Building the World's Most Trusted Driver
38:36 Share

Hello everyone, welcome back to the asic cy podcast. This is stuff now. One of my favorite podcast be recorded since I during the team was just about this time last year. That episode was on autonomous vehicles, but I was actually also in an antony spec that was my first ride in a self driving car.

And over the last year, i've seen so many others have their first, as we will, has expanded the public and phoenix and temperature o while also placing its roots in Austin and aleg. In twenty and fifteen, we protested its first fully driverless ride on public roads and then open to the public in phoenix s in twenty twenty. But IT wasn't until twenty two that autonomous drives were offered in seven years ago.

And by the end of tone to me, three, eight clock in over seven million driverless smiles slowly than all at once. So with this space moving so quickly, we want to skip on update on where this industry is today, crossing the pon to properly introduced. Episode here is our very own air revolution host and is extremely general partner, Sarah one.

As a reminder, the content here is for informational purposes only, should not be taken as legal, business tax or investment advice or be used to evaluate any investment or security, and is not directed at any investors or potential investors in any asic sense y fund. Please note that asic and e and syphilis ates may also maintain investments in the companies is discussed in this podcast. For more details, including a link to warn investments, please see a six inc. Outcome slash disclosure.

Hey, guys, i'm throwing general partner on the sixteen z growth team. Welcome back to our AI revolution series. In this series, we talk to the geni builders who are transforming our world to understand one, where we are, two, we were going and three, the big open questions in the field.

I guess this episode is demetri dog v the coco of wam o dmitri has LED wao to solve some of the biggest chAllenges in bringing A I to the real world. And after tens of millions of miles of testing, wy mos vehicles have shown themselves to be safer and more reliable than human drivers, myself included. The metro has a unique perspective given that his work has span multiple AI ml development cycles across decades. He was an early pioneer in self driving cars, working with toyota and stanford on darpa grand chAllenge before joining google self driving car project, which then evolved into a way o in this conversation from a closed door event with a sixteen z general partner, David George, demetri talks about the potential of embody AI, the value of simulations and building training data and his approach to leading a company focused on solving some of the word hardest problems without further due. Here is the metro in conversation with David.

Maybe to start take us back to stanford, if you will. And that was when you first started working on the dark project and maybe give us a little bit of your history of how you ended up from there to here.

My introduction to autonomy's vehicles was when I was doing a postdoc um as as stanford you just mentioned, David uh this was a during to pray a lucky with the timing of IT. This was when the darpa grand chAllenges are happening. Darpa s that defence advanced research project agency that started these competitions, uh, with the goal of boosting field of anonymous vehicles a and the one uh that I got involved in was in two thousand and seven that was called the darpa urban chAllenge of the set up.

There was it's onna look, a toy version of know what with working on since then, uh, was kind of supposed to mimic the union in an urban environment so they can create a fake city on an abandoned and they populated with a bunch of autonomous vehicles, a bunch of human drivers, and they know they have them to various tasks. Uh, so that was kind of my introduction to this whole field. And IT was a bit of A, I think in a dark, these chAllenges are often by people in industry considered kind of a foundational, pivotal moment, uh, for this whole. And IT was definitely that for me I was like a you light bulb, light switch moment that my really got me hoped what .

was like the hardware and software that you guys had at that point, two thousand seven?

yeah. I know a very high level, not unlike what we talk about today. Know a car that you know has some instrumentation so you can not tell you what to do.

And you get some feedback back then. You have about the post system, much of, uh, internal measurement system. Maxim geoscope kind telling GPS tells you how you're moving through space.

And IT has sensors and writers writers and cameras, the same stuff we still use, say and then there's a computer um know guess the sensor data in and then tells the car what to do. And I want software and you know software head perception components and decision making planning components and some AI. But of course, everything you know that we had, like one of those things, are over that.

How long has been eighteen years? One of that is change drastically, right? And talk about A I today, vers A I.

We had, you know, back in two thousand seven, two thousand nine, nothing in common and similar. Everything else has changed. The sensor are not the same. Computers are not the same.

Yeah, of course. So then, okay, so take us. So at that point, that was the pivotal. That was like a light, bold moment. And then at that point you said, OK, i'm at stanford. I wants to make this my career right is that and then and then was toyota where to go from there?

