's now about driving
19:43adoption up, right?
19:44And for driving adoption up, at least on enterprise level, we see certain needs and requirements
19:50that need to be fulfilled and need to be in place, yeah?
19:53So, while everyone is excited and everyone has plenty of ideas what to do, companies need tooling,
20:00companies need, you know, they need platforms, they need something to really make these LLMs part
20:08of their products and processes, and one essential thing is that every use case in the end is somehow
20:14still individual, and you somehow have to account for these, for these very individual
20:20characteristics of use cases, right? Um, there often is not really a value to simply use an
20:26off-the-shelf model and that's it, yeah? Models are, uh, uh, the essential components and the essential
20:32fuel in these LLM AI systems, but you really need full systems, you need more complex applications,
20:39you need ways to connect the models to your data, for example, yeah? This is like always a prerequisite to
20:44really get value out of it, and, um, and that will often be very individual, not just by
20:50company by company, but really use case by use case, and this is very essential that
20:53these companies have these tools and these processes and, and platforms to do that,
21:00and the second important thing is around transparency, so we see that people,
21:08differently said, we, we cannot really understand large language models, right? So, um, I have a
21:14technically background in mathematics and we can hardly explain them, you know, the way mathematicians
21:20like to explain the world and do proofs on stuff, yeah? Um, so this means, this means, um, this means
21:28there is some limitation in full understanding how they work. Still, people want to get as close as they
21:33can to understanding what is happening there. This is, for example, why we see open source models still
21:39being of high interest in enterprise setups, right? Um, there's still, you know, not, I mean, there is
21:47appetite also for proprietary models, but a big appetite for open source models, especially for
21:52sensitive applications. Whenever you really want to have to make sure, hey, the outputs are, are, are right,
21:59the quality is right, I want to avoid, I don't know, hate speech or wrong facts out of my model,
22:04then usually we see that open source models are still the choice. And this transparency is something
22:09that also needs to be ensured, um, outside of, you know, the license of the technology, open source
22:15proprietary. It's a lot about, you know, we see enterprises asking themselves questions like,
22:23well, how good is this now really working? And is this really reliable enough? If I'm a bank and I
22:30want to
22:31use LLMs and risk management to extract relevant information of companies out of their company
22:36reports, I have to make sure that these LLMs get the facts and I have to prove that out somehow.
22:44And
22:44even if there are things where these LLMs are having challenges with, let's say they are very bad with
22:50numbers. Yeah, that's fine. But you need this transparency. And once you're in production,
22:56and once something is running, you know, you still need transparency because things will change.
23:01Your data will change. The things people will use the LLMs for might change. You always need to ensure,
23:07you know, that the quality is in place, that you get value out of it. And I think this is
23:12pretty much,
23:12um, this is pretty much really essential to go out of this 1% adoption today and move to 100%.
23:20And look,
23:21for us, um, yeah, that's to make the link. That's our future at DeepSet, right? To provide this tooling
23:27and provide these capabilities to give enterprises, um, the tools to build custom systems and to give
23:33them also the comfort, yeah, to use them, to run them in production and, you know, to extend their usage
23:38over multiple use cases. Yeah. It's really interesting hearing you describe that because in some ways,
23:44even though we're at a time now where DeepSet is at much larger scale than when we first invested and
23:50when
23:51you were early building the company, the way you're talking about the technology is actually the
23:55same in the sense that the, the, the technology is advanced, the company is advanced, but so has
24:02the uncertainty. And actually, I think that's never going to be different. I think as companies go
24:08through their life cycles, the state of the world is highly uncertain. The state of technology is highly
24:13uncertain, but what's very refreshing. And I think as a really interesting guiding North Star for some of the
24:20founders in the audience is the way that you always really focus on the use case and solving the
24:27problem. And I think one of the things that we see in, in speaking to lots of amazing founders out
24:31there
24:32is that the, the state of the technology changes, the state of the world is very dynamic and very uncertain,
24:38but what really is a sort of a, a, the guiding North Star and the thing that, that gets us
24:44very excited
24:45is when founders and companies are focused on solving real problems for people. And it's sort of,
24:53and actually I think there is a slight distinction in solving a problem in the absolute sense and
24:58solving somebody's problem. And I think you've got to do both. You've got to solve a, an important
25:04problem, and you've got to do that by solving somebody's important problem. And, and so I think
25:08what's really interesting is actually that you're, it seemed it's been a, it's been a few years since
25:14we first spoke, but actually, and, and, and the state of the world, the state of the technology
25:19and everything else has, has evolved significantly, but the way you're talking about it is the same.
25:24It's really focused on the use cases, focused on solving problems. And I think really that is the
25:30guiding North Star for all technology, not just in AI, but more broadly in, in building an interesting
25:36startup and also getting a lot of attention from investors, customers, and, and users, ultimately.
25:44There are some founders in the audience who are just starting out on their journey,
25:48thinking about starting a company, thinking about doing their, their first, their first funding
25:52round perhaps, or, or even earlier than that. Do you have any advice for people that are just starting out?
26:00So, as I, as I said earlier, you know, pick, pick, so pick something where you really want to become
26:08an expert in, a problem domain, and then work on, work on solving that problem, right? Be very clear
26:16about how do you solve it? How does the product you're building solve it? And take the time for this.
26:21You don't have to rush it. Yeah. Um, you shouldn't take 10 years, of course, but even there, you know,
26:26there are examples of companies that did 10 years of services or 15 years of services and then
26:31translated that into scalable businesses that, you know, grew to multi hundreds of millions of ARR within
26:37five years. Also that can be fine, but really take this time and make sure that you have as less
26:43overhead as possible in this time, you know, really be always free to move. It's, it's, it's a vanity
26:50thing, you know, um, to say, oh, I have a big team or I'm hiring or this, you know, you,
26:55you don't have
26:56to do it. Solve the problem. If you need people, hire them, but really try to stay as lean as
27:01possible
27:01until you really feel you have it, you know, and something you're, you're onto something and you will
27:07feel it, you know, the outside will give you confirmation. Your gut will give you confirmation,
27:13information, but really take the time and focus on that and get an accountant early on,
27:18I would say, because this, this would have been well invested money and well invested
27:21time from my side. Probably. Yeah, no, I, I agree. I agree with all of that. I agree. Make,
27:27make your life as easy as possible. And I think it also sort of thinks about the way that we
27:31think
27:31about fundraising in general. I think you're, you're raising a round of capital to run a set of
27:37of experiments to find out what the truth is. And I think, um, you've done that really well.
27:42It's been, it's been a privilege and an honor partnering with you as an investor, uh, and, uh,
27:47look forward to everything that deep set and yourself will, will achieve over the future.
27:51Thank you everyone. Thanks a lot.
28:07Thank you.
Commentaires