Latest / What's Your Problem? / Mapping the Unmappable

Transcript

Mapping the Unmappable · What's Your Problem?

0:15
Speaker 1: Pushkin, what do you understand about maps and the world
0:22
Speaker 1: because you have been, you know, studying it and working
0:25
Speaker 1: on it for so long.
0:26
Speaker 2: I think the reason I personally love maps is just
0:30
Speaker 2: because there's a super super complex problem and very very
0:33
Speaker 2: hard thing to solve. That's why why I'm so passionate
0:36
Speaker 2: about it.
0:37
Speaker 1: This is Philip Kundall. He's the chief product officer at
0:40
Speaker 1: a company called Grab. Grab is huge in Southeast Asia
0:44
Speaker 1: and its main businesses are delivery and mobility, kind of
0:47
Speaker 1: like a combination between Instacart and Uber, and as a result,
0:51
Speaker 1: maps are at the core of its business.
0:53
Speaker 2: But the fascinating thing for me is like, once you
0:56
Speaker 2: get a digital representation of the real world, you can
1:00
Speaker 2: basically the vision that I always had for long, long years.
1:03
Speaker 2: If you haven't solved that yet is imagine when he
1:05
Speaker 2: went from Yahoo to Google. Right. Yahoo was like the
1:07
Speaker 2: static index of web that you could say, for something
1:11
Speaker 2: that just leads you to a website. Right if you
1:12
Speaker 2: would search for sports and it gets you to a
1:14
Speaker 2: sports side and then you need to navigate your way.
1:16
Speaker 2: And then Google you can say you search for a
1:19
Speaker 2: very specific thing and it brings you exactly to that page. Right, Yeah,
1:23
Speaker 2: and that's what we haven't solved with with maps yet.
1:26
Speaker 1: Right if you say, if we're still in the Yahoo
1:28
Speaker 1: era of maps, is what you're telling me.
1:31
Speaker 2: Maybe slightly beyond the Yahoo era. But let's say like
1:34
Speaker 2: when you say, where do I find the in Southeast Asia?
1:36
Speaker 2: Durian is like a super popular fruit? Right say where
1:39
Speaker 2: I said, where do I find the freshest Durian for
1:42
Speaker 2: this price? Right now, right now, the real world, right now,
1:46
Speaker 2: in the real world, right like, which store stocks a
1:49
Speaker 2: Durian right now? Show me where it is and then
1:52
Speaker 2: guide me to that store. There's nobody who has solved
1:55
Speaker 2: that problem. And and that's the fascinating part for me
1:59
Speaker 2: that I want what Google has done for the web.
2:01
Speaker 2: I want that for the real world.
2:03
Speaker 1: I mean, that's a wild problem. When you formulate it
2:06
Speaker 1: that way, you would have to know everything all the time.
2:09
Speaker 1: There's an information problem there that is quite hard.
2:13
Speaker 2: I mean, the hot problems are the fun ones and
2:15
Speaker 2: a yeah.
2:23
Speaker 1: I'm Jacob Goldstein, and this is what's your problem? The
2:26
Speaker 1: show where I talk to people who are trying to
2:28
Speaker 1: make technological progress. Philip Bilton sold a mapping startup before
2:32
Speaker 1: he wound up at grab in twenty nineteen. Today, he's
2:35
Speaker 1: based in Singapore, which is one of several countries where
2:38
Speaker 1: Grab operates. Others include Indonesia, Malaysia, the Philippines, Thailand, and Vietnam.
2:44
Speaker 1: And in addition to doing delivery and mobility, they also
2:47
Speaker 1: do payments. It's what they call a super app. And
2:51
Speaker 1: I wanted to talk to Philip about maps in particular
2:55
Speaker 1: because I mean, I guess I just kind of thought
2:57
Speaker 1: maps were solved or assumed that without really thinking about it,
3:01
Speaker 1: and then when I started learning about GRAB, I realized
3:04
Speaker 1: I was very wrong. So there's that dream of real
3:08
Speaker 1: time mapp apps that Philip mentioned a minute ago. But
3:11
Speaker 1: there's also something that's really interesting and kind of more
3:14
Speaker 1: immediately relevant, and that is this The online maps we
3:18
Speaker 1: use in the US and other developed countries just don't work.
3:22
Speaker 3: Very well in a lot of other parts.
3:24
Speaker 1: Of the world. And when Grab launched several years ago
3:27
Speaker 1: in Malaysia, that turned out to be a big problem.
3:30
Speaker 2: I mean, from day one, you needed maps, so you
3:31
Speaker 2: can't run the company without maps, and grabs started using
3:35
Speaker 2: a third party service. But then what we've realized a
3:38
Speaker 2: lot of these services are built for like a developed
3:41
Speaker 2: market like the US, and very built around the mental
3:44
Speaker 2: model of like cars, and Southeast Asia is really different.
3:49
Speaker 2: We're operating primarily on motorbikes, even two thirds of our
3:53
Speaker 2: transport trips on the back of a motorbike. And then
3:56
Speaker 2: if you know Southeast Asian cities, then you have these
3:58
Speaker 2: narrow alleys and sideways and so on, and traditional maps
4:02
Speaker 2: just don't cover them because interesting, basically how maps are
4:07
Speaker 2: made traditionally is with these big mapping vans that I'm
4:10
Speaker 2: sure you've seen driving through cities. One hundred and fifty
4:13
Speaker 2: two hundred thousand dollars basically look like a way more right,
4:16
Speaker 2: That's kind of like how they look like, and they
4:19
Speaker 2: don't cover the roads that we need to deliver our
4:21
Speaker 2: services and really reach our customer because they expect to
4:25
Speaker 2: be picked up in their home. They expect that the
4:27
Speaker 2: food get delivered to their front door, and not just
4:30
Speaker 2: to the nearest street that might be two hundred meters
4:32
Speaker 2: away in terms of like a big car drivable street.
4:35
Speaker 1: So lots of life, lots of people live, lots of businesses,
4:39
Speaker 1: earned streets that a van couldn't even drive down if
4:41
Speaker 1: it wanted to.