And if I thought about those terms also, like this is the future. I want to make that happen. I want to be building this thing career OK you know but IT IT was that was the next that there was the next big step is um a number of us uh from the darpa a chinese competitions a started the google self driving a project was about a dozen of us in two thousand nine. I came together. Google will support an exam larian sergey uh to see you know if we can take you to the next the step and um that we worked on IT for a few years and that project then became way more in two thousand and sixteen, and we've been on this path since then.

okay. So we have this new big breakthrough in general. Vi um someone say it's new is someone say it's seven years in the making. Um how do you think about learning advances that have come from generate AI to what many would describe as more traditional AI or machine learning techniques that were kind of the building blocks for software technology up to that point?

Great question. So maybe can generate is abroad term. So you take a little bit of step back and talk of the role that A I plays in the vehicles and kind of how we saw the various breakthrough in A I map to the space of our task. Example, A I has been part of self driving or total most vehicles from the earliest days, as back when we started IT was a very different kind of A I M kind of classical and decision trees.

Classical computer visions with a hand engineer features kernels instead uh and a then now on one of the first really important breakthroughs that um uh happened in A I computer vision, but really was important for us for our task was the uh impassable in convolution neural networks right around to twenty twelve, right? Many of you are similar with the alex net and know the image competition. This is where alex not can blew away a out of the water or other approaches.

So that obviously has had very strong applications for our to like how you do computer vision and not just on camera, right, how you're on now you can use kind of nets to interpret what's around you and do a about detection and classification from ca, from light our data, from your imaging raters. So was kind of a big boost around that. A twenty, twelve, thirty time frame.

Um and then we are played with those approaches and you try to extend the use of companies to other domains. You just beyond perception with me some interesting but limit success. Uh then another a big uh very important breakthrough having around twenty seventeen one transformers I came around had a really huge impact on language understanding, language models, mission translation so forth.

And um for us IT was a really important breakthrough that really allowed us to take a mouth N A I uh to new areas well beyond perception. And if you think about and if transformers in the impact that they had on language competition is that they're good at uh understanding and predicting uh and generating sequences of words, right uh and uh in our case, we think about in our me about the tasks of you know, understanding and predicting what people will do. I get other actors in the scene or the task of decision making, in planning your own ject orties or in simulation of generating generate A I our version of generate A I generating behaviors of how the world will evolve.

Uh, the kind of these behavior, like these sequences, are not unlike sentences, right? Can Operate the state of objects, kind of local continuity, but then the global context of the scene really matters. So this is where we saw some really exciting breakthrough in behavior prediction and decision making and simulation.

And then since then, we've been on this trend of your models getting bigger. The people start building foundation models for multi tasks. And most recently, of the all of the last couple of years of the breakthrough is in large language models and in a modern state, modern day, uh, general I visual language models where you kind of a line a image understanding and language understanding. And there's been a most recently one thing i'm pretty excited about is kind of the uh the intersection of combination of the truth so that that's what we've been uh uh very focused on that way more most recently is taking kind of the AI backbone and all of the AI the way more A I that is uh, over the years with build, uh, that is really professional of this task of autonation driving in combining IT with kind of the general world knowledge and understanding of these elements.

One of the things that you just mentioned is a the role of simulation um and how that has been. You guys have had major breakthrough in the use of simulation and this idea in the recent breakthrough s in generate A I um around synthetic data in its useful. This is somewhat in question. I would say in your field, this idea of synthetic data and simulation is extremely useful and you've proven that. So maybe you could just talk about the simulation technology build, how is allowed you to scale, you build that real world understanding, you know, and maybe how it's changed in the last few years?

Yeah, definitely. IT is super important. And in our field and largely, if you think about this question of now evaluating the driver, know is that good enough? It's know how to answer that. There's no a lot of metrics and a lot of a data as you have to build up.

And then you you uh, how do you evaluate the latest version of your system or you can just h throw out on the physical world and then you see what happens you have to do in simulation. But of course the new system behaves differently from what you might have happened um in the world. Otherwise you have to have a realistic close loop simulation to give you you know confident so that that is one of the most important needs for the simulation.

You've also mention synthetic data. Is that another um uh area where simulation allows you to have very high leverage and you just going to explore the long tail events, right? Maybe there's something interesting that you have seen in the physical world and but know you want to modify the scenario and you want to come to to turn one event in two thousands or tens of thousands of variations of the scenario.

How do that? You know this is where the simulation comes in. Uh and then you know lastly um um if you know sometimes want to I A evaluate IT and a train on things uh that you've never seen, you are a very convent experience.

So there this is where purely synthetic simulations come in. There are not based on anything that you've seen in the physical world in terms of technologies that they go into place. And uh, it's a lot. And that is like a huge general eye problem.

