Waymo’s Journey to Full Autonomy: AI Breakthroughs, Safety, and Scaling
44:30 Share

Hi listeners. Welcome to no players today. We're hanging out with to me. Treat dog up cosey wao wao started as the suffer project within google back in two thousand nine and eventually spin off as its own company. Now IT provides over one hundred thousand paid rides each week across san Francisco, L A, Austin and fano. I love taking emos and am regularly campaigning for Better south bay coverage, were excited to dig into all things robo t taxes, self driving, what IT takes to deploy this technology on a mass .

scale and what's next for way now. Now we can start up with just a little bit of the history of self driving google, how you got involved and how thinking of all over time.

I've been doing this for quite a few years. Think, uh, about eighteen. No, I got started in uh around two thousand six.

This was the time of the darpa grand chAllenges um this is when darpa organized uh a few competitions. They called the grand chAllenge in robotics for the with the purpose of advancing research and autonomous vehicles. And so the first competition that had was the first grand chAllenge.

The chAllenge there was to create a card that could drive autonomously in a desert, I just like this, that environment, but drive for about one hundred miles. Nobody succeed, but there's a lot of great progress that was made. And they repeat the the chAllenge and a few team succeeded on the hills of the created another chance that called the darpa urban chAllenge.

Now where the set up was I got about mock a city that was supposed to imitate and we're driving on public roads like and that's the one uh that um I was involved in. I was and this was kind of my you know moment where clicked for me the future and the benefits and i've never look back. That's what i've been doing every sense. And then we started this project at google in two thousand nine in a small group of us. And then I grew into what now he, we started the company very begin of twenty thousand.

Seems like a lot of the lining ager history of this field all traces back to a handful of loves. Ethra b is stanford and a few others. And IT seems like the founders of a lot of the companies that ended up eventually existing in the ecosystem .

all came out of team and the stanford team. Uh, a few people who came to start this project at google and two thousand nine came from those teams.

I you think when you started working on this, this was considered a little bit a crazy thing to do, right? IT was early on. A lot of parts of the waves of deep learning hadn't really quite happened yet um in terms about applications across all smart bears like alex and existed like .

all these are these are things that nothing yeah and have look the rate people you know we heard a lot about as being crazy and never going to work in poll honesty ourselves. We're not exactly sure if you know we are just a lib IT crazy completely you insane when we're trying to go .

problem as an research project. Are you a you had an idea in your mind of like an end point date where this would .

be viable on public grow more of the former but was he was research chAllenge IT was a research project. Uh um then when we started google, IT was under the belief that we can make IT work. Uh and if we can, then the impact, the positive impact of this technology on the world and the mission is worth but IT was the early days.

So we actually I have very little day a to long the time that you start. That was actually the first question that we post, not build the product right in the first a couple of years or so, you wouldn't have like a product in mind or a target date of my the first sort of business was to explore the space. So we to that end, we created some milestones for selves, are with the goal of prototyping and learning and just understanding.

So after those two years, is that oh, okay, you know there's something there. Start talking about know what the product could be. And actually our first product that we thought was going to be viable, uh, was what doubt you would call him and advanced drivers, the system, right? We have some expections small number of years that would take for us to uh, when you know we after working on IT for a while and make more progress on kind of the core of the technology, we decided that was not part for us that we want to go after full autonomy service on twenty thirty.

You are now doing something like one hundred thousand rides a week, so five million rides a year b sort of annualized. That which is incredible. What was the inflections point of what suddenly caused sort of volume to happen or all these things to come together? Because he feels like .

a reasonable recent phonenet some sense. yeah. So we rewind to talk a little bit. I think my mind there are a few of generational discontinuous steps on kind of that progression from that point in twenty thirteen when we said let's go for IT to where we are today, just exactly one hundred thousand trips per week, more than a million miles per week end growing. So some interest is in two thousand and fifteen, those kind of zero to one moment.

