Tech CEO: AI Is Not As Smart As You Think! - Ashfaq Munshi

The fundamental problem I have today with all of these large language models and what we call AI is that AI doesn't know how to learn. You have to retrain it. Well, now you have to retrain it.

it. Well, now you have to retrain it.

You spend millions and hundreds of millions, God knows how much money. We

got to build models that can learn, that can observe, that can see things, that can understand things. And that's really how I think the real world stuff can actually come in.

You've seen like so many different ways you met different people. What is a story which has stuck with you from a motivation perspective today? Access to

capital is so easy that people blow it very very quickly. They don't value the money that an investor puts into a company. Young startup founders say, "Oh

company. Young startup founders say, "Oh yeah, we, you know, we raised $50 million or $15 million. Oh it

didn't work. Okay, $15 million has gone down the drain." There's no feeling of, "Oh my god, I blew somebody's $15 million." But if you value money, spend

million." But if you value money, spend it properly and spend it wisely, success does happen.

So Ash, welcome to the podcast. Thank

you. So for folks who don't know, he's had a prolific career. Uh he did his undergrad at Harvard, then went to Brown, and then Stanford. He's worked on

like multiple companies, had had successful exits worth hundreds of millions of dollars, being the CTO of Yahoo, and I was fortunate enough to be a co-founder with him for a very small exit called something company called

CryptoMax, which ended up at Snowflake.

So, Ash, very excited to have you here and like, you know, chat about your story, chat about GPUs, optimization, China, US, and a lot of these things.

>> You want to pick an old man's brain?

>> Yeah. No, I'm very excited. I think I think the type of stories you have and like the the things you've seen from you know the dotcom era pre- do era mobile wave AI cloud all of those things I

think be very helpful for our audience to understand but before I get into like technology I think one thing which really stood out for me when I met you and started learning about your journey

and started learning from you was you were 18 you were studying at Harvard and you got married and you had a kid I think in your second year or like first

year and you were studying school, do a night shift at the guard and you told me that story as like hey that was just pretty normal and when I look at today's

tech culture especially you know the SF Valley culture people are like hey man got to hustle we got to build companies forget about relationships forget about you know even getting married

are a far thought for folks just want to get your perspective like how are you thinking about that >> yeah So, so I did get married when I was

19. I had just turned 19. Um, and uh, we

19. I had just turned 19. Um, and uh, we didn't have a kid quite that quickly, but yeah, you know, I was in my early 20s when we had our our first child. Um,

but I have to tell you, it was probably one of the smartest decisions I ever made. And the reason is because it gave

made. And the reason is because it gave me stability and something that anchored me. Um, which was really, really helpful

me. Um, which was really, really helpful for me. And I've always been very much

for me. And I've always been very much of a family person. Um, I always saw for family first and then everything else

sort of came second. Um, you know, give you some examples. Um, uh, when I was at Silicon Graphics, I was probably the poster child of the company for work life balance, right? I I went home at a

certain time every day. Um, I worked my ass off, but it was very clear that there were certain boundaries and then I would be home for dinner. I'd be home to spend time with my kids. And uh you know

even when I had to travel I take red eyes so I could make sure I could put my bed kids to bed before I went. Um I made sure that my jobs were such that I never had to commute too much so I wouldn't

wouldn't uh miss opportunities with them and I tried very hard not to miss any of their events. Um so you know that was

their events. Um so you know that was extremely important to me and for me you know the legacy that I leave behind is not the money that I make. It's not the

companies that I build. It's the people that I touch. And the very first set of people that I touch is my family. That

legacy is far more important to me than anything else. Because realistically,

anything else. Because realistically, even if you're massively wealthy, people just you're just a check. I mean, it's the it's the people who interact with you, the people who love you, the people

who interact with you and understand something that you teach something.