What's really important that, that simulator is realistic, IT, has to be realistic in terms of your you know sensor or perception realism and as you um IT has to be uh realistic in terms of the behaviors that uh you see from other dynamic actors if the other actors are not behaving in a alysia way, like if pedestals are not walking the way to do in the real world, you need to be able to quantify the kind of the scenario that are you create and simulation to the realism. The rate of occurrence in the physical world, right, is a very crazy to sample something very easy, a to sample something totally crazy and similar. But then what what do you do with that? So I think that that raised me to the d point of a realism, is that IT has to be kind of realistic and quantifiable at the macro o level, at the statistical a level.

So there is any commission. There's a lot of work that goes into building the simulator that is a largest. One has know that that level of realism across those categories, and they have to be intuitive.

Ly, think about IT you to build a good driver, you need to have a very good simulator. But to have a good simulate actually have to build models of like realistic pedestrians and cyclists and drivers, right? So it's kind of do that either yeah.

of course. And then by having this simulation software that is very good at making real world, uh and very usable in the sense that you can create variables in the scenes, you can actually give the driver multiples of the amount of experience that they have on the road.

all in real miles drive that exactly driven tens of millions of miles in the physical world at this point, we driven more than fifteen million miles and a full autonomy. We all, you know, write our only mode, but we have driven to tens of billions of miles simulation. So you get, know what is a magnate .

of speaking of multiples of miles driven? Uh, one of the hotly debated topics in the AI world today is this concept of scaling laws. So how do you think about scaling laws as IT relates to autonomous driving is IT miles driven is IT certain experience had is a compute like what what are the ways that you think about that?

Um so model size matters. So where we are saying um now scaling law ws a blid, a lot of um typical old school models are used severely undertrained yeah and ah so if you have a bigger model, have a data that I should does help you, I just have more capacity that generalized Better. Uh so we are seeing uh scaling laws apply there.

Uh date of course usually matters right um and but it's not just in accounting the miles right for hours IT has to be you know the right kind of data that you know teaches the models or trains the models to be in good at the rare cases um that that you care about and I know there is is a bit of a than a wrinkle because then have to you can build those very large models, but in our space that has to run on board the car, right? So you are someone still IT into your on board system. Um but we do see a trend in a common trend and we see play out our space where you're much Better of training a huge model and then distilling IT into a model model and just training small models. Yeah.

i'm going to shift gears a little bit and i'm going to do A A sort of simplifying statement, which is probably going to drive you crazy. Um but the darpa school of thought is um you know the sort of A A rules based approach, right a more traditional kind of AI base approach um with a massive amount of volume and you document education and then the model then learns how to react to those. Um the more recent approaches from some other large players and startups would say, hey, we just have A I from the start make all the decisions and and you don't need to have sort of all that pattern recognition learning, you know like the indian driving that, that is kind of a tag line out there. Um what is your interpretation of that approach and what elements of that approach have you taken and applied inside of waa?

Yeah no, I think it's kind of again, sometimes you know the way people talk about IT is kind of this weird. The economy is this or that course, but it's not if that and then some right. So IT is your big models. IT is and turn models yeah IT is uh general AI and combining these models with violence. Ms, right.

But the the problem is it's not enough, right, right? Something like we want to know the the limitations of those models, right? And that's and we've seen you through the years a lot of these breakthrough s an eye right continus transformers, you have big entin foundation models, their huge boost to us and know what we've been a doing at that woman through the history of our project is uh constantly applying and pushing forward these state of our techniques.

Also comes some cases, but then applying them to our domain. And what we've been learning is that they really give you a huge boost, but they're just not enough, right? So in the kind of the theme has always been that you can take, you know, your kind of latest and greatest of knowledge of the day and is fairly easy to get started, right? Know like the curse this look like that.

And they have like of the curse of shipping, but the really hard problems and the remaining points, year, year, year one percent. And there is not enough. So right, and then you have to do stuff on top of that, right? So yes, you can take you know now, no, you can take you know n to n model, go from sensor to injectors or our actuation typically ate on build them in one stage.

You build them on stage, but you know you can do like back problem through the whole thing. You know the concept is very, very valid. Uh, you can combine IT and you know whether the O M, and then you add close simulation some sort.

And you know you're after the races, you can have a great demo like homo of the box, you can have ads or a driver system, but that's not enough to go all the way to full autonomy. So that's where really a lot of the hard work happens. So I guess in question is not as this or that this and then what else do you need to take IT all the way to have the confidence in? You unction remove the driver and go for full autonomy.