This is one year for the first time we put a car in the road and that was that what call the third generation uh of our system was the a third generation of our cell driving hardware sweet sensors computer and we put IT on a custom design vehicle that we called the firefly, but we took a few rise, but nobody behind the wheels, zero to one moment, uh, then the next, uh evolution the kind generation or skip was our four generation of our driver was there the pacific minims ah with the four generation of the way more hardware weet and we deployed those uh um in a full autonomists mode in in arizona and in chella and we actually opened IT up to the public in twenty twenty but at that point uh we the focus was on doing IT uh repeatedly right and the focus was on yeah matching the of the technology the building of the driver, the evaluation of the driver and of doing releases in a regular ebs. Getting in out to real customers, hearing from the customers, understanding the feedback g you know iterating so that was the focus of that force generation was not you know to grow and scale and capture the market at and then at that point, we made the decision to jump to what you know, and I call the fifth generation of the way the driver is on the jail or I basis, is what you see. The feet today in in four city that .

was very smart, actually started is on a mrs. In california. I think, you know, for example, I think crews ran in this in issues and Simon isco, where there was activists like putting comes on the cars and trying to stop them any of the things. And so that seems smart to start in everyone this what occurs, one of the criteria that like to just think that is a of a .

test base depends on the different time horizon. The first generation we picked, uh, the development area there was kind of medium complexity, uh, and the goal there was much used to can go end to end. So we picked in an environment where we thought that was that we have enough of the boxes to help us learn the most important things that we want to learn in the risk while and that was the deployment.

Uh, and then there the development of the system suffered, the development of the system, you want to go after the hardest st problems, the environment wether so been doing that. So we had decision to no in china and IT crates way, you know to learn from the end to end system wide, you know, pushing on developing the system. Then when we have a you know learn enough and we made that this continues jump to the fifth generation of our driver and I said, okay, like that's the platform we believe that we wanted take to scale.

What's the hardest environment for a self driving car?

So dancy matters, speed matters, uh, weather matters OK so I can see where those come together. Uh is um where most complexity from the winner really bad? And no no. But that is why we picked for sisco. I good for learning vancsik driver IT is a very interesting commercial market.

Uh we also the same time went uh after downtown photo x IT has no a different maximum of the higher speed roads and h that gives us the uh the way we think about IT at terms of the development and evaluation of the driver is kind of in the space of the Operating, not necessarily areas. You take cities and new map to the of the and deploy night. And i'm looking forward is the kind of the in terms of how we think about the future cities in a few months, IT is market.

I know this is there are a good market from the commercial perspective. And what is that? The technical complexity, you know what is the regular for the environment? That's how the land apply. yeah. Well, with the big technology breakthrough .

that got you to the fifth generation driver, that is really the one that you think you can scale.

So the biggest wasn't not not surprising, uh, and there is generation breakthrough with every generation of the hardware of art of the driver. There's a new hardware. So it's getting capable. It's getting simpler as getting, uh, more expensive and a lot of simplification.

but a boost really. A P, N, A, I, was the shifter change. Those important was that moving and does what of? And A N, D, L, for everything. Was that the transformer backbone .

with understate? curious. That was a missipipi though. So for that last jump, the model, we had the big breakthrough before alex, around twenty.

So that gave us a big boost. But replace and then IT was a few of those things coming together. IT is transformers is in bigger models in my compute, coupled with of the whole evaluation.

We often talk about the architectures and what's really more important surrounding architects. The architectural is an neighbor. But really, to make IT work at the level that we care about, you need, like everything around IT data engine, that a fly wheel of training, the system of valuing IT.

And you are going to have to think about the problem of evaluating the driver, intend them of building IT, right? And you need simulate data. So all of those coming together, I think what leads to and the brakes to continue that you are seeing very today.

Can you explain how you just think about evaluation internally? And then also you know how that might defer from how regulators evaluate this from .

the safety case perspective? So that has a big question um but um I think glad asking about because IT is a super important and constantly. right? We always talk again about the building of the driver, but it's the two problems, the building of the driver in the evaluation of IT and they go, you know, hand and and so internal starts with you have figure that out what metals you care about, then bringing the data to support the valuation of those mechanically english, which you know have hundreds uh then you need all the infrastructure uh, and including things like the simulation, some things you can evaluate an open loop know about something you need close to simulation for.

