Scale like Amazon: How Jeff Wilke is Rebuilding USA Manufacturing

Most robotics demos never ship.

Deployed is about the ones [music] that do.

This is deployed where physically AI gets real.

So, super excited today to be talking to to Jeff Wilke who's been a long time, you know, role model, peer, executive leader, manager, advisor,

investor, like it kind of every And and admirer of Brad Porter.

Well, thank you. Thank you.

Uh Uh but really wanted to like go back a little bit. So,

you know, a little bit of your career arc, right? You you started in

arc, right? You you started in manufacturing and then you brought a lot of that capability 20 year 22 years at Amazon, but then really saw the transformation coming way before

everyone else did as to why it was so imperative the US bring back manufacturing. Um and so give us a bit

manufacturing. Um and so give us a bit of that arc and that context and then we're going to go deeper in in a few of those areas, but really wanted you to kind of set the frame.

Sure. Well, again, Brad, it's uh it's it's awesome to be uh with you. I've

enjoyed our conversations for what 15 years uh together and this one's going to be no different. I'm sure it's going to be

no different. I'm sure it's going to be wide-ranging.

Um you have a huge array of interests.

So, uh you know, I woke up thinking about what could Brad ask me today and there's no way I'm going to guess.

Um so, I grew up in Pittsburgh. And, you

know, I grew up in Pittsburgh in the '70s and I watched the steel mills rust.

Uh and more importantly, I think I I watched the impact that the degradation of industry had on the communities that were around me. Um you know, people lost

their job parents lost their jobs. Kids

and families were impacted. Um

unemployment was very high in the area.

Uh other you know, other than the Pittsburgh Steelers and one great season of the Pittsburgh Pirates, there wasn't a lot of hope. Um and I think that kind

of story played out across the US over 30 or 40 years where we were optimizing for short relatively short-term financials.

It was harder and harder to, you know, convince investors to support long-term investments and property plant equipment sort of capital investments.

Uh it was easy to find partners in Asia and elsewhere who were willing to do things at much lower cost. Uh and then eventually include engineering and design so that, you know, all you had to

worry about was a brand and a market and everything else would be taken care of.

And and that seems like a very thin layer of uh expertise and and core competence, you know, for for companies and ultimately for industries. And

I think what's played out is that we we've just hollowed out these really important core capabilities that allow us to make the things that that we depend on for wellness, for

energy, for transportation, for entertainment. And it uh it makes the

entertainment. And it uh it makes the society and ultimately democracy more brittle. And it I just growing up in

brittle. And it I just growing up in Pittsburgh cemented to me the importance of having a robust industrial economy.

And I spent, you know, I spent a couple years as a software developer as a a bit of a detour. Uh and then I just started reading all this stuff about in the late '80s about the decline of American

productivity uh and the impact that would have on the the wealth per capita, basically the resources for our society.

And um I decided to go back to school to study manufacturing and work there for 6 years and then Amazon called and I did a little detour myself.

Yeah, let let's go let's go a little deeper in that kind of Amazon history.

So, Jeff Bezos tells this famous story about how in packing books he he thought they should get knee pads and some engineers said maybe we should get maybe we should get desks, which he thought was a

brilliant idea. So, how much how much

brilliant idea. So, how much how much beyond desks packing books was it when you arrived? And then you obviously

you arrived? And then you obviously brought a really um you brought a transformation vision to Amazon of how you thought logistics really could work. And and so tell me a

little bit of that story. Well, look, a couple things. The the company had a

couple things. The the company had a bunch of the raw capability that uh that was going to turn out to be really important beginning with software.

So, the idea that all these physical processes have some instantiation or twin in the digital world was present even in the even back then in the architecture that Amazon was building

and it had very competent computer scientists compared to most people in the physical world. Um what it didn't have was a point of view that would

match the ultimate sort of factory physics, the challenge of moving all these things through a warehouse especially at peak times like Christmas.

Um and and that's something I happen to have by accident because of my focus on manufacturing and the work that I had done for 6 years in the chemical um metals

electronic components industries, I learned a ton about how you manage bottlenecks, how you uh sequence work to get the most out of your uh whatever your most

capital intensive bottleneck piece of equipment is. I learned a lot about

equipment is. I learned a lot about ways that you can improve quality and lower cost at the same time including things like statistical process control and even Six Sigma was popular at the

time. But ways to look at variation and

time. But ways to look at variation and characterize it and reduce it.

And and these things were and also the how to run a complex plant that's more complex than a plant that simply has and I say simply with no due you know, disrespect to people who operate complex

warehouses, but when you're moving pallets around with fork trucks or maybe cobots, it's a simpler challenge than when you're assembling orders from multiple

items, maybe millions of items and you have a very short time to get the first and the last item into a slaughter of bots before you plug up the entire system with half completed

orders, which was really the challenge that they had at peak. Uh and I just walked in with a set of tools that had been honed over decades by really really smart people working in places like

Toyota.

Um and we didn't have time for me to do first principles redesign of this whole thing. I had a playbook.

thing. I had a playbook.

My instincts were that it would work and we were very fortunate that the principles uh scaled from a billion to, you know, now, you know, essentially a trillion dollars going through the the network.

I want to I want to get back to some of that kind of innovation in the network and how you think about that and then ultimately want to kind of rebuilding what you're doing now. But I I want to stay in some of some of this early

Amazon lore as well. I think Jeff Bezos not too long ago um told a story on stage crediting you for for coming in and saying, "Jeff, you're the number of

ideas you're producing could destroy the company." Uh I think, you know, we have

company." Uh I think, you know, we have a lot of founders who are who are tuning into this. I think

into this. I think I think they're interested, you know, at the same time Alfred Lin sent out something to kind of sort Sequoia portfolio pointing out that um you you

have to be kind of scanning for opportunities and maybe be willing to position your company to to get lucky with new opportunities or to seize new opportunities. And so, you

know, how did uh I'm sure there's much more to that story with Jeff. Like how

did you guys find the right balance between you know, exploring and then focusing on, you know, exploiting the opportunity that you had in front of you.

