Mapping the world

Welcome to another episode of Practical AI. This is Daniel Whitenack. I am a data scientist with SIL International, and I’m joined as always by my co-host, Chris Benson, who is a principle emerging technology strategist at Lockheed Martin. How are you doing, Chris?

I am doing very well. How’s it going today, Daniel?

It’s going great. It’s been a podcasting day. Actually, I don’t know if you saw, but right before this one I recorded with the Changelog, who is our sort of sister podcast. The Changelog Podcast is all about software engineering and open source and other things; they’ve been going for a long time, and the creator of Elixir was on there, talking about their new Numerical Elixir library and Axon, which is a neural network library for Elixir, which is really cool. I really enjoyed the conversation. I don’t know anything about Elixir; I don’t know about you, Chris…

I used to follow it a lot; since I have not been doing as much focus strictly on programming languages, I haven’t… But I know it’s a really cool language. I know Jose, the creator, is a pretty talented person.

Yeah, for sure.

So now that he’s doing these data science libraries, I’ll have to dive into it. So I’m gonna listen to your episode and see what he has there. I’m excited about it.

It’s really cool and impressive. One week they released the numerical library for Elixir, the next week (seemingly) they released the neural network library, and not long after they released their own version of Notebooks… So it’s pretty cool.

Okay. Watch out, Python, because Elixir is right on your tail there.

[03:55] [laughs] One of the features listed on their Notebook library was sequential execution, which for anyone that works in Notebooks and struggles with state, that’s pretty cool. Anyway, I was pretty excited to do that. I’ve already had a good conversation about all sorts of AI things today, but I’m really excited to hear more about some things related to autonomous driving, and also mapping.

I know we’ve talked about before, but I think talking about the specific mapping side of that and some of the data along with that, and also the human-machine interface that goes along that is really interesting, and that’s what we’re gonna talk about today with Ro Gupta, who is CEO of Carmera. Welcome, Ro.

Hey, guys. Thanks for having me on.

Yeah. Before we get into all that, could you just give us a little bit about yourself and your background?

Sure. I’m co-founder and CEO of a company called Carmera. We’ve been around for about six years. Before that, I was at a different startups in the web 2.0 hype phase of the internet. Before that, I’ve also been with bigger companies, like Disney… But really, my academic grounding and what we do with Carmera started back in the ’90s, when I was an undergraduate in an operations research program at Princeton, and got exposed to the early forms of everything we do now.

I actually designed my first autonomous mobility system in 1998, I think; very theoretical, early forms of ML, computer vision… The non-deep-learning kind of computer vision. The old school. Neural nets - everyone in all the faculty thought they were gimmicks that they just kind of had to tell us about…

Little did they know…

Yeah, well… It took about 15 years for it to get real.

A lot of people saw cool vision, but it definitely at that time – I get what you’re saying; it was a neat toy, for sure.

Yeah. So that’s a little bit about my background. Going back farther, I was born in India and moved to the U.S. when I was young, but as we can talk about, I think, spending time in developing countries also definitely has influenced how we see infrastructure - infrastructure is the hot word right now - but particularly roads and digitizing roads, which is what we do.

I would love to hear about that at whatever point you wanna dive into it, and just kind of understanding what kind of perspective that gives you…

Sure.

…as well as what you’ve just mentioned in terms of getting started back in the last winter for AI, before neural networks evolved back out into the field of deep learning as it is today, and having gotten into autonomy at that point. I’m pretty interested in how that’s also shaped how you approach the deep learning side of it.

Sure. I’ll say this - for me, as a tool, as a means to an end as opposed to an end in terms of my career, it’s fascinating, as you noted; not only are we knee-deep in it in our current jobs, but I was exposed to it at a pretty impressionable time in my development… But for me, actually, I think what has been kind of hardwired into me in terms of from a very young age, and again, growing up, I was born in Calcutta, India. One of the most densest, chaotic urban settings you can imagine. I live in New York now, and we think that’s pretty dense and chaotic… It’s nothing like a developing country’s big city.

I think when you grow up in an environment like that – I mean, I’ve gotta think subconsciously… Well, not even subconsciously. I think even consciously - I realized this later in life when I was older and also spent some time in some developing countries, like in Africa for example, in Mozambique, with some NGO work I did - you really don’t take infrastructure, specifically mobility infrastructure for granted, because it is so much at the forefront of everyday life; you have to take into account to get somewhere much more than you would – there’s many cases where there isn’t even a real drivable road.