They need to build there a uh realistic scalable simulation to support all of that, right? There are all of these metrics that guide are the development of the system that help us improve and kind of train the way ee way my driver. And then IT falls into equal validation and evaluation, the aggregate of the evaluation validation methodologies, what we call the national .

safety framework, uh, miles driven, an urban setting or you know some compares into human drivers in a given city. The Operator, what is a relative safety level of what we must doing verses a human driver at this point.

we are pretty proud of of our record we have now that we've driven tens of millions of miles as we all right, the only mode uh and driving know today, more than a million miles uh per week, think we can with pretty good confidence in especial ata say that we are than you so publish some b um the latest data point that we shared IT was based on twenty two million full autonomous writer only miles and uh we compared our performance uh versus human benchMarks uh by different severe levels of contacts of collisions, different severe level.

So we see depending on the severity level, we see that you know for the low severity, more security outcomes were about the factor of two Better than the humans background. And as you look at more severe outcomes, the gap increases. So it's the most what we want to see, right, for airbag development effect six Better than human drivers.

And that's without any notion attribution of cause or four. If you bring that into the me, most of unavoidable this really can do. Yeah, we have done a study with swiss ari global, I think the biggest global pressure, and we partner with them.

We share the data, they're on analysis. And they found that for damage claims, we had about a four x reduction versus the human baseline. And for bottle injury claims, we had in one hundred percent reduction. IT was a small early was a one less than four million miles but also um um IT was from their point of view statistically significant but again.

where pretty to that yeah it's amazing given what do you think the regulatory .

stance should be? We enable uh this a technology and service of the mission of making roads safer. And and we've been engaged with regulators for eight years and have a dialogue. And so AR, we've had good success getting all the necessary premise and all of to a lot of scale. The way we think about IT internally in this, how kind of we have the dialogues with the regulators, with computers, with reuters, IT needs to be based on transparency and IT needs to be a responsible iterate, gradual processes. And because the thing is very new, technology is very new, the products very different.

If you're doing a million miles a week now, what prevents faster roll out? In other words, that seems like you proven out that this is a safe solution. It's working very well at scale.

E, you have your generation five driver. That seems to be a quite performance. Why not go bigger.

faster? Who we are, scale ing explanation. Ally, I took us about three months to get from fifty thousand to one hundred thousand, right? So we are moving a good rate. But the mean thing, uh.

what the number of miles that are driven in the U. S. Player, these .

calls.

Again, you all the way we think about IT is that it's important for this to be an an interview process where we earn trust. That is not a thing where you know build than you turn on the switching as rait needs to be as news. Different needs to be a dialogue.

IT needs IT to be, like, know talked about establishing 呃， the safety chair record, right? And we needed to build up to that, right? We collected in s of millions miles.

Now we feel pretty about that. We know that gives us confidence that turns us trusted. Ah so you have to be transparent about where we are.

And then we the biggest shift that's coming from the technology basis going forward, we had him under a from the on our potest a couple weeks ago, and he sort of contrast what he viewed as the text approach with his more than his user of software driven to the a very motivator which was more sort of hardware centric. Um you think that's a correct character zone? And also how do you view things sort of changing over the next you know uh, a couple years or year two in terms .

of your right now? No, I don't think I been so about full time. I we talked about you know a few uh, big breakthroughs that allowed us to get to where we are today, company transformers bigness now the most recently and of combining a the way A I was the congenial knowledge of know the alms.

This is at the core of IT. And you know hardware matters right today. I have to Operate in the fish world.

Uh, so the hardware do we have um on our cars gives us but time to think about vanity like you have to see well you so you know a human driver, right if you can see you, if you close you, if you your vision is not twenty, twenty, you don't have your glasses. I still drive right will just going to drive. But the point I can, the way I see IT is is all about A M.

