0:00
Starting a company is a roller coaster.
Search this video
0:01
So, there's like low lows and there are high highs.
0:03
And [music] my view is that you should just ride the roller coaster.
0:08
Like people try to mitigate it, but just ride it. And try to have fun.
0:16
>> Please welcome Eric Landau who is the co-founder of Encord who is building the data layer for physical AI, right? >> Yep. >> Welcome Eric. >> Thank you. Thanks for having me.
0:28
>> [applause] >> So, as I was going like preparing the the chat and looking at your background, I was actually pretty surprised.
0:36
Like it's kind of like a very surprising that you did big data particle physics.
0:43
And then you became a quant for 10 years. >> Yep.
0:46
>> Before creating Encord.
0:46
Okay, like lead us through that. Like why?
0:48
Like is there a common thread here?
0:51
>> Yeah, the the common thread is almost the kind of trajectory of AI development in the time that I've had my career.
0:55
So, it's it's quite interesting cuz do you know the the bitter lesson? >> Uh remind me.
1:01
>> The bitter lesson is this idea that AI you know, it used to be that to build an AI system you would think very carefully about it.
1:09
You would do all the feature engineering.
1:10
You'd try to induce as much domain expertise into the the system before kind of wrapping up in machine learning.
1:17
And they found that actually it's better to not do any of that and just to scale the the system.
1:23
Just add more data, more compute into it.
1:24
And when I started in physics, I was doing particle physics.
1:28
The kind of systems in the the strategy was based off of the physics itself.
1:31
So, you'd filter the data based off like the way the particles you know, you know, were were moving the momentum and and then in high frequency trading, it was thinking about the market, what factors actually move the market one way or the other and then um really crafting that and then putting it together into machine learning model afterwards.
1:51
And now it's just throwing data into a big pile of machine learning and then uh seeing it inside. Yeah. >> Excellent.
1:58
Uh before speaking about Encord, um I realized we were like chatting quickly yesterday about that, right?
2:04
Uh I mean, it was not the best ideal bet to become wealthy.
2:11
Uh I mean, you've been 10 years as a quant.
2:12
I guess you were making way more money as a quant than you did as a founder. Is that right? >> Yes.
2:18
Uh and in fact, when I quit, I quit during COVID and uh everything just kind of I quit to go start a company and that month uh the market went absolutely crazy.
2:28
It was the month where oil was trading negative, extreme high volatility, people were making a lot of money on uh Wall Street bets.
2:35
My desk made more money in that month than I think they made in the previous 3 years combined.
2:40
So, I just quit um and I was sitting on my couch fighting Python dependencies in a machine learning uh library and I was like, "What have I done?"
2:47
Uh but obviously, uh I didn't do it for the money.
2:51
>> Yeah, you left a lot of money on the table.
2:52
Well, you didn't know you left as much as that.
2:56
Um and so, what did you do it?
2:56
Like, what led you to uh actually quit a very well-paying job to become a founder?
3:03
>> Uh so, one, my co-founder and I really believed in AI and uh to us, it felt like this was the technological paradigm shift of our generation.
3:14
This was the early days of the internet.
3:16
This was the early days of computing.
3:18
And whenever I I I was a quant for quite some time and for the last several years when I would go to sleep at night, I would like hit my head on the pillow and I would think, "What am I doing with my life?"
3:29
You know, and uh when I started a company, even when all the craziness happened in the markets, uh I would go to sleep at night and just think about the problems of the company.
3:37
I wouldn't have the uh existential fear anymore.
3:39
It was all like more very >> very concrete ones.
3:42
What we are doing actually was providing value to someone. Well, did it at first.
3:48
>> Uh no, it doesn't at first, but you believe you're part of something which is bigger than you and it's very rewarding.
3:52
It's extremely rewarding to be in a company and seeing it grow and seeing it the impact that can it can have. >> Okay. And so we're in 2021.
4:03
Is that that that's when you created Uncode AI?
4:06
Um how long did it take for you to could be convinced that there was something here?