That's the lasting value. And the

reality of it is after two or three generations that value disappears. So

your value on this earth is very limited. Might as well make it

limited. Might as well make it interesting. Think about what a lot of

interesting. Think about what a lot of people today, young people do. Hey, they

don't want to buy stuff. They want to buy they want to do experiences. What

are experiences? Memories. Memories that

I do with my family. Memories that I leave behind are the memories that are the most important things. Those are the things of durable value. If I teach you something and you turn around and do something with it, that's

extraordinarily valuable. Right? That to

extraordinarily valuable. Right? That to

me is far more important than me giving you money for your startup.

Does that make sense?

>> That 100% makes sense, right? Like I

think especially the centering point and like the memories and and the experiences and really having the meaningful impact. So what's your

meaningful impact. So what's your perspective like today there there are a lot of memes or there's a lot of tweets about like the 996 culture you know like in China and especially that coming to

Silicon Valley. You have been in the

Silicon Valley. You have been in the valley for almost 40 35 years.

>> Long time.

>> Yeah. a long thing. So what's your perspective on that? Like how what advice would you give young founders especially in their 20s, early 30s? So

you know I work a lot but I love what I do. So it's not work to me. So you know and do I count it as

to me. So you know and do I count it as 996? Do I work as hard? Probably. But

996? Do I work as hard? Probably. But

what I do is I try to interle that with time for my family with time that I take out to be able to go do things. because

I I love what I do. It doesn't ever feel like work and my family doesn't see me like grinding at something and coming back all tense and all that stuff. They

see me enjoying what I'm doing. And so

I'm not bringing back, you know, a a huge thing on my on my back saying, "Oh my god, today was was oppressive." Or

any of that kind of stuff. So my biggest advice would be work as hard as you need to, but enjoy what you do. If you enjoy what you do, the rest of the people around you will see it. they will

understand your enjoyment and then you will then turn around and the time that you spend with them will become quality time because you're enjoying it. You're

not you're not uh you're not you're not under a whip. That's the most important thing here. Enjoy it and then you know

thing here. Enjoy it and then you know it works. Look, I still pull

it works. Look, I still pull allnighters, right? I still wind up

allnighters, right? I still wind up working. I mean, there are days when I

working. I mean, there are days when I start my day at 5 in the morning and end at 9 or 10 o'clock at night. That's not

unusual for me. But do I think about it in terms of 996 culture and I have to go do this? No. I the way I look at it is a

do this? No. I the way I look at it is a bunch of things I need to get done. Some

of them have a time bound and I want to go do that. A lot of my time is not just spent working but it's spent learning right I spend a lot of time reading and

learning and learning comes in two things. One is through reading and

things. One is through reading and finding and then the second one is talking to smart people that can actually teach me things that can actually give me a different perspective on the set of things that I'm doing. So

what happens is it becomes social as much as it is the learning thing right it's like I go to something I go meet somebody I'm learning and I'm having a good time learning I enjoy their company and at the same time I have an extremely

high bandwidth conversation from which I'm absorbing a ton of information that I can use to do everything else that I do. That is great advice and I think

do. That is great advice and I think like a great way to start the podcast.

Now to get into your other kind of perspective right because you have spent time building in very massive technological ways like pre-com era.com

era cloud mobile and now AI and you know each technology has built on the previous one and made it easier for us to launch like all of these things. I think like AWS kind of made

things. I think like AWS kind of made infrastructure as a service pretty easy.

I obviously didn't see like what starters for before AWS, but AWS is much easier and like now I'm building a company in the AI space. I'm like, holy this is so much easier than the than than building a company.

>> We built the plumbing of four years ago.

>> So from your perspective, what are some positive things you've seen across these technological waves which keep which keep give you a lot of optimism for like the future? you know,

the the waves that I've been involved in um have generally been driven by people with a vision, right? People who

want to make a difference, people who want to invent, make something better.

Um and that story hasn't changed, right?

So, you know, you think about Netscape came out of SGI, right? Um that was like, oh, we're inventing something new.

At SGI, we actually in 19 in 1994, we did the first video on demand to a thousand homes in Florida. I mean, think about that. 1994, we were doing

about that. 1994, we were doing streaming, right? It was like we were

streaming, right? It was like we were inventing all kinds of interesting things. It was the invention that drove

things. It was the invention that drove us and got us to where we need to where where we needed to go. Then we had the dot thing obviously because of the internet and a bunch of things happened.

