How to Be a Fearless Engineer with Cerebras Systems Founder and CEO Andrew Feldman

When you do fearless engineering, um, you, you're not afraid of, of screwing up. You're not afraid of breaking something. You're afraid of not learning, not getting better.

What does it take to create something that could be a total game changer? What if we told you that it's going to take a whole lot of bravery and a pinch of paranoia? Andrew Feldman and his team at Cerebras Systems were told that it was impossible to build a computer chip that could deliver the same performance as hundreds of graphic processing units. But they tackled that challenge head-on and have created something incredible. The CS-2, the fastest AI computer in existence, which is being used to tackle the world's most pressing problems. But none of this would've been possible without a bit of audacity in what Andrew Feldman calls fearless engineering. IT Visionaries is brought to you by Salesforce Platform. If you love the thought leadership on this podcast, Salesforce has even more meaty IT thoughts to chew on. Take your company to the next level with in-depth research and trends right in your inbox. Subscribe to a newsletter tailored to your role at salesforce.com/newsletter. Welcome everyone to another episode of IT Visionaries, and today we have a special guest. He is the founder and CEO of Cerebris Systems, Andrew Feldman. Andrew, welcome to the show.

Albert, thank you for having me.

All right, right out the gate, we ask all our guests the same question, but we all want to know: what is Cerebras, and what do you guys do?

We're a company based in Silicon Valley here with, with offices in San Diego and Toronto. We build a new type of computer, a type of computer optimized for artificial intelligence work. Artificial intelligence is one of the most exciting domains in our economy right now. AI algorithms are recommending cartoons for my 8-year-old granddaughter, and my 88-year-old mother-in-law is asking Alexa to play Frank Sinatra, and Alexa's putting together a playlist that includes songs that she's forgotten she liked. So it's, it's impacting us at every age group, and it presents very interesting challenges to compute. It's computationally intensive, uh, it's hard. And so at Cerebris, we saw an opportunity to build a new type of computer from the ground up. We built the chips, we built the systems, and we built the software that allow AI practitioners to do their work vastly faster, often hundreds or thousands of times faster than on traditional equipment.

So we've had a lot of companies on this show, a lot of guests on the show, and they tend to be on more of the software side or the modeling side. I believe you're the first person that's more on the hardware side of this, this, this challenge.

Albert, we have better show and tell in hardware. Our show and tell in hardware is phenomenal.

That's true. That's true. I don't know. I don't ever know what I'm looking at with software. Is it vaporware? Is it real hardware? Hey, can this process this? But give our audience an idea. You, you mentioned before, and, and like it says on your website, uh, Cerebras is revolutionizing compute for deep learning, and you mentioned for AI. What is it that makes the resource requirements so intense for a computer? Why is it that, like, for example, if I don't have the right hardware, that maybe I'll never be able to unleash the power of AI because of whatever it takes. I'd like to understand the difference between what a normal computer can do and what you guys are building.

AI is an interesting, uh, type of work. It, it is really trying to find insight in data. It's sifting through huge amounts of data, sometimes billions of images, in order to teach an algorithm that is general purpose to teach it insights. So years ago, for example, Albert, if we wanted to teach a computer to identify a cat, we would write what a whisker is. We would say, this is what whiskers look like, this is what paws look like, this is what tails look like, and we coded those in very specific algorithms. and then we would look for that in an image. In about 2014, some folks came along and they said, uh, instead of doing that, we're going to take a general-purpose algorithm and we're going to modify it so it, it improves, it learns, and then we're going to show it millions of pictures of cats. And each time it gets it wrong, we're gonna show— we're gonna have a little mechanism for it to learn what it got wrong. And what we discovered was that was far better at identifying cats when it was said and done, but it took a long time and took a ton of data. And so what we've done is we've recognized this new type of work this work that is about finding information and data and sharing that information with an algorithm, improving the algorithm, showing it more data, this iterative process. We found out, we worked hard to understand what is exactly computationally challenging in that work. And it turned out what was hard wasn't the calculation. The calculations were pretty simple. There were a huge number of them that required tens and hundreds of thousands of compute cores, but it was the communication between them that turned out to be very hard. And we solved that problem by building a chip that was bigger by 56 times and 2.55 trillion transistors larger than any previous chip. I happen to have one here.