4:09
Like the early years were not that easy, right?
4:12
>> No, we we were in the desert for a couple years.
4:15
Um so >> Wait, like what made you continue and not give up? Or pivot?
4:20
>> Uh we no, we still believed in AI and you know, if you look in the long long horizon, the power of the technology, the generalizability of it.
4:27
Um but it was hard to convince people uh that it was as important as it was.
4:34
Uh and then ChatGPT happened and the kind of tenor of the conversation changed, which uh definitely helped us quite a bit.
4:41
>> So the ChatGPT release actually knocked the market?
4:45
Like and is that because people needed more data or because people starting to realize uh that AI was important?
4:52
>> Yeah, it was that more they were open to the conversation.
4:55
So it's funny because product market fit for some folks, it's uh so product market fit is both product and market and these things have to come together.
5:03
And so um some people they get it they achieve product market fit by setting a market constant and then compounding their product to hit that market.
5:11
Sometimes it's the market comes to the product, you know, the world changes and suddenly the thing that people didn't care about now care about.
5:18
Uh but most cases it's them kind of uh both moving at the same direction.
5:20
And for us our product was just continuously getting better and better and better.
5:25
And then after ChatGPT, I wouldn't say there was like a big binary threshold change, but the market started coming to us faster than it was before.
5:31
And so um a lot of people a lot of companies you talk to them they they did a product launch, they changed their product and then they suddenly felt immense product market fit.
5:40
That never happened to us.
5:41
It just felt like it was a little bit that better every day.
5:44
Just kind of was slowly slowly compounding.
5:46
And looking back on it, um my co-founder, he thinks the time when we had product market fit was uh when we saw, you know, a a sale close in the Gong channel, and we didn't know what the company did.
5:56
It's like, "What is this company?"
5:58
Um and that was very uh satisfying because we were doing all the sales ourselves.
6:03
We were like intensely involved in every single part of every company that we were trying to sell to.
6:08
And when you see something, they're giving you money, you don't know what they do, uh that's like, "Oh, we made it."
6:13
>> You you are not involved in the sale. Is that what you mean? Okay. When was that? >> Uh uh 2 months ago.
6:17
No, um it was it was some some years ago.
6:20
But, it was only looking back.
6:22
There was never a single moment where like, "Oh, we have got product market fit."
6:26
>> Yeah, I mean, uh it's kind of like actually one of the topics here about I mean, you were a quant physicist, and you had to build the sales org at some point.
6:34
How were you able to do that?
6:34
How did you hire the first sales people, and when?
6:40
>> Yeah, um it was I wish we had a more elegant answer to this, but it was trial and error.
6:46
So, [laughter] >> How many did you have to fire?
6:48
>> Uh yeah, so we had a we brought in a sales team, it didn't work.
6:50
We brought in another sales team, it didn't work.
6:54
By the third sales team, it started to work.
6:56
And it's funny because, um you know, the the advice that they give you in YC, it's one of the things that we found is like, it's almost always right.
7:04
So, >> I mean, the advice of YC is always right.
7:07
>> Yeah, it's almost always right, but you still make the mistake.
7:09
And I remember very distinctly, there's a particular hire that we were about to make, or actually we made.
7:14
And we talked to you, this was maybe a couple years after we had finished with YC, and you were you said, "Oh, for this role, you should hire you should look out for this and this and this, and don't do this and this and this."
7:27
And we're like, "Oh, no."
7:29
Uh because we had just hired exactly what you said not to hire.
7:31
And uh we realized, "Oh, we made a big mistake."
7:36
Uh but, we still brought them on, and we had to learn that it wasn't a fit.
7:39
Um we we should just cut it there, but um, we don't need to keep them.
7:44
>> That one was um, I I think probably 9 months. >> It's okay.
7:49
>> And then you Oh, 9 months.
7:49
Okay, not 9 days or 9 weeks.
7:53
>> We've gotten um, a lot more efficient at at Learnings here, but uh, yeah, this was something that we learned the hard way. >> All right.