[07:59] So I think that probably had a big impact on me… And for me, when I say AI is a means to an end, for me the end has always – I think what’s always been interesting to me is… You know, I see road as sort of the circulatory system of the planet, the IRL version. It is just how things move around, how the nutrients get to us, whether it’s to our house from an Amazon delivery, or whatever… And I’ve seen also what happens when there are clots all over the place; it sucks, and it’s just bad for everything.

So I’m not gonna pretend like when I was four years old and I was moving from Calcutta to United States I almost had some epiphany, but I think that hardwiring got me very interested in the concept of basically packetizing roads, the way that our digital lives are and our internet is. So first, before you packetize, you’ve gotta digitize, and that’s exactly what we in our industry and Carmera is doing.

So I’ll kind of pause there, but I think – I’ve actually not really thought about it in that way until you’ve just asked me that, but… I mean, I sort of thought about it disparately like that, but I think – I’m gonna stick with that answer.

It was a good answer. It was an interesting answer.

This is a little bit of a shrink session, actually…

[laughs] No, it was fine… I’m here, I’ll send you my bill at the end.

[laughs] I’m fascinated by the way you described that, and your thought process around roads. I think when there is also – maybe cities in the U.S. have a certain perception of roads, and our expectations around how they operate, which is of course very different than if you go to the developing world… I think also there’s varying degrees of perceptions around maps and what we expect those to be, and contain in all of those… So maybe you could just talk for a second about in terms of the autonomous driving world, what sorts of maps are involved in the development of autonomous driving technologies. Maybe some of those are different than the sort of maps that we might think of right away in terms of “Hey, I look on my phone and here’s a route from my house to the restaurant down this road”, and that sort of thing.

Yeah, sure. I think it’s probably good to abstract that a little bit initially, and then speak a little more specifically to maps, both for machines and humans. For machines, for autonomy - basically, for the autonomous driving use case, which is the vertical that we really started focusing on when we came out of stealth a few years ago - we serve other use cases as well, but let’s just start with that, since this is an AI podcast… For those of your listeners who are Bayesians or Bayes theorem fans, it’s really best to start off with that maps are priors, basically. I was talking about my past, and when I was an undergrad in the ‘90s - basically, it’s just statistics, and probability, and then we had to invent more and more impressive terms for all this stuff; ML, and deep learning etc. But that’s basically – I was taught Bayes theorem in the ‘90s, and that’s what maps are, it’s priors for Bayesian reasoning, and that includes a robot car.

So the starting point is “Okay, what is the role of maps?” It has more actual utility than that, actually, and many people feel like the role of it as a prior in the pure sense may actually – it probably won’t ever go away totally, but may be less of the emphasis, whereas more of the value of maps being in kind of the foresight value, so in telling you what your sensors don’t have line of sight into, but for planning purposes.

But really, much of the industry is still very reliant on maps and high-definition maps as priors for the initial localization and perception decisions that the AI-based vehicle is making. So that’s really what it boils down to. The question then is “What really moves the needle in your map prior?” and then there’s a big optimization problem for that, versus cost, speed of update, and then of course scalability is part of that big equation as well. That’s kind of the initial premise of what everyone’s working on, but it’s also a bit of a moving target, which I can certainly talk about as well.

[12:28] I’m curious, you mentioned a couple times high-definition maps, and I see that as verbiage on your website and what you’re working towards… So when I think of high-definition, I mostly think of photos or videos that have a certain resolution, or something like that; when I’m watching Netflix, or whatever it is.

In terms of the data that’s needed for whether you’re using this data for priors, like you were talking about, or maybe you’re using it in other ways that people are developing, what does high-definition mean in terms of the map data that’s used in autonomous driving and other related technologies? Is that also images and video, or is that high-definition in terms of information? Could you expand on that a little bit?

And if I can add to that just a tiny bit… For those who don’t have your experience, differentiate what that is versus what the map that most people think about in their head is; obviously, that’s not enough, but maybe how do you have to step forward to get to where you’re at…?

Sure. I haven’t come here to plug our blog or anything, but we just happened to write a piece about this that’s really relevant, so if anyone’s curios, you can go to carmera.com and click on that.

We will definitely include it in our show notes.

We are curious. Tell us about it.

Well, it also has a very pretentious title, which is my fault… The name of the title is “The mapping singularity is near.”

