I sat down with Trevor Blackwell to discuss investing in robotics startups. Trevor is the cofounder of Y Combinator, which is the world's best startup accelerator. He also founded Anybots, which is a pioneer in telepresence robotics. He's invested in dozens of robotics and hardware startups.
We discuss founding Anybots, why iteration speed matters, what will happen in self driving, vertical vs horizontal platforms, and the true cost of regulation.
Here is the video, with extracted times for bits below, along with the full transcript.
1:34 The thesis at the start of Anybots
3:27 Wheels vs legs for telepresence robots
6:50 Why did Anybots make legged robots
8:04 What do you think of Boston Dynamics Spot Mini
9:50 Vertical vs Horizontal applications in robotics
12:24 A proof of concept proves to everyone what is possible, and emboldens copy cats
13:37 What would have happened to Cruise if GM hadn't acquired them
16:15 What does it takes to launch self driving car service like Waymo or Cruise
18:04 Benchmarks for autonomous driving performance
19:48 Why it's better to have customers hungry for a solution
21:06 How to overcome the challenge of rule based behavior planning, like Tesla doing end to end ML
23:42 How early stage startups should build hardware by focusing on iteration speed and tight customer feedback loops
26:58 Vertical vs Horizontal applications and their impact on iteration speed, like building 3D printed houses
28:43 What should hardware companies get done during the YC batch. Ideally cheat with teleoperation if you can
31:17 Why Tesla's approach to getting data from real drivers is very effective path to full autonomy. You need to figure out a business model that isn't bleeding money while you're getting data.
33:47 What Elon Musk gets wrong about end to end autonomy
35:38 What robotics companies would you like to see applying to Y Combinator?
36:36 Drone delivery is working well where people are more desperate. Generally technology is developed where need is greatest, and then moves more broadly to other applications.
38:50 We get accustomed to high quality, like the speed of delivery. The standard can always improve. How many more iPhones would people buy if delivery were every 15 minutes.
40:19 Speedy delivery for auto parts is a clear application.
41:38 How telemedicine can be much more efficient
42:32 Regulations in medicine make developing products far more expensive.
43:15 Regulation is like a fog of war which dramatically lowers iteration speed
44:50 Inaction causes harm, including an example of the cost of $1 parts and the tradeoffs for human life.
46:06 What do you think of Marc Andreessen's "It's Time to Build". San Francisco doesn't feel like the future. We need more experiments.
Note: the transcript below was made by a robot. I was able to filter out filler words, but I'm keeping the rest, because sometimes it's hilarious.
Welcome to tango WC. I'm Yvonne Kerrigan. And today I bring you an interview with Trevor Blackwell. Who's the co founder of Y Combinator and the robotics startup, any bots. It's a really fun interview. We cover legs versus wheels, what it takes to win. And self-driving the most important thing when building a hardware startup, and also the kinds of startups he would like to see apply to Y Combinator.
So without further ado, I bring you Trevor Blackwell. All right, today, I'm joined by. Trevor Blackwell, one of the co founders of Y Combinator and the founder of any bots. Thank you so much for taking the time, Trevor. My pleasure. So today I was hoping to dive in into something near and dear to both of our hearts on the robotic side.
And you have a few vantage points here. So you, you are roboticists like a very technical founder of a robotics company. And also as an investor you've, at, through Y Combinator backed. Probably a few dozen robotics companies at this point and seeing different stages of development. And so there's a lot of different, interesting angles on that that I would love to try to Peck at, to try to understand the problem space.
and so I'd love to start by going back to any bots, that was founded in, early two thousands. Is that right? Yeah. 2001 even, I, I, I've been at Yahoo for a few years and I left. I left Yahoo on a Friday and I started any bites on Monday. in retrospect, I should have taken a couple of months and, you know, hung out by the pool or something before, getting back into full time foundership where you just interested in robotics and you thought this telepresence, you know, we wanna, we want to build robots that are like people or what, what was the thesis at the start?
And how did that evolve? You know, I wasn't thinking about telepresence much at the beginning. I was thinking let's, let's build. Autonomous robots. and you know, that's really hard. but, but, but, but sort of tele operated as a good, stepping stone to it, and it at least gives you, you know, lets you build something that you can see working.
You can make sure the hardware is capable of doing the thing you want to do with, you know, with the human controlling it. and so, so that's a logical first step. toward building a, an, you know, an autonomous robot. so I built these, these, Tel operated robots, and, actually found it was pretty useful sometimes, you know, being able to remote into the office and drive a robot around and pester people instead of trying to get them on Skype.
it's interesting time. And right now it's, late may and, a lot of people have been in lockdown. And so the. The office is depopulated. And at the time where you expect a lot of people to be remote, there's nowhere to remote into. And so you're seeing just a lot, a lot of video calls. but the idea of some remote presence is now is going to be like permanently on top of everyone's mind.