Is this the, uh, the wafer that's on your, uh— I think I saw it on the website, the wafer.

This sad little chip here, this was the largest chip before us. That's how big a chip is. They're the size— this is a big processor. This is the size of a postage stamp.

All right, I'm gonna— for the audience that's not watching the video, Andrew just held up a chip. I would say it's like 8 thumbnails. That's, that's— just imagine your thumbnail, multiply it by 8 times. Okay, he is now currently holding something that is clearly much bigger than the palm of his hand, maybe two palms up, two palms across.

Yeah, it's bigger than a dinner plate, Albert. So what we were able to do is solve a problem.

That probably makes bad podcast to show pictures, but no, I described it, it's good.

We went from something the size of a postage stamp to something the size of a dinner plate. It allowed us to keep all the compute resources on a single piece of silicon, allowed them to work faster, to spend less time in communication, to spend less time going and getting information. And the result is we're doing work hundreds and thousands of times faster.

Okay, so I gotta ask this question because I used to work with a guy who would oversimplify everything, and he was It's almost annoying, you know. And so the solution can't just be to make a bigger chip. Like, that can't— is that, is that literally a solution? It can't just be to make a bigger chip.

Making a bigger chip was a problem that people had failed to, to, to do for 70-year history of computing. So it wasn't just that, that, oh, just make it bigger. It was that nobody had ever built a chip bigger than about 800 square millimeters. And we come along and say, yeah, we have this idea, guys, we can build one 46,000 square millimeters. And people look at us like we're crazy. And it involves solving a huge number of technical challenges that had never before been solved. Once we had the chip, we had to solve a new set of challenges, and those were how you build a computer around it, how you deliver I/O to it, how you power it, cool it, put it in a form factor that's easily deployed in anybody's data center. And our saleable unit is 15 rack units tall, so it's about 26 inches tall, and it fits in a standard data center infrastructure. And yet it delivers the performance of hundreds or thousands of graphics processing units.

Yeah, so what was your background prior to developing this chip? Because, So you have this idea that, hey, listen, the only way to make these, to solve these deep learning AI ML models faster is to have more, like you said, more processing speed, more compute power.

First, Albert, I have no ideas. My co-founders have ideas. Early on, I buy pizza and raise money. My co-founders, Gary Lauterbach and Sean Lee, and Michael James, JP Fricker. They had the big ideas. Um, they'd all— we'd all worked together at my last startup where we built a new type of computer, this time optimized for energy-efficient compute. So we're computer architects. What we do is we think about, can we build a new type of machine for a new type of work? Has something changed in the world that creates a new type of work? Like when people suddenly wanted lightweight compute at very low power, that was a new type of work, and that's what the cell phone processor was. It was very different than the processor that had traditionally been in a server or in a CPU. In the same way, we look for these changes in what people want and what, what workload is being used and we ask ourselves, is that market large enough? Can we build something that is fundamentally advantaged? And we saw this in late 2015. We saw the rise of artificial intelligence as a key workload, and it was— it's been an extraordinary ride. You know, the market went from approximately zero to tens, soon hundreds of billions of dollars.

You hinted at it earlier, AI will be a part of just about every That's right. Compute function out there.

That's right. It has nibbled its way into things. You get in the car and your phone tells you your usual ride home isn't the most efficient path, right? You, you think of what, what's the number of times a recommendation engine— you go on Amazon and a recommendation engine is pretty darn good at recommending things you might like.

When, when you were— when you and your team were developing this, were you all Were you con— were you already thinking, hey, we're gonna have to build the whole system around it? Because it's one thing to build— because even like a modern computer, if like I were to buy a Dell, right? Like there's every, every part on a Dell or somebody else made it. Someone else made it. Yeah. They're just putting it together.

You know what I mean?

