Two centuries ago, probably 80% of us worked on a farm. A century ago, probably 40% of us worked on a farm. Today, 2% of Americans work on a farm. Did we become obsolete? No, we just found other things to do.
Most robotics demos never ship. Deployed is about the ones that do. This is Deployed, where physical AI gets real.
Awesome, Alfred. It's so great to be doing this conversation.
Thanks for having me on your show, your podcast.
Yeah, this is super exciting. Is this the first time you have a guest or because it's usually just you?
Yeah, that's right. I've been doing my own kind of weekly videos and doing that as a podcast. And so, now we're trying to bring some folks together and really right now folks that I've worked with and known a long time. And obviously, we've worked together maybe more years than we want to admit.
All the way back at Tellme Networks. Tellme Networks. And you joined in '99, 2000?
I want to talk about AI, right? I think physical AI in particular, robotics. Obviously, your background as both, you know, Tellme was an early AI company. So, I want to kind of go back to some of the things we maybe learned there and lessons from that. But also, you know, operating at Zappos, operating at Amazon, and then moving over to the investment world, DoorDash. Like you've seen this kind of scale and growth and what it takes to apply technology into new operational domains at scale. So, I really want to kind of tap into that today. But yeah, maybe give me a little bit your journey and maybe with an AI focus. I mean, people probably heard your story, but give the kind of your touch points with AI and where you started to get some of that experience.
I mean, maybe start with a line from one of my partners, Constantine, that AI is kind of a misnomer. In some sense, we've been working on AI since the 1950s when Alan Turing came up with the Turing test. And things that seemed magical for a period of time was what AI was. So, there's a time when people thought just being able to do sort of that level of computation was magical. Then there was user interfaces, that was magical. Then there was, all right, well, that's not that magical. Like how do you do something that's natural language processing? So, when we met at Tellme, the frontier of AI was speech recognition. And you know, Tellme did put servers in the cloud, built, you helped design the architecture for basically putting servers in the cloud. We had to rack them. We had to build our own data centers. There's no one else's cloud to use.
There's no one else's cloud to use and we answered 800 numbers and we automated them.
So, a lot of what over the years I've seen like AI, whether it's voice recognition, machine learning, natural language processing, computer vision, we don't call that AI anymore. It's like we just keep moving the goalposts of what really AI is. And now we're in the world of large language models and those things are magical. And so, now we're thinking, okay, well, what's AGI? And AGI in the digital world is not AGI in the physical world. But I go back to our time at Tellme and that was pretty magical to be able to automate. When we started, it was 20% of calls just by changing the UI and asking a question or two to narrow down a listing for 1-800-555-1212, which is like 800 directory. And it was just information. All we did was instead of saying, 'How can I help you? What's your listing? Here's your number.' We said city and state. And then the listing and then that's sort of narrowing things down. So, some of the AI was just purely asking the question the right way and then obviously, recognizing what you were saying, adjusting for accents, things like that that over the years, we just don't think that's AI anymore.
Well, what I remember with Tellme, you know, Tellme was really this application layer of speech recognition, right? Was how important it was to build trust. Right? If you wanted to automate, so like one of the things that we became insistent on and our voice design team became insistent on was that you be able to get to an agent, right? Because if you could get to an agent easily, then you were much more willing to trust the automation system and give it a chance. I know, I'm curious how you think now about kind of trust and AI and how we're doing at kind of building that trust with customers. And really how that's going to play into obviously, at Kiva, we think a lot about trustworthiness in the experience of interacting with robots.
I think in every generation of technology, there's a scariness factor. Like what does this thing do? Is it actually going to be helpful? Sometimes the technology is a little kludgy. It doesn't do everything. And trust has always been an issue to get right with technology. And if the companies that sort of think about trust and figuring out the human in the loop elements, what makes you trust this new technology, it's really, really important to sort of just make sure that that is done properly. And so, today, when we started looking at just digital AI, it was all about human in the loop and making sure there are no hallucinations. And if there were hallucinations, how do we suppress them and give feedback into the system that this is a hallucination. Back then in voice recognition, if you couldn't do it quickly enough, you passed it on to an agent. And that was important because you knew that it was going to get solved in a short period of time. It wasn't just going to think in an infinite loop and you're going to be in that loop 5 minutes later and not get your answer. And you know, when we left Tellme, you went to Amazon and I went to Zappos. And one of the things that we needed to do was to do automation because picking, packing took a lot of time. And processing returns, there's a very human element to it. But moving shelves from point A to point B versus a human going point A to point B, that was pretty interesting of a problem. And so, we found this company called Kiva Systems and we implemented the Kiva Systems. We needed to have a place where there'd be trust because we're moving a whole shelf from one point to another and it was interacting with humans. Humans would walk alongside Kiva moving around. And eventually, we figured out, okay, well, that can run at a certain speed. Can we run this where no humans are in around a certain space? Can we run it a little faster? And eventually, humans stopped going from point A to point B. They were just moving the shelf from wherever it was to the human. That was an element of developing trust over time. And then the magical thing about the intelligence of machine learning there is like you move the inventory around in sort of logical ways that you move the slower moving things to the back of the distribution center and you move the things that are moving very, very quickly to the front of the distribution center. This is this natural flow that happened. And then the other thing I would say about whether it's digital AI or physical AI, the human in the loop element to fix problems because exception handling is incredibly important. Whether it was at Tellme Networks or at Kiva Systems and where you're doing now at Kiva, when something doesn't go perfectly, how do you handle that exception? When some of the best companies, when something goes wrong, it's their way to recover. And in some ways, when we see something that doesn't work as a consumer, we're a little disappointed. But when we see the company and or the technology or the combination of human and technology and the company fixing the problem, you turn someone who may have just ended up being a detractor into an immense promoter if you're able to correct that problem quickly.