7:58
So, most of you are not there yet.
8:01
Uh, well, there is no world in which you keep someone 9 months that you know is bad.
8:06
>> Yeah, and it's never um, the thing is it's never Sometimes it's like bad.
8:08
It's like, "Oh, man, this is bad."
8:10
But usually it's one of those things where you kind of know in your gut.
8:15
Um, and or you know you have the feeling, but it's just difficult.
8:18
And there's never a strong for There's never a strong reason.
8:22
You just know uh, it's probably not a fit, but there's nothing that pushes you to it.
8:27
And what we've gotten a lot better at over the years is just following your gut and making the right decision faster.
8:35
>> Yeah, I cannot uh, emphasize that more.
8:37
Um, let's go back to the product now.
8:40
Uh, because you didn't start with Physical AI.
8:41
Physical AI is kind of like the It felt It feels like you've pivoted a few times.
8:45
Still always around data for AI, but the target applications have changed over time, right? What did you start with? Was that healthcare?
8:54
>> Uh, we Yeah, we started in vision and our thesis originally starting with vision was it was the hardest modality to start with.
9:00
And so, we always wanted to be multimodal.
9:02
And because our thesis was that AI systems would be multimodal naturally because humans are multimodal, the best way to make that you make the best decisions is to take as many sensory inputs as you can uh, you can accept.
9:13
Uh, so, vision was a very dense data type.
9:16
We developed a lot of expertise into it.
9:18
And then we moved into multimodality.
9:20
And what happened with Physical AI is that uh, that's one of those cases of, you know, your product goes up and the market comes down.
9:26
Um, the market started really coming in our direction.
9:28
And the biggest application of multimodal AI now is in Physical AI.
9:33
So, now we work in a lot of robotics, autonomous vehicles, logistics, manufacturing.
9:36
Um, but we started in other use cases which were less less popular now.
9:44
>> It feels like uh in some sense physical AI is still nascent as a market.
9:48
I don't know like you say that I see in core bots and I don't see robots every day yet.
9:52
Uh but what you're saying is that the market is already there.
9:56
It's already here for you.
9:58
>> Um so there are a lot of robots that are working in the real world, but um the opportunity in physical AI is enormous.
10:06
80% of economic activity is manipulating or moving things in the real world.
10:11
So we live in the real world.
10:13
And if you talk to a lot of the robotics folks, they think in 5 to 10 years there will be more robots than people.
10:21
So in this room all the other seats will be robots that are listening to the talk and then transmitting it back into some centralized server.
10:29
>> may not need the robots for that.
10:31
>> Yeah, yeah, but the YouTube feed >> Yeah, you will >> This will be a world a room or a world packed with autonomy, let's say. >> All right.
10:40
>> Your company is doing great.
10:42
And you are betting the future of the company ahead of the curve on a new space. Is that fair?
10:48
And like how did you come up with that decision?
10:49
Did you see new customers coming from physical AI or was that the result of a deep conversation with your co-founder or something?
10:58
>> Uh it's yeah, so with AI you always have to you can never be comfortable.
11:01
You always have to go try to think about the next thing.
11:05
And the best way to do that is just to talk to people.
11:07
Like talk to as many people as you can. Talk to smart people.
11:11
Um and have conversations with your co-founders.
11:13
So we're constantly constantly uh even talking to prospects, to customers, um people that are ahead of us in the space, these at these YC events, uh talking to other co-founders in AI, and really just trying to absorb what's going on the What are they What problems are they thinking about? What are they doing?
11:28
And then synthesizing those things into the the next bet that we can we can make.
11:31
Um so it's not us sitting in a room quietly trying to imagine the future.
11:35
It's really being out there and getting as much data as we can. >> Okay.
11:39
So you shift the narrative, you shift how you speak of the company itself.
11:44
What about existing customers?
11:47
>> Uh well, because we >> Did you still have a lot of like health care models and >> Yeah, we still have a bunch of health care companies.
11:52
Uh we have companies in basically every application.