Um, what was interesting is I personally like engineers inventing things because they're developing something brand new. The MBAs coming in

doing things, their only thing they're interested in is making money and and you know defining a market or anything else and that's great, that's wonderful.

uh and you know it needs to be done but there's not this passion of changing the world from either a structural point of view or from some technological point of view or anything else. There's none of

that passion. It's all about how fast

that passion. It's all about how fast can I grow this thing so I can get money and I can get my bank account to get bigger. Right? I personally don't like

bigger. Right? I personally don't like that. Right? I'm not interested in the

that. Right? I'm not interested in the the bank account inventor or the bank account CEO. I'm interested in the

account CEO. I'm interested in the person who's going to invent, who has a desire to go make a big difference and change the world, right? Fundamentally

change the world. Um, and you know, right now we we always talk about disruption disruption disruption.

That's all wonderful, but it's a bunch of empty words to be honest. Very few

people can actually do real disruption, right? If you look at hardware, there's

right? If you look at hardware, there's invention in hardware all the time. How

often is it disruptive? Not very often.

But what happens is every new bit of technology every something new happens you know whether we go to fiber or you know something else now all of a sudden you know networking can be done a lot faster there's a fundamental change that

happens so there's lots of changes that happen in those pieces of infrastructure that nobody celebrates but without those none of the other things that people are doing could actually happen right think about you know the switches that are

there the power infrastructure there the chips that are there all of that stuff has to be present in order for us to do any of the AI stuff and then all of the the groundbreaking work that people have

done to build systems, right? This is

hard work. This is not something where you kind of put something together.

Think about something like Kubernetes.

Kubernetes took a long time. It was, you know, started at at Google under a different name, came out, became Kubernetes, and now runs a lot of infrastructure, right? That kind of

infrastructure, right? That kind of invention takes time and that becomes pervasive. That's the stack that

pervasive. That's the stack that everybody builds on, right? You started

off with the 8080 microp process, right?

Then we went to the Pentium, then we went to, you know, sort of other things and now we suddenly have discovered that Nvidia chips are really good, right?

They were nothing more than a graphics processor and a floatingoint unit that people used to think about. But now we moved our computation to that and that becomes important and that has become important. But that is built on a whole

important. But that is built on a whole lot of infrastructure that was done. If

you look at the Nvidia chips, it's not just floating point. It's also a lot of networking. It's also a lot to do with

networking. It's also a lot to do with memory. Getting all of the things

memory. Getting all of the things actually working together is a non-trivial systems exercise and you have lots of people that contribute to it. What happens is we get people who

it. What happens is we get people who have these great visions to come out.

Unfortunately, as I said, a bunch of them are are MBA types and doesn't quite work. But then you have inventors. I

work. But then you have inventors. I

mean, as much as I kind of, you know, have a issue some issues with with Elon Musk, the guy is doing some pretty interesting stuff. I mean, you know,

interesting stuff. I mean, you know, reinventing how we do space. You know,

he didn't exactly invent Tesla, but okay, he he took it and he took it somewhere, right? There's a bunch of

somewhere, right? There's a bunch of things that he's doing which are inventions. And he's not doing it to

inventions. And he's not doing it to make money. He's doing it because is a

make money. He's doing it because is a vision that says, I want to be able to fundamentally change the world. And

that's the kind of inventor that actually makes a difference in the long run. Everybody else is sort of

run. Everybody else is sort of incremental. And fundamentally, let's

incremental. And fundamentally, let's face it, 99% of us are going to be incremental, right? Only the 1% are

incremental, right? Only the 1% are going to actually make fundamental changes. And those breakthroughs can

changes. And those breakthroughs can occur in many different ways. Whether I

discover something fundamental in science or I wind up having some way of putting a system together that's different. I mean, think about what

different. I mean, think about what Steve Jobs would with the with the iPhone. That's that fundamentally

iPhone. That's that fundamentally changed how all of us do things, right?