It's all about building the system, building the I, building the software and being able to evaluate. And that's what we have today right so far. I need we have they are achieved the driver. We have built all the machinery to evaluate and no take. Now for us, it's an optimistic. It's an optimization simplification and we have that the the thing that works, you have a good mechanism to valued at, then you know that really is drastically different in terms of how fastest, I guess, i've been in, in the other mode for years where we have not solved the kind of the can cru t.

And you're kind of hypothesizing what I would take and like, you know know how what is the yield of this technical breakthrough and you're kind of climbing up health, right? And this this is big and that's what I felt like for us for many, many years as kind of the trend of each industry, usually false. And I just minded that it's a quantify different place to be in when you cry not have the valuation and .

you can optimize and scale as part of your optimization. Do you think of like reducing or simplifying the sensor sweet .

as an important priority? Every generation increase, keep body, but also simplify drastic, and then the cost comes down every generation. So that was true not for the previous generations, uh, made a big jump from the uh press generation to two days and then from the four to the fifth and then now going to to the six generation.

But was the primary focus simplification and bringing down the cost as well as know with a vehicle making up about the use experience, experience and of course, economies of scale, right? There's nothing like fundament at the component that we use our there's just all of them are faze the scale this that you can get to write. The typical of making a rare is to be expensive in the past, right, put all the cars to bring down the cost. Computers.

I guess, qualitatively, we know that uh, people can drive well with like a very simple vision system. And right. And obviously the sarrion world, we have dramatically more in different types of sensors. Is there any sort of and a logical framework that you will use in terms of the amount data slash text need to collect in order to have a performance system relative to the curve on the a eyes side? In other words, can can you follow some sort of scaling curve on eye that you can be predict the sentimental that you can .

do without what we have now, uh, is now now that to jump to the core answer to your question, now that we have cracked the out of the driver and the valuation, we can answer that question with da. So we talked about different start with humans have yeah uh cameras on a pivot right then we can drive up and a few vision is not very good of you like you maybe not drive as well as computers can, right? Um so that's why we have vision.

We have my cameras, we have lightness, we have writers fighting. They will give us news, some enemies uh and we can talk about you know what the present that they of how they they bring to the table, the data bring to the table and how they are is the complimentary I then what you get from the different sensing model is just by the kind of the physical world. But you again, in terms of, know, how do we answer that question? Know for years IT was more, therefore, al and the hypothesized, you know how much of your certain thing you need.

Now that we have a thing that works and we can evaluate, we bring data to the tape more answer that question and what what happens to the way and driver. If you take away something right, maybe you add noisy system, maybe you d degrade, maybe you take away a full sensor metal, maybe you take away later and you take rater away completely, and you just drive camera drive via. I can use a human drive with one I or you know with buler revision via are you that acceptable performance? What you get a license? No same thing for us.

We can just take some way and we can answer that question. Like is the performance can I still drive of course, is the performance ance good enough, uh, for full autonomy and is a good enough for our bar of redness and safety and oh uh and but again, this isn't the context of the autonomy is in a context a of a uh scale and IT is in the context of the responsible deployment and the high bar for safety and redness. The way of herself.

If you change you know some of those inputs and say you talk about in a small scale where you talk about, you know not full autonomy, talk about the drivers system, then the answer changes like you might have a different if you still have a human in the loop and the responsible for safety, a different configuration, where there is, you get rid of sensor modalities altogether. Or in, you know, you maybe use all three model ties, but you pick a different up because it's cameras, cameras, their solutions range, cleaning letters. But the answer of your Operating point, which change based on at the bar product .

that if you look forward a year or two like now, you know crack the nut is looking at scale and um probably more about the business. Like do you think of robot taxi at scale as the the near term business plan? Are there are their modalities or like deployment? If I think about this is just like a capex problem, like other other of news that are .

important for you guys, explore fox. So we are focused on technology, focus on the product. We are learning for our years every day in trust, and you are setting up the ecosystem of the Parkersburg this space。

So is a primary focus, right? We are very excited to buy a the commercial opportunity there. It's not the only one, I think of way more, as you know, technology company who are building a journalizing way more driver and the mission to build the world's most trusted driver.