4:42
Speaker 2: Yeah, no chance. I mean, when I do the immersions
4:45
Speaker 2: like on the back of a motorbike. There's absolutely no
4:48
Speaker 2: chance that with a man you can get there. It
4:51
Speaker 2: was a scooter centric world, and Southeast Asia updated so quickly,
4:56
Speaker 2: like I mean, there's like entire new neighborhoods springing up,
4:59
Speaker 2: and it was traditional maps are built in a way
5:01
Speaker 2: that these big vans collect once every one or two years,
5:05
Speaker 2: and we need to refresh the maps in like days.
5:08
Speaker 1: So you have this problem, like what's the first step?
5:11
Speaker 1: So you realize the maps aren't working. You're a mobility company,
5:14
Speaker 1: that's that's not gonna work. How do you It seems
5:17
Speaker 1: impossible to think, oh, well just map everything.
5:19
Speaker 2: How do you go?
5:20
Speaker 1: How do you start doing that?
5:21
Speaker 2: Yeah? Exactly right. It's a crazy problem, right yeah. I
5:24
Speaker 2: mean the numbers that were out there when Google started
5:27
Speaker 2: mapping the US, they apparently spent like anywhere between half
5:31
Speaker 2: a billion to one billion dollars and it clearly didn't
5:34
Speaker 2: have that amount of money to map it. And it
5:37
Speaker 2: seemed crazy, right like people when we started doing this,
5:40
Speaker 2: I mean, I got like so much feedback that people
5:42
Speaker 2: thought we were completely insane because God cost us and
5:46
Speaker 2: Southeast Age are right, I mean, we are home to
5:48
Speaker 2: like close to seven hundred million people, which is like
5:51
Speaker 2: a lot larger than the US. So you can imagine
5:53
Speaker 2: like how much it would normally.
5:55
Speaker 1: Cost, yes, two x exactly and crazy dense cities, traffic,
5:59
Speaker 1: tiny alleys. So how do you even start to undertake
6:03
Speaker 1: this project? What do you do?
6:06
Speaker 2: Yeah? No, so this was this was really the fun part.
6:08
Speaker 2: That's such a fun challenge yourself. But basically what we
6:12
Speaker 2: started is we started taking it from first principles like
6:15
Speaker 2: why was it so expensive to map? And then we
6:18
Speaker 2: tried to solve those problems and the problems why it's
6:22
Speaker 2: so expensive normally to produce a map are a few things,
6:25
Speaker 2: the ones that I just said. The mapping vehicles are
6:27
Speaker 2: extremely expensive. Then the people sitting in the mapping vehicle
6:32
Speaker 2: are extremely expensive because you sent them around driving in
6:36
Speaker 2: the country, sleeping in hotels. So the cost like to
6:39
Speaker 2: send these vans with people around cost a fortune, and
6:44
Speaker 2: like operating that is just super super costly. And that's
6:48
Speaker 2: what we had a unique advantage because obviously we had
6:51
Speaker 2: our drivers already crisscrossing the city in like insane amounts.
6:57
Speaker 2: I mean our drivers. Any city is like crossed by
7:00
Speaker 2: our drivers like one hundred times or more in a day,
7:03
Speaker 2: so there's like no roads that our drivers don't see.
7:06
Speaker 2: So we thought about two things that we needed to
7:08
Speaker 2: drive the radically down. One is we build our own
7:11
Speaker 2: collection hardware. So instead of building these like massive rigs,
7:15
Speaker 2: we build small GoPro like cameras with an echip in
7:19
Speaker 2: there that we can give to our drivers. And instead
7:21
Speaker 2: of costing one hundred two hundred thousand dollars for full
7:24
Speaker 2: mapping van, these cameras are under order of like a
7:26
Speaker 2: bunch of one hundred dollars, so orders of magnitudes cheaper.
7:30
Speaker 2: And then the second thing is we said, the driver's
7:32
Speaker 2: drive around the city anyway, can they just have the camera?
7:35
Speaker 2: And then we just need to pay them a little
7:37
Speaker 2: for them. It's a great extra income. But they get
7:40
Speaker 2: their main income from doing food deliveries or doing mobility trips,
7:44
Speaker 2: and they get an extra income from that. And that's
7:47
Speaker 2: because we could deploy like we have in Southeast Asia
7:51
Speaker 2: probably one hundred times at least more cameras than anybody
7:54
Speaker 2: else deployed because they're so cheap. And then we don't
7:57
Speaker 2: need to send drivers traveling everywhere.
7:59
Speaker 1: So let's break it down a little bit like building
8:02
Speaker 1: your own hardware, like the driver part is kind of obvious, right, like, oh,
8:05
Speaker 1: we've got these guys already out there, could we get
8:08
Speaker 1: them to do the mapping. It's not at all obvious
8:10
Speaker 1: to me that you would need to build your own hardware.
8:12
Speaker 1: So tell me about that, Like, first of all, why
8:15
Speaker 1: do you need to build your own hardware? Why not
8:16
Speaker 1: just dumb question, buy a camera?
8:19
Speaker 2: No, great, great question. Actually, I mean that's what we tried. Okay,
8:21
Speaker 2: So I think when we went and looked into this,
8:24
Speaker 2: because we didn't want to build hardware, and so basically
8:26
Speaker 2: we tried two things. So they were these go pros
8:28
Speaker 2: a few hundred bucks cameras, and then they were the
8:31
Speaker 2: more professional mapping Greade cameras twenty thousand bucks. So the
8:35
Speaker 2: problem with go Pro was a bunch of things they made,
8:38
Speaker 2: usually as action cameras, right, like you can go skiing
8:40
Speaker 2: and so on with them, but they don't have great
8:42
Speaker 2: GPS in them, and basically because you don't need it,
8:46
Speaker 2: right Like, if you go like somewhere mountain biking, it's
8:49
Speaker 2: typically open sky. It's not like a dense urban environment
8:52
Speaker 2: with big skyscrapers that reflect GPS. So they're decent for
8:55
Speaker 2: what people usually use. Goal pros, but they're not the
8:58
Speaker 2: media level precision we need for high density urban center mapping.