But it was the right time to pick the you know think about the wherewithal it takes to say I can take this technology for a screen. I can take this technology

for for uh you know a CPU. I can package it all together and I can deliver it first as a you know the killer application was not that it was just a phone but it also had GPS in it so I could do mapping right and sure I could

do email and all the other things but it just changed the way everything was done. There were only four basic

done. There were only four basic applications at the beginning, right?

But even those four were massive.

>> Yeah. People still use mobile GPS, you know.

>> That's exactly right. So all of these things are sort of fundamental changes that these people did. They had the vision to be able to do it. By the way, look at both of them. They're they

didn't go after saying I want to become the richest person. They basically said I have a vision of how the world should be. I understand these technologies even

be. I understand these technologies even though I don't understand them necessarily at the the the deep down level well enough to be able to integrate and I can I can assemble the team to be able to do that. That's the

other thing that's very important.

Great builders can attract great people to build stuff for them. Right? One of

the things about if you think about uh um you know somebody who's leading a war, right? They can command the troops

war, right? They can command the troops to walk into the worst possible situation, right? essentially the

situation, right? essentially the equivalent of walking through a wall.

But if there isn't a leader that says you can you can do it, I'm going to be there with you and I we you can do it.

Nobody would even try, right? And so you need these leaders that can pull you to basically do something that doesn't come naturally and then put the right set of team together where everybody plays off of each other and said, "Wow, oh, we

could do that. Oh, I didn't realize we could do this. Here's a different way of doing." And then all of a sudden you

doing." And then all of a sudden you invent something that is really, really breakthrough. But you need a leader that

breakthrough. But you need a leader that pulls assembles the team together and gets things going. Right? One of the critical pieces of being an entrepreneur

is assembling a great team, right? I

don't care how good your idea is. If you

have a second tier team, it'll never go anywhere, right? You know the old the

anywhere, right? You know the old the old thing about, you know, a B team with an A idea will screw up, right? An A

team with a B idea will turn into an A idea and be successful. And that is absolutely true. It starts and ends with

absolutely true. It starts and ends with the people around you. End of

>> story%. No, I think you you said few things, right? So I would love to double

things, right? So I would love to double click on that. I think one thing you talk about the team that that I think 100% agree with you. You talked about like foundational building blocks which are contributing to our wave today. So a

lot of people who are listening, they may not even understand like what Silicon Graphics did or what are some of the key fundamental building blocks. if

you had to like in two minutes break down some of the key technologies or key foundations which are very critical for the AI wave today what would those be so I think you know we could start at the

sort of the lowest level of the architecture right we can start with chip technology you know um back in the 80s and 90s it was hard to get very high

density uh chips then what w up happening is we started getting to you know two micron kind of line widths a line width is basically how small a wire I can actually go

The moment I can do that, I can pack a lot of circuitry on there. The denser I can get, the more compute I can actually put on there. If I look at fundamentals of computer architecture, they were

actually invented by IBM in the 1960s, right? So almost all the things that we

right? So almost all the things that we think about with with um uh with registers that was actually done in the 1960s. IBM even had virtual machines in

1960s. IBM even had virtual machines in the 1970s, it was called VMCMS, right?

Most people don't realize it, but they were the original inventors of that, right? So you have the chip architecture

right? So you have the chip architecture which really really matured and the equipment that actually allowed you to make it. So you think about companies

make it. So you think about companies like um applied materials that enabled people like Intel and TSMC to make these semiconductors and then you have all the design tools that were sitting on top of

that. All the design tools that allow

that. All the design tools that allow you to say okay how do I actually put all of this stuff together? How do I actually go and make all that work?

Those kind of things are really really important. Then there was a language for

important. Then there was a language for actually being able to design right. Um,

so you could actually now turn the design problem into a software problem where lots of people could be working on things together to be able to do that.