And that's a tony work. That's a tony work through the entire and of life cycle of these models and the entire system right? So starts with training.

Like how do you train? How do you architect these models? How do you um you know I value them then yeah if you put a bigger system, that model themselves are not enough.

So you have to do things around them. You have to know they have, uh, modern genera. I is great. There is some weaknesses and of uh goal enter uh planning and policymaking and kind of understanding this new three spake Operating in the three d spatial world, I say have to add to or something on top of that.

We talked a little bit about this simulator that's a really hard problem of and then, you know, once you have something, you know, once you deploy IT and you learn how do you feed that back? So this is where all of the really, really hard work happens. So it's not like end to end versus something else. IT is end to end and a big foundation models. And then like, and then the hard .

work and then all the hard work. Yeah IT totally makes sense. Uh, that is a great segway into all of the progress that you guys have made.

Writing in the way more for those who have done IT is an extraordinary experience. It's not to say that you have solved all of these complex test, but you you solved a lot of them. What are some of the biggest AI or data problems that you still feel like you're facing today?

The short answers going to be you are taking IT to the next or of magnetite of scale, multiple or magnetic scale. Will that come additional improvements that we need to make A A great service, right? But you know just to level sudden make a uh entrance where we are today, know we are up your driving and um all kinds of conditions driving a twenty four, seven and o in hoenig a little bit.

Those are the most metro markets but also in a ay and an Austin. And all of the complexity that you see go driver around the city rate guys of weather conditions, whether it's no fog or you have storms, does storms or rain storms down here like all of that of those are conditions that we do Operate in, right? Uh, so then I can think about, you know, what makes IT A A great you know customer experience, right? What does that take if you grow by of magnus de? There's a lot of improvements that we want.

So that becomes a Better service for you to get from point a to point b right? Like ah we asked for feedback from our writers. Uh, a lot of feedback we get is has to do with the quality of your pick up and drop off locations, right? We are learning from users like we take a magical theme was delightful experience from the time you start the APP on your phone to when you get an decision so that that's a lot of the work that we're doing right now.

Yeah pick up and drop off for what it's worth, is an extraordinary ly hard problem, right? Like do you kind of block a little bit of a driveway if you're in an urban location and then have a sensor that says, oh, actually I just saw somebody opening a garage door. I need to get out of the way um you know how far down the street is acceptable to go poor or if you're in a parking we're in the parking line, do you go like this is an extraordinary ly hard problem but to your point is huge for user experience.

This is right. And just I think that's a good example. Like just hate just one thing, one of the many things that we have to build in order for this to be an awesome product, right? Not just like a technology demonstrator.

And I think you just like you hit exactly the uh um on a few things that make you something that is uh the face bit might seem fairly straight forward, right? Okay, you know but I know there's a place in the happening to pull lovers like how hard can to be right? But really if it's A A dense urban environment there is a lot of these factors, right?

Is there like you another vehicle that you are going to blocking? Is there are a garbage or this opening, right? Like what what is the most convenient place for the user to pick up? What is so is really gets into this yeah the depth and the subtlety of understanding the the the the semantics and the dynamic nature of the driving task. Doing things that are safe, comfortable and predictable and IT leads to a uh a nice thing as pleasant, delighting customer experience.

of course. okay. So you've mention the step, but fifteen million miles, I know the numbers, probably a little big, bigger than tuesday yeah it's growing by the by the day fifteen million autonomous was driven.

That's incredible. Um even more impressive, you didn't share the stat yet. IT results in three point five times fewer accidents than human drivers. Is that right? And I think .

three point five access the reduction in injury and that's about to extradition in the police report .

table and lower severity instrument. This sort of comes to a question of um both kind of regulatory and a business or ethical. What what is the right level that you want to get to oba see you want to constantly get Better but is zero level at which you say, okay, we're good enough and that's acceptable. The regulators yeah .

so there's no um simple super simple shorting answer, right right. I think it's starts with that IT starts with those statistics that you just mentioned. Yeah I can know the day we care about t is that rose are safer. You look at those numbers yeah ah we when we Operate today uh and we have strong geral evidence that uh our cars are in those areas safer than human drivers.

So on baLance that means a reduction in collisions and in the harm uh um then in action on top of the numbers and publishing this is according the latest numbers that we shared yeah consistently you're sharing uh numbers of as our service sales up and uh grows. Uh if you can also bring in an additional lens of what company did you contribute to a collision and we actually published I think that was based on about four million miles for twenty million miles we publish to join study with a swiss R E. This is uh, largest global reinsure in the world.