9:03
Speaker 2: So that was one problem. And the second problem with
9:06
Speaker 2: Goal pros was they just have no tooling that they
9:08
Speaker 2: kept operate twenty four to seven drivers upload automatically data,
9:12
Speaker 2: so they're not made They usually made for like a
9:14
Speaker 2: one two hour recording and not this heavy duty recording
9:18
Speaker 2: for the back of our motorbikes for twelve hours in
9:22
Speaker 2: blistering heat in Southeast Asia, tough rail and so on.
9:26
Speaker 2: So that's basically why both of these cameras that existed
9:29
Speaker 2: didn't fit the needs that are very harsh environment.
9:32
Speaker 1: So would the mapping camera have worked, but you didn't
9:34
Speaker 1: want to spend twenty grand per camera. That was presumably
9:36
Speaker 1: the simple problem with the mapping camera.
9:38
Speaker 2: Yeah, twenty grand was the problem. And the other problems
9:40
Speaker 2: are they're also quite heavy and bulky, So those cameras
9:43
Speaker 2: are close to ten kilograms, which is not super easy
9:47
Speaker 2: to operate. So they were too bulky and too costly.
9:52
Speaker 1: So what do you just get on a plane dest
9:54
Speaker 1: engine and tell them what you need, Like, how do
9:57
Speaker 1: you build your own camera?
9:59
Speaker 2: So this was really certing that we actually had a
10:02
Speaker 2: small team and Scenzan already building building off our battery
10:06
Speaker 2: swap blockers for scooters. And back then I talked to
10:11
Speaker 2: our Cito at that time and he is like kind
10:14
Speaker 2: enough to say, like I feel it, like we really
10:16
Speaker 2: need to find something next for this team to do.
10:17
Speaker 2: So you can have them all. They can build your cameras,
10:20
Speaker 2: so amazing, We're very lucky.
10:23
Speaker 1: So tell me about the camera they came up with.
10:25
Speaker 2: Yeah, so the first camera that we built, we called
10:27
Speaker 2: it a Karda cam, so like after like carda in
10:31
Speaker 2: like some language, was like map. So basically those cameras
10:34
Speaker 2: where the first version was mounted on the helmet of drivers,
10:37
Speaker 2: predominantly it was two hundred and fifty grams mounted on
10:41
Speaker 2: the helmet of drivers, and it would basically be a
10:44
Speaker 2: camera with an AI chip and really highly accurate GPS,
10:48
Speaker 2: like that's all we needed. And so they would just
10:50
Speaker 2: mount them on their on their helmets and go about
10:52
Speaker 2: their day and just at the end of the day,
10:55
Speaker 2: take it home, take it on their Wi Fi, upload
10:57
Speaker 2: the data and yeah, we would get data from tons
11:00
Speaker 2: of cameras.
11:01
Speaker 1: And then you made a cardacam too, right, tell me
11:04
Speaker 1: about Carteracam too.
11:05
Speaker 2: Yeah, we made a bunch of iterations of our hardware.
11:08
Speaker 2: So Karda CAN two is our latest generation, which is
11:12
Speaker 2: basically a three sixty camera, so it has basically four
11:15
Speaker 2: camera lenses and all directions. Has also like more advanced centers,
11:19
Speaker 2: like allied our sensors in there, so we basically made
11:23
Speaker 2: the setup a lot more professional that we can capture
11:25
Speaker 2: maps an even higher level of accuracy to get things
11:29
Speaker 2: like lane level navigation for our drivers or advanced safety features.
11:33
Speaker 2: So the latest camera is basically a great iteration that
11:37
Speaker 2: has taken just a step further to capture more details
11:39
Speaker 2: in the map.
11:40
Speaker 1: And where does it go? Where does the camera go
11:42
Speaker 1: if somebody's driving a scooter.
11:43
Speaker 2: Yeah, so now we've built a mount on the back
11:46
Speaker 2: of a motorbike. So we've built actually the cameras that
11:49
Speaker 2: you can either mount them at the back of your
11:50
Speaker 2: motorbike that it doesn't disturb you and in a pole
11:52
Speaker 2: so that's high so that you can see because the
11:55
Speaker 2: camera obviously needs to have a good view even if
11:57
Speaker 2: there's cars around, so we've built a special mount on
12:00
Speaker 2: the back of a motorbike. We have mounts also for cars,
12:02
Speaker 2: so some drivers also mount them on cars. And then
12:05
Speaker 2: we have a backpack that you can carry if we
12:07
Speaker 2: want a map indoor. The other thing that's really important
12:10
Speaker 2: in Southeast Asia is malls are gigantic. Do you have
12:14
Speaker 2: like these malls which it takes you fifteen minutes to walk.
12:16
Speaker 1: Around, and so people will get a delivery to whatever
12:20
Speaker 1: the noodle shop in the middle of the mall, and
12:22
Speaker 1: you've got to put that on the map.
12:25
Speaker 2: Well, the other way around, people get a delivery from
12:27
Speaker 2: the noodle shop of course, of course, and our driver
12:31
Speaker 2: needs to find their way there, and if they get
12:34
Speaker 2: into the wrong entrance of the mall, it might cost
12:35
Speaker 2: them ten to fifteen minutes extra to walk back and forth.
12:38
Speaker 2: So we need to precisely not only tell them go
12:40
Speaker 2: into the mall, but we also precisely need to tell
12:43
Speaker 2: them where to park their motorbike. So what we emphasize
12:46
Speaker 2: when we build our maps that we build them like
12:48
Speaker 2: and to end, and we say, here's where you park
12:51
Speaker 2: your motorbike, here's where you walk, here's where you pick
12:53
Speaker 2: up the noodles, and then you can do the same
12:56
Speaker 2: in reverse.
12:57
Speaker 1: How many of these cameras do you have out there
12:59
Speaker 1: like today? How many people are driving around mapping with
13:01
Speaker 1: these cameras today ish by end of this year.
13:04
Speaker 2: We have about twenty thousand cameras in the field. And
13:07
Speaker 2: to give you a sense, like professional mapping companies in
13:10
Speaker 2: Southeast Asia, to my best of my knowledge, have about
13:13
Speaker 2: tens of cars, but not tens of thousand tens of cars.
13:17
Speaker 1: So does that mean they have basically seeded it to
13:19
Speaker 1: you like that you have won the mapping Southeast Asia fight.