So now you could build compute chips that were really good. At the same time, we had a whole bunch of stuff happening on the network side, right? It wasn't

just that we we send stuff over copper anymore. We could send it over fiber

anymore. We could send it over fiber optics. We could compress it much

optics. We could compress it much better. We could do a hell of a lot

better. We could do a hell of a lot faster job to be able to do it. Then we

could interconnect these chips uh much much faster. we could actually create

much faster. we could actually create racks that were interesting. But then we realized, hey, that generates a lot of heat and requires a lot of power. So we

invented cool ways to being able to do that. And so all the all the physical

that. And so all the all the physical infrastructure sort of got built up this way. Then as the physical infrastructure

way. Then as the physical infrastructure started getting built up, we created uh guys who put this network together around the world, right? So fiber was laid, machines could communicate with

one another. You have people like Okami

one another. You have people like Okami who could place things at the edges so that things were happen to be much much faster. Now this infrastructure gets

faster. Now this infrastructure gets built up around all of that and then there was a software stack that came on top of that. Unix was probably really really important in being able to do this. Unix kind of unified a lot of the

this. Unix kind of unified a lot of the the the disparity that was around in the world before that and then you started now having saying okay now I have a fundamental systems infrastructure I can go do. People started building

go do. People started building applications. One of the first things

applications. One of the first things was databases and in fact relational databases. IBM had databases before but

databases. IBM had databases before but not relational ones. By the way, the other thing most people don't realize IBM invented relational databases as well, right? So, so the the reality that

well, right? So, so the the reality that >> that's absolutely true, right? IBM

invented relational databases. They did

not patent it. Oracle came along, took all of the stuff, re-implemented it and created the market for relational databases, right? So then you have sort

databases, right? So then you have sort of storage mechanisms and all the things that you can go do for storage. Then you

have distributed architectures. Then you

build sis then you build software layers on top of that that take advantage of it and say oh I can I don't have to just use this one machine right I can now use virtual machines that's where VMware

came in big big step ahead in being able to go virtualize all the resources that were there then we started doing massive scale distributed systems all of that stuff built up then all the UI work that

got done right Apple did a lot of the seinal work that actually was based on stuff that came out of Xerox Park so all of this stuff kind of comes together in very nice ways. So you have systems for people to now use their imagination and

build applications >> package system is going to like now get >> the package comes together and the reality of it is every time we make a big change in the packaging some fundamental change can happen in what we

can expose and what kind of application we can build right that's a pretty key point right like in in this AI generative AI wave like we are seeing that right like suddenly we have given

models a capability to like ask a question and for a user what we've exposed is like hey you in a natural language suddenly ask a model to like help you write code or ask a model to

generate an image or like ask a model to like generate a song which obviously before we models had some of these capabilities but they were not exposed one also LLMs were just not at that

level so we suddenly have these two shifts so from your perspective what are some opportunities over the next 12 to 24 36 months you see maybe

two three ideas which we can maybe we can you know do something different here like when we spend two minutes talking each about these ideas for listeners who thinking about like hey what can they build maybe get some >> so so I think I think one really cool

thing and it's a hard problem right now is we have these these these models right these language models what we really need is world models right how do

we see and interact with the world appropriately right that's a very hard problem how do we get a robot to actually put a hand around solve a problem to pick something up and move it

Right. Um how does it feel? What the

Right. Um how does it feel? What the

things are there? All of that kind of stuff is probably the next decade I think of really really interesting work.

Right. And the interesting part about that is how do you collect the data for that? In language we had it easy. Right.

that? In language we had it easy. Right.

We had all this data that was sitting in the machines.

>> Yeah. Right.

>> For this you have to simulate, you have to watch, you have to do all these other kinds of things. So it leads one to to ask the question how are you going to build these world models? Where's the

data going to come from? A lot of people saying it's coming from simulation. Yes,

there's lots of good simulation that we can do to be able to go do that. A lot

of it is from watching real world things happen. Um and and you know, I'm I'm

happen. Um and and you know, I'm I'm sure we'll we'll we'll wind up uh we'll wind up doing that. But you know, the fundamental problem I have today with all of these large language models and

what we call AI is that AI doesn't know how to learn.