And we went to deploy the driver, not just the right healing. There's more than three twenty miles in the there's trillion miles worlwide. Uh, so the vision and the missions is to deploy the way driver different commercial products in different applications and be different you know many ties across all of. So that includes uh things like deliveries, those things like a long hole trucking and includes uh and things like personal only the but right now, we are very focused on right.

Can do you think you license that out? Do you think you would actually build the vehicles for these different these cases?

Uh, we're not ah we've never been uh uh in the business of building vehicles partner. And I guess this is how like this is. The mission is important.

The opportunity is so massive that alone. So think about me, and even together, the ecosystem to pursue all of those different martial and applications and all of those. And you see you doing that in the right healing. We don't build our own vehicles, partner with partner, partner ams, partner the companies for the .

next decade two. So if you know it's it's very exciting to see this ramp in terms of you know driverless rights that are happening. And one could ardo at some point some proportion of the population, just like people fled to uber to um you know ferry them around in major cities versus driving or taking taxes, rather things that could be a full period of to autonomous systems versus owning cars.

You just be able to order something on demand, have IT show up at the right time and you just get in and IT takes you where everything to go. D view that is uh ten percent use case, thirty percent year is like proportion of miles. You think we'll convert over time.

I think over time will see more and more. I know with the logy matures as IT is deployed and more of these different products, and I think you're starting some of that in right? hello.

And that's not uniform, right, but ties. Look at people who live in the people who live in your k even before autonomy, right? There was a shift.

Fewer people want to to see those areas on cars. This is the Younger generations, right? It's a not not what they are expect about, right? Uh, and I think with me a autonomy in those that I am most excited about this and all of those models will be bringing the safety benefits from this technology to the one of the reasons I ask is I remember .

um talking with people in the self driving world. I don't know how long go eight years ago and years whatever was and at the time everybody thought this weight was coming, I think a lot of people were off in terms of the time frame, in terms of what actually happened. And there was sort of a flurry of startups all getting up and running at the time.

And a lot of the conversations were around how urban environments would change over time. Where do you actually put a parking lot or where do you park your car versus having like a lot outside of the city that would then the cars would come in and pick you up at the right time and versus everything needed to be central ized. And so one of the reason was asking that question was a trying get a sense of your review of the word of magnus de, and also time frame by which some of these transformations may occur.

I don't want to speculate that from, I think, the vision that is absolutely the crack up about the safety benefits of does the primary one one there is accessible once you have now you get those benefits, once you can deploy scale, right? And once you can deploy a scale, you can do things like, you know, you know, use land in a Better way, right? Instead of you're taking up so much space for parking loss, having been personal vehicles that just sit around ninety percent of the time, you can do Better than a society.

We can do Better, right? All of those. But the thing with those uh benes that come from scale and uh again, safety is the primary one that were very focused on. Now we've talked about as kind of the north star as the mission, as the vision for you know many years. Uh, but no, he was always kind of with once we get the scale.

So I think today we're starting to actually earn the right to talk about a starting to realize that mission, tens, millions of miles that behind him well, but more than a million per week. We and the benchMarks of safety, benchMarks that the just significant, partly ambiguous, and talk about actually a real tangible safety benefits and like reducing, reducing harm and injuries that are happening on the road today. So that's a primary focus on. I think that's the primary benefit that will see. And then beyond that, I think there is going na actually most like the .

ones you what do you think is the role of uh, traditional arms and in this world um when you know I say functionally like a car, tex you point to point right hilling takes you from point to point with like you know different tears of comfort level. But you know a very large industry has been built around people buying passenger vehicles for like you know industrial design or brand or all these other things that uh uh you know you really need a ford drap tor to run around the barrier, right? Do you think when .

when you .

know perhaps one of the more primary um drivers of value is now A I and the ability to do this autonomously to know we think .

of what we're doing as building the driver and and the driver .

could still drive and .