13:25
Speaker 2: I think it's still so so early. Fair. I mean,
13:28
Speaker 2: there's so many things we want to do.
13:30
Speaker 1: Interesting. I love that. So let's talk a little more
13:32
Speaker 1: about what you're doing now and then let's talk about
13:34
Speaker 1: what you want to do. So you have an incredible
13:37
Speaker 1: amount of data coming in now, right, I mean, if
13:39
Speaker 1: you have twenty thousand cameras driving around every day like
13:45
Speaker 1: that is a wild amount of input. What are you
13:48
Speaker 1: doing with all that data? I mean there's basic mapping,
13:50
Speaker 1: and I'm sure you've got that, but what's the next level?
13:52
Speaker 1: Presumably have AI? You said you have light? Are like,
13:54
Speaker 1: tell me all the things you know because you have
13:57
Speaker 1: all this data.
13:57
Speaker 2: Yeah, you're right. I think the basic vision that you
14:00
Speaker 2: have is get an accurate representation of the real world
14:03
Speaker 2: in real time. That's that's the that's the goal of
14:07
Speaker 2: mapping in real time?
14:09
Speaker 1: Is crazy? In real time? It's crazy? Yeah, that's wild.
14:12
Speaker 1: But fine, that's just like right now, what do you know?
14:17
Speaker 2: Yeah, so a bunch of things is I mean, first
14:18
Speaker 2: of all, everything that you need for navigation roads, how
14:23
Speaker 2: our road's drivable? Is the road safe to drive? Does
14:26
Speaker 2: it have a lot of potholes and things like that,
14:28
Speaker 2: and that signage obviously is it a one way road?
14:31
Speaker 2: Is it not a one way road?
14:32
Speaker 1: Can you worry me about a particular pothole? Could you
14:34
Speaker 1: be like, be careful, there's a pothole on the left
14:37
Speaker 1: coming up, try that right lane.
14:39
Speaker 2: And we're doing actually very cool things we're doing right now.
14:41
Speaker 2: We've already launched a safety navigation that tells you, hey,
14:46
Speaker 2: is the road safe to drive? Based on potholes? And
14:48
Speaker 2: does it have street lighting? Is it safe? Because if
14:51
Speaker 2: the road at night is well lit, it's a lot
14:53
Speaker 2: safer to drive, but it's not. So that's kind of
14:55
Speaker 2: like a practical use case that we that we already
14:57
Speaker 2: can do with that.
14:58
Speaker 1: Interesting as I'm sure you know, like ways in this
15:02
Speaker 1: country tells you where there's like a speed camera or
15:06
Speaker 1: like a speed trap. Do you do that?
15:10
Speaker 2: Yeah. Absolutely. So we have an AI voice reporting that
15:15
Speaker 2: drivers can share anything they're like, oh, the right lane
15:17
Speaker 2: of this road is closed, and then it gets processed
15:20
Speaker 2: and basically used for all the other other drivers. So
15:23
Speaker 2: we've launched that and it's been really successful as well.
15:26
Speaker 1: So when you say it's AI, does that mean the
15:28
Speaker 1: driver just says it and it gets integrated into the
15:31
Speaker 1: maps or what does that mean?
15:33
Speaker 2: Yeah, that's that's exactly right. There's the future. We actually
15:36
Speaker 2: work closely together with our friends at open Eye. So
15:39
Speaker 2: the driver natural language reports an issue and then it
15:42
Speaker 2: gets processed and if you're not clear, we ask your
15:44
Speaker 2: follow up question. If you say, hey, the right lane
15:47
Speaker 2: is closed and he said like, oh, do you mean
15:48
Speaker 2: this road? And then he says drive but yeah, that's
15:50
Speaker 2: exactly the road I'm talking about, and then we process
15:52
Speaker 2: it and basically one all the other drivers accordingly.
15:56
Speaker 1: So you mentioned your friends at open ai. It seems
15:58
Speaker 1: like good friends to have a good place to have friends.
16:00
Speaker 1: I know you have a deal with them, Like what
16:03
Speaker 1: is the nature of your deal with open ai and
16:05
Speaker 1: more generally, what's the work you're doing with them?
16:08
Speaker 2: So MAP I think is one of the key things
16:10
Speaker 2: we do together with them and use their models to
16:13
Speaker 2: improve our maps. The voice stuff that I shared earlier
16:16
Speaker 2: is like one of our one of our highlight features.
16:20
Speaker 2: And in general, we're just embedding like AI and everything
16:22
Speaker 2: we're doing. We're having we're having like over a thousand
16:26
Speaker 2: different AI models that we're that we're working on, so
16:29
Speaker 2: it's basically deeply, deeply embedded in many many of the
16:31
Speaker 2: things that that we're doing.
16:33
Speaker 1: Tell me more, like, what are some more examples of
16:36
Speaker 1: how you're using AI.
16:38
Speaker 2: So one of the latest things that that I really
16:41
Speaker 2: like that's really cool is the other thing that has
16:44
Speaker 2: really huge impact on our marketplace is weather. In Southeast Asia,
16:48
Speaker 2: the rain is insane. It is like pouring down, the
16:52
Speaker 2: roads are flooding, so and for our services for mobility
16:57
Speaker 2: for deliveries, that has a tremendous impact on that. So
17:00
Speaker 2: knowing when it rains early is super super critical so
17:04
Speaker 2: we can adjust the marketplace. We've launched AI based rate
17:08
Speaker 2: detection that a bunch of things, So we deploy sensors,
17:13
Speaker 2: other sensors in cars. So we have basically a device
17:15
Speaker 2: that's called a card to dongle that we deploy and
17:18
Speaker 2: the cars that we own and it plugs into a
17:20
Speaker 2: port in the car. That's called OBEDI two port, and
17:21
Speaker 2: that reads when the windshield viper's going.
17:24
Speaker 1: Oh genius, such a simple way, such a simple way
17:29
Speaker 1: to know if it's rating. Does the car have the
17:31
Speaker 1: windshield wipers on? It's so second order. I love it though.
17:35
Speaker 2: Yeah, and how fast it is?