Yeah, I mean this is the the balance is the keyword and you and actually might a better word might be tension. You want these to be in

be tension. You want these to be in tension, not necessarily balance. Maybe

implies too much comfort.

Um so, the first thing is I my brain tends to think in terms of process in most of the things that I encounter in the world.

So, I think about um you know, flow maps, um uh topography, uh and, you know, how things move uh

together and sort of systems dynamics of how those movements occur. And so,

anytime I encounter a bottleneck, whether it was in a a plant, in software, or in an organizational design, I would apply the same kind of thinking, which is how do you

maximize the use of the bottleneck and prepare everything else to be subservient basically to that to whatever you've chosen to be the bottleneck. And by the way, one of my

bottleneck. And by the way, one of my one of the principles that that you and I talked about a lot in the early days is that I think you have an operating plan when you choose the bottleneck.

When you decide where you want the bottleneck to be, if the bottleneck reveals itself to you like in a surprise, you know, month after month, quarter to quarter, year after year, you don't really have an

operating plan. You have a reaction

operating plan. You have a reaction plan. Those two things are are pretty

plan. Those two things are are pretty different. So, I found myself, you know,

different. So, I found myself, you know, uh focusing more and more on the bottlenecks of the company after we you know, started to make progress in improving the plants. And uh it was very

clear that Jeff was an idea machine.

Like he just which is what you want. And

and I think Alfred is correct that you want to have people uh especially if you're fortunate enough to have them at the top of the company who are generating really great ideas. Now,

there are people that generate a lot of ideas and most of them are terrible.

Um in Jeff's case, he generates a lot of ideas and a lot of them are really good.

And he's, you know, his brain is constantly working on, you know, sort of innovating around challenges in the world. And I loved having that stream,

world. And I loved having that stream, but I watched the stream get like it would hit the the bottom and so it hits the bottleneck of like the company's ability to implement any of it. And it

stuff flying all over the floor cuz basically what happens is the when the org starts to drown, it builds a shield and then the ideas just hit the shield and just just scatter all over the place. Um and

nothing gets through. So, what I think that comment did for Jeff and later in that little thing that a couple people sent me and I I sort of listened to what he said, he he's

correct he identifies the other part of the conversation I think is is even more important, which is you could say, "Well, I I have to slow down the pace of ideas.

That would be one approach for Jeff to have taken. But fortunately, Jeff didn't

have taken. But fortunately, Jeff didn't take that approach. He said, "Well, what do we need to do to relieve the bottleneck so that we can process all these ideas?"

these ideas?" And and what do we need to do organizationally? What systems do we

organizationally? What systems do we need to build? What constraints do we need to put in place?

Um and we worked on that for for years together and with the S-team to figure out how to organize mechanisms and uh you know, and a human hierarchy

so that it could absorb many more ideas.

And I think that the trick is an operator a great operator is always going to be in tension with a a great innovator. Now occasion very

innovator. Now occasion very occasionally they're in the same person, but I've there's almost nobody I've ever met who is world-class at both.

And the trick is that the operator while they're going to feel like the innovator is throwing too many things at them, the operator has to catch enough of them

to be better than the competition.

And the innovator has to have enough awareness to not drown the operators who are doing things that are way more complex and hard than they appear to be

when you're an innovator. And this is one of the the asymmetries of this kind of stuff is that you know, operators generally know how hard it is to to innovate. They might put put it

innovate. They might put put it sometimes, but they actually know they're you know, they're not great at it.

Uh sometimes innovators gloss over the hard details of operating that make it really hard to instantiate these things in the real world.

Um and so hopefully you get that tension at the exactly the right strength and force so that you build an amazing company.

So so I joined your organization I think in about 2014 um and it's about the time

uh that Amazon unleashed one of its boldest ideas uh on 60 Minutes um with the drone program. Uh

give us as much as you're willing to kind of the backstory of how that came about cuz I don't think it's I don't think it's told very much. Um everyone

just assumes it's like Amazon working backward, but like there's a little bit more to it I think.

Well, one thing that's really cool is I I just happened to check in with a the the team um um a week ago or so and it sounds like uh they're making real progress. Lots of deliveries are

progress. Lots of deliveries are happening. So you know, you probably

happening. So you know, you probably have a lot more information than I do, but I keep I keep up with those folks as well. Yeah, I know it's

well. Yeah, I know it's it's amazing.

it is super cool. Um

so that's one of those things where uh a a person Gur Kimchi uh came to me in 2013 and said, "Hey

uh somebody's going to build a drone delivery system because we now have basically the onboard compute that's fast enough. I

mean that's what kind of we were waiting for, right? If you if you're not going

for, right? If you if you're not going to fly a dumb bird you want something that probably has some sense and avoid capabilities you want something that uh has

autonomous flight control on the on the vehicle in case it loses coordination.

Um you you have you have to have enough compute and that computer has to be light enough.

And it kind of took into the 2010 time frame before you know, maybe a little bit before that, but before we could actually think about doing this at a reasonable cost.

And he saw it and said, "I think we could do it. I

think we could build a you know, a drone system for delivery. Um we can't use a quadcopter.