Naming is important, as we’ve found out doing a podcast… If you look at statistics, it turns out the name of the post is definitely important. I mean, content of course is what grabs people, but I totally get it.

Well, thankfully the content has been received really well, actually across our industry… Even quite frankly people who’ve had to [unintelligible 00:14:17.25] Anyway, go read the post… But to answer your question, what the post is really helpful for is really defining these things. And the problem is these have been very nebulously defined, actually. High-definition - I’m not sure who actually coined it. I don’t know if it was Navtech, or someone else; maybe it was from the DARPA days… I’m not really sure, but it was just sort of this useful catch-all term for something higher-definition than a normal map that humans have been using for research and navigation…

Like a static image, essentially…

Yeah, exactly. And really, even those maps only started to really become digitized in the 2000’s, really… So it’s still pretty recent. So the difference is they’re actually – what we’ve written in that post is the differences between what we call STs, so standard definition maps (that’s the maps we’ve known forever) and high-definition maps (HD maps) has felt pretty binary for the past 5+ years. SD maps have been for human use; typically, it’s like researching a restaurant, or giving you some basic directions, but you’re still the one driving, or riding your bike, or walking. And again, that’s for you the human to consume. HD maps for the machine - and like I said before, the primary need is in the form of priors, with some other added benefits for localizing and path planning.

And to answer your question in terms of what are some specific differences in the data - again, the post has some nice sort of examples. We show you an intersection mapped in SD and then in HD, and to give you an example, that SD map for that intersection - I’m doing this off the top of my head, but I think there’s only maybe like nine features that you need to represent that intersection. If you had it in a database, it’d only be like “This many lane lines, this many signs etc.”

[16:17] The HD map - it’s maybe nine features and a few dozen attributes of each feature. The HD map, you’re talking hundreds of features and thousands of attributes. So every turn line, and every traffic signal, but not even just the traffic signal, but knowing about every bulb, and the phasings, and which lane of traffic each bulb controls…

So part of it is what we would call feature and attribute granularity, so you’re talking like 100x difference in that. Part of it is also in spatial accuracy. An SD map - typically, you can be off by tens of meters; that’s very common. Whereas HD, typically the absolute accuracy bars have tended to be in the tens of centimeters. There’s a bunch of variants; some people are in the lower end, some people are closer to a meter… But definitely sub-meter and in the tens. So it’s both of those things.

And ultimately, the point of that post though is we’re actually seeing trends towards a convergence. We’re actually already seeing it shift from a binary situation to a continuum. So as a company, we build high-definition maps from the bottom-up as we need to, and typically we do the full-stack more for like the urban robo-taxi, mobility-as-a-service type deployments, where you have a more contained geography, but much higher granularity in accuracy requirements, to the level I’ve just mentioned.

However, the biggest unsolved problem really in mapping for all levels of fidelity is change management. When things change, it’s still really the old methods of map creation or map maintenance are just too slow or expensive for this to work, especially for machine-first uses. So that’s what our technology and our focus has been on… Either for our own base maps, or even for some other mapping company’s map, we can modularize our change management technology which uses camera-based crowdsourcing to do that really efficiently at scale.

So for that change, what we call change-as-a-service - that we actually do in a layer of fidelity internally called medium definition (MD), because that’s a really useful state to keep data in, and only upgrade it to HD quality when you need to. Otherwise, oftentimes it’s overkill and it’s needless cost or time to do that across the board.

But what we’re seeing, this trend we’re seeing is actually convergence, where the AI is actually getting so good that it’s not asking as much of the HD map as it used to even just a few years ago. They’re saying “You know what - actually, for those features, if I just know that it’s in this block, or at least in this hundred-meter or ten-meter stretch, that’s fine. You don’t need it to be placed perfectly to ten centimeters. But these other ones I do.” So basically the lines are blurring…

So the point is that there’s some convergent layer in between the standard definition and high definition, this so-called medium definition, and it’s already where we’re seeing the technological needs go to… Because the real-time perception and controls technology are getting good enough where they don’t need as many crutches in the form of priors as they used to. However, they definitely still need some, and are still very reliant on them… And the trade-off though - what they’re saying is what we need this is in way more places; just mapping a little urban area in Vegas, or just the highway network isn’t good enough; we need this to be way more ubiquitous, and we need you to update it way more frequently than the old maps used to be. If you look at Google Street View, for example, you are lucky if the last Street View image was within the last year. Sometimes it’s three years ago, depending where you live, or even more.