11:54
And the most fun one um Uh people always ask what the most interesting one is.
12:00
It's facial recognition for cows.
12:03
Um so it turns out that um humans are are really bad at recognizing cow faces, but algorithms can do it quite well.
12:09
Um >> Probably better than you.
12:10
>> Yeah, so um we've seen every we've seen everything in AI.
12:13
Um and that's the one of the the >> So who is annotating the the data? >> Uh >> Humans though.
12:19
>> Yeah, humans that are experts in Well, you can there's other ways of tracking cows through through the barn.
12:23
Um and but you want to know their health and things and um yeah, humans are are um are only a part of the annotation process as well.
12:33
>> Yeah, once you know who is the cow, like you tag it, then you know you can recognize its face many times or something. >> Yeah. That's that's cool. That's cool.
12:41
Um And what would you say are the biggest use cases today then in physical AI?
12:47
Like you said autonomous driving as one.
12:50
>> Yeah, autonomous driving, that's one that's um further along in the maturity curve.
12:54
So it's closer to actual production.
12:56
>> Uh actually in production in some places.
13:00
>> Yeah, I'm close to kind of mass scale production.
13:02
Uh and one of the bigger the biggest growth for us now is in robotics.
13:05
So uh humanoids, consumer robotics, uh manufacturing, there's so many different applications for for robotics that we're seeing.
13:14
>> Um And how do you collect data?
13:18
>> How do we collect data? >> Yeah.
13:19
>> Um so we do have a a facility in the Bay Area that uh we have robot systems and operators that are are collecting data for various tasks.
13:27
Um a lot of the data is collected in production as well.
13:29
So, our system uh ingests production data from the customer and then uses >> Is that video stream or is that >> It's yeah, video stream, sensor data, audio, uh language.
13:40
It's It's multimodal data.
13:43
>> So, how does that work that facility in the area?
13:45
Like people come with their robots and you have like a some setup >> That's right.
13:51
Yeah, we have It's almost like film sets where you kind of build a thing, you have a robot do a thing in a specific environment, collect data in that um that environment.
13:58
But uh data collection is usually for the companies that are earlier in the stage.
14:03
They're collecting enough data for pre-training.
14:05
A lot of companies already have production loops.
14:07
>> So, what's the rest of the product?
14:09
>> The rest of the product is doing the management of that data, uh the curation, so selecting the right data.
14:14
It's We say it's like um finding a million needles in a billion haystacks.
14:19
So, there's a lot of data that you need to use, but even more data that you have.
14:23
And then the annotation and enrichment of the data and finally the evals.
14:28
>> Do you have a Can you share with us like maybe some sense of the scale of the operation?
14:32
I don't know if you share the revenue or team whatever.
14:36
>> Uh well, we have hundreds of of customers.
14:38
Um we deal with more data than was used to train uh at well, from what we know GPT-4.
14:43
We don't know the the latest amounts of data for for GP uh GP 5.
14:47
Um yeah, we're we're dealing with multiple petabytes of all different types of multimodal data.
14:54
>> So, you just mentioned that you built a facility in the Bay Area, but you're based in London.
14:59
You're European and you still expand your team, build that facility there. Why?
15:05
>> Yeah, so my co-founder is in San Francisco.
15:07
I'm I'm still based in London.
15:09
We started the company in London.
15:11
And our general philosophy is uh you should just be where the important people for your company are.
15:16
So, for us it's customers and then talent and then investors in that order.
15:20
Um but different people have different uh weightings of those categories.
15:24
So, if you're building something that in in deep tech, uh you don't care about customers, you might over-index on just the talent bit.
15:31
And then you can just be close to where the good talent is.
15:33
For us, most of our customers are in the US, and of those, the vast majority are in the Bay Area.
15:40
Um so, my co-founder moved there, we built a facility there, we built a lot of operations there, uh just to be close to have those conversations. >> Huh?
15:49
>> Uh talent is is make but I think um London has some great talent for for AI and engineering great um uh universities, um people from uh from other companies.