the driver exactly you put the driver to the car and different fourth factors and whether it's a car is good for right healing and certain where are a different vehicle you need for a good transport or truck or that you want to take on longer trips will need different cars。 We will need different form factors. And I think it's very, very complemented tary we are doing and what the you know the the car industry, this building .

perhaps to a laws point about how cities will change the answer has currently been given the efforts to change both drivers and cars and the environment to just change the driver right um which is what you guys have done. Do you think there are arguments still to change the infrastructure like you can make, for example, in the public transport space, right? There's other form factors that require the participation of the public sector in order to deploy a absolutely .

a sustainability is your import for us. Safety is a primary thing. But I think all of those modalities can coexist. Uh, in fact, uh just in the last couple of days, we announced um uh something that we're doing where we are incentivising people to take meals in um the cities where we Operate. Two public transit bs and everybody sure how do .

you think about the form back to the carts? Sofi know that there is companies like zip s that amazon bought where they kind of hold out the inside of the car because you no longer needed the stern column everything they put seats facing each other like a london cow ah you have any thoughts on what that experience will look like in the future? Is more and more things, moto autonomous self driving, right?

Healing systems as the designing the car around the passengers makes sense. I in the past with designed cars around primarily the driver, right? Uh, if it's the way more driver, it's all about the the rider experience.

So we have a done a quite a bit of work on the six generation of the driver and the park and the car is designed with the passenger in mind. So is more facials and IT is all about the user experience. You know you have flat floors, have lower floor, france have doors, slide side.

So it's all about so absolutely, there's different aspects of we don't have cars facing each other. I think it's problem question, like some people get no nauseous when you do that, like you kind of want to know their benefits facing forward. But you know all of that I think we will be for us is and instructed figure out is being forward. But I think that the key point IT becomes much more like the design is around the writer, not just like you .

could do very addressing excuse to where you know, I always wanted a car with like a pilon on the back or something as your commuting. You just kind of.

I like, I like, I want to be a bike, but I want my bike to be inside .

of a car. Might be .

specially alive. Good.

good. I want to have a much less exotic form factor. I just want to go to take a zoo. M, stable internet.

and .

not, not.

but we do. We do increasingly, you know, team members calling to meetings from way. Mos, so this is, but IT is IT actually we talk about, but IT, uh, IT gives the point of patache.

IT just becomes in this, if you don't have another human car, right? You can, you can do a work meeting. You can do cold. You can. I listen to your favorite music, full volume, not worry about action of like having another human that you are sharing the spaces. So we are seeing that was one of the hypotheses of the benefits are of of our product, and we are seeing very positive feedback from our areas today.

I am so I just somebody excited to see you this technology expand coverage range is the blocking factor to let's say, you know a billion miles a week. Is that like putting more cars on the road from a capital perspective? Is that just Operationally this can only happen so fast? Is your view of um like what you want to see from a safety and trust perspective, like consumer trust perspective?

What's the bottle that is the letter. We always know our playbook has been to go about IT response and gradually and earn trust every step of the way and have this transparent dialogue again, this is a very new thing, new technology, new product, very different from what people used to. I think IT has to be the sitter and again, your trust, the thing that's hard to burn but very easy to lose.

So that's the main thing, right? Um and use we see that you see that like in places where know we Operate and we've ve engaged with communities and there's writers who have used way most IT is there's a lot of trust. There is little giant people will use the word magic, uh a lot about the experience.

I, and they go to a different place where people have not experienced that, and there is more anxiety of trust. So you can just get there in one step. You have to do IT kind of response. Operators rid oats .

in the elevator to close the door for you and push the button control. That experience is interesting to see the evolution of different types of technology over time or people's interpretation of um how do you think about genius billies you mentioned your building a gentle proper driver, the opening port and other type of vehicles.

Do you think there are other extensions in the other forms of robotics with your building? Or do you think those are all more specialized models? Or how do you think about where this could go from that .

perspective on the driving part? Um we've going to design IT to be journalizing and were very happy saying with the fifth generation driver and like they, I generalized really well based on again, we've been using data from a very raw dd to build IT even for deploying responsibly and gradually once we believe that we achieve the level of performance that we require for a certain quality which maps to areas.