That’s exactly the problem we’re solving, and kind of paying attention to how these goalposts are moving towards this MD steady state that we think is emerging.

You had mentioned when you were starting out talking about how you viewed infrastructure, how you viewed roads… You were talking about this sort of perspective of also thinking about infrastructure in developing countries. Of course, a lot of the focus on navigation systems and technology has been focused in U.S. or European cities, and that sort of thing… If we’re able to start bringing down the fidelity of what’s required for navigation and even some advanced technology, like autonomous driving, what does that mean for the developing world? Do you see that sort of change being impacted as well?

Absolutely. Totally. That’s exactly my hope and our hope. That’s why we’ve been really delighted, actually, to see that these trends may be – they really do appear to be merging, because that post was written on the backs of a lot of pattern recognition across the entire automotive and AV industry, and maps industry, and a lot of big companies I can’t really mention on the record… But everyone has been sort of – especially, interestingly, in the last… What are we in - April? So maybe in the last 8-9 months we really started to see that pick up.

And by the way, I think some of that has been almost a forced discussion, given where Tesla’s going with so much of this. You could actually kind of argue Tesla’s already essentially using MD maps, they just don’t like to call them that. They kind of eschew using higher definition maps altogether, but they are using enhanced map data that a normal human map wouldn’t have… It’s just that they wanna be rid of it altogether and have a super-purist, AI-only approach.

But I think the good thing about Tesla is it’s really forced a lot of the old guard, so to speak, the incumbents to think about future realities sooner than they otherwise would have.

I think MDs - one example of that, and I think that it does get us excited… You know, a big kind of ethos of Carmera is you look at – if you look at our website, and even one of our internal inclusion groups, we use the term “for all” a lot. One of our long-time partners, Toyota, “Mobility for all” is their tagline. So that’s always been very much the ethos of Carmera, is to sort of liberate and democratize. We’d always used to use those terms in the early days of “How do we liberate this data that only some small handful of really large companies are able to collect and make it accessible to all?”

[24:06] My hope is actually what we kind of saw with something like telecommunications. When I was in India as a kid, and then even when I’d come back, it was like a big deal and pain in the ass to make a call, anywhere really, and especially internationally. You’d have to go to these government booths on the side of the road… Whatever, it was a pain; it was not a connected country. Then mobile came along, and they didn’t have to follow the same decades-long [unintelligible 00:24:36.13]

Yeah. They skipped that whole step there.

Exactly. So that’s exactly what I’m hoping this - and many other things - could mean for developing countries getting access to this type of technology sooner.

And what do you think – just to follow up on that… You mentioned crowdsourcing as well was a big part of your change management of maps… I’m thinking back to when I visited Russia, I visited Yandex. So if you go to Russia, and least when I was there - and maybe it’s changed, I don’t know - when you’re navigating on the streets, everybody uses these Yandex maps, which looks amazing. It’s like Google Maps, it’s beautiful… But it’s like, they were telling me “Hey… If you look at the fidelity that you get from Google, they’re not putting in the effort to map all of these roads in Siberia, but we’re gonna put in the effort over here”, and that’s why people use it, is because they have that fidelity.

It’s cool to see localized companies do those things, but it seems even more powerful if you put some power in the hands of users and people that are users of products to help crowdsource that data. What does that scenario look like to you as we move forward?

Well, there’s two thoughts I have on that. First off, I’m optimistic – both of thoughts I’m optimistic on other countries, especially developing ones, kind of not being second-class citizens when it comes to having their roads mapped at the same standards we see here.

First off, maps have been a huge money pit for a long time, and there’s a good reason Google has such a lead - it’s because they were the only company who had that will… First off, that’s a big part of it - the founder-led will to do it. There’s an old story about how Larry and Sergey were convinced to do this… And of course, just the dry powder to go spend an inordinate amount of money on these maps around the world. And that’s part of the reason we exist, is because that’s not tenable at a certain point, especially – it’s one thing to spend a huge amount on creating the map once around the world, but even for Google, or for any company, you can’t keep doing that for maintenance.

But the other thing that has been useful is - you know, in the early days, maps weren’t really directly monetized. They were more just this absolute killer app that set Google so far apart from others in terms of using their other consumer products, whether it’s search, or Android, or whatever. What does appear to be happening, if you kind of read into some of their investor statements and things like that is they’ve been able to connect those dots much more closely, of what they’re spending on maps actually monetizes from their other products, like search, for example.