Right? The models don't know how to learn. You know, if you think about I

learn. You know, if you think about I have grandchildren, right? I see my little granddaughter going off and do something. She figures out how two

something. She figures out how two things can be moved and problem solves, right? It's not like she was trained on

right? It's not like she was trained on massive amounts of data to be able to go do all this stuff or very small life.

>> She hasn't had access to the she's had very constrained stuff, but yet she can solve very complex problems. >> But every time she does it, she learns how to experiment. She learns how to learn from all of those things. She

learns how to learn from the failure.

Then we guide her to be able to go do that. Models don't do that. Models don't

that. Models don't do that. Models don't

do that at all. Models, think about every time you want to do something new with a model, you have to retrain it.

Well, now you have to retrain it. You

spend millions and hundreds of millions, God knows how much money to spend and say, "Okay, now it's going to get incrementally better." But it doesn't

incrementally better." But it doesn't know how to learn. I think the core of this problem is we got to build models that can learn, that can observe, that can see things, that can understand

things. And that's really how I think

things. And that's really how I think the real world stuff can actually come in because they can learn by experimenting.

>> So which companies you think are focused on that or like which labs or like labs?

>> The only one that I've seen that actually has articulated that this is the kind of thing they want to do is is actually thinking machines labs right they just recently I forget who it is at thinking machines lab did a TED talk

recently that basically says hey we need to teach these systems how to learn. I

don't, you know, I I don't see um I think Meta, if they got their virtual world stuff and all that stuff together with this, I think could be doing interesting. Um, you know, Niantic is

interesting. Um, you know, Niantic is doing interesting work.

>> Yeah, like I guess Niantic came out with the Pokemon Go you guys can find. So,

that's some really interesting data >> that if I'm like, you know, in my 20s or early 30s and I'm like, hey, I have like good math background, CS background. I'm

like, hey, I want to work on a problem over the next 10, 15 years. I don't want to chase this AI hype. What should I do?

Like should I go like do a PhD or like go try to join a company? Like how would you approach that if someone wants to learn this and like really be involved in >> I'll go back to what I said about people, right? Where are the smartest

people, right? Where are the smartest people you can find?

>> If that's at a company, then you don't need the PhD to do it. You're getting a PhD level and incredible PE people in one place. You know, in the old days,

one place. You know, in the old days, the only way to get that kind of expertise was to go to a university, which was world renown, and you could go do stuff there. No longer is that true, right? You could go, I mean, honestly,

right? You could go, I mean, honestly, you could go to MSR, or you could go to Google, Google Brain, or you could go to OpenAI potentially, or some of these other guys where you're surrounded by extremely smart people that are pushing

all these boundaries and things of that sort. You don't necessarily need to get

sort. You don't necessarily need to get a degree to go do it. Um, the the advantage. So, what does what does a PhD

advantage. So, what does what does a PhD get you? The PhD is just a proof that

get you? The PhD is just a proof that it's a piece of paper that proves that you can do independent work that you can generate your own independent idea that's deep enough for some you know

that that says you have enough knowledge in one place that's deep enough that you can generate a good idea that does that that's not but that notion is not restricted to a university it could be

anywhere right I know some extremely smart people that uh actually never went to college but have unbelievable deep understanding of stuff that would

far exceed most professors that I can meet today, right? But they learned it, they understood it, and they worked and they were surrounded by incredible people to be able to make that happen.

>> That is great advice. So, we talked about an infrastructure problem or like a very deeply technical problem. What's

like maybe an application level problem which you think is pretty interesting?

>> You know, look, everybody's working on code right now. So, I think code is interesting. um you know my experience

interesting. um you know my experience with um the uh the various uh agents that are out there writing code they still need a lot of work um there's a lot of guidance they need and things of

that sort they don't really know how to abstract properly so I think building systems that can abstract and be able to really back off and abstract something

and really understand sort of a conceptual what's going on that by the way applies in law it applies in medicine it applies in all these other things Right? Medicine is a little bit

things Right? Medicine is a little bit of um is an interesting uh sort of sequ non sequator if you will. Um in medicine doctors learn how to pattern match.