Like they're just kind of, they're kind of like Lego masters. Like, oh, this guy gave me the chips and this guy gave me some motherboards, transistors, diodes. I'm putting it all on the thing.

Sure.

But I don't make every single thing. I'm sure they make a lot more than they used to, but I know that IC components are like every manufacturer's making something. So were you automatically, are you, were you thinking you're beginning like, hey, this is our unique IP. The chip is our unique IP. We're going to build the computer for this because no one else is thinking this way. Is that kind of like the genesis of the business?

Yeah, that is the genesis. That when you, when you see a problem you can make 100 or 1,000 times faster, that's interesting. And if lots of people think that problem is important, you might want to make a company around that. And we like building systems because when you build the system— my analogy is, if you were to build a race car engine, you wouldn't want someone else to put it in a Volkswagen. Yeah, right. If you want it to perform like a race car, you got to build a race car. You have to think about how each part interacts. Yeah, with every other part. I know, Albert, you probably drive a 911. No, the 911, the engine sits over the rear axle, and every decision about that car, every decision is informed by where the engine sits, the distribution of weight. You couldn't just build a great engine and say, here, anybody take that engine. They built every aspect, they tuned every aspect to build a near-perfect car. And it's the same in computers where you can build a processor and give it to somebody else. Yeah, but if you want to build something, if you want to build a race car, if you want to build something that is head and shoulders above, you, you build the whole system. Remember, that's— Apple went the other way. Apple was building the system for a long time and not the processor, and they said, I need the processor too. Yeah, and now they're building the processor.

I mean, they only just started, right? The M1 chip, I think, is their first chip.

That's exactly right. First chip for the computer. They built the ones for the cell phone for a while. Yeah.

So there was their first chip for the computer that they only, they just started going into it. Probably, you know, it's probably similar to what you experienced, which is they, you know, they clearly ran into this, hey, we have these ideas, we have these like requirements and then manufacturers weren't willing to do it. So it sounds like you guys came in with the idea like, hey, we're going to build this chip, we're going to build a system around it. So this is a big investment. Anything in hardware is materially more expensive, capital intensive than, than, uh, software, right? So you get yourself to a working prototype. Now you gotta get someone to use it and say, Andrew, this thing works. Talk, talk about how that first experience was like getting this thing into the hands of someone that wanted to model something.

I'll tell you a couple stories. First, we, we have raised a lot of money. We've raised more than $730 million. Um, and so yeah, it's expensive. We do, it's a lot of money. And we do big hard projects and we like to talk about it as fearless engineering. You know, we— the first time we got the wafer to work, the founders sat there and stared at it. I mean, for like an hour. I mean, nobody in the history of compute, nobody in 70 years had solved this problem. And like, we're sitting in bad real estate in Los Altos. you know, in a jerry-rigged lab, and we'd done it, and it was awesome. It was an idea that came from my co-founder's head, and we had worked with vendors for years. We'd solved extraordinarily hard problems, and to see it run was breathtaking, and it was just such an awesome moment where your every ounce of what you put in sort of is now alive and is now working. Our first customer was Argonne National Laboratory, and the very first use case was in the largest cancer study in the U.S., a project called CANDLE. And they immediately found, using our system, drug interactions that they'd never seen before. They immediately found fundamental insight, and that was awesome. You know, you don't— you know, we, we didn't start the company to, to to identify a better fitting t-shirt on an 18-year-old. That's really cool and important. Body image matters, but we'd like to solve problems like cancer. And we have pharmaceutical company customers who are using our system to accelerate drug design. The guys at GlaxoSmithKline are a big customer of ours and they're doing extraordinarily interesting science with our work. Our system's been deployed at, as I said, at Argonne National Labs, at Lawrence Livermore National Labs, at the Pittsburgh Center for Supercomputing, in Europe. At TotalEnergies for energy exploration and clean energy development. Um, it's enormously rewarding when you see sort of 6 years of your, your passion, your work, your every ounce of your soul solving important problems. And that's why we do this.