Um you know, that we could start with a quadcopter, but we're going to have to probably build something that's a little bit more complicated than that and eventually it's very clear that we we needed to have wings of some kind or

it's just really inefficient. But uh he was already thinking about all this stuff and and uh he said, "You know can will you fund it?" And I said, "Well, you know, this is this is Amazon.

The way that we think about new ideas is we write a press release and working backward documents and we uh we try to imagine the product at the time of launch and encode all of that in the in

the working backward documents.

And so he wrote a document about the experience that we might envision and then we did then some cost analysis to say is it even feasible that we could do this.

And to my surprise it looked feasible. And

feasible means that you could imagine a trajectory where the cost of a delivery by drone would be less than the cost of a delivery by truck.

And you know, it it wasn't it didn't require fundamental science to get there.

So you know, next step was to uh kind of fund a team.

And I didn't I didn't talk to Jeff. I

didn't ask the CFO. I just funded it.

And uh Gur hired uh a group of people uh to get started on work and they worked in in one of the guys' garages uh

a garage for for a couple months to put together uh a drone and prove you know, kind of proof of concept. And then

Jeff and I were having a conversation at at uh at an S-team meeting and he said, "Uh you know, I've been thinking a lot about drone delivery." People had started to

drone delivery." People had started to speculate about it, but nobody had come out with anything. He said, "I I actually think you know, this is going to matter for us in the long run and we should get started."

And then I said, "Well, turns out there's this very small team that's been working on this and uh you know, they've made a lot of progress. Way more to go, but uh you

progress. Way more to go, but uh you know, we have some idea kind of of what it might take." And he said, "Well, you know, I think he said

"Can we uh get something flying and uh have a video in like I don't know, 6 weeks or something?" Right? I mean it was [clears throat] if we had started from nothing there would have been no chance.

Uh but we were far enough along that we we were able to put together the background and the um and the experience that was featured in 60 Minutes and the rest of history.

No, that's awesome.

I and yeah, I mean your contribution to that that was really the uh an opportunity for us to work very closely together um on an operating thing.

We'd done a lot of architectural work together on the software side, but this was like a different kind of experience for us and you know, your your maturity, your understanding of

both the physical instantiation of these you know, algorithms and capabilities and then the software side that was going to be necessary to orchestrate them uh was incredibly valuable for for that

team.

Um and you're a great recruiter of talent, too. So you brought a lot of

talent, too. So you brought a lot of great people to that. I mean it's still I mean I've run people who have that on their LinkedIn and they're there's an amazing team of of you know, graduates of the primary

program.

No, I'm super proud of the the team we built. In fact, I I took the recruiter I

built. In fact, I I took the recruiter I worked with there and and brought her in as employee number four at Cobalt. So

uh she's built the team here as well and um You know, Yeah, that that if I if I think about that if if there's there's one thing that I I just I wish we had been able to

move faster.

Like and that's that's on me. I mean I I uh you know, there are there's this tension this is another time there's tension between getting the design to be you know, right. And when we had the when we

know, right. And when we had the when we ultimately got to that design that had the foil for lift and a much kind of simpler approach to

uh meaning no need to actually move the angle of of propulsion. We could do all that with uh with software control. Then

we had a machine that I knew would scale.

Um but I think we we were slower than we should have.

You know, it's fascinating cuz I I took a lot of lessons from that. I think the um cuz the first time I was really integrating hardware, software, and AI,

right? And and then operations. And the

right? And and then operations. And the

it's obviously a systems discipline, right? To bring all of that together. Um

right? To bring all of that together. Um

and I think one of the lessons I learned from that was that um when you're trying to bring the whole system architecture together

more brains don't actually help you. You

need you need vertical you need like these T-shaped people who have kind of broad systems breadth and vertical depth and you need you know, maybe a dozen of them, right?

Who can kind of cover each of the different functions from aerodynamics to you know, to the AI, but um but we did get that team pretty big fairly quickly and

now when I you know, when I move over to robotics and and even at Cobalt when we're doing that early systems work I actually try to keep the number of the number of folks smaller

um to really iterate really fast on the system. And then once you get the

system. And then once you get the system, then you can explode it. And so

I I think we we almost went too fast early, which causes go a little slower in the middle. I think you're right. I

also think and I think about this a lot now. If when you think about work design

now. If when you think about work design for small new things and even for you know, established things um I I always ask uh designers to think about two main

principles related to how humans and machines work together and one is for every team every team I think you have to ask the question like so if you're running an established business and you have teams

that are not using AI for much workflow. First question is, you know, what is the agent that would enhance the capability of each team

fastest and in a most powerful way. And like

build that agent, you know, in a week or less. And I think what we would have

less. And I think what we would have done is said, "We don't need as many human engineers. We need these amazing

human engineers. We need these amazing human engineers and then at least one amazing agent that's a machine for specializes in each area where we have a

human team.

And that would have have allowed us to I think scale faster without you know having to add as many people which just introduces management challenge.