And so it will be interesting to see how this evolves over time, because there are some cases where, you know, a phone call is telepresence, right? Where you can talk to someone remotely and you're not there. There's also a case where you wants to be around where a bunch of other people are. And I think that that use case of, like a mobile eye, not just one, you know, a couple to my laptop here still makes a lot of sense.
There's so many things you need to solve though, to get that working really well. I want to ask about the. Sort of the form factor of the mobility. Cause you've worked on both, wheeled robots and legged robots. And one of my critiques of robotics that are remote, like this is that, you know, stairs are hard.
And so it does make sense to go after legged robots, but it's so much harder technically to get that done. And maybe now. we're sophisticated enough, but, you've built both. And so part of that might've been experimenting. So can you talk through a little bit about, the experience of building each and sort of the why behind them?
Well, legs or legs are enormously hard. they're, very failure prone. Like if any actuator fails it's robots probably going down. they're, loud. They, they, you know, rattled the floor. The people working below you will hate you. I I've heard, this was one of the reasons why, the MIT and a lab on the fifth floor of the, of the computer science building, I always had trouble keeping their funding.
and, And so it's, it's basically a bad deal for, for almost every case, you know, except for some very extreme, you know, exploration or firefighting scenarios. so I think wheels are the way to go. that's a practical thing to do. there's this, you know, with, with telepresence robots, you know, We figured stairs, wouldn't be a, a big issue there.
because if you're actually using these things at scale, you're going to need several per office building anyway, probably several per floor. if a good fraction of the people are remote. And so it makes more sense to just log into a robot that's on the right floor rather than trying to navigate it up and downstairs.
Yeah. I think depending on the application, I mean, this is kind of the point where it's like, you have a user that has a problem in order to solve it. You need to build a product in order to do that too. You have to understand the constraints about problem and then fit a solution to that. So if you are trying to deliver, Chipola to my house, there's probably some stairs, cause you're going outside, you know, maybe that you need to go up and down curves, and, or maybe you go all the way and deliver it to a doorstep.
and so as you get to the delivery use cases, you start to see more hairy outdoor, unknown environments that are very tricky. and you know, like not every sidewalk has a wheelchair access ramp. So it's one of these things. Where can you assume that your have the ability to do this? Or maybe you can strain the problem?
Like, well, the person just has to go out that wearables going to stay in the streets and like, this is the only place it's going to stay. and he brought up military and firefighting where. Obviously you might have like the whole house coming down in a firefighting situation or who knows what's going on in the military as far as being able to go up and downstairs.
I know in my experience that I robot with the pack bot, where they had these flippers and then treads, that was in order to write the robot and also to be able to go up and down stairs. And so it was like a, in that context, a very basic requirement is you need to be able to do it, but in that case, they use these treads that.
as far as like noise and like the smooth design or like antithetical to what would fit in an office. So that mobility platform, it like is determined by the product experience. and so, so when you made like a robots had any bots, is that just an experiment? So you, like, you want it to go after this?
Or what was the, or did you, is that how you learned that that's not the case that you want to go after? yeah, I would say it was, it was. Learning experience. I mean, it's a, it's a sort of grand challenge problem. and so that was one of the reasons I thought it was a good place to start is, that if I had a, you know, machine learning programmable system that could get, you know, two legged walking, working, then it was probably good for all kinds of things.
so you know, one of the problems with, robotics is there aren't any benchmarks. You know, someone could build a robot that does something and like you look at it and is that impressive? It's hard to tell, you know, how much has that environment been cooked for that particular robot to, to work well in?
so, so, so at any rate, you know, walking on two legs is at least a thing that enough people have tried that we kind of know how hard it is and something that succeeds at that must. It must be good for something. So, in the news, you see a lot about Boston dynamics, which, I've been tracking the company for a long time.
I worked with them when I was at I robot in 2007. so they've been at least a 15, 20 year old company. and, been focusing on this mobility platform for a very, very long time. And I'd love to actually just get your read on, Spot mini. It's the name of the product right there. And it's, it's four legs, not two.
So there's a difference there. And then also it's a developer platform, which is interesting. And that's well, I have, I might have some opinions on this, but I love just your take. Like how, like on, technically on how that product is doing and also the business side of this, of what do you, what do you predict happens with that platform?
Well, technically it's, it's really impressive. you know, it seems to work really well from the videos I've seen. and, they seem to be able to mass produce them at not ridiculous costs. so that's great. I mean, that's, that's really the first thing like that. Yeah. and, Do you know how many, they have no idea what their business case is like, why, who needs that?
You know, people are like all the, all the sort of videos and user stories I've seen. Like, I don't, I don't really buy it. You know, and so enforcing social distancing and parks. Well, I mean the little wheels, do you find for that? Yeah, there's I saw in one of our demo videos that carrying a cinder block and so it's like, it's yeah, it can carry 20, 30 pounds.