Right? So the doctor is as good as the experience that they've had. In other

words, the number of cases that they've seen that's a classic example where I think uh the machine learning what we have today where we just give it lots and lots of data um it's able to do

really really well. And so, you know, you see um especially in imaging diagnostics and stuff like that, the machine learning models can often beat what the what the doctors can wind up doing precisely because they just have

more information that they've been trained on. They understand a larger

trained on. They understand a larger corpus than the human does to be able to go, you know, make that make that change happen. Um so, they don't know, you

happen. Um so, they don't know, you know, they don't need necessarily a lot of abstraction to go do it. The same

thing is probably true in law and maybe even in in um in accounting and things of that sort. And by the way, those are exactly the set of things that people are building sort of verticalized AI applications on today because it's

really a function of the right type of data being able to make sure that the right constraints are set up and you can actually get pretty damn good results out of that, right? Um I think for the

next few years, uh the ideal thing for me or I would see is um uh AI assisting humans, right? So human in the loop I

humans, right? So human in the loop I think is probably the most important thing that one can do over the next few years because we have problems with hallucinations. We have problems with a

hallucinations. We have problems with a whole bunch of other things of that sort. Human in the loop can actually

sort. Human in the loop can actually solve a lot of those problems and accelerate how fast the human can work.

So if you can now suddenly 10x the human, 10x the engineer, 10x whatever it is, now you could make a substantial difference in the number of cases you can see, the number of things you can go

do and your velocity increases and and uh and and you know your output increases and a result you can actually contribute more positively. So I think taking the human experience and

enhancing it using AI is the set of applications that I would work on because those the good news is the guardrails are the human being right they can see oh that's hallucination I don't have to worry about that if you

try to make it all automatic I think it's going to be problematic >> that's a great point right so like really think about applications where you can enhance a human experience there's some human in the loop and that kind of augmentation really helps you

get a get a good >> correct I mean you look at you look at the legal stuff That's exactly what it is, right? You look at the the um the

is, right? You look at the the um the the other stuff that people are doing.

That is really a a really rich vein.

>> Yeah. 100%. I love like there's so many legal AI stars. I'm I'm actually surprised like there that many. One

thing I wanted to ask you is that you also like have worked as an exec in some large companies. You were CTO at Jahoo.

large companies. You were CTO at Jahoo.

You've built startups. One of the one of the challenges a lot of folks have is sometimes deciding between a large company and startup. And and one of the motivation for a startup is hey I you know there's a lot of learning there's

fastmoving things there's less bureaucracy while at a large company you always get into processes systems which which delay things from your perspective

how does innovation get done at a large company which is very similar to a startup and which people may have a wrong conception about >> well so um large companies actually can

innovate and do innovate Microsoft's a great example right Microsoft probably the best example, right? Think about the changes that have happened since Microsoft formed, I think 1979.

>> They've written every single one of those waves. They may not have been

those waves. They may not have been first out of the block, but they've done a really, really good job of taking that innovation, understanding it, creating a product set around it, innovating on that, and going really deep to

understand that, and then having the account control to be able to go, you know, sell it appropriately. That's a

great example of a company where, you know, they can innovate. uh they might not innovate as fast as a as a startup, but you don't have to, right? Because

one of the things that a big company you can do is you can watch to see what other things are going on. You can see it's a self- selection process.

Something is doing really well, you have one of two choices, imitate it or buy it, right? You always have that option.

it, right? You always have that option.