And then the other thing of course is that every manager I think should be we should think of them as a hybrid manager meaning every in every management job on a box there's a human

and and a machine sort of complement to that human that makes the human manager way more effective and allows them to have larger span of control a flatter organization move

faster and so I think I think it's so And yeah it's it's tricky to bring hierarchy in too early when you're building a system on the other hand you

need that kind of like lead engineer chief engineer whatever to to help kind of I do remember yeah we we started to go a lot faster when we started to build

the simulation tools and the AI tools to to look at different morphologies of of drone because that you know that ultimately was the the key to getting the kind of range and and performance

and you need that range to make the economics work and Well let me let me fast forward a little bit so and I'm going to tell the story as I

remember it but then I'm kind of curious your your take of the story Oh it's amazing plus there's it's not only is it a different perspective but there's so much time which I'm finding more and

more that time really distorts stories but please let So so here's what I remember so I remember some time in

I think it was either late 2016 or early 2017 I think it was maybe like January 2017 is what I remember I remember a meeting with your leadership team that you called and you would you would asked

your finance and economists to to build a long range forecast to build kind of a 10 year forecast and what I recall is that that forecast

said by 2021 Amazon would have a million employees and as I recall it was about 200,000 at the time and we weren't sure how we were going to scale and

I just remember like the whole room not believing it right I remember everyone being like the the data must be wrong right how are we going to have a million employees

and then I remember the obvious takeaway was are even are there even a million people who want to work for Amazon right and and how could that even be possible

and therefore how do we how do we bring automation online faster and uh you know that that was kind of like my first like peak of like oh this could be a a

meaty enough problem to get really excited about for myself but I'm I'm curious there's probably a little bit more backstory there's probably a little bit maybe I don't know if you were trying to set up that

that like inspiration light bulb for folks or if that was kind of organic like I'm curious if you maybe even if you remember that whole meeting I do in fact it was

it's this visioning thing that over the years I've tried to use as a way to think about preparing for scale and it's just like it's very simply

you know compound out growth in what whatever is the core measurement of the thing that you're working on and and get it to the compound it to the point where it's

you know substantial enough that it causes that kind of a reaction and then ask some questions like is there enough capacity in the world you mentioned are there enough people that might want to work at Amazon it turns

out there were you know you you have great benefits and you created a nice environment and and you know in places where the alternative is you know a much worse job like you can you can get lots of great people to

be excited about working there but but what else might limit you and it could be physics it could be engineering capability and it turned out I I just started asking like

what and this started in 2014-15 but I I worked in this with Melissa Eamer who was my technical advisor at the time and we just started saying so

are there enough cardboard boxes at the scale we're going to be turns out there was plenty of cardboard you know is there enough steel for shelving yes

you know are there enough places where you can put warehouses yes one of the interesting things was you know is there enough capacity in the US delivery system the answer was no

so so there were two or three things that came out of that analysis that changed my view of of our strategy one was we had to build a a much more significant transportation

capability or we were going to run into the slower than necessary pace of deployment of capital by the player the incumbent players in the industry that's

why Amazon built you know it's it's air fleet that's why it it does you know you get deliveries by Amazon drivers and then all this stuff that

I think Dave Clark drove brilliantly to scale came from that analysis which was like wow we we are going to you know we're not choosing our bottleneck if we if we let things play out

the you know the incumbent transportation carriers will create a bottleneck for us which is unacceptable and so we have to choose the bottleneck and we're going to choose to vertically integrate into transportation because

otherwise we can't support customers you know and we made that choice and I was super super proud of the way the team executed there

but the other one was was management and one of the conclusions that I reached from this thing is if we're going to be a million people we're going to have roughly like 10,000

managers in the retail part of the company managing the whole thing and at the time we had between 1,000 and 10,000 it was probably like 6 or 7,000 or something but but you could see it

growing to 10 and then who knows [clears throat] from there and I started to ask the question which you remember of my team regularly which is why would we ever need more than 10,000

managers at any scale to operate this thing this thing being you know a retailer and my conclusion was we we have to

invent a way to not need more than 10,000 managers and maybe need less so that they can humans can work on other things than simply exchanging ideas that

already exist from one person to another and compressing mean you know you if you if you wrote down the things that that management bureaucracies are great at that you know it tends to be like memory

compression you know some kinds of certain kinds of computation these are all things that LLMs got really good at but in 2017 or 18 you

know I wasn't smart enough to read the transformer paper when it came out and then leap to okay the singularity or whatever is actually getting closer I

was just saying to the team if I just look at first principles we're going to hit a bottleneck that is management and we're going to have to reinvent management science if we're going to continue to grow and now we know what

that reinvention is it's learning how to work side along machines so alongside machines so that we don't end up with you know bureaucracy that overtakes everything So not too long after that meeting

I remember Dave Clark kind of assigned a goal against himself to hire a robotics leader cuz he said there were kind of three big bottlenecks transportation

management and and automation and and I remember coming to you and saying hey I I might be interested in that um but I'm I'm curious I mean you you have

a kind of a way of looking at talent that's a little bit different and I'll admit like you know I was a distinguished engineer at Amazon I had a lot of credibility and authority but my background was was large scale

distributed systems the e-commerce platform architecture the you know migrating the company off of Oracle databases like that was kind of my like wheelhouse and so I'm curious

like what cuz I'll admit like you know I you go into those meetings if you're interested in the job you you you put all the imposter syndrome aside right and just raise your hand and say I think

I can do this but meanwhile like I'm certainly sitting here being like I you know I haven't been in the operations side of the house I'm I'm curious like again from your your side of the story

when I came into your office like what did how did you look at that what what were you seeing maybe in in me that would give you the confidence I might be able to to tackle this um

yeah I'm curious kind of the other side of that conversation Well what I try to do is is understand the mental models of the most talented

people in an organization that I'm responsible for and and then think about how because those those mental models when they're rare really a superpower and then what what mental model is

likely to have the most impact for a particular problem with particular resources and in this case I'm talking about you I mean if I

described your superpower which I wrote to you on multiple reviews you know I think you have the the ability to build

the the best mental model of an architecture of a complex architecture and and especially to rotate that model around

and aim your attention at whatever thing needs focus without losing the context of the things that are around it a lot of people can kind of spin up on something navigate to a place work on

you know that part of the topography and then you know work sorry the topology and then work on another part of it cuz this is kind of like imagine that we have a graphical

representation and then work on another part of it.