I think. So as you carry one center block at a time and all the truck drivers out there, like. You know, like forehead slaps, like, and who puts a sort of like on like, how does this work? so actually this touches on an interesting point, which is, vertical versus horizontal. and actually I've seen a few different companies that are taking on very difficult technical challenges and it could be in computer vision or in natural language processing or other areas.
And they are going after horizontal because. they think there's enough of a developer ecosystem to build on that. And if you had vertical actually patella presence in the vertical category, especially office base telepresence, which would be different than. No pipe inspection for oil and gas. that kind of, even for the exact technical use case of telepresence, you'd have different robots that come from that different capabilities.
and so, I don't know if you have any system of thinking about this on when to go vertical or when to go horizontal because you have a developer platform at the spot. Many that is obviously incredibly capable and really, really impressive what it could do, but being a platform, they don't have to pick the vertical application, which means they.
They don't need the specialization to understand what the use cases are, but then maybe those cases don't exist. So it's like, where's the market it's really difficult to know as a startup, whether that platform is there. So how do you think people should decide to go after one or the other? It's that's really hard.
the problem with going vertical of course, is that. You need to have this enormous range of, of capabilities within your company from the very deep technical stuff to the sort of consumer focused marketing sales, that's really hard. the problem with not going vertical, is that, you know, you make some platform and then you sort of put it out in the world and hope someone figures something cool to do with it.
and you know, that might happen or it might not. And then if it does happen, you have to worry about them, you know, shopping around for a cheaper version of that platform. so it's, it's hard to make the case, for, for a sort of horizontal platform play in, in, in robotics, hard to make the business case that you're going to end up with a longterm monopoly.
Cause, cause it's not clear what the lock-in is. You know, there are already starting spot many imitations in China. You know, that, that, I don't know, it looks like their videos work too hard to tell. but, but you know, sooner or later that technology is going to become copied and, and, and widely available.
And then, you know, what's what do you get for having spent 15 years developing it? Yeah. Yeah. I mean, there's a. The proof of concept that something is even possible is actually very, very valuable. And so in terms of imitators, when you know that, something is doable, you can make the bats that your team could figure it out.
Whereas in a lot of cases, you don't even know if it's possible. Like, you know, so you saw in computer vision in the last few years, this unlocking of a huge number of capabilities so that, you know, a computer can reliably find a pedestrian or a cyclist and a video feed. And so every single. Self driving car company knows that.
So they they've seen these demonstrations of it. So they know while we could probably get the reliability up high enough that the system would work really well. Whereas if you're developing that without knowing if it's possible, now you have some other applications, like, let's just invent something.
So can you replace a card dealer at a casino with a robot? So while I'm not sure I've ever seen a robot successfully shuffle a deck of cards actually sounds pretty hard. Of course there is the automatic shufflers, but it's like. Hands and dealing with sounds maybe a dedicated hardware could do it, but you know, it's one of those things where like, I don't know if that specific possibility is, is the capability as possible.
And so, there's a lot of risk in doing it, but once that's demonstrated, then the competitor at competitive landscape gets much more dangerous for others to do just the, if you're building the platform to be able to replicate just that platform capability. What do you think would have happened to Cruz?
Had they not been acquired. it's, it's so hard to know. you know, I I'd like to think they would move a little faster, but had, you know, it. I'm sure they get some benefit from GM and being able to integrate really deeply with the car. cause a lot of the earliest systems were basically just retrofits on the cars.
I mean, Cruise's first thing, they had this giant actuator, you know, pushing the brake pedal and clamped onto the steering wheel and there's like duct tape everywhere. and yeah. You know, so if even if you had perfect software that could identify every other car perfectly and you know, there's still a lot of things that can go wrong with that car.
So to get to the level of reliability that you need, which is, you know, a few crashes per billion miles. That's an extremely high level of reliability. And so, you know, you have to think through all these things that happen, like, you know, certainly you got to deal with tire failures. You've got to deal with the engine shutting down.
And I mean, there's hundreds of things that can go wrong over a billion miles, and that the software has got to do something not catastrophic in all those cases. so I, I think. Ultimately, you know, self-driving does have to be integrated with the vehicle company. You can't just bolt it on. Yeah. So ultimately I think they did the right thing, but, maybe I wish they'd stayed independent a little longer.
Yeah. You can say the, the need for a large amount of cash to do the deep integration that they need. that that's inevitable. So that could have been. You stay independent and raise $5 billion from SoftBank or something. And, then you can see the development that way. and it's also interesting in terms of it's many, many billions.
I mean, if you're talking about, if you're talking about being able to claim. That you've, you only have a few accidents per billion miles. You got to drive several billion miles, you know, on the last iteration of the software, you know, the ones before that don't count. and, so that's a lot of, a lot of billions of miles by the time you've, you've really proven it.
and, you know, miles costs 50 cents, according to the IRS, you know, profit is probably what they cost if you're running a fleet. so. Yeah. Some experience looking at those numbers. It's many billions of dollars. Yeah. It's, it's very, very expensive. And, my intuition is that actually both Waymo and GM are fairly conservative and probably could have launched in a limited area.