I heard some stat that 43% of Google search results right now return a map. I mention that because there’s now a much – not just Google; anybody who has a business like Yandex, of course, the Google of Russia, they are much more able to justify financial investments in mapping assets, because I think it’s becoming clear, for example for location-based advertising, things like that, the connections are becoming clear.

[28:02] The second reason is what you were hinting at, Daniel, which is around the power of the consumer… And interestingly, in developing countries, especially in Asia – you know, there’s an interesting genesis story for Carmera, where I was still at my previous startup in 2013, and randomly that year… I don’t know if you guys remember will remember this, but there was a bunch of viral videos of meteors that year; they happened to be spotted in Russia. And these meteors happen very quickly; even if you see it as a human, it’s usually too quick to whip out your phone and fumble with it and get the camera going. And I learned that the only way they were captured is because some people were driving by with their dash cams on. And I was like “What’s a dash cam?” I’d never even heard of a dash cam back then.

Now, what I realized is in certain countries, like Russia, like China, like Korea, because of the additional risk of driving, and also maybe the lack of maturity in some of the insurance industries, dash cams are actually very common there, and they have been for a while, because of just personal security and being able to prove liabilities, and stuff like that. And of course, that’s right around the time when cameras being so cheap, connected, IoT, all the other buzzwords… So I think that’s also interesting, because these incredibly cheap, connected cameras everywhere, clipped to every moving thing, is not just a rich country thing. In fact, if anything, as I told you, we were seeing it in less developed countries, even before the U.S, because of that insurance need they had.

So I don’t think we mentioned this, but one of the ways that we do what we do at Carmera is, in addition to getting data from car cameras - for example, you can see some of the public work we’ve done with Toyota, where we’ve used data from Lexus production-grade cameras, to doing maps and change management… We have never wanted to rely solely on the automotive companies for our data, so we’ve also developed our own partnerships with commercial delivery fleets. These include one of the largest ones in the world, but also small and medium businesses who do pick-ups and drop-offs every day for storage, or installing signs, or whatever that might be… And they are using basically dash cams, sort of evolved versions of these dash cams that were first popped up in Asia years ago, that we provide to them… But they have our technology running in the device, [unintelligible 00:30:30.12] technology in the device, that also then passes it off to the cloud at a certain point… That is crowdsourcing; but for us, we’re actually using – we kind of coined this terms “pro-sourcing”, because we focus on professional fleets, as opposed to just any consumer… Because for us it’s more efficient.

These are delivery drivers who on average are driving 10 to 100 times more mileage than just maybe you would, Chris, going from home to work and back… But what we’re seeing is there’s kind of an ecosystem forming as well on the telematics and dash cam side, where there’s a bunch of companies who sell these devices, and maybe they have access to some of the data, or maybe it’s the delivery fleets themselves, some of these really large logistics companies; they’re able to collect all this data, but they don’t know what the heck to do with it… And that’s where we can come in and structure fairly low-grade, raw data into very high-grade, up to high-definition or medium-definition, depending on the need, data at scale. And having a mix of both consumers crowdsourcing for you and professional fleets crowdsourcing for you makes for a really good portfolio of sources.

One of the things that you said a little while ago - and it’s been kind of tickling my thinking ever since you said it - you mentioned earlier about something that I’ve heard in other conversations, in other venues as well, and I love your perspective on it… And that is you kind of mentioned the way that you guys are approaching it and the business bets that you’re making, and then you have Tesla and other people out there that are kind of doing a pure AI approach, and there’s a legitimate conversation going on within autonomy about how to approach that…

[32:11] I work for another company with interests in autonomy, and whether it’s ourselves internally, or other organizations, everyone’s talking about the different approaches… And I’d love it if you could lay out what that conversation looks like. Clearly, you have a bias in that you have bet your company on a particular set of ways to do it, but obviously, Tesla has that, as you said, pure AI… Can you define what that would mean from their context, and then maybe differentiate a little bit about the business bets that you’re making, and kind of lay the conversation out?

Yeah.

I’m not trying to pitch you against anyone, but I’m just trying to capture what the industry is talking about in that way.