Um but they they can't they can't remember enough about the connections between the pieces of the topology to to optimize them in the way that you can. And I I

always like that's why I I thought you were a great sort of chief architect with me as we were thinking about moving from the old systems to a you know a uh

a native a native AWS uh system. And um and in this case it

uh system. And um and in this case it was the same kind of thing. We I thought that the problems would not be um

let's call it the atomization of the automation. So, you know, in the

the automation. So, you know, in the instantiated in a in some kind of robot.

I actually thought our our largest challenges would likely be orchestration.

And um and we were going to need both.

We're going to need the clever ability to you know, disaggregate work and uh and then reaggregate it in different forms.

Preferably less complex than the forms that they replaced and more flexible.

More precise, flexible, fast, lower cost. Um and

cost. Um and and then have an orchestration layer on top. And you were the only person in the

top. And you were the only person in the company that I thought could do that with a partner like Dave who you know, has a a different way of looking at the world,

uh a different kind of communication style. Um but the two of you together I

style. Um but the two of you together I thought would be a really powerful team.

And I as you recall, I counsel you that there would be times when you drove each other nuts.

And uh and I would ask you to just you know, listen because the insights you'd get from Dave would be pretty you know, remarkable. And I asked him to do the

remarkable. And I asked him to do the same thing and I told both of you guys if you if you really can't agree, you know, come find me we'll go to lunch and figure it out.

Now, I I think that ended up being a really obviously super productive partnership and I think the yeah, Dave's ability to kind of see

the both the network level process flows but also just what was really going to work in moving um ho- hopefully we'll get him on one of

these as well. He uh I'm sure he will.

You know, I I do the same thing. I think

about people's mental models and that you know, uh maybe even a little more reductionist in that like my mental model for you Jeff just ended up being like the closed loop feedback diagram. I'm like just just

feedback diagram. I'm like just just assume that everything going through Jeff's brain is the classic you know, closed loop feedback diagram and I just I would teach as many people as I could in your organization to use that model

and it helped them tremendously to figure out how to communicate with you.

And then Dave, you know, Dave tells a story where he uh he would help move instruments around for a band in like high school and college. And so he was he was always just thinking about like

how am I going to get all this equipment from here to here with the smallest truck and pack it all in and he just scaled that to like the globe, right?

And um and so that was that was amazing to work with him for.

And then yeah, I I was able to look at hey, this system applied to this problem with this software integration uh with this AI capability. And I think you know,

another one of my superpowers is like to really understand what's in a short reach of where the technology is right now, right? Like I I never like to take

now, right? Like I I never like to take on the like impossible problem that's 10 years out. Some people love the like I'm

years out. Some people love the like I'm going to go work on quantum computing or or fusion energy and it's like We need We need like those things are going to work. So, we need people to be thinking

work. So, we need people to be thinking about them. That's right. Um but for me

about them. That's right. Um but for me I'm looking for the like how do I make impact fast, right? And so and I can do that by solving technology problems. Um and so I'm looking for that like what

are the hardest technology problems that I know can be solved fast, right?

Because if I can solve them faster than anybody and get them out there. And so

the robotics was just an a tremendous playground for me, right? In that there were huge number of opportunities to apply um and that's obviously the the thesis

behind behind cobot was that you know, these type of robots can um can solve real problems in in hospitals and logistics and manufacturing

right now. It wasn't a a massive leap to

right now. It wasn't a a massive leap to go do it and then they get better. I

think this is the other side looking toward the future and you know, I I do a similar thing of kind of rolling into the future and trying to understand what capabilities are going to come along and

do you get better as those capabilities come along.

Coming to automation though, one of the things we've talked about since we left is that I think both you and I have gotten to see more operations that aren't Amazonian operations

um since then and I think we've both been surprised that the level of automation, the level of software control, the level of you know, operational rigor is is not

at the levels that we kind of expected it might be. And I'm I'm curious if you have a thesis as to why but also you know, we're obviously working a lot with

customers to try to transform, right? To

try to embrace and adopt you know, how do you do a transformation to automation, you know, digital physical process. But I'm curious like where you

process. But I'm curious like where you see some of those bottlenecks. This is

becoming a complicated question but there's another piece to it that may be tied in which is I think one of the things Amazon did really really well and I'm curious if this was really intentional or just kind

of cultural um was hired operators who were excited about change cuz not every operator is excited about change. So

anyhow a meaty question but I'm curious your take on like what's it going to take to bring more automation into um

into kind of North American supply chain and manufacturing base. Let me start with with the last thing you said which is you know, hiring people that are excited about change and then what kinds of

people like that. Um I I think that's going to be vital. I I mean we are we are going to be uh changing existing systems. So, the way I think about the the sort of

physical world right now is uh of industry is that there are just too many deployed assets. Yes, some

costs are sunk.

But we we cannot and probably from a financial perspective should not rebuild absolutely everything in the world with completely new technology. That that's

just it's it's it's not going to be efficient and we're going to we don't have the resources to do it.

So, we're we're going to actually have to use some of the things that already exist and improve them over time

alongside things that are new uh new technologies, new new robotics, new AI.

And the the the sort of hybrid portfolio I think is is likely to be the best way for an established firm to proceed.

And the the people that um are going to be most successful with this are the people who are not afraid of change to your point and who have enough

technical understanding of the process that they can work alongside you know, uh a um cobot

uh coder as as well as cobot machines to instantiate the understanding that they have into you know, better and better and better designs for process and

ultimately for product. And so it's like so you want you got to have people who have really good mental models of of the inputs and outputs of the thing that you're

using to drive value for your company.