I mean, when you talk about launch, it's always more nuanced than that. It's not, it's like as a daytime, nighttime good weather, bad weather, like San Francisco versus the peninsula. These are all very different areas and where you want to launch. You know, Phoenix has lanes that allow you to stay in your lane in a way that just San Francisco doesn't, in terms of construction sites and cyclists and narrow roads, parked cars, double parked cars, it's so much more aggressive and where you need to go.
and so it was such a power move for cruise testing in San Francisco. I mean, that's the, that'd be the, one of the artists places in America anyway. Yeah. I mean, they they've written about this and I love it. the idea that. if you're driving a mile, cause miles are expensive and they're relatively, equally expensive in any area.
So if you're gonna go out to Phoenix or leaning out in the middle of nowhere, freeway driving, or, San Francisco, you want that marginal mile to have a faster learning rate where you encounter more weird things. And so. You know, San Francisco is full of weirdos, but it's also full of, really weird driving scenarios.
Like you have immunity is above ground. You have all these cyclists pedestrians, of course, it's a very difficult environment. and so the thinking, which I think is very correct is while they test their, then they, they're going to make faster progress. and so I do wonder if you were to put their vehicles in areas where Waymo has been, how the relative performance would be.
Cause they're almost not. You know, you don't get to see the side by side comparison because they're operating in different cities. so a guy, and to your point before, it's like, who's the head, like, it's kinda hard to tell, like, what is the benchmark here? You know? Cause the stats get pretty. Complicated pretty fast.
You know, the main benchmark you hear about is accident rate, but a, and or disconnect rate at any rate. but you don't, you don't hear about how nice it is to drive in them. and at least rumors I heard from the early Waymo experience was that it was really annoying to drive in it because it went, you know, exactly the speed limit and stopped fully at every stop sign.
And it was like being with the most timid driver ever. Yeah, hopefully I don't have to pay attention to the driving. That's another dimension I've heard where it's super interesting at the start and eventually in a, in a good way it gets boring or like, yeah. You know, I trust the machines and I was slowly like, some, you know, geriatric driver, like just, you know, gingerly driving along.
It's exactly what you wanted the robot to be super boring. Like you don't have to pay attention to it. And so that, that experience, I mean, I've told this to my wife who is, very reticent to do anything in self-driving. It turns off all the autonomy of our cars. Like doesn't like any of it. and what I was trying to explain to her is like, Yeah.
Once it's risky, if you don't trust it, that's super nervous situation. Like you don't know what's gonna happen, but then if it gets to the point where everyone trusts, it that'll be a totally different psychology when you're in it. When you, when you do look at the accident rate and you're in it for a while, and it seems normal, they just stopped thinking about it.
And so it's really hard to predict the, like the user reaction to. The autonomy, because if you look at it autonomy today and just project that out, it's like, that's not accurate because you know, you stop too fast or go too slowly or too, you know, not aggressively enough on some unprotected left turn. but if you have, If you need to predict what the experience would be like after it's better and able to launch, which is going to be boring.
And you got to figure out who the early adopters are like. even though I'm an early adopter of tech generally, like I'm a pretty good driver and I have a. Model asks and it's kind of fun to drive. so, why don't, I don't need it. but you know, people who've, too young to drive, who've lost their license, who have some disability that makes it very hard to drive.
There's a few different companies there. One of them that just reminds me is voyage. So they're operating in a retirement community and that means that you have people that do struggle with mobility generally. and it's a relatively constrained environment. So it's touching on a few points here because it's.
easier environment to which means you could launch faster. It's a very hungry community for this kind of ability. And then they can scale with that to more aggressive areas. So it's interesting seeing these different companies like Waymo, generalists in a relatively easy area, voyage kind of narrow area in an area that could launch fast and then cruise just going as hard as they can for the hardest area, because they think the general case matters a lot.
So it's not the first city. It's the. The speed to your 20th or 30th city when it comes to self-driving and this is true for other areas, perception works pretty well. The computer vision is now very capable and so you can identify things really, really well. So, you know, here's a pedestrian, there's a cyclist, but when it comes to the behavior planning, it's a lot of rules.
And that means it's like brittle in terms of how things are moving around. And this is true for manipulation as well, where. You might be able to detect the object in his pose, then deciding how to pick it up is often a lot of rules or these, a lot of constraints to make that work well. and so what do you think is going to happen there?
I mean like there's the Tesla end to end self-driving with behavior planning, baked into that one neural net, but then Waymo and cruise in my understanding are both very rules-based. So when it comes to, for example, your example for how fast something slows down that is. It's governed by a magic number in the code that then they can tune to make it more human.
But there's still that bit of code that says here's the rule and how you approach a stop line and how you should stop it. so what's your take on this sort of more difficult behavioral planning problem in the autonomy side? I, I think that's one of the great questions about robotics generally is how do you.