So, to add more to that story, obviously for folks who may not know, so Amazon acquired Kiva Systems in 2012. And then when I took on robotics at Amazon in 2017, that team, the Kiva's was now Amazon Robotics, Boston-based team. And we scaled that from about 50,000 robots to half a million robots in 5 years. Amazon's over a million robots now. But you were telling me that it almost didn't happen. Yeah, yeah. And so, tell a little bit more of that story.
Well, so, Amazon acquired Zappos in 2009. And during the due diligence process, we showed the team at Amazon how we ran operations. And they were a little surprised that how low our cost per transaction or per order was. And like, how are you doing this? And we show them that we were using all this automation. We wrote our own warehouse management system. We wrote a lot of the back end. We were routing people and Kiva in very efficient ways. And what I was told is our cost per order was lower than that of an Amazon fulfillment's order. And part of that is because we were only processing shoes and handbags and clothing, much simpler orders. But they were very surprised and they were like, 'Oh, what else are you using?' And we showed them that we were using Kiva systems. And I think the team and Jeff Wilke was like, 'This will never work at Amazon.' Part of it is because Zappos' spike in the holiday season is like four times that of the regular rest of the year, but Amazon's spike in the holiday season is like 25 times or 50 times. You can't have 25 to 50 times more robots during that time. So, I was like, 'All right, whatever. It's not my problem. You guys figure it out.' They saw how we operated. They took some of the robots and put them in other places. I heard that they were being used for Amazon Fresh for a period of time. And then a few years later, Amazon showed up to Kiva Systems and bought them for $800 million because they wanted Kiva to only produce robots for Amazon. And you were put in charge because you were a distinguished engineer that built much of Amazon's retail systems. And you were so good at it that they put you in charge of scaling the robotics team. And you became VP of robotics for Amazon. So, that was an amazing run for you.
Yeah, I mean, my experience at Amazon was really really incredible. Really fortunate to, I mean, this the scale and the expansion. Now, I want to take a slight detour because we've been talking about... So, I started my career at Netscape. Netscape was acquired by AOL. Tellme was acquired by Microsoft. Zappos was acquired by Amazon. Kiva was acquired by Amazon. Then I was sold to Scale AI or mostly acquired by Meta. How as an investor, how do you think about acquisitions and the return and because everyone's always focused on the IPO, the kind of, and yet a lot of companies and a lot of companies we work for end up, you know, selling over time. Like how do you view that as success, failure, middle? Like how do you think about it?
Maybe this is a non-answer. I don't really think about exits as an investor. I never really thought about exits when I was an employee of those companies that eventually exited. I think what you want as an investor or you want as an employee is a founder that you trust. And you want a founder with big dreams. And so, sure. Like when you start out on day one, you're not building the company to be acquired. I don't think any of us would join a company that was just meant to flip. I don't think we would join a, you know, you want to invest in a company or join a company that's just like, 'Okay, if it hits my threshold of a price, I'm just going to sell the company.' That's not the ambition. And keep in mind in this day and age, even back then, you get a high-paying job in technology, why take a pay cut and join a startup unless you really believe in the mission? And so, most of the time what we're looking for as an investor are founders that want to put a dent in the universe. Our mission at Sequoia is very clear. It's been the same mission is to help the daring build legendary companies. We're not helping the daring build a company to flip. We're not helping, you know, sort of those who want to buy low and sell high. And so, I don't see it as a success or failure. It is just the path. And some people find that it's much easier to sell a company because they realize the market is too small. They realize that the... And so, that's perfectly fine. We took a risk. My job is to take risk. Founders' jobs take risk. The risk doesn't always work out. And it's perfectly fine to sell the company if you realize this is not working the way I thought. Sometimes you sell the company because you realize that the acquirer can take it to another scale. I would say that's probably true with a variety of companies like WhatsApp was taken to another scale at Meta. Instagram was taken to another scale at Meta. And YouTube and Google. Your experience with Kiva. I don't think Kiva had more than $8 to $11 million of revenue at the time. It was acquired for $800 million. But it didn't have the scale that Amazon eventually scaled it to. Because of you. You scaled it to millions of robots. And so, the impact that a platform can have on your company, your idea, your product can be immense. And you know, sometimes you go all the way and you take the company public. And it's extremely extremely rewarding either way because you're on a mission. You try to change the world. And that's why it's fun.