And the more technical the understanding in total, not everybody has to be an engineer but the more technical the understanding in total the better. If

you've outsourced all of your technical thinking it's going to be way harder for you to make this transition.

And uh so as we started to build this company Rebuild Manufacturing which is at the very beginning we talked about you know, my journey back to manufacturing and um the goal is that

uh we want to help ensure that the next generation of important products are made in the US.

And this starts with great engineering.

It starts with great process and product engineering.

Uh when I run into when I do a plant tour or run into people who are running manufacturing companies I'll talk to them about their process and they usually want to talk about

electromechanics.

And you know, if they're if they're technical at all, they want to talk about how we sort of move things, make things, transform things.

Um and then they talk about orchestration and software and stuff as a as like a you know, an add-on.

And a simple question for them is who's the most which I used to ask. Now it's a little more complicated the way we're writing code these days but I would ask who's the most senior computer scientist in your organization and to whom do they

report?

And you know, at Amazon it was the CEO.

At you know, at most operating companies it's somebody who hasn't written code in a long time who doesn't work for the CEO. And you know, the next easy

CEO. And you know, the next easy question is like so do you have a chief mechanical engineer? Well, of course

mechanical engineer? Well, of course we're a car company. I'm like well at the at this point a car has as much compute on it as it does electromechanical. So, why on earth

electromechanical. So, why on earth would you not have a senior computer scientist uh architect who's sitting on your senior team alongside your chief engineer who's almost always a

mechanical engineer.

Um and and that's kind of the problem. I

I think you you have to not everybody's going to be able to write the kind of software we wrote at Amazon for orchestration.

But I think the the companies that are going to make this jump in light speed to wherever we're going are going to be the ones that are really good at characterizing in an engineering way. So, the

way. So, the the inputs and outputs and the state and I'm going to reveal myself thinking about this sort of control close up control thing, but really understand what is observable and what's

controllable about the current state.

And then how would I take advantage of AI and robotics to morph from what's honed now to what

can be way more flexible, productive, fast, robust, and low cost.

Awesome.

Yeah, tell me tell me a little bit more.

So, you you know, you've talked you just talked about this kind of hybrid model of building off of what we have. The

have. The you know, Rebuild has taken a strategy of bringing together not just engineers, but also some of the manufacturing components to kind of start to build products end to end. Give

me a little bit more kind of the thought process on it's not it's a non-traditional like venture way to build a business, right? On the other hand it's it is a model that that other build you

know, private equity and other places.

So, give me a little bit the thought process on how you guys approach building the company that way.

Well, the first thing I wanted to do is make sure we weren't constantly fundraising. Um I wanted to

constantly fundraising. Um I wanted to actually be constantly building.

So, we were fortunate to to launch the company in 2021. We we got access to a lot of capital, which was just luck.

Um and we took on a lot of it including a lot of dilution up front, which I was totally okay with and that put my own capital in and then people that that I knew and trusted. And we took a very long

uh duration point of view. I mean, it's it's it almost echoes Jeff's letter to the shareholder Bezos's letter in 1997 where he said like if we have a choice between long-term cash flow and quarterly earnings results, we're

taking the cash flow every time. And

that's the message I gave to the investors was we're not optimizing this thing the way a private equity firm would where we're buying things, we polish them, we improve them, and then we try to sell them, you know, in a few

years.

We we buy them expecting that we probably won't sell them. Now, if the right thing to do is sell them, of course, we would love to you know, dumb, but but we we decided to buy enough capability so that we would have a

complete portfolio of engineering talent. So, we have

talent. So, we have electrical engineering, mechanical engineering, material science, chemical engineering, um computer science, AI. Uh we actually have a relatively large tech team of of

computer scientists and AI for our size.

And um uh and we it's a it's a complete set of those capabilities that we built process around so that people that were in a small standalone

company that only were good at mechanical engineering have great connection to a world-class aerospace design company that we have in Denver called Answer Engineering. And together

they can build structures that neither of them would have been able to build at arm's length in the past because as you know, software is still the only engineering discipline that has hardened APIs.

The interfaces between every other engineering discipline are way more squishy.

You know, we have P&ID diagrams. We have ways to describe interfaces, but they're they don't have the sort of clarity that software APIs have.

And to me that's why you need collaboration and among the other engineering disciplines that that the market and arm's length transactions don't actually achieve. So, we decided

build a complete base of engineering, buy some production plants because the instantiation of our idea, the success is going to be that we build more plants

making things that we need in the future in the United States. One of the reasons that that's such a key mission for for me or way to describe it, I I think that um y- plants have this advantage of

concentrated capital.

They tend to not be all clustered around the population like they're not all in LA and New York. You want to go to places where you have land and you have, you know, highways and stuff. So, so you end up with plants likely to be

dispersed.

They they're hard to move because you have a lot of heavy things that are expensive in the ground.

You can move them, but they're hard to move. And so, you have enough time

move. And so, you have enough time for communities to of humans to grow up around those plants.

And like the most important thing I think we could do for the stability of our democracy is to have, you know, hundreds of communities around the country have plants that they can

depend on for, you know, at least a large enough number of years that you can get, you know, from first grade to, you know, to graduate kind of thing.

Um you you can't guarantee a plant's going to be there for 50 years, but if you can if it can be there long enough, people can have stability and build the kind of communities that that, you know, we sort of um

uh we dream about from from the past.

And um and we're not going to recreate the past, but uh but I do think that we if we combine all of these capabilities, uh and Rebuild won't be the only company

doing this.

We have a real shot in the US of kind of rebuilding the capability to control our own destiny in the things we make.