Yeah. And for those people listening and trying to, you know, this is just something I know of I'm sure you have more use cases than this, but I remember some of our previous guests talking about like how large electronic medical records are, right? So it's well beyond, well beyond like a, um, you used an example earlier of like photos of cats and dogs. They're saying EMRs are way bigger, intensive files. The, the, the identifiers are way smaller. So like the amount of processing and compute to identify, like you said, at a cellular level, How can we better detect, for example, if we, if we wanted to answer one question, how do I detect cancer earlier? Because I want to help people. We all know that the sooner you detect it, the more treatable it is. But for most of us as humans, we actually won't get a detection, uh, result from a doctor until they see something. So they have to actually see it with their own eyes. And of course we know our eyes are not as good as what AI machine, like, and they're talking about the future at, implication that this can be is just phenomenal.

I mean, there are lots of ways to, to, to look at what you just said. I mean, you have a big hospital system, they've got patient records on hundreds of thousands of patients. Um, those records are text, they're images, they're, uh, lab results. There, we now have models that can run through them all and find patterns. And we have a customer called inference that, that has partnered with hospital groups and is going through medical records, finding patterns to improve care. We have partnerships with those who have MRI data and mammogram data and pathology data, where you've got these images and the images are very large. and where they're looking at developing models that can identify things that the human eye misses— early detection of cancer. Um, and so we are using computers to— what AI is doing is using computers at this class of tasks that, that humans sort of struggle at: finding patterns in large amounts of data, um, identifying, uh, in a probabilistic sense what a, what a tiny shadow smaller than a human eye can recognize might be on a mammogram. These are things that computer imagery, and then with AI on top of it, can do really well.

So when you, you mentioned those first waves of customers, they immediately were recognizing how much faster it is. For our audience, if you had not been to Cerebras's website, he's got some interesting examples. One of them is like AstraZeneca saying he used to take 2 weeks plus to train AI models. It was now down to less than 2 days. It's even got a direct quote, 52 hours to be exact. Pretty bananas. Uh, when you think about it, how many times, how many systems did you have to build before you had one that was like, you were willing to like show and demonstrate its capability?

How many mistakes we made? Yeah. A lot. Yeah, when you do, when you do fearless engineering, um, you, you're not afraid of, of screwing up. You're not afraid of breaking something. You're afraid of not learning, not getting better. And when you're working on really hard problems, each time you, you build something that doesn't work, it's an opportunity to learn. You go back, you check your process, you make sure that the hypotheses you had, which were proven out which weren't proven out, and you get better, and you just rinse and repeat, and you do this again and again. And when you're building things that nobody else has ever built, there's no book you can read. Yeah, nobody has data that you can go online and dig into. You can talk to experts and people have ideas, but nobody had ever seen a chip as big as, as what we were working on. Nobody had ever delivered power to it. Nobody solved any of the problems. And you write them out one by one and you work on them. And if you get stuck and you keep going, you do it again and again. And sometimes it takes months and solve the problem. You keep moving forward.

At any point were you worried that, hey, this just might— we might not— it might not be possible? The reason why it hasn't been possible—

oh, every CEO has got to be paranoid as can be. I'm worried every day. Um, I, I think, uh, you know, paranoia is a, is a core competency of I'm a good CEO.

I remember the first time, you know, the company has since now folded, but I, I remember getting my, uh, getting to see a juice, a Juicero once for the first time, and someone was like, yeah, it's great, it does this. And I was like, this is the slowest machine I've ever been around. Like, why is it so slow? Like, normal juicing is faster. When you first saw the first results, where was it? When you— I guess you probably didn't bring it to a customer until it was actually validated that it was computing faster.

In hardware, from the time your system is first taking its gasping first breaths, it's usually a year before you get it to a customer. And yeah, it does. It's like anything else. It begins and it's crying and it's not moving much, and then it's crawling and you're wondering if it'll ever walk and run. And you work on it and you work on it and it takes its first steps and then it falls and you got to go back and you got to figure it out. And Pretty soon you're walking comfortably and then you're running and you're slowly working performance, you're working your— the breadth of the capabilities and you're tuning it, you're working on it and working on it and you get to a point where you show it to customers and it— they get excited and they ask you for 50 things you hadn't thought of and/or things you had thought of but you thought you could do later. I think that's the process of breathing life into hardware.