And the way they look at IT is, you know, who contributed to an event? And there we saw, uh, like the same theme, but the numbers were a very strong that you fear seven six percent reduction in uh um police a property damage collision and IT was in one hundred percent reduction in claims around bottle injury. So if kind of bring in that lens, I think the story becomes something more compile compelling right? But there are some collisions.

Where does the bulk of events that we see would be light? Uh and somebody just plays into you, right? sure. So h but then um I think we do know it's it's a new technologies in new product. So um is IT is held to a higher standard.

Uh so we when we think about our safety and our readiness framing methodology, we don't stop at just the race, right? We build over the years one of the uh uh huge areas of investment and experience over the years like how you know what else do you need. So we have done and we have done a number of other different things, publish some of our methodologies.

We've shared our readiness framework. Uh, you know, we do other things like we actually not just statistically, but on your specific events, we build models of a an attentive, very good human driver like not distracted human a such right, but doesn't want to compare our driver too, right? And it's it's a model.

It's one as in particular scenario, we evaluate ourselves verses that model human driver and we hold ourselves to the bar of doing well compared to that very high standard. And then know you pursue other you know validation methodology. So that's my answer is that is the yeah of the aggregate of all of those methodologies that that that we look at to decide that yes, you know the system is ready enough to be deployed and scale.

I love free to talk about what do you think maybe today in in the future about market structure, competition um and what kind of all you envision we are playing.

So the way we think about new amo in our company is that we are building a generalized boat driver. That's the core and that's the core of the mission of uh making uh uh transportation safe and accessible right? And um we are talking about right healing um today that our mean most mature primary application.

Now we envision a future where are the way the driver will deployed be deployed in other commercial applications, right? There's deliveries, there's trucking, there's personal own vehicles, right? So in all of those are guiding principal would be to uh think about the gold market strategy and way that uh accelerate uh access to this technology and gets deployed as a broadly um well of course doing IT gradually and deliberate things safely, you know as quickly and um and broadly as possible.

Uh so with that as they are getting principle, we are going to explore different um commercial structures, different partner ship struck, for example, in photos today uh we have a partnership uber um and right healing both in the right healing in uber eats where so in phoenix we have our own APP. You can download the way more APP and you'll take a right on our vehicle will show up and take you where you want to go. Uh uh that you one way to experience our product.

Another one is through the uber APP, we have a partnership where you can get through the uh APP matched with a our product the way to driver the way of vehicle and is the same experience right? But this is a another way for us to accelerating give more people to experience for autonomy and that gives us a chance to kind of have think about different uh go to market strategy, right? One is you know uh as you know, having you know more of our own other one is more of a you know drivers of service somebody else s so it's early days but will iterate and but all know in service of that .

mean principle. That's amazing. Yeah, that's gonna. That's gonna exciting. Um maybe on back to the vehicle, what about the hardware stack that you use? You and i've talked a bunch about said like going all way back to arpa.

Um you know it's kind of the same stuff, right? It's it's sensor. They have advanced quite considerably. But you know you still use readers and right or um do you think that remains the future path for autonomous driving? White are .

specifically, um yeah no way. I mean, uh the sensors are a physically a different right. They have each one cameras, bitters, rather the half there benefits.

Each one brings their own benefits, right? Cameras, obviously you color and they give you you know very high resolution a lighters give you a direct treated measurement um of your environment and their negative sensor right kind of bring their own energy pitchy dark when there is no external lights. Ce, yeah, you still get a seek just as well as they do during today, Better in some cases.

And then you know rather uh is you very good at like punching through just physics and different wavelengths right to if you build an imaging a rater ish we do uh ourselves um IT allows us to know give you an additional regnant ency later. And IT has benefits also negative sensor IT can directly measure adopted the also other objects and if IT degrees differently and more Gracefully, some weather conditions, very dance fog, our very uh so you know they will have their benefits. So if you are our approach has been to you know uses all of them.

And yeah, that's how you have redundancy and that's how you get an extra boost and capability of the system. Um and now we are on if today, deployed and fifth and working to deployed the six generation our sensors. And over those generations, we have a improved a reliability, we improved new capability and performance, and we've brought down the cost very significantly. right? So ah yeah I think the trend you know for us know using all three minutes just makes lot of sense uh again, you know you might make different tradeoffs if you are building a drivers I system versus a full autonomous vehicle where .

you know that last point years one percent really yeah absolutely um what are the observations that we have from the very early days of this wave of L O M S is that there has been sort of early a massive uh race of like cost reduction. And many would argue that it's sort of a process of commoditization already. You know it's very early days.