Totally. I’ve seen a few different ways of framing this. Some people frame it as – they use a dichotomy of AI versus rule-based, or something like that. I also more recently saw it, which I kind of like, as someone using sort of nature vs. nurture… Maybe let’s start with that. We were talking about priors before, and that’s where I’ve seen – actually, the nature vs. nurture dichotomy I saw because we’re connected… So one of the several academic institutions that we have pretty close ties with is NYU here in New York. NYU is known for Yann LeCun and his work, but also Gary Marcus… And in some ways, they’re sometimes friendly at odds with each other, and I’ve seen them even debate this concept of nature vs. nurture for AI… And I might be morphing it a little bit simplistically or off what they mean in their arguments, but when it comes to what we think about, and what Elon’s thinking about when he’s saying what he’s saying about Tesla’s eschewing of things like maps is “Can an AI get to where it needs to get purely on just learning, and essentially nothing else? Or is there a need for there to be a certain–” I think Gary Marcus uses the term “innanteness”. Again, I’m kind of morphing it for this conversation a little bit… But in our case, that might be analogous to the use of priors. So I think a lot of these debates boil down to that.

Actually, yeah, we kind of have a dog in the fight, sure; we are a mapping company, and that’s used for priors, but actually, for me, I’ve always felt like what is actually gonna solve the problem both now, but where can you future-proof yourself, including on the business side? It’s really important to future-proof yourself, so that if and when certain trends materialize, you can sort of seamlessly ride that wave, as opposed to completely flipping your technical approaches that you’ve taken. This SD/HD to MD is a perfect example of that. Right now, I don’t care who you ask - especially for really high levels of autonomy, and also where the driver is truly not in the loop, and also in complicated environments, no one is able to do that without priors, and no one thinks that will be possible from a safety case, from regulatory and societal acceptance rate case for several years at least, if not more than that. The but is what if that changes?

The thing is, we always have to be humble, because these things change in a very nonlinear way, and our lizard brains still really struggle with nonlinearity in predicting things, because we just don’t know where we are in those [unintelligible 00:35:38.08] So that’s why we always exercise humility there, and kind of think about what-if scenarios… And I think the what-if scenario of allowing this AI to be more nurture than nature, so just purely whatever you expose it to, it learns, and it just gets better, and it needs less and less of what it was hardwired with from the beginning, I think it would be a great thing. For us, it would actually allow us to focus on higher-level problems, where you’re switching from certain problems on the lower rungs of the hierarchy…

[36:17] Again, I don’t mean to plug our blog, but another post we referred to in this last post was this thing we call “The mapping hierarchy of needs.” It’s sort of a take on the Maslow thing. Over time, there’s higher-order problems that the data that we create still are really important for. It’s just that it’s stuff like user experience, or compute efficiency, or economics… Whereas the first-order problem that everyone’s really trying to get over the bar with is safety. So that’s where, as I said right now, everyone really wants to use good priors for that.

But in the steady state, you could totally envision maps being more used for things like comforts, and monetization, and things like that. If you think about aviation or other industries, there’s certain datasets that were much more critical for safety, but are really now much more for comfort. Say for example turbulence, or weather data. I’m old enough to remember when we did worry about poor-taste jokes about TWA, and stuff, about their safety.

You and me both.

Remember those?

Yeah, I’m about the same age, I think, so I agree.

Yeah, yeah. Remember, we actually used to think about that when booking a flight to somewhere; we never do, because – do they still use weather data? Yes. But you don’t worry about it for crash safety, you worry about it more for “Am I gonna have a smooth ride and not spill my Martini that I ordered, with my Delta bucks?” I think you’ll see a similar progression, and I think the usefulness of the data that we inject into the AI - it’ll just change in nature… And that’s okay, that’s a good thing.

So we’ve talked a lot about driverless and automation, autonomous driving… But one of the things that you talk about both on your website and in that blog post about HD and MD maps is consumer maps… I thought it was interesting some of the things you were talking about in terms of possibly us in the future seeing enhanced functionality in consumer maps that could be driven by MD maps, where maybe in autonomous driving we’re able to use less fidelity, but we have this higher-fidelity that’s available for consumer maps. How do you see that evolving? …and maybe some of the AI capabilities that might be able to be built within consumer maps because of higher-fidelity data that becomes available.

[39:52] Yeah. I mean, that’s why we gave it this high falutin title of “The mapping singularity is near.” I do think that we’re already seeing those trends. Even if you look at – let’s take Apple Maps; they were way behind Google for a long time. But if you – I don’t know if you guys use Apple Maps. I try to constantly sample all the major ones, just to see what’s going on…

I bet your phone has so many map apps on it… You kind of have to try them all, right?