17:36
Speaker 1: Oh? How fast? It's how fast the car is going,
17:38
Speaker 1: because the slower it's going, the heavier the rain. Is
17:40
Speaker 1: that the inference how fast.
17:42
Speaker 2: The windshield wiper is going. Because you kind of like this,
17:45
Speaker 2: are this right?
17:47
Speaker 1: So okay? So so and what do you do? What
17:50
Speaker 1: do you do with that data? That's the input, what's
17:52
Speaker 1: the output?
17:53
Speaker 2: The output is basically we know the moment it rains,
17:56
Speaker 2: we know demand will go up and supply of drivers
17:59
Speaker 2: will go down. So what we do We try to
18:01
Speaker 2: activate more drivers, so we would send out to all
18:04
Speaker 2: the drivers, Hey, it's starting to rain. There's a fantastic
18:07
Speaker 2: earning opportunity. If you're not worth now's a great chance
18:10
Speaker 2: to get on the road and make extra money. And
18:13
Speaker 2: then we try to actively get the supply of drivers
18:17
Speaker 2: up so that we can keep our reliability at the
18:19
Speaker 2: levels we need.
18:20
Speaker 1: We know people get all worked up about sarge pricing.
18:23
Speaker 1: Do you use searge pricing in that setting? It sounds
18:25
Speaker 1: like a classic use of search pricing. Everybody wants a driver,
18:27
Speaker 1: nobody wants to drive. What you need is a higher price.
18:29
Speaker 1: Do you do that?
18:30
Speaker 2: That's exactly right? Or like you need to motivate the
18:33
Speaker 2: drivers to come on the road.
18:34
Speaker 1: Yeah, no, there is a market. I'm pro searge pricing.
18:37
Speaker 1: That's a great example. What's another example of the way
18:40
Speaker 1: you're using AI?
18:41
Speaker 2: So we used AI to translate in many scenarios. For example,
18:47
Speaker 2: when I go like to Jakarta, I obviously don't speak
18:49
Speaker 2: any Bahasa, but I can message our drivers in the
18:52
Speaker 2: real time, translate any messages, send them a chat to
18:55
Speaker 2: the driver and he can reply back to me in basade.
18:58
Speaker 2: I see it in English on my side, he sees
19:00
Speaker 2: it in basade my site. But the more important one
19:03
Speaker 2: for me was like food menu translations. So we invested
19:05
Speaker 2: quite a bit because whenever I go to Thailand and
19:08
Speaker 2: then like I mean, I can't read any tie script obviously,
19:11
Speaker 2: and I in the past I look at like some
19:14
Speaker 2: things on the menu and I have no ideas that
19:16
Speaker 2: like for me, always are worries that ultra spicy can
19:18
Speaker 2: I eat it or not, and I didn't even know
19:21
Speaker 2: what the item is, Like can look at the picture
19:22
Speaker 2: and kind of guess, but sometimes merchants don't have pictures.
19:26
Speaker 2: So we used AI to to translate all these menus
19:30
Speaker 2: in all kinds of languages, so that that has been
19:32
Speaker 2: a super impactful one as well. We'll be back in
19:39
Speaker 2: just a minute.
19:50
Speaker 1: So I'm curious about grab maps enterprise, right Like, you're
19:54
Speaker 1: selling something related to your maps to big companies right
19:59
Speaker 1: to to Microsoft, to Amazon and to government. It's like,
20:02
Speaker 1: what tell me about that business?
20:04
Speaker 2: That was quite an interesting one, right Like, we would
20:06
Speaker 2: have never imagined early on when we built our maps
20:08
Speaker 2: that this business will be created. But we've got people
20:11
Speaker 2: like once we publicized our maps and people have seen
20:14
Speaker 2: that they work generally quite well, we've got people approaching
20:17
Speaker 2: us say hey, can we use them as well? And
20:19
Speaker 2: that was like the genesis of the enterprise business. And
20:24
Speaker 2: then we started working very closely with AWS where any
20:29
Speaker 2: developer on their platform can use our maps now, so
20:31
Speaker 2: we have over one hundred different developers already using it.
20:36
Speaker 2: We partner with Microsoft, as he said, a bunch of
20:38
Speaker 2: other large tech companies that in Southeast Asia started using
20:42
Speaker 2: our maps because they've seen i mean, all these things
20:45
Speaker 2: that I shared earlier that just really serves their needs
20:47
Speaker 2: in Southeast Asia a lot better than anything else out there.
20:50
Speaker 1: What's an example if something someone has built, you know,
20:53
Speaker 1: on top of it.
20:55
Speaker 2: A really cool startup that I've seen that used our maps.
20:58
Speaker 2: What they do They go from merchant to merchant and
21:01
Speaker 2: collect like the old oil that they use for cooking
21:04
Speaker 2: and that basically recycle it. So they send somebody around
21:07
Speaker 2: going to all these merchants collecting that. And in the past,
21:11
Speaker 2: and a lot of these merchants are like small neighborhood
21:13
Speaker 2: mom and pop shops in all these like little side
21:16
Speaker 2: rolls and alleys, and that had a hard time for
21:18
Speaker 2: the for the person going around collecting all of this
21:21
Speaker 2: stuff finding that. And that's one of these examples where
21:24
Speaker 2: they use grab maps to make the navigation of the
21:27
Speaker 2: person going around and collecting all of that a lot
21:29
Speaker 2: a lot easier. And there's many of these kind of
21:32
Speaker 2: stories where people use it for things that that are
21:36
Speaker 2: very similar in those kind of finding the last mile
21:38
Speaker 2: has been really really hard before. So I'm curious.
21:43
Speaker 1: What you're working on now, Like, Yeah, what what are
21:46
Speaker 1: some of the things you're trying to figure out that
21:47
Speaker 1: you haven't figured out yet.
21:49
Speaker 2: I think the one thing that we're really passionate about
21:52
Speaker 2: is solving more of the indoor problem. So that's one
21:55
Speaker 2: thing that we're really mapping more so. We have a
21:57
Speaker 2: decent amount of malls already map, but there's still so
21:59
Speaker 2: much more to do. So hopefully we can find that.