Um I would say the observation from autonomists driving over many, many years now is kind of the opposite thing. There's been a thinning of the field. Um you know it's proven to be much, much harder than expected. Can you just talk about maybe why that's the case now?

Always have this property that is very easy to started, but it's very inseam ly difficult to get IT you know all the way you too full autonomy so you can remove the driver. Um any other there's a maybe a few factors that contribute to that. Um one is compared to the l ams and kind of A I in the digital world, do you have to Operate in the physical the physical world is messy and IT is noisy? No, I can be quite humbling, right?

There's all kinds uncertainty, noise of that yeah can pull you out of distribution, if you will. So that's one thing that makes us very difficult. Um and uh secondly is a safety.

I sure uh these AI systems in some domain you know this is creativity and is great in our dome. The cost of mistakes are your lack of accuracy has very serious consequences. The bar very, very high. I don't know me.

The last thing is that IT is, uh, you have have to bring real time and you you're putting you these systems on fast moving vehicles and you have to the second matter, right, you have to make the decisions very quickly, I think is in the combination of those that really not together IT lead to you. The the trend that you have been saying is that like IT, it's an end, right? Have to be excelled on this.

And then and then right is all of the boy, the bar is very, very high for, and every component of the system and how you put them together. But now there there is big advances and they boost you and they preferred the system forward. But there no silver bullets, right? And there's no more cost.

And because of that lack of tolerance for errors, you have a very high bar for safety. You have a very hybrid from regulators. You know it's it's it's very costly to go through all those processes. Um and so IT IT makes sense in i'm very grateful that you guys i've seen IT through despite all the humbling experiences that you that you had along the way.

So it's been a long journey, but it's a enough for me and uh many people that way more IT is super exciting and very, very rewarding to find me to see and become reality. We talk about safety ni in many context, right? It's a big question, right? But you know, here we are in this application of A I in the fiscal world, in this point a pretty blessed and increasing body of evidence that you we are seeing like cancer safety benefits.

So that yeah, I always of people IT was a long journey and very costly along the way. But this is probably the most powerful manifestation of A I that we have available to us in in the world today. And you can get in a car without a driver and it's safer than having a human. And that's just remarkable. What are some of those humbling events .

along the way in those really couples? Oh, I sorry. I remember one there's one route uh um that we did that started and I started on view than once through people to through the mountains to highway one and took highway one to sano and around the city a little bit .

and like actually finished a long .

street two thousand one .

hundred years in human drivers would so uh.

you are doing IT one day and then we're driving and omit through the mony pootle park. Were driving through the mountains and its foggy early morning and then we're like to seeing objects and you our objects is like random stuff on the road in front of us. There was like a bucket and like a shoe.

And then there was like, at some point we can cost like a, you know, a rusty bicycle. I, okay, what's going on there? And eventually, like the card, you know doesn't know know handles IT. Okay, no, maybe not very smoothly, but we can get stuck and we cash up to like this a the dump truck that has of stuff, period, losing things, that person obsta to the car.

This is like a cartoon. You know continuation of anonymity is being wrong and you guys that's that's pretty cool. Um okay, last question. I'm going to see you up to do some recruiting probably. But um if you are in the shoes of the audience here and just kind of seeking your first job, i'm going to take something that you said, which is like I can see your passion and excitement for doing the start up thing right and like you know kind of longing back for those days is so cool um what advice would you have for these folks in where to go, whether it's type of company, type of role, industry or anything else?

会 吗？ 你。 Find a problem. We're talking about A I today. But to say fine, a problem that matters, problem that matters the world, problem that matters to you. Chances are gonna a hard one yeah many things we're doing have that property. So don't get discuss by yeah not known by what others might tell you and you start building and keep building in dog back.

A huge congratulations on all the progress you guys have made. And as a very happy customer, thank you for building IT, and we really appreciate you being here.

All right, that is all for today. If you did make IT as far, first of all, thank you. We put a lot of thought into each of these episodes, whether it's guess the calendar touches cycles with amazing at a gar, Tommy, until music is just right. So if you like what we put together, consider dropping as a line at rate this pocket com splash asic sensitive and let us know what your favorite pisos it'll make my day and i'm sure comes to will catch you on the flip side.