I mean, how do you get the right behavior? do you, I mean, you can, you can try to program it explicitly with, you know, feedback loops in code, or you can try and use some kind of learning. but in the learning, there has to be a, you know, a reward function or, you know, some kind of there's you, you, you, you have to specify what you want and some other way.
and, and for something like, you know, how to approach a stop sign, it's not just like. Did you cause an accident or not? You know, there's lots of ways of doing that badly, like so badly that it'll terrify people. but, but it's hard to quantify that. So I think, you know, that's an example where it's actually easier to specify, you know, sort of a normal behavior as, as a, you know, as a function of speed over time and distance, than, you know, a general reward function for how good a particular stop was.
Hmm. But I think there needs to be both. I mean, some things, like vision problems seem to work much better to have a huge data, set, a huge label data set and train on it. Right. and, control problems seem to work. The only good way we have of doing that now is to, you know, explicitly have control theory.
People think about how the feedback loops work. No, one's gotten very good results, trying to do any kind of a tabula rasa learning for that. So I wanted to ask about sort of the timeline for development here for startups in the earliest stage, because when you're there, there's like building an engineering prototype that something starts to work.
And then there is a starting to design for manufacturability and that end use case and like the mass production. And so this gets more and more expensive and you really want to make sure you have a good use case, where you, you really understand like the product and the user experience. And so when it comes to patterns you've seen and the different kinds of companies, have you noticed anything as far as like what works well here?
Because it's so expensive to develop hardware and making many of something is also expensive and other ways where, you know, the manufacturability and reliability of it has to be far, far higher, but the earliest age companies. Like the, the seed round, that a software company might have, would get burned up much, much faster.
The capital requirements are higher. And so maybe the answer is sprayed a lot more money. but then that's kind of begging the question cause it's like, well, what kind of validation do you need to be able to be able to do that? and so I just want to ask him the general space of the cost. Of getting to something at higher scale and the validity you need or the validation you need to be able to raise that money.
What patterns you might've seen here that you think are effective? Well, even if you magically have a lot of money, it's still very slow to iterate on hardware. unless you basically make that the number one priority every day is to be able to iterate, You know, the, the best companies I've seen have been able to consistently do like a weekly iteration cycle where they change something.
Like they make some meaningful design change, they whip up some hardware in their lab in the back. They show it to a customer and it gets to use it and they act, and they get feedback. And they, you know, if he can turn that around in a week, which is an amazing feat on hardware. I mean, in software, you turn that around and afternoon.
But, you know, if you really focus on it and you're able to turn that around in a week, that that's, that's, that's, that's great. And that that's seems pretty correlated with success, that might imply not a consumer focus, right. In terms of just like the cost point for our consumer being far lower and the scale being far higher.
And so is this, is this like a thing you've seen with Raleigh? I think you can do it on consumer stuff too. You just have to have some captive consumers. I mean, you can't, you can't, it doesn't work to do like a Kickstarter and then ship, you know, a hundred of them all over the world randomly and hope, you know, a few of them actually get unboxed and, and played with and you get some feedback.
it's you gotta, yeah, you gotta have some. Friendly consumers who are going to try out your slightly mass broken thing every week. Hmm. Yeah. The iteration speed is especially for industrial, you know, for, for, for, for industrial things then. Yeah. You certainly need to iterate, you know, in a real setting that also brings up that platform versus vertical side to it where, I've seen some companies that are going after.
Very broadly, some manufacturing robots that have new capabilities and that might be considered a vertical, but then are they making the end product? A good example of this recently was a three D printed house company that I saw. And, you know, it might be far, far cheaper to build out the house if it's three D printed, but then your buyer is especially in construction that very traditional in terms of how deals get done as far as, you know, people meeting people.
But then also. Relatively slow absorption of new technology. And so they, the answer might be, you have to go and build the houses yourself, and then you can iterate very quickly on the houses that you build. the answer might also be that, you know, you just have to break into that industry and sell to it.
Part of what you should do in the beginning is design your business model around being able to iterate, you know, the fastest you can. and. Ah, that's hard with a platform. you know, maybe you can make it work, but for something like a, three D printed house, I think what you'd want to do is you'd want to make a house and then have someone move into it.
And then. Well, you know, watch them like, notice that you forgot to have a door into the kitchen or something, and then you make an X one next to it. You just apologize to them and make another one next door and keep going. If you're making a platform, it's because you don't even know what the real application is.
You know, something like something like spot mini, you know, who knows, you know, I assume they're hoping that someone will. Figure that out for them. If it turns out that there's some change that they need to make, you know, it would have been nice if they could have designed that in five years ago when they were designing the hardware.
So for YC companies that are doing hardware, what do you hope they get done during that three months? Because. I mean, normally if it's a SAS company, I'm just gonna, draw caricature where you have some B2B SAS, and you have a certain number of leads and you want to dial up that lead source, get better at closing excellent customer support.