22:02
Speaker 2: Butever you go into these malls, we can exactly help
22:05
Speaker 2: you to find whatever you're looking for, specific store or
22:09
Speaker 2: general a general kind of like shop or so. So
22:13
Speaker 2: that's that's one thing that I really want us to crack.
22:16
Speaker 1: So it's the general shop idea that like, if I'm
22:18
Speaker 1: just not working for grab, I'm just in the general
22:22
Speaker 1: public and I'm like, I want to buy a pair
22:24
Speaker 1: of shorts, I just type that into grab maps, and
22:27
Speaker 1: grab maps tells me where to go.
22:28
Speaker 2: Is that what you're thinking of there, Yes, but we
22:30
Speaker 2: always think about it with a hyperlocal twist.
22:32
Speaker 1: Yeah.
22:33
Speaker 2: So, and what that means. As an example, let's say
22:36
Speaker 2: in Indonesia, a large part of the population is Muslim,
22:40
Speaker 2: which means they generally eat halal, and this is like
22:45
Speaker 2: all the mapping platforms they don't support that. You can,
22:47
Speaker 2: of course for every search for you can say restaurant halal,
22:50
Speaker 2: this halal, this halal, but you cannot make it part
22:52
Speaker 2: of your user profile. But it's it's like, it's not
22:55
Speaker 2: it's not something that you switch every day. Well, like
22:57
Speaker 2: either your preference is halal or is not. And and
23:03
Speaker 2: a lot of the mapping platforms support putting in your profile.
23:05
Speaker 2: I only want halal restaurants because I don't eat anything else.
23:09
Speaker 2: And those are the kind of things that we see
23:11
Speaker 2: in Southeast Asia that we need to really solve because again,
23:15
Speaker 2: nobody else would solve those kind of problems. So capturing
23:18
Speaker 2: this data accurately and knowing all these details, those are
23:21
Speaker 2: the kind of things that we really put a lot
23:24
Speaker 2: of emphasis on.
23:26
Speaker 1: Are there technical problems you're working on, Like are there
23:30
Speaker 1: things where you haven't figured out the right tech or
23:33
Speaker 1: where you need to build something, or where AI models
23:36
Speaker 1: aren't quite where they need to be, but you're trying
23:38
Speaker 1: to push them.
23:39
Speaker 2: I think what we are trying to do is what
23:41
Speaker 2: I said earlier, like you want to have a real
23:42
Speaker 2: time accurate model of the world.
23:44
Speaker 1: Yeah, so let's talk about that. Like real time is
23:47
Speaker 1: a crazy phrase in that context, right, Like when you
23:51
Speaker 1: say real time a model of the world. What are
23:54
Speaker 1: you thinking of?
23:55
Speaker 2: Yeah, I mean a simple use case would be nowhere
23:59
Speaker 2: there's parking right now? Yeah, not like where there's parking
24:02
Speaker 2: in general, but I'd say a roadside parking. If you
24:05
Speaker 2: had another crap card driving past and say, hey, fifteen
24:09
Speaker 2: seconds ago there was a freak parking spot on the right,
24:12
Speaker 2: that's that's extremely useful.
24:15
Speaker 1: That is extremely as well that I know people in
24:18
Speaker 1: Brooklyn and San Francisco who would love to have that functionality.
24:22
Speaker 2: Yeah. I think for us, really the goal is always
24:25
Speaker 2: what we know is like shaving off seconds of every delivery, right, Like,
24:29
Speaker 2: that's what really really makes a delta. I think that
24:34
Speaker 2: I really loved Steve Job's old mental model. But he
24:37
Speaker 2: convinced people to make the macboot like ten seconds faster,
24:40
Speaker 2: and he said like, well, if you take this, I
24:42
Speaker 2: don't remember the exactly now mark for him, but he said, like,
24:45
Speaker 2: if you make the macboot ten seconds fast and there's
24:47
Speaker 2: fifty million people using it every year, you save like,
24:50
Speaker 2: I don't know, it's like ten lifetimes of people's time
24:52
Speaker 2: waiting for the Magic Boot something like that, and fast
24:55
Speaker 2: right Like across I mean across I calculated to share
24:59
Speaker 2: this always with a team. Across a billion deliveries, two
25:02
Speaker 2: point five seconds saved across every delivery is roughly one
25:05
Speaker 2: lifetime that you can save. So any second we can
25:08
Speaker 2: shave off by getting cash to park faster, by getting
25:12
Speaker 2: motorbikes to park faster, park at the right space, at
25:15
Speaker 2: the right time. That's kind of the problems that we're
25:17
Speaker 2: really passionate on solving.
25:19
Speaker 1: How long does it take you to do a billion deliveries?
25:21
Speaker 1: How is it a billion per per way?
25:24
Speaker 2: We don't publish exact data, but it's the auder of
25:27
Speaker 2: magnitude of it, like a year or less.
25:29
Speaker 1: Oh wow, okay, that's a big number. So how do
25:34
Speaker 1: you get real time parking data? I mean, so, is
25:36
Speaker 1: the constraint now just getting more cameras to more drivers,
25:40
Speaker 1: like the what's the rate limiting step?
25:42
Speaker 2: Yeah? Great, great question. I think more cameras is one constraint.
25:46
Speaker 2: The other constraint is doing all the smart processing on
25:49
Speaker 2: the act because obviously you cannot upload all these data
25:53
Speaker 2: because that would be extremely costly, and mobile networks in
25:56
Speaker 2: Southeast Asia aren't quite that powerful that you could upload
25:58
Speaker 2: millions of video streams to the cloud. So we're working
26:02
Speaker 2: a lot on what is called like ATGYI that we
26:04
Speaker 2: can run all these models that are powerful but not
26:07
Speaker 2: in the cloud but on the app or at least
26:09
Speaker 2: on a mix of boths.
26:11
Speaker 1: So in this case, is the edge the actual camera?
26:13
Speaker 1: I mean, what what is like is the dream the
26:16
Speaker 1: camera itself is doing the work and just uploading something
26:19
Speaker 1: very simple to the network.
26:21
Speaker 2: That's in many places already happening. We already have quite
26:24
Speaker 2: powerful AI chips in the cameras. But I mean, of course,
26:28
Speaker 2: like no mobile phone, no edge camera right now is
26:30
Speaker 2: as powerful as let's say CHET GPT with GBT four O,
26:33
Speaker 2: that's sure.