So, now you went to Amazon. You went to Scale. Now you've started Cobot. You're trying to take all of your learnings from all these different places, from Netscape to Tellme to Scale. And you bring it to Cobot because you understand software, you understand operations, you understand AI, and you understand robotics. And that's why it's a rare combination. Often we're looking for founders who have problem market fit. And so, you have founder problem fit here. You have founder market fit here. It's really interesting to have you lead a robotics company because you bring all these different disciplines and combine them and bring them to work.
Well, I think robotics is such a complicated space, right? Because you're bringing AI, you're bringing software, you're bringing hardware, you're bringing operations together in a system, right? When I think about that kind of application layer of physical AI, right? How are we going to bring physical AI into the real world? It's, I mean, in some ways like what Cursor did with bringing coding models was they built a whole scaffolding around the developer experience and workflow. And it's amazing they've done that. To do that in robotics, there's so much more complexity you have to bring in order to transform a hospital operation or a warehouse.
I want to come back to physical AI, but one thing I want to touch on before we get there. So, fascinating kind of thing that I remember from Tellme was that, you know, Moore's law, right? And you know, Sequoia famously invested maybe the first or early check in Nvidia, right? And now we're watching that scale. What I remember at Tellme was, you know, every couple years we got faster, we call them rec machines, rec speech recognition machines, right? And I suspect you were signing off on a lot of those, you know, investments. What it allowed us to do, and this is kind of a subtlety of the algorithm at time, the kind of Gaussian mixture models. To your point earlier, we had to return the speech recognition result quickly, right? Which means we had to do the computation very, the inference now, you know, I don't think we used the term inference as much back then, but inference on the model. And in order to do that quickly, there was a pruning function in the algorithm. The Viterbi search would prune branches that were less likely. And what we found, every time we got new speech recognition machines, we could turn down the pruning, which meant we could explore the model space more completely. And the speech recognition got better every time. And I think about that now with inference and with compute and how parallel we're searching the model space and just how like Moore's law has really enabled that and how much now we've gone back to kind of optimizing software all the way down to the GPU layer. I don't know. I'm curious what you're seeing. Obviously, there are new chips, you know, the Groq acquisition by Nvidia recently. Do you think this Moore's law keeps playing out this way and we keep going through these cycles or just AI is driving something completely new?
I think maybe the other way to look at this is like when something works, you just keep pushing it. And sometimes you push, it requires someone on the outside to push things that other people don't believe. So, Intel pushed Moore's law. And Nvidia pushed something different, which is not the CPU, but the GPU. And in either in whatever case in technology, you are always trying to figure out more efficient ways of computing, more efficient ways of sending data from point A to point B, more efficient ways of processing. It's not just compute that you need faster and faster and faster. You need disk to go faster and faster. You need input output to go faster and faster. And it requires a lot of people who find these technologies fascinating to keep building and building and building. I mean, I don't know, I don't remember how many times you had to re-architect the Tellme system, how many times you had to re-architect the Kiva system, how many times you had to re-architect the hardware so that it can operate faster or cheaper. Better, faster, cheaper is the advantage of technology. And Silicon Valley is basically built on that. Cost advantages, speed advantages. And you know, it is fascinating to me that technology makes things cheaper and more powerful at the same time. The laptop that you and I carry is cheaper today than the laptop I got when I was at my first job. I don't know how many times more powerful and yet it cost less. And you know, technology is a great deflationary thing. And it just makes all of us much more productive and much more efficient. And it lowers the cost of doing business.
Yeah, I was thinking about this recently. I think like I think Jensen did those of us in robotics a huge favor about a year and a half ago when he put that chart up that showed, you know, language models, agentic AI, and physical AI as the like because what it did is it caused a lot of really top AI scientists, ML engineers to see that problem and want to go after it, right? And you know, what I'm really excited about is you know, 18 months later we're now seeing some of the research paper, seeing the diversity of kind of work in world models, in VLAs, in reinforcement learning. Obviously, we have our own thesis as kind of what some of the critical problems are in physical AI, but I'm curious kind of what you're excited about, what you're seeing, and how maybe you think this plays out.