15:59
Um so, there's a comparative advantage for for talent uh right now in in London Europe.
16:04
Although, there's now this London maxing term, which is um kind of arbing out the uh the talent pool in in London.
16:11
>> it's that the uh yeah.
16:12
Like the competition for talent is slightly less intense in Europe than it is in uh in SF. >> Yeah. >> Okay.
16:18
Um um you are competing with like kind of crazy good companies too.
16:24
I mean, uh when we speak about the data collection, I'm thinking about scale, uh and uh and Merkle and many other all are US companies.
16:33
Like, how do you stand out?
16:37
How do you win against them? >> Yeah.
16:39
Um so, one when we started the company, uh the the main um pushback that we got was like the space is way too competitive.
16:49
There's too many companies.
16:49
And then over the years, a lot of our competitors actually they pivoted or they consolidated or they ran out of money.
16:55
And for a long time, there actually wasn't much competition in in the field.
17:00
Now, it's getting competitive again.
17:00
And our view is actually we want it to be competitive.
17:04
We prefer to have really good competitors because they they push you, they make you better as as a company.
17:10
And uh standing out, we all have different kind of flavors within the the ones that that you mentioned.
17:14
Our focus is really on physical AI and scalability of the data.
17:19
Like being able to operate on the multiple petabyte level of data, which is hard for people to do, and which we spent a long time, um, you know, building the foundations for.
17:30
>> If uh anyone here is building in physical AI, uh should they try, like and they're small, and they're just starting, what should they do?
17:37
Should they use such a product like yours?
17:39
I guess you have gone up market over time, and your product is maybe not affordable for them.
17:45
>> Uh yeah, so usually where people come to us is when they've built internal tools, they've tried POC, and they're at the inflection point where they need scale.
17:54
So, their their their models are going from POC to production, or they're just getting a big influx of data.
17:58
And I I usually I I don't want to like oversell our product when it's too early.
18:02
It's like when you're ready, you should use a product like ours, um but you can start with, you know, open-source tools. >> Like what?
18:10
Like any any product Anything to recommend?
18:13
>> There's a there's a whole host of them.
18:14
I don't want to undercut my own my own book here. >> That's okay.
18:19
Um before uh concluding here, uh any uh as we're on the topic of advice, any advice you'd give to founder in that space, physical AI?
18:30
>> The advice that I think is important for basically all founders, and uh one of these came from one of the YC retreats that um the CEO of Reddit was talking about, which is uh starting a company is a roller coaster.
18:42
So, there's like low lows and there're high highs, and um my view is that you should just ride the roller coaster.
18:49
Like people try to mitigate it, but just ride it, um and try to have fun, uh because whenever you're at the low, you know that a high is coming.
18:55
Whenever you're at a high, you know a low is coming, and just accept that the that you're on a roller coaster.
19:01
>> I sometimes I I speak about like hacking your brain.
19:03
If you can hack your own brain to enjoy the fires, and then when there is a new fire, a new challenge, you just run to the fire.
19:10
Run at the fire like because you it's exciting.
19:15
Then it's it becomes so much more enjoyable as a journey.
19:16
Because yeah, you're going to go through crazy highs and crazy crazy crazy lows.
19:22
>> Yeah, roller coasters are fun.
19:24
>> Did you think that when you started? >> No.
19:27
>> When did you start thinking that?
19:28
>> Uh uh only relatively recently actually.
19:31
I when you go through enough of these these signing sort of fluctuations, you feel like, "Okay, I can just make this fun." >> That's interesting.
19:40
It's rare that people think that it becomes easier.
19:44
Do you think it become it became easier?
19:46
>> Uh you you become more you realize that you have more control than you think over your own mental state.
19:50
And then you can kind of accept what the reality is and how you how you react to it.
19:54
And that's something that has helped a lot in seeing the fluctuations in increase over time. >> Awesome. Thank you, Eric.
20:03
It was great to have you with us today. >> Thanks so much. >> [applause]