I recently did a bit of a purge, a little bit of a cleanse, but it’ll get back to where I was… But yeah, some of the natural language directions, for example, that you’re seeing from Apple Maps - and I think others now as well - that’s a good example of like… It’s kind of giving us some super-human machine-like qualities that we didn’t use to have with normal master humans… So it’s like instead of just “The next turn is right-hand turn on this street”, and maybe even giving you the amount of meters away it is, it’s being a lot more precise and allowing you, the human, to almost feel a little bit more like a machine, to feel more automated by saying “No, [unintelligible 00:41:00.21] sharp turn at the McDonald’s…” You know, really starting to get more granular…

Let’s just sort of play that out… Let’s just say a package delivered today. That driver, whether they’re working for Amazon, or for any other company - they might be using a standard, off-the-shelf mapping/navigation app, or they might be something… You know, in certain cases, like Amazon, they’re savvy enough to actually have their own routing and navigation capabilities on top of general consumer maps. So even today, that driver would greatly benefit from being guided by their maps app precisely to the right part of the curb, that’s already pre-optimized for safety; making sure it’s not a dangerous part - especially if it’s a city, or something - to pull over. Proximity, obviously, to the doorstep, or wherever they have to go. The probability of it being vacant, whether you have statistical data or other data from, say, the city - you should be able to incorporate that.

There’s a lot of things that could make that delivery driver superhuman, as opposed to just letting him wing it, which is what we’ve done for a long time. That’s kind of what we’re talking about. If you compare that to the inputs that an autonomous delivery vehicle would need, in that [unintelligible 00:42:24.13] future it’s not that different. It’s all those things I’ve just mentioned, plus - yes, the robot is gonna want more vector information on the map, on where precisely to be able to pull into the curb space… And for the human you don’t really need that, because they can handle that just fine, doing the steering themselves. But that’s what we mean by that MD-fidelity layer, where humans are now being empowered to be superhuman and almost more like machines, and then machines are basically just being able to make more human-like decisions, and just smoother, slicker decisions that a really experienced, attentive human would make in terms of pulling over in that drop-off space for that curb… That’s an example. As I said, we’re seeing into that now, it’s pretty cool.

I’m curious, as you’re much closer connected to this industry, where you see the trust levels of both companies and users of these technologies in terms of their capabilities. Because it’s one thing to enable capabilities, it’s a whole other thing to ensure that humans adoption is one thing, but also building trust and understanding how humans should and shouldn’t operate with these technologies.

[43:44] I’m thinking back to like – it’s probably a meme in your industry, but that one Office episode where the Garmin or whatever tells Michael Scott to drive into the lake, and he trusts the Garmin and drives into the lake, or whatever… And of course, that’s getting at the fact that “Hey, at that time people generally knew that there’s a lot of flaws in this, be careful what you think about it…” So how do you view people’s general trust in these sorts of mapping and navigation technologies, and how do you see that evolving as we move forward?

Yeah… I don’t wanna go off on a tangent, but there’s actually also – part of the origin story for Carmera, believe it or not, was a Curb Your Enthusiasm episode, which I can talk to you about later, but kind of related to…

Yeah, for sure. I’m sure that there’s all sorts of memes posted around your office, of various things…

Season eight, episode eight is called Car Periscope. Look it up.

Alright, we’ll link it in the show notes… [laughter]

Anyway, but to answer your question - yeah, it’s a big, big deal, and that’s why we’re very involved… Like, the trust is the biggest deal, actually. And we try to get really involved in that… For example in shows like this, and really just educating and being super-honest about where we are today, and where we’re going… We’re on the board of the main educational body of the AV industry. It’s called Partners for Automated Vehicle Education (PAVE) I think. So yeah, we’re big believes in that…

And look, I’m not actually a Tesla hater at all. In fact, I actually give them huge credit for whipping much of this industry into gear, including on the EV side… But we are very worried about lasting impacts to trust with stuff like FSD (the so-called Full Self-Driving). You’re seeing a lot of the incidents that are happening, and what’s it gonna take to set us back because of losing trust for training Tesla’s neural net.

So I think the best way to trust is all the things we’re doing, which is transparency and education, but ultimately, it’s making the technology boring. Again, going back to aviation… Aviation is incredibly, amazingly safe, and it’s boring, and no one things about those things really anymore… And it took time. It didn’t happen overnight. I don’t think autonomous driving needs to take as long to really industrialize at scale as aviation did from the Wright brothers. I don’t think it will take that long. But it’s definitely taking longer than 2019-2020 predictions everyone was saying back in CES 2015, of when we’re gonna have robot taxis everywhere.