All right, so lightning round. So, three

questions for each of us. Um I'll go first and then Jeff, you shoot one and we'll we'll go back and forth. So, all

right, Vlad, coding. What are you excited about? It's my New Year's

excited about? It's my New Year's resolution.

That's awesome. Awesome.

Shoot one back at me.

Cobot or humanoid?

Oh man, you stole my second question, but I I obviously I'm I'm well on the record of thinking that you know, the the the thing I've been thinking about lately is that like humanoids are this like least common

denominator, right? They are not going

denominator, right? They are not going to be the best at anything. And so, you know, when you build a robot for all the logistics in a hospital, why would you use the least common denominator? Why

wouldn't you use a robot that's that's really suited for that? You know, I I get why you won't build two or three of a robot, but if you can build a few hundred thousand robots for hospitals you're that beat a least common

denominator robot, you're going to do that. And so, I think, you know, I think

that. And so, I think, you know, I think humanoids are going to have to compete with some great hardware and they're just going to lose most of those competitions for many domains, right?

There may be some domains where, you know, the generic kind of helps, but um anyhow, I'm I'm on the record thinking that we're going to have really, really great capable robots that are um

that beat human performance on multiple dimensions.

You agree [clears throat] that they'll they're probably going to be humanoid robots and some and cobots.

I think so, but I think like I still struggle to Okay, so I have a a framework for robotics. I think you you first of all, you need to be able to do the work, right? You need to be able to

perform the process path, right? Um and

second, you need to be able to do that with unit economics and costs that provide a a positive ROI. But third, you kind of need to be the best solution at

that in that domain, right? And so, you know, I look at like storage systems. Obviously, you know, things like Ocado or AutoStore or others. What happens

over time is initially they both get a bunch of sales, right? But over time people start to figure out which one performs better and the market shifts toward that one, right? And it you know,

it can take five or six years. These

days timelines are getting compressed, but it can take five or six years to figure out which one's the best one. But

by the time everyone figures out which one's the best one, then one of those is getting a lot of orders and one of those is getting very few orders. And so, you know, I think yes, humanoids are going to get used in a whole bunch of places

initially well, if the AI is good enough and the safety is good enough And and they're robust enough. I mean, the the more sensors and actuators you have on board, the more they can break.

And they need to hit some the unit economics where, you know, you can actually produce a large number of them.

So, I think I think they get outcompeted though in almost every in almost every space. I think there are some like

space. I think there are some like entertainment style like if you're going to have a robotic bartender, you kind of want them cracking jokes and like, you know, you want an attractive bartender that looks nice, right? But

cobots cobots don't crack jokes?

Oh, they do. No, they do. They do. But

they're like for doctors and nurses and surgeons, you know, not not uh not good bar jokes right now. Um

so, tell me like drones and robots. We've talked about it. Like what is there a big difference

it. Like what is there a big difference with if let's let's say that it's important to our national security that we be good at drones. Do we Do we if we get good at that, do we also get good at

at robots or is there something fundamentally different?

This is a really interesting question. I

I think um the difficulty in these kinds of physical systems is the is the variation. And this is why there going

variation. And this is why there going to be probably humans in plants that are trying to be lights out for 30 years even if most of the stuff is automated in 10 because there are transients,

startup, shutdown, things break, like stuff happens in a complex physical system. And when stuff happens, you

system. And when stuff happens, you know, if if you haven't seen enough of it, you can't automatically fix it. You

need, you know, a human to to kind of be involved or or some other, you know, agent that can work in a different way from the main line. And um

line. And um I would just say like flying has fewer sources of variation than, you know, operating on the ground.

And you know, and of course, that's a very general statement and there's complex environments that, you know, I've never been in because I'm I'm not an ex-military person, but um you know, I think in terms of

controls, we have we've we've had automated flight control for decades versus you know truly automated

terrain you know independent vehicles that can kind of navigate the physical world. And so I like I think

physical world. And so I like I think that getting autonomy in the drone space the work is going to be in the kind of orchestration of the you know what you

want the drones to do and how you want them to behave in in battlefield conditions and how you want them to act when something doesn't work out. So how

are they self-healing networks and that sort of thing?

All of that is going to matter. It'll be

helpful to have that in the ground-based stuff. But you're also going to have to

stuff. But you're also going to have to have you know a lot of the data that I was talking about before collected about the environments that ground-based

things would operate in before they're going to be as robust as things that operate in the air.

What what shoot another one my way. So

when do humans build the first lights-out plant? And I mean like a

lights-out plant? And I mean like a manufacturing plant that literally has there's no humans necessary to operate it maintain it it just it just runs.

One funny thing is we say lights-out and then I've had some people ask like why would we turn the lights out? We're

not actually going to turn the lights out. Lights are for quite useful for

out. Lights are for quite useful for [laughter] for robots to be able to perceive the world. But

world. But I I used my like AWS customer card to to get a tour of the Gen 12 plant in

Shreveport Amazon's Gen 12. Have you

seen that one? No not yet. Yeah

try try to work your way in cuz it's it's everything we had on the design table in you know 2020 when COVID hit you know Amazon finally built it and I

mean 25% productivity improvement 25% speed improvement and it's funny because when we did that analysis and we automated we said what's going to be left? When I get mostly going to be

left? When I get mostly going to be people pushing carts like handling these kind of secondary material flows right?

And sure enough as soon as I walked in the building the first thing I saw was someone pushing a cart full of totes right? And I'm like wow we we were right

right? And I'm like wow we we were right we projected into the future properly. You know I I want to go tour

properly. You know I I want to go tour the Lego plant because like the world just fills up with Legos and crayons. I

feel like those are almost fully automated already.