So you have a low churn, a excellent product as well. And you make that number go up into the right. And then when you're fundraising, it's like a debut and it's like, look at this amazing graph. That's every everything I just said as much harder with hardware. So in robotics is like a autonomy layer on top of the hardware that makes it even harder.
and so, for all the hardware companies coming into YC, like how should they spend their time? Like, what have you seen work really? Well? I mean, almost nothing works, works very well. you know, I mean, w what can work is if you're going for an application, like self-driving that everyone agrees, you know, if you can make that technology work, it'll be great.
but I think it's, if you're going for an application that, you know, is just your hunch, that this is something people would want. it's very hard to prove that out, you know, before spending a few years and millions of dollars, you know, building something, you know, cheating with teleoperation for example, Is is, is a good start.
So if you're, if your product is going to be like some robot arm that does something, you know, fold someone's laundry or something like that, you should build a tele operated version and put it in someone's house and have a guy sitting outside in a van with the controls and fold the laundry for them.
and see if that's something people actually want make my guess, is it probably isn't like folding laundry. Isn't anyone's. Biggest problem. yeah, I think I saw a demo from Willow garage a few years ago where, they debuted a robot that was, you know, it has these big shoulders and these pinchers.
Yeah. and they cleaned the whole house. It was amazing. and so. Only at the end, do you learn that was just tele operator. It probably took like 400 hours, right. It was actually moving pretty fast. And so tele operator was the key there. It wasn't a autonomous and I think their point was like, the platform is fine.
Like, it doesn't matter that, you know, you don't have as much mobility in the shoulders or at the hands can only carry so much or that they're just pincher. It's not, you know, fingers. they're able to get everything done. and so that's like a. It's like a call to arms for the software folks, like the perception planning.
Yeah. To really get it together and solve this problem and relate it to that. I think an interesting corollary to your point about teleoperation is training data. So, you going back to self-driving, I'm very. I mean, I'm just bullish on the industry overall, but Tesla's approach, means that they have these human examples.
It's not tele operated, it's, operated by a human in place of the autonomy, to get you the training data to solve the problem. And so I think teleoperation has that pattern where that robot folding the laundry, however much you have somebody, a human doing that. It's all data for your robot to do it eventually.
And so there's this nice evolution from. Yeah. And building the platform you could use with a human on top, you get the data and there's more and more capability about what the robot could do it on its own. It's like a path to full autonomy. So I think a lot of things, like that that's the right way to start as some kind of tell operation or, you know, or, or like, a ride along system that's, that's observing a human doing it somehow.
And a problem is it's, it's been, it's rarely economically viable to do that as a business on its own. you know, I found that with telepresence robots that, you know, it, it in principle, You know, you can save some labor costs by using people in lower cost areas. So you can, you can hire people in Tennessee and they can work on an office in California and maybe you got a 50% labor savings, you know, but that's not, that's not going to pay for the whole rest of the system.
You got to figure out a business model that at least doesn't lose massive amounts of money while you're collecting all this data. And so, you know, the Tesla thing. I'm sure they're spending a lot on putting cameras on all these cars that people aren't, you know, that don't have the self driving feature.
they, they have a positive unit economics on the car itself. and even when you look at the operating cost of having data piped back to Tesla, that can't be so expensive. And you compare that to the cost of paying a driver in a car, which is the alternative. right. So that they're not paying the driver is the key.
so they have margin in the car itself and, there's no such margin. If I were to say, take a Lyft. Car and add a whole bunch of expensive sensors on top of it, which are going to be just capital costs out of the way. And then expensive operating costs on having a fleet go day in, day out, and then also the data pro cost and everything you need to be able to do as a Tesla's in a very good position, just in the, their ability to not pay for a lot of that part, to make it worthwhile.
It also means they don't have very expensive sensors because you know, the car would have been much more expensive had it had LIDAR on it. And so their autonomy problem is harder because they don't have it. I actually remember seeing a video from Elon Musk talking, I think it was with Rogen, maybe, maybe Lex on, you know, some of these awesome podcasters I'm asking what the hard part is.
And he said the localizing, every object in your environment, like all the other cars and the bikes, that's really hard. And then the behavior is easy. And I, this to me was as much as I love Tesla's approach the biggest red flag, of their robotics capabilities, because. They don't have LIDAR. And so they've like structurally made it so that they don't know the exact position of where everything else is.
And like, that's a it's. So it's easy for everyone else that has LIDAR to know exactly how far in the velocity of these other vehicles that is. and everyone else in the industry thinks behavior planning is hard where like that's the hardest part to do, knowing where everything is. That's relatively easy, knowing what to do, given where everything is.
That's hard. So it's a really interesting take on this because like, of course, when you have cameras and radar, they're able to, like it's harder for them to localize on each little bit. And so I I'm worried about their self-driving as a result of that. on the other hand, the amount of data they're going to get is so much larger that's.