26:34
Speaker 1: I mean, you have this sort of spectrum where you
26:36
Speaker 1: have a whatever one hundred million dollar data center on
26:39
Speaker 1: one end and one hundred dollars camera on the other
26:42
Speaker 1: and some things in between. Right, So what can you
26:44
Speaker 1: do on the camera now?
26:46
Speaker 2: We do things like privacy, so we blur people's faces,
26:50
Speaker 2: we blow license plates because we never want to upload
26:52
Speaker 2: this information. We do weather detection what I said earlier,
26:56
Speaker 2: with rain, so we detect on the cameras if it's raining,
27:00
Speaker 2: things like that. We detect traffic signs and see if
27:02
Speaker 2: they've changed. So we actually already run large part of
27:06
Speaker 2: processing on the edge and then only if something has changed,
27:10
Speaker 2: then we upload some data into a validation on the server.
27:13
Speaker 2: But that already allows us to reduce what we upload
27:15
Speaker 2: by more than ninety percent to.
27:18
Speaker 1: Just tell the server if something is different.
27:20
Speaker 2: Basically, yeah, exactly.
27:22
Speaker 1: So I know, you know, we've talked mostly about maps,
27:25
Speaker 1: but obviously GREB is doing a lot of different things.
27:28
Speaker 1: Are there other parts of the business that we should
27:31
Speaker 1: talk about?
27:32
Speaker 2: The business is so fascinating, so if I have time,
27:34
Speaker 2: we should talk about all of our business.
27:36
Speaker 1: Yes, just tell me what's one other sort of frontier,
27:39
Speaker 1: one other thing you're working on.
27:41
Speaker 2: I think the other thing which we've really invested deeply,
27:44
Speaker 2: which is also quite closely connected to our maps, is
27:47
Speaker 2: when we look at across the delivery journey. The other
27:50
Speaker 2: part where a lot of time is spent is in
27:53
Speaker 2: the merchant cooking and preparing the food. And that's an
27:57
Speaker 2: area where you spend a lot of time optimizing together
28:01
Speaker 2: with the merchants to make sure that the food is
28:04
Speaker 2: prepared in the right time, that the driver arrives in
28:07
Speaker 2: the right time. So you've done lots of lots of
28:09
Speaker 2: cool things. So, for example, we've built a data science
28:12
Speaker 2: model that accurately predicts at every time of the day
28:15
Speaker 2: how long the merchant needs to prepare an order. So
28:18
Speaker 2: we can detect the busyness of the merchant and say,
28:22
Speaker 2: if we know all the merchant has gotten a lot
28:23
Speaker 2: of orders, normally they take to prepare, They prepare to
28:28
Speaker 2: bud me in like three minutes, but when they're very busy,
28:30
Speaker 2: they take seven minutes. And the merchants don't want that
28:33
Speaker 2: our drivers crowd their store and weigh in troves, So
28:36
Speaker 2: we typically allocate the driver only when we know, okay,
28:40
Speaker 2: the food is ready in seven minutes, and we don't
28:42
Speaker 2: allocate the driver immediately, but we know all there's a
28:45
Speaker 2: driver two minutes away, we just allocate them like five
28:48
Speaker 2: minutes later so that he's in the shop just in time.
28:52
Speaker 2: And that also cuts the delivery journey, makes a lot
28:54
Speaker 2: more pleasant for the merchant not having like a lot
28:56
Speaker 2: of people crowd their store, and allows us to offer
28:59
Speaker 2: more affordable price to the consumer because they don't need
29:01
Speaker 2: to pay for all these minutes of the driver waiting.
29:03
Speaker 1: So just all these optimizations, all these different margins where
29:07
Speaker 1: you can optimize. So if we think about whatever five years,
29:11
Speaker 1: ten years out, when you think about this sort of
29:14
Speaker 1: medium to long term like what's your dream, Like, how's
29:18
Speaker 1: it work? What's going on?
29:20
Speaker 2: I think like for me, the world is changing so fast.
29:22
Speaker 2: Five to ten years prediction is really really hard with
29:25
Speaker 2: all the AI advancements.
29:26
Speaker 1: How far you want to go?
29:27
Speaker 2: Two years? Four years?
29:28
Speaker 1: You tell me, like what you just give me a
29:30
Speaker 1: dream for the future.
29:31
Speaker 2: I mean the obvious one that I'm very passionate about
29:34
Speaker 2: is robotics that make so much sense in our marketplace.
29:39
Speaker 2: So if you can add robotics to our marketplace, that
29:42
Speaker 2: will be a huge, huge change and something we're actively
29:46
Speaker 2: working on to make happen. So that's I think probably
29:48
Speaker 2: the thing I'm most passionate about to change.
29:51
Speaker 1: Tell me what you're actively working on with robotics.
29:54
Speaker 2: We haven't shared much which we do on the delivery side,
29:57
Speaker 2: but for example, I mean things like autonomous cars. We've
30:01
Speaker 2: signed an agreement with a bunch of like autonomous car
30:04
Speaker 2: providers to come with us to Singapore and so on,
30:07
Speaker 2: so we'll do a bunch of things in that space.
30:09
Speaker 2: But basically you can imagine that it will be in many,
30:12
Speaker 2: many parts of our delivery chain.
30:14
Speaker 1: You need to build an autonomous scooter based on what
30:16
Speaker 1: you've told me, right, I feel like the analogy to
30:19
Speaker 1: the map story is the autonomous scooter that can go
30:21
Speaker 1: down all the alleyways.
30:22
Speaker 2: Great, if you ever want to have a job in
30:24
Speaker 2: our product team, to join us, build autonomous schooters with us,
30:28
Speaker 2: because you're exactly right, but you need to build products
30:30
Speaker 2: that work in our region. So that's exactly the right
30:33
Speaker 2: mindset that we're trying, like what we say always and
30:36
Speaker 2: we want to build hyper local products that work in
30:38
Speaker 2: Southeast Asia.