Yeah. And one of the things that we've, you know, we've, we've heard, and it's, it's not just in the tech industry, it's every line of business, but once something is proven that it's pretty good, people generally start talking. So I'm sure people that were testing models on your systems were starting to talk. Give me an idea of what was the— I guess, what was the demand like? Were people skeptical to you, or they're like, uh, I don't know if this guy actually can do what he says he can do, I don't know if the bigger chip is going to matter, or were people I guess here, because here's what I'm thinking, right? Every, the, the public clouds offer AI services, right? But I'm sure everyone has used, or at some point probably caps it out where it's like, hey, it's not working as efficiently as I want it to. They want a better system. So everyone's got like something that they're like, I wanna, I want, I need this because whatever I'm currently using just isn't delivering the result on that. And it's not like, and like you said, it's no longer a software, it's not a software problem, it's now a hardware problem. It cannot throughput this much information.

That's exactly right. The, the, the way the software world feels a hardware problem is when it takes forever to do your work, right? When you have to wait a month, like some of the large models today, to get an answer.

A month.

A month.

Yeah, no one's used to that.

On 1,000 GPUs consuming dozens of megawatts of power. Yeah, megawatt hours of power every, every day. Um, I, I think that's how you feel pain, that that's when the hardware isn't getting the job done. And, and you read some of the quotes, you know, some, some guys have found us, you know, 160 times faster, others 120 times faster, more than 100 times faster. Some, in one case, there we were 200 times faster than a supercomputer. And what that does is, is it enables the, the ML practitioner or the software professional to test ideas really quickly, right? If you can do work that used to take 2 weeks in a day, if you can do that, you can test more ideas in 2 weeks, right? And if you test more ideas, odds are you're going to get a better one. And that's a sort of a fundamental tenet in, in artificial intelligence work, is that if you can shrink the amount of time it takes for the machine to learn, to do training, you can show it more data, you can get better results, you can test more ideas.

So then this is, um, I mean, I guess this is maybe speaking more to the future strategy of your company, and if you, you know, if you don't want to answer, I totally understand. So it sounds like this has to be an on-prem solution. Like, I gotta buy your machines, I gotta get my own racks, I gotta do all this, connect it to my data sources. Is it ever going to be a service at like the clouds, or do you get nervous that the clouds want to recreate your chip, or I don't know?

We always— come on in, the water's nice. It's a hard project. Um, we, we have solutions on-prem and off-prem and as a subscription and in Cloud. Um, these are our big high-performance systems, and some people want that, but not every day. Yeah, I want it one day a week or, or a week a month. Um, others want it continuously. Um, some want to buy it by the hour, and we have partners and customers and purchase options regardless of whether you have a facility or you don't, whether you want to get it in a cloud format, whether you have data that resides in, in a cloud and don't want to pay egress fees, we, we have, we have options for you.

There you go. I mean, you thought of it all. Dang. I was thinking, I was thinking like, I— because I— when you were— the way you were described— when you were describing it, you know, all I know is what I know. And I was thinking like, man, this, this has to be only an on-prem solution. So that's pretty interesting that you already thought about all the different ways that people can experience.

We want people to to use it. I mean, that's your passion when you build infrastructure for a living, you know. You want, you want people to use it and put their ideas on top of your ideas and, and be a foundation for, for watching their ideas fly. That's the beauty of infrastructure.

Which brings up actually another thing that I'm sure you guys are planning on and always developing towards, which is that people are— since the beginning of time, people have kind of been able to push the limits of what is available to them. So I'm sure the requests and the ideas now that they're processing things even faster, it's probably unlocking new ideas maybe that you didn't even imagine. Like, what, what kind of the requirements that people are now probably coming to you with? Or like, you're like, dude, I just, I just, I just got here. Like, why are you asking me for this now? What kind of requests?