First off, everyone knows by now it’s way harder than people thought. It’s the classic 90/90 problem; you’re 90% of the way there, but then you realize the last 10% is actually 90% more…

But I think the boring part is really key, because for example when Waymo won, they really have been the first to truly take the driver out in Phoenix… And you know, not just them, but other AV companies we’ve worked with or know really well, they always say that same thing. And by the way, my smoothest AV ride has been when I forgot that autonomous mode was engaged. This particular one was in downtown Detroit, and I got out and I told them “I am so impressed, because in the other AV rides I’d done, I had been thinking about it, I was remembering it, and I remember I kind of zoned out for a second.” So that, whether you’re thinking about level four, the mobility-as-a-service model, or you’re talking about kind of the more Super Cruise, autopilot on a highway - once the user really feels like they’re at the comfort level where either the whole ride is such that it’s kind of mundane and they don’t even have to think about the technology, or in the case of human assistants, where it’s more highway assists (like I said, like Cadillac Super Cruise), where they’re still confident that the system is really good at bringing them back in the loop, that is what it’s gonna take. And the good thing is those are parallel tracks that are happening. They just need to play out in the right way.

The problem with boring, using kind of a Tesla approach that people fear is it’s really hard to square boring with “Oh, but you still need to be in the loop.” It’s really, really hard to do that. It doesn’t work very well.

[48:21] I think you’ve called out a truth there that goes beyond just this use case that we’re talking about… It’s technology in general, and that is that when things become boring, it is really that point where you see acceptance, you see people move on in their thinking… Sorry, Apple, but none of us think about our iPhones, and for Google it’s just there… And I agree with you; I think that’s the secret, is when people just don’t care too much. It’s just part of the fabric of their life.

That kind of makes me think forward… While we have you here, if you could kind of finish up with telling us what you think we need going forward. If we abstract – and I’m using the term “map” loosely in this case, so it can be whatever you want it to be… But what kind of map needs to exist in the future to move us toward boring, to move us toward daily acceptance, to make life evolve in that way everyone’s comfortable with it? What needs to happen to get to that next level? What does that “map” look like for tomorrow, that doesn’t exist today?

Yeah, the rule of thumb we always use, and was given to me actually by… I believe it was someone who was associated with Google Maps in the very early days - it’s 100x. 100x on ubiquity and 100x on temporal density, so freshness. Basically, one order of magnitude isn’t enough on either of those dimensions from where we’ve been. it’s two orders of magnitude to get to this [unintelligible 00:49:46.15] steady state that we’ve all been talking about during this podcast. All the fidelity, and cost, and speed, and humans in the loop, out of the loop - everything ultimately is this massive optimization problem to get to that 100x.

Awesome. That’s a great perspective. I really appreciate the work – you can tell that you and your team put a lot of work into that blog post that we’ve been referring to, and sort of the clarity that it brings around some of these things… So I appreciate you being willing to put in time to that sort of communication, because I think it is very helpful. I encourage our listeners to check that out, and to check out all the things that your team is doing.

Thank you so much for taking time to chat with us today. It’s been a pleasure.

Yeah, thank you for having me. If I could just also plug, for folks – especially in communities like yours, whether it’s for maybe working with us one day, or even just riffing on some of these things… The reason we put them out there is because we get really interesting feedback back, and oftentimes we’ll publish a follow-up… So please do come to carmera.com, or hello@carmera.com, or LinkedIn, or Twitter. Let us know if you have thoughts on some of these topics. Also, if you go to carmera.com/join you can see some of the things that we tend to look for for team members.

I think one thing that we didn’t have time to cover, which is totally fine - we can save that for sometime else - what’s kind of cool about what we do is we use AI to make AI. We use AI to power AI. So everything we talked about in this episode was about the output of the data we’re injecting into AI… But for the other geeks who are interested, all the [unintelligible 00:51:35.09] we use to actually create those maps, that data, is pretty cool stuff, too. So check it out.

It sounds like we’re gonna have to have you back for another episode…

[laughs] It’d be my pleasure, yeah.

Yeah, for sure. I definitely hope that happens. Thank you so much. We’ll put some links to those you’ve mentioned in our show notes. Definitely check those out. Thank you again for joining us.

Thanks, guys. Take care.