30:42
Speaker 1: We'll be back in a minute with the lighting round. Okay,
30:55
Speaker 1: let's finish with the lightning round. So I read on
30:58
Speaker 1: your bio you write that you mostly travel between Singapore,
31:03
Speaker 1: San Francisco, Berlin and Cluge. Tell me about Clue the
31:09
Speaker 1: other three I'm familiar with Clues. I don't know anything
31:12
Speaker 1: about How does that wind up on that list?
31:14
Speaker 2: As a city in Romania, in the heart of Transylvania actually,
31:19
Speaker 2: and we have a we have a small engineering center
31:22
Speaker 2: there in our maps team, So I've spent a lot
31:25
Speaker 2: of time there because for my own startup, that's that's
31:28
Speaker 2: how I wound up Inclusure. Originally for my startup, Scobbler,
31:31
Speaker 2: which was a maps and navigation startup. I built a
31:34
Speaker 2: engineering team and cluge, so I've been going to Clusion
31:38
Speaker 2: working with engineers there for the last fifteen years or so.
31:42
Speaker 2: What's clue Like, it's fun. It's a student town, so high,
31:47
Speaker 2: high energy, young population, very smart, the computer science department
31:52
Speaker 2: very good. They used to clone the IBM mainframes for
31:54
Speaker 2: the Soviet Union, so a bunch of hardcore engineers.
31:58
Speaker 1: You also write in that bio you wrote in Berlin,
32:01
Speaker 1: I experienced some of the best parties ever. Tell me
32:06
Speaker 1: about a Berlin party.
32:08
Speaker 2: That was prior line, That was before I moved to
32:11
Speaker 2: before I moved to Southeast Asia. But Berlin was a fun, fun,
32:14
Speaker 2: fun journey for I lived there for five six years.
32:17
Speaker 1: Maybe he didn't tell me anything about any party. So
32:22
Speaker 1: and then you also say you lived and worked in
32:24
Speaker 1: Singapore and Berlin and San Francisco, and I'm curious, like,
32:27
Speaker 1: how is sort of whatever professional culture, work life different
32:32
Speaker 1: in those places. What are the sort of striking differences?
32:35
Speaker 2: I mean, Germany is extremely direct, but like that's where
32:38
Speaker 2: I'm originally from. I'm originally German, and Germans are known
32:42
Speaker 2: to be super super direct, which in Southeast Asia doesn't
32:47
Speaker 2: work super well. If they're not accustomed to it. So
32:50
Speaker 2: I think for me, the thing that I needed to
32:52
Speaker 2: adjust most is to become a lot more indirect, to
32:56
Speaker 2: become a lot more like have those conversations and like
33:00
Speaker 2: smaller groups, not in front of everybody. So I definitely
33:04
Speaker 2: had to just my style quite a lot when moving around.
33:08
Speaker 2: But I found that always incredibly fun. I always loved
33:11
Speaker 2: learning about new cultures, So after some adjustments, I really
33:15
Speaker 2: really like it here.
33:16
Speaker 1: That is super interesting. Was there, as you were saying
33:21
Speaker 1: that I was wondering, like, was there a particular moment
33:24
Speaker 1: when you realized, Oh, I'm not behaving correctly in this
33:29
Speaker 1: cultural context?
33:31
Speaker 2: Yeah, I mean this predate grap So this was when
33:34
Speaker 2: I was first doing business in China. So I sold
33:37
Speaker 2: my startup to a Silicon Valley company, but I also
33:40
Speaker 2: got a two hundred people team in China report to me,
33:42
Speaker 2: which was the first time that I ever managed a
33:44
Speaker 2: team in China, and I was extremely surprised why they
33:50
Speaker 2: wouldn't tell me about all the mistakes, all the things
33:52
Speaker 2: that I did wrong. And I said, like, that's very unusual.
33:55
Speaker 2: Normally I get a lot of like pushback from engineers.
34:00
Speaker 2: And then I started like asking people is why don't
34:03
Speaker 2: they do this? And then people said like, well, it's
34:05
Speaker 2: kind of root in a public forum to say the
34:08
Speaker 2: boss's wrong. And then I realized, like, oh, that's that's
34:13
Speaker 2: why I just need to like change how I'm asking questions,
34:16
Speaker 2: how I'm managing teams. And that's when I first managed
34:19
Speaker 2: teams in China. Took me quite a while to figure out, honestly.
34:22
Speaker 1: And how does San Francisco fit in relation to both
34:26
Speaker 1: working in Berlin and working in you know, China and Singapore.
34:31
Speaker 1: Where's the US on that continuum?
34:34
Speaker 2: Do you ask for me? I think the thing that
34:36
Speaker 2: I really loved in San Francisco was just the craziness
34:40
Speaker 2: of ambition. When I spent like April back in in
34:44
Speaker 2: SF four month to like vibe code and hack on
34:47
Speaker 2: a bunch of hobby projects, and like you go in
34:49
Speaker 2: every random coffee shop and there's somebody who said they
34:51
Speaker 2: work on a billion dollar idea and they they're just starting,
34:55
Speaker 2: they have nothing yet they're just like, Okay, I'm going
34:58
Speaker 2: to make it big. So I think that that ambition
35:02
Speaker 2: and that willingness to openly declare it even if people
35:05
Speaker 2: know it's super unlikely. There's not thousands of people who
35:08
Speaker 2: start billion dollar company, but in the Bay Area that's
35:10
Speaker 2: thousands of people who say they will and this conviction
35:13
Speaker 2: and this like optimism. I think that was for me
35:16
Speaker 2: one of the most striking things in the US, and
35:20
Speaker 2: that I still love the energy whenever I go there.
35:22
Speaker 3: Basically, Philip Condall is the chief product officer at GRAB.
35:33
Speaker 2: Please email us.
35:34
Speaker 1: At problem at Pushkin dot fm. We are always looking
35:38
Speaker 1: for new guests for the show. Today's show was produced
35:42
Speaker 1: by Trinamanino and Gabriel Hunter Chang. It was edited by
35:45
Speaker 1: Alexander Garreton and engineered by Sarah Bruguer. I'm Jacob Goldstein
35:50
Speaker 1: and we'll be back next week with another episode of
35:52
Speaker 1: What's Your Problem.