And of course and you the other part of the question of you going beyond autonomous vehicles, some of the stuff by the nature of the problem and the complexity and uh some of the research that we do uh is pretty foundation tional uh when we talk about your perception, right yeah can be in a car, you can be in a different modality by Operating in the physical world. Uh, a lot of the research and we publish a lot of work to done, I think can be benefit those community as well. Talk about you know A I uh, being deployed in a kind of real time system and a safety critical system.

A lot of the work that we have done, I think translate I S talk about the evaluation of the system and body applications in beyond time. Vehicles need a good, realistic, scalable, simulate fundamental work translates so, present so on. We are very focused on the is of miles where we can have the positive benefits. So for us, I focus I I think focus is very, very important. So we are being very less yeah they are focused on a can I .

ask go back and I ask perhaps a technical question here like a while back he said, um you wanted to know focus on the full autonomy problem. There are there are many other teams who actually have some lineage in the leg. You know we most suffer google programs that chose, uh, a use case that looks like IT was going to be easier trucking, like long'd trucking, uh, deliveries is not clear that much easier. Do you think there's a lesson to be learned here? Or or at least you know there are more miles being driven autonomously on the road and passenger vehicles by way more than in these other applications today.

Yeah, I think a great question. Big differentiation that I would draw and like ords of matilde and the difference between full autonomy versus scale versus uh you know I drivers the system um and that's the big like that's the cases to your question um of the different vehicle platforms, different Operating to and you can uh slower speed up applications where you know do local deliveries or you can have no a trucking application, a freeways. Um and then there a little bit different but before talking about full autonomy, maybe they're second order differences, but the first story of complexity is still there.

You can uh like that if you think about the core, the heart of the problem of your building, a journalizing and safe driver and are being able to value IT and the incredibly high bar of safety, the complexity of the noisy mesa physical environment and the long tail of yet people are doing all kinds of no where things uh and uh the necessity of making real time decisions where a milliseconds matter and like how hard that A I problem is, the distribution, the counter is change a little bit if you are going to talk about ways or spit. But the fundamental aren't don't get that there is no silver as you don't get to sp the complexity, for example, freeway ste and ominous case there are a bit more structured, I tigers. But uh you still uh encounter with lower frequency, but at higher speed, where are the severe that you encounter, all kinds of things you encounter construction zones you encounter, you know gills and matters is and what kind of stuff falling off of the cars in front of you cars um you know having getting into accidents and kind of spending out and from view, you encounter you know people driving record.

No cars are more ycl encounter pedestrian is j walking. You encounter, uh, all kinds of things, right? And IT happens much less frequently.

So this is a way that might be 2 if IT happens once per million miles。 None of us, i've seen new examples like that in our world. So you can you can lead to the early stage optimist is a simple fiction. If you want to do IT full time and you want to do IT a scale that complex.

And why is that breaking from, like, you know, uh, let's say, advanced striver assistance that IT seems to work in more, more scenario versus that. So full autonomy.

what's was the delta? Yeah it's the number of nights. And is the nature of this this problem, right?

If you think about you know where we started in two thousand and nine, uh one of our first yeah a milestones one goes。 So we set for ourselves was to drive, yeah uh, ten routes. Each one was a hundred muslim all over the barium know freeways.

Uh, downtowns of tisa o around my total, you know everything. And you had to do hundred miles with no intervention. So the car had to, in the drive, autonation, from beginning to IT.

That's the goal we created for about a dozen of us, maybe eighteen months, which you does. Two thousand nine. No image, no continent, no transformers, no big models, tiny computers, no right.

Very easy to be. Just started. It's always been the property. And with every wave of technology, all we have very easy uh, to get started. But that the hard problem and is kind like that the early part, the curve has been getting you know in deeper and deeper, but doesn't where the complex is. The complexities in the long tail of the minimum. And you don't see that if you go uh, for a prototype, if you go for you know the drivers, the system uh, and this is where we spending all of our that's only hard part of the problem. And I guess you know, nowadays is always been getting easier with every technical a cycles nowa you can take with that, all of the advances of N, A, I and special, the genera lamps and s you can take, and almost off the shelf transformers are amazing.