You know coffee processing seems to like just show up at Starbucks with so you know I think there's more that is like highly high automated than than people

realize. But the connections the the

realize. But the connections the the next connections are some of that like ground-level material movement the exception handling paths and the replenishment paths right? And so you

know we're working with a big manufacturer right now on handling a bunch of those replenishment tasks.

So when those start to come out and then the transportation right? The load and unload of the truck and the truck itself being automated. I actually think like

being automated. I actually think like for a number of industries bottling or other things where like it's really been these complicated machines that really automate everything

we might not be and you know AI-driven and statistical process control the computers are driving a lot of it. Like really once we get the

of it. Like really once we get the secondary material flows handled which is what you know Cobot's heavily focused on and then the autonomous load and unload and vehicle movement we're not we're not that far. So you know 10 years

7 years 12 something in that window. All

right I I give you a very specific question though about lights like truly lights-out no humans anywhere around it.

You really think 10 years?

I think for some some very dedicated manufacturing capabilities I think it can be completely inputs in and and outputs out autonomously. I do.

autonomously. I do.

And and machines maintaining the machines?

So you mean no no technicians ever coming in the building?

no maintenance people.

that's further out. That's

Okay. That's that's a good 20 30 years.

All right and we we agree as always.

Yeah yeah yeah.

The okay next question for you.

You think a lot about you know bottlenecks and where things are coming.

Obviously we're in this AI explosion.

What's the bottleneck you're thinking about that maybe people are everyone's talking about data centers and energy but is it is there another bottleneck you're you're thinking that people might be missing?

Well I don't think they're missing it but I think it's a hard one and that is it's data. Like the the the data for

it's data. Like the the the data for physics models and for you know any kind of physical AI is not in the corpse of the web. And I'm talking about like milla

web. And I'm talking about like milla and milli and and microsecond you know frequency data.

That kind of data is what it's going to take to actuate in really precise environments in the physical world. It's

why we you know an LLM can't drive a your car. And we've spent you know now

your car. And we've spent you know now almost two decades collecting enough data and labeling that data in order to be able to train these models that can operate you know a a vehicle in the

physical world. And this is going to

physical world. And this is going to this is going to be true in every single instantiation of physical AI. We are not going to be able to use the compressed

data that humans have observed and recorded in tokens on the web and you know images video and you know in text as the basis for

physics-based models. It's just there's

physics-based models. It's just there's too much other information that's going to be necessary and I think it's going to be a big bottleneck. And you know a different way of framing that is

uh that you you might ask which is is sort of like what what is physical AI going to look like in 10 years? This is

the question you and I were talking about over email and and I actually think it's I think it's very hard to predict for sure. I think

these you know at an abstract level these things that I've kept talking about sort of flexibility and precision robustness speed and cost those are the elements of the

you know the the robots that will build in the future that are going to be really important. But how those those

really important. But how those those elements sort of turn out how each of those vectors looks for for different applications is going to be highly

dependent on the data that we're going to collect in the next 10 years.

So I think it's really hard to predict what the the total ecosystem of physical AI is going to look like. What I know is going to happen is we're going to collect a lot more sensing and actuating

data in the real world and in specific applications and they'll be you know a lot of that data will be proprietary for the people that work in those spaces

and that will allow us to build both more atomized you know things that can come together in in manufacturing swarms or

transportation swarms to do complex tasks with great orchestration that might be really hard for us to predict today.

You are such a great software architect and and you know probably don't write that much code alone these days. Like I'm not sure anybody does.

days. Like I'm not sure anybody does.

But how are you thinking about the you know if if people that watch this have you know are are young people who maybe just graduated with a computer science degree

or you know graduated a few years ago.

How how is how is that skill going to be useful or not in the world that Cobot is helping to create? Yeah I think I mean I think

to create? Yeah I think I mean I think it is a very dynamic time. I think it's a very dynamic time for young people right? I have uh uh

right? I have uh uh 17-year-old and a 15-year-old and so I'm I'm thinking about this like other other parents as well right? I think the um

you know I I think what's going to happen is the distance between our idea and being able to realize that idea in the world is going

to get shorter and shorter. And so you know for the Jeff Bezos's of the world right? Who are producing lots and lots

right? Who are producing lots and lots of ideas the speed between those ideas and making them happen is going to get less and less. And so I think you know I think it's important

for for young people to think about like be creators right? And and the creation tools can be music they can be videos

that AI is giving us kind of great creation generation tools. Some of that can be code some of that can be other things but really kind of how do we

inspire young people to create and to kind of get excited about these creation tools. I think with social media and

tools. I think with social media and things we created this consumption right? Created consumption economy we

right? Created consumption economy we created this information consumption just consuming all the time. You know

the value is accruing to people who create or people do and so those creation tools are getting better and better and so I'm I'm encouraging young people to

to use AI to think and I I get a little frustrated that schools have you know been a little bit discouraging people to create with that. It's like how do I how

do I create the poem I always wanted to write right? You by partnering with AI

write right? You by partnering with AI versus you know trying to get it out of my own brain entirely right? And I think you know that kind of creation

is is very exciting and the tools are only going to get more powerful and that gap is only going to get smaller and smaller. So you know entrepreneurialship

smaller. So you know entrepreneurialship taking you know taking ideas creating I think it's going to be you know that that energy that threshold

of like hey I have this song I want to write right? And then it's just you know

write right? And then it's just you know do you really have the activation energy to get over it? Well, now

now the tools are there and the energy you need to do it. And so, just getting in that creative loop is the thing I encourage people to do.

It's awesome. It's awesome.

It's good to see you, Brad.

Yeah, this is awesome. Thank you so much.