Sometimes it feels like that, Albert. Like, well, I just solved a 70-year-old problem in the industry. I'm 100 times faster. And you want what?

Yeah, exactly. It's like a continuous— like, it's not good enough, right? It's not fast enough, it's not big enough, it's not, you know, whatever that— whatever the it is, people always want to go faster. And I know in this industry it's— that won't change, right? So the day you've solved, you know, it's— someone wants to go even faster. What are some, like, the projects people are trying to— trying to solve for now that you're like, wow, we— we have to keep improving, we got to keep doing this so that we can solve these problems that are on the horizon.

You know, that, that's both the bane and the joy of our industry, is that as fast as you can invent technology to build infrastructure, somebody else has an idea how they could use 10 times what you built. And right, I mean, that, that's, uh, as soon as we got, got Ethernet, we were dreaming of fast Ethernet, a gigabit Ethernet, 10 gigabit or 100 gigabit Ethernet. 400 gig— I mean, what if we can go faster? What— I mean, as soon as, as we, we had computers, we were dreaming we'd go faster. Um, the— there is an area of, of artificial intelligence called natural language processing, and there was a group in, in San Francisco called OpenAI, sort of pioneering researchers, and they published some very interesting results that showed The bigger the model, the more compute it used, the better the results. And almost without bound. Yeah. And what that meant was, you know, what we used to think of big took, you know, let's just use number of GPUs as an example. It used to be 4 or 6 or 8, and now they're using 10,000 and they want to use 40,000. We come along and say, well, we can do all that in 4 boxes. What about in 16 boxes? We can do work that, that you can't imagine yet. And they take that challenge and they say, yeah, we can, we can imagine, we can build a model, give us, give us, give us a year, we will do— we, we will want more from your system. And that's the great interplay between, between sort of infrastructure and those who use us.

No doubt about it. The next request is going to be they're going to want the the compute power of your wafer, but they're going to want it back to the size of—

I mean, dreamers were not reasonable, right?

Andrew, Andrew, can you put that in my phone again?

Can you write exactly— I, I want to do this in my phone. Like, well, last year that required a supercomputer.

When you think of, when you think of the business, I guess, where it is today. Give me an idea, is it mostly in the United States right now? Have you already reached international or like people—

We have, we have customers in Europe, we have customers in Asia. We're deploying equipment at a really healthy clip. It's fun to see other people's imagination working on something you built.

Yeah, no doubt about it. Andrew Feldman, it was super fascinating hearing your story. The solution, like you said, sounds super simple, but there's a reason why it was never done before, right?

Yeah, it's hard. And building chips, uh, normal chips, traditional chips, is no easy project. And innovating on the physical design, making a chip 56 times larger than the largest chip anybody else had ever built, uh, having a chip that's got, you know, 2.5 trillion times more transistors than your nearest competitor. 2.5 trillion times more transistors than your nearest competitor.

Well, now when you say it that way, that sounds a lot different.

That's what we did. Um, we didn't just say, oh, can you Make it bigger.

Oh man, it's a lot of fun. You know, Andrew, one of the— you know, I've really enjoyed having you on as a guest for the show. I think our audience is going to love hearing this episode. They're going to like hearing about all the— really, because, because that's really what it is, right? Like you said, hardware— you're, you're, you're building something to be used, but it's really all these practitioners who are now going to come up with really fascinating solutions. You know, me, I always say this in every episode, but like I have a personal interest in the medical, you know, longevity, removal of chronic diseases. These are— this is the kind of stuff that's going to go into it. And that's, uh, you know, we're gonna— it'll be amazing to see what comes out the next 5, 10— 5, 10 years from, from what people are discovering today.

There's a huge amount of data in AI and data in, in the life sciences. I think that the genome, whether you're doing genetics, whether you're doing epigenetics, whether you're doing mitochondrial DNA, whether you're doing, uh, the, the data that accumulates through our lives in, in our medical records, um, there's insight there that we've yet to harvest. And AI is a really good methodology for, for gaining some of that insight.