If William are amazing, you can take kind of a violent uh that can accept h images or video and is you know has a decoder or you can give you in a text problems in a book text and you can find tunit was just a little bit of you know data to go from A A camera data and a car to instead of words to trajectories or you know what the decision take the thing because of blogg box. You take me for a let me find you a little bit and like I think if you asking good Grace in computer science to build, no, today, this is what they would do. And out of the box, something that is amazing, the power of transformers, the powers will, is mind blowing, right? Just a little bit effort.

You get something on road. And worse, you can see your drive and tense hours of miles, and we just will blow your mind. But then is that enough? Is that enough to remove the driver and drive in a millions of miles and have a safety record? Now that is, that was really Better than no right? I guess this is know with every technical evolution technologies in a breath for in a eye that seemed like that .

rating is the right way to think of the um the iteration cycle for amo. Now still like many, although A I companies where in eval some set of cases comes up that you don't handle as well as you want and then you collect more data and you put them into the pipeline, you retrain and you deploy? Or are there still architectural changes that are happening even past this point of cranking the night?

So the first thing you mentioned where data collection and I know understanding where performance not gonna kind of building the whole you know data fly wheel and the values to fly well, that at the heart of IT.

But I think there this is, this is where I guess you know a bit unions, you know, what do you do? What is the architecture and what is the training methodology in particular? Uh, kind of the simplest thing you can do is, you know, an end model that is trained just an immigrant ital.

Some drivers, very easy sensors, pixel going, driving in a behavior that you have examples of and you just trying to to imitate human drivers and you can run this a fly wheel and kind of run this circle that you described while Operating. I know that paradigm and you will make progress right very well understood of approach, right? Um you baLance your data, you find some examples you know where you are not performing as well as you would have. Like if you figured out .

you know how to .

value IT look you simulator how started to a mult turn cry this whole machinery even without having a similar do look to the you find more examples where, you know, do you like more, for example, things you wanted imitate and you reduce the number. You find examples where you know you must do something you don't like. So you know, you know, this is not your, and you continue to improve.

You might put two in the wrong place, might put two in the right place for a drivers system, will to a wrong place for a full autonomic system. So then you to build that machinery, and at a high level, that principle of new data and the augmentation. Uh, halle t but you know really go the business to for autonomy, you need to do other things, right? You need to do synthetic data.

You need to do close simulation. Maybe sensor simulation is not enough because only do that scale that is highly vision is not practical, right? Just italy in through the whole thing.

So then you get into like immediate representations. Can you simulate in that place and you still are doing of at a very high level? S, I think IT depends on exactly, exactly.

So I guess, where are you ready today and .

where do you think we're going? I think we are in reflecting an journey that's been quite a few years. I find myself and more excited than ever about where we are, the momentum and the future.

right? I've been doing this well for close to two decades. Uh, the vision was always there, but we had this big existent al question.

Can we build the thing? Can we can figure out was good enough and how to evaluate? Uh, can well people want to use IT? Can we do IT in a way that in a commercially viable? no.

Uh, it's also on uh and you can be go the distance like now where we are today Operating at the scale we are in scaling, we have demonstrated that we can do the thing. We are proud of our safety record, figure out how to evaluate IT. We see that people wanted use IT and give a very positive feedback and people excited about IT.

We see that we can do IT in a way that cost efficient and completely viable. Uh, so I am super excited about what the future holds. Started talk about realizing the mission of actually making, uh, realizing those safety benefits. So now it's all about your optimization, scaling and bring this technology to more people.

amazing. Now, very exciting. Thank you again for joining us today.

Thank you for him. gra. You the through progress over the last year, small on the time.

Thank you.

Thank you. Find us on twitter .

at no players pod subscribed to our youtube channel if you want to see our faces fall the show on apple podcast, spotify or where if you listen that way, you get a new episode every week and set up for emails or fine transcripts for every episode at no dash fire stock com.