If you're going to train the system end to end to include that behavioral planning, I I'm confident that they can get there. It just going to require a hundred times more data. So while it might be cheaper for them to acquire, it's also going to be. like they need a lot more of it to make this work, which is, I think the general consensus that these vision systems can work, but it's a lot of data required, you know, a hundred billion miles, something on the order like that in terms of companies that you'd like to see, start applying in a robotic space to YC, like some of the applications you'd like to see there.
like, is there, is there any application you'd love to see? I think delivery is going to be one of the biggest ones. and there's so many different kinds. you know, there's. Multiple kinds of drone delivery, like fixed wing and quad-copter yard, you know, entirely different kinds of technical problems.
and, the sort of rolling food delivery robots, I think are going to be big. you know, efficiency is there, like if you can build the, the amount of money you can save by having a little thing, you know, this big moving around, To deliver food instead of a car, like a 3000 pound car with a human in it.
is it, the cost savings are so big that if you can make that work, that'll be great. and you know, the, the drone delivery stuff is working pretty well in a few places where they don't have cost constraints. Like zip line is operating in Rwanda and doing lots and lots of deliveries. and. No, they're the alternative is like, you just can't get vehicles through at all on some of the muddy roads and the rainy season.
so, you know, they're delivering life saving medical stuff, you know, that, that couldn't be delivered otherwise. So, you know, they can, they can absorb the initial high operating costs. so I think he got to find a place like that. So, you know, you don't want to start with delivering big Macs, by, by quad-copter.
First of all, because any self-respecting dog is going to be able to get it. And second of all, it's just like, how much are you saving? Like a dollar or two. and so even just the logistics combination of delivery for high value things, where, where people really value getting it fast, it goes back to your earlier point about finding a user that is, very interested in this solution so that they can work with you more on developing your solution.
It's also really common. And it's always the case that for every particular, like fairly broad application, like if, if your goal is to eventually be able to deliver lunch to every house in the country, you know, you still got to start with something where, where the, where the value is a lot higher and the tolerance for mistakes is higher.
Yeah, this is true broadly in technology. I think because you have, for example, plastic surgery, where people are very motivated to not look their age, and that has brought applications in other areas like, you know, reconstructive surgery for a burn victim or things like that. And people who've been in accidents.
and you have, a lot of things that are really going to be like this, including, some neural deficits you trying to overcome. And that will be how we get. Brain interfaces and the computers, you have these narrow applications that people care greatly about, but then the technology enables you to do so much more, more, much more broadly.
And so eventually then delivering the big Mac is worth it. even though the, it wouldn't be the first thing to attack today. And so delivery, it's also interesting because we get so accustomed to how good things are. And so if you think back to e-commerce 10 years ago, like how fast you expected to get something.
And now if you don't get it the next day and a covert aside like a day or two, seems like a long time. And I, I expect that to be true for this pattern will continue. So you'll have, at some point in the future, you can imagine the moment I think about something it's in my hand or very close to that, like it's wherever the closest version of that is like it can come and be in my hand right away.
and there's so much to do, to be able to enable that to cause this just gets harder and harder and harder to make these systems profitable and viable, technically capable enough to be able to do it. And so I think autonomous delivery, just in terms of, you know, most commerce being in this category of a consumer says they want something and then they have it.
there's a, there's a lot of fruit there, but it's really hard technically to be able to get every last detail working. but it's also, there's a lot large prize. The other end. And, you know, you can prototype that in one neighborhood, you know, maybe even with human runners, but you'd like to find out like how many more iPhones would people in this zip code buy.
If we could get them, you know, whatever they clicked on in 15 minutes, or maybe it's negative. Like maybe people buy things in advance because they know if they break their phone, it'll take a long time to get a new one. so. I think it would be worth going and quantifying those kinds of things. One of the applications I think is exciting is in, autoparts, there's this thing where you put a car up on a S on a stand and you look under it and you think, and you realize, Oh, it needs a, something or other from the warehouse.
and currently it takes like an hour for a courier to bring that during which time either you've got your hoist in use for an hour where you put the car back down and bring in another car, and. So there's an example where if you could get something from a warehouse in five minutes, ought to be possible, you know, to get five miles in five minutes, by air, you know, it changes the equation dramatically.
There's enough of these. There's just enough of these, high value use cases that I think, you know, delivery, you can be con confined a big enough niche to start there. Yeah. And that requires you understand the space too. So when it comes to. Construction projects and, you know, different work sites where people are trying to get something done where, you know, the example of like the Bay for, maintaining a car, that's like a very sensitive part of the business where if you can get more cars through, then the whole operation becomes a, that much more efficient.
and so I, how many things are in this category where if I just had what I wanted quickly, the business would be that much more productive. let me think about this recently in a slightly different sense, just in terms of telemedicine. Where, during lockdown, an appointment has to be remote. but then that highlights how much time is wasted, because like, in terms of, what it takes to see a doctor, and, I haven't seen this yet, but even breaking that even more to be asynchronous so that a doctor can really see all the requests that come in and then maybe schedule a call with some, but then also just have a reply.
like if everything needed to be in Slack, if somebody invented email, that'd be very exciting. And so today we have worse than Slack. It's like everyone needs to be in the exact same room to see your doctor. And if it could only be digitized and then asynchronous where appropriate, there's so much more can get done there.
and so, yeah, I think generally there's, there's so much more technology if you just think about the industries and how things work today, and, understand those verticals and then build an application to, to meet that demand medical industry is hard. You know, I explored telepresence, for, for a medical diagnosis.