No doubt about it. I want to say thank you once again for joining us, but before you go, Andrew, it is time for the lightning round. The lightning round is brought to us by Salesforce Platform, the number one cloud platform for for digital transformation of every experience. Andrew, this is where we ask you questions outside of the world of work so our audience can get to know you a little better. You ready?

I'm ready.

You used the 911 earlier in the conversation. Are you a car fan? And specifically, are you a 911 fan?

I'm not a car fan. I think of them as a way to get to work, but I think they have a great deal of parallels with compute. You build a car for a job, right? One car is built for taking kids to soccer practice.

Another car is built— My sick minivan, I got one.

That's right, your sick minivan. But what we know, Albert, is that it's really bad at moving 50-pound bags of concrete or 2x4s. Yeah. It's terrible. And if you want that, you get a Ford truck, you get an F-150 or something, right? And what's perfectly designed for one thing getting those kids to soccer practice, is terrible at another thing. And it's the exact same in building compute hardware. What's really good at, at doing graphics is not good at all at, at doing AI work. What's really good at one type of problem is very rarely good at something else. You know, you— no matter how much you love your minivan, it's not a two-seat roadster.

Yeah, it's not fast, it doesn't handle well in the corner.

That's right. That's exactly right.

So you're not a car guy. We got to ask, what do you do outside of the world of work for fun?

I dance Argentine tango with my wife.

Oh, really?

We met in a tango class at the Mission Cultural Center in San Francisco 30 years ago, plus or minus. And we still like to dance together, which is a huge testament to her patience with me.

Well, you've got— did I hear you right? You said 30 years?

30 years.

So now here's a fun question. Are you good at it?

I, I, I, I suck less than I did before. 30 years in, I'm less bad than when I started.

I can see that. I could— not that I'm knocking you, but like physical skills are one of those things where it's like, yes, you put in the hours, but there's definitely limitations to where we all are. I'm an avid surfer, but if anyone says, are you good?

I was like, I'm not good, I just like Yeah, I'm better than I was 4 years ago, but I'm not good. That's exactly, exactly right.

Now, a lot of people that work in the hardware space tend to also be builders outside of work as well. Do you build things?

I, I don't build mechanical things. I, I built a small backyard orchard.

That's good.

Fruit trees, olive, citrus, stone fruit. I— that was new, and I got fired up, and that was really rewarding. And turns out I can tell you how many, how many wheelbarrow loads in a cubic foot of, uh, cubic meter of dirt now. I mean, we— it's sort of fun to, to get out there and do some physical work and see what you made. Uh, in spring they bloom. It's sort of a really nice thing.

Yeah, I'm, I'm a novice gardener myself, and I've yet to meet— you know, we interview a lot of smart people on this show and a lot of technical people, and a lot of them do gardening as well. No one can answer this question, so I gotta ask you. Why is it that when you dig a hole and you have dirt in a pile, and then when you take the dirt and put it back in the hole, you always have not enough dirt? Someone needs to figure this out.

Look, I, I don't know. I don't know why I grow tomatoes for about $60 a pound. Is about what they cost when all is said and done. You go to the farmer's market. Um, it— I— something magical happens, I guess, when you dig a big hole and you use the hole digger and you're wailing away, and some magical thing happens and some of the dirt disappears and moves into the ether. I think that's the answer there, Albert.

Andrew, it's been awesome having you on the show. Thanks for sharing a little insight about your world outside of the world for the work, and thanks for sharing what you're doing at Cerebras. It's super fascinating because on the, like, to the outside, it's again one of those things where it's like, oh, when someone has solved it, it's like, oh, why weren't other chip manufacturers going down this path? And it's, it's always fun hearing from the people that are doing it, and it's like, dude, the reason why no one's done it is because it's really freaking hard.

Amen, brother. Amen.

Thanks for joining us today on IT Visionaries.

It was a pleasure being here. Thank you for hosting me and hosting Cerebrus. Cheers now.