Boy, the regulations just make it a very painful, you know, there there've been a bunch of emergency use authorizations for, various kinds of telemedicine, during the pandemic that may help people get comfort with the fact that it works perfectly well, diagnose a lot of things over the phone and, you know, cause people can take their own temperature and for a lot of it, like for an infectious disease type stuff, I think it's very rare that there's some complicated thing that only the doctor can do with their own fingers.
Right. Yeah. You know, it's like temperature, blood tests, all do all that at home. Yeah. I think it's hard to appreciate when you're outside of this, how much we were talking about iteration speed before. And if you add a regulatory burden on top of that, and it's not, it's not a burden because if it were just like a weight to lift, I think that would be easier.
It's like a fog like this fog of war and the cost of feeling out in that space. lowers your iteration speed and makes it very much harder to have a successful product here. Because from the regulator's point of view, they think, Oh, you know, it takes like a few years to develop a product. So if we add a another year onto that, we're only adding like 30% or something like that.
But no, it's not because it's it's it meant like originally during that three years, you could iterate a hundred times, many hundred times. And now you can only iterate once over this, over this four year cycle, with, with, you know, all the way to end user testing. like back when people could just sell patent medicines out of their, out of their, you know, wagon at the, at the, at the fair, you know, they could, they could try new ingredients every week.
Certainly learned something. I mean, that's the hard part it's like, I mean, it goes back to, automotive as another example too, because. when something is so on the nose about health, then we become very cautious. Like this precautionary principle comes into play. It's like, Oh, we don't want to do any harm.
but everything causes harm. tens of thousands of people die a year in car accidents. And so it causes harm. It action causes harm. And so you're making a choice one way or the other. and when you break it down to it's, it's incredible. So I remember discussing with some automotive friends about.
So the way a break works is they have a disc and it's squeezed by a caliper. And then the car slows down, right? Like a disc brake and a car. And at caliber, you're making four per car and you make millions of cars. So you have, you know, tens of millions, hundreds of millions of these devices. And so if you're a car company and you spend 10 cents more or less, that is millions of dollars more you can make or not.
And so the cost of these calipers is like a dollar to three. And the $3 caliper is far, far better at stopping and you will stop the car faster and you will get in fewer accidents. And so a car companies are accustomed to this, where they have an engineering question where they know the capability and they have to map that to, something that people really, really care about, like the chance of getting in an accident.
And so, there's, just highlights. These decisions are made constantly. whether or not we, we are all aware of them. you have to have to understand like every choice, including inaction is, which is a vote for the status quo, essentially. Like we want to keep this going for longer. but that, that iteration be a long time.
Yeah. I mean this, this whole, w there's been a call to action recently. I actually love, love to hear you take on this. So did you see the Mark Andreessen it's time to build, that was making his rounds? Yeah, I mean, I can predict what you say, but what was your take on that in terms of why we're here or how to get out of this funk in terms of how fast we can make progress?
He's totally right. We haven't been building the things we need. and it's embarrassing seeing, like, you know, I remember in the nineties when I lived in Boston and came out to San Francisco, like, it felt like the future here and now it doesn't seem like the future. Compared to Taipei or, or Singapore, it feels kind of run down and stagnant.
and, yeah, you gotta fix that. And you know, like the what's frustrating is seeing these basic things like, like building houses, you know, you figured out how to do really efficiently on time ago, you know, that kind of peaked. Certainly by the eighties, you know, you could, you could crank out lots of housing and we can't do it now anymore.
And it's all regulatory burden. So we need more experiments. I mean, like it's fair enough to worry that, Oh, if we just took away all the regulations, there'd be terrible consequences. You know, people would build shantytowns and, and buildings would fall over. Like there would be problems, but. We've gotten way too far in the other direction.
And so we've got to try some experiments of saying, well, let's, let's build a small city here and let this company just do it as they best they can figure out and see what happens. And then actually go and say, look, this city costs, you know, one 10th as much and got built 10 times faster than, you know, building another suburb of Phoenix or someplace.
And so like, God, let's, let's fix this. Yeah, that's awesome. so I think we're just about out of time and we can wrap it up there. So thank you so much for taking the time. My pleasure. I want to say, thanks again to Trevor for taking the time and also to you for getting it this far in the video. If you have any questions or comments, please leave a comment below and please do subscribe to the channel.
It really helps immensely. And also if you want to hear from anyone in particular, any other person in tech that I can get ahold of and try to ask some questions, please let me know in the comments or email me at Avon at tango DVC. Thank you. .