Um it will be the case that for those people who are itchy, you're not going to be happy at a big company. Just not

uh you're going to be happy at a small company. But I would argue that you need

company. But I would argue that you need both experiences, right? What needs to happen is you can be itchy at work at a startup and learn a lot really fast. But

then what happens is the rate at which you're learning doesn't have any structure around it. So you're sort of putting up the structure. You're doing

everything. And sometimes what needs to happen is the raw data needs to actually be put into proper structure. So you can say, "Oh, this is how I should think about this. Oh, this is how I should

about this. Oh, this is how I should think about this." There's a reason why big companies are successful because they've structured things. That's the

time you go to a big company. You go to a big company and you go learn what's the structure. How do I think about

the structure. How do I think about this? Why did they create this

this? Why did they create this infrastructure? Why did they do this?

infrastructure? Why did they do this?

Spend time learning that. So now you learn that. Once you've learned that,

learn that. Once you've learned that, you can now go back to a smaller company, apply that and your itchiness to go to the next thing. I would argue the best thing to do is ping pong between them.

Right? Ideally, ideally what happens is you go to one of these companies that's quote unquote small company, but you're at a high enough level and the thing grows nicely and now you can go set the infrastructure yourself, right? That's

kind of the ideal scenario. But you need to understand why structure and and and uh discipline are really important as opposed to let's get out there and shoot

with, you know, both guns.

>> 100%. And I think like then there's a balance between chaos and discipline and structure and you know like large companies obviously >> have that >> otherwise they'll be pretty pretty big issue like you know when you're doing

quarterly earnings or things like that >> correct but you also asked the question is how do large companies do this right so large companies are are interesting sometimes large company says here's a division go and do whatever the hell you

want you don't have to be part of the infrastructure go build go do whatever the heck you want it doesn't happen all the time and in fact In in standard industries, I don't think it happens very often, but certainly in tech, it

happens. Go build this thing out. Maybe

happens. Go build this thing out. Maybe

because I acquired a startup and said, "Go run, do whatever you want. Build

your own salesforce. Do whatever the hell you need to do. I'm going to put some protection around you to be able to go make that happen." When you start seeing that, you start seeing velocity go up in terms of invention inside the

company. But then you get duplication,

company. But then you get duplication, right? These two divisions are doing all

right? These two divisions are doing all these five things that are underneath it. Exact. you know, they're doing it

it. Exact. you know, they're doing it differently, but they're now spending a ton of money doing it, right? So then

what happens is management comes in and says, "Oh, we're spending too much money on this thing. Let's centralize a bunch of these things so we can squeeze efficiency out, right?" So you learn,

you run to be as dynamic as you possibly can be for market share. You squeeze

efficiency out, you spend a little bit of time squeezing efficiency, and then you go back to the other model, right?

The ideal is you just go back and forth.

Now, the whole company doesn't necessarily work that way, but you always have some things that are going independently, some things that you're squeezing the stuff out of, and you're constantly cycling between these. Those

are the biggest, most successful companies in my opinion. Apple's a great example of that, right? Apple runs

things very independently, right? You

have to sign an NDA from one group to the other to be able to talk to to other people in order to do that. I'm serious.

>> So, what that means is there's, you know, you have your world. You're

executing. Now only if somebody at a very senior level uh right like the exec team level decides okay I need infrastructure to be able to go do that they'll make that call and they'll put

that that that thing in place but until then you're running on your own right and literally I mean I have I I know people at Apple who work at Apple who've signed NDAs with several different

divisions because they all work together and you can't say what this guy's working on and what the other guy's working on >> Apple's iPhone like there's a secrecy even between like people who are working on the same I for sometimes don't know like what the other team is like

building uh around like some other hardware or like >> so so you know you look at that and go how can that possibly work it works right there is a there is a living example it actually works right so you

know there are many many ways to slice the slice this thing uh you know in the case of Apple and companies like that where you're generating huge gross margins and things of that sort efficiency is not the most important

thing for you to do velocity is more important right but at some point you're going to wind up squeezing efficiency some point you're going that balance. So

like a company >> balances over time.

>> So the best companies are trying to find that balance and like you know sometimes people who are itchy may not find like hey this this company this balance is not working for us we should leave but at some points you know you may find a company which is like hey this is a

perfect balance of like giving me the chaos but also like the structure which I need to succeed.

>> Correct. So that actually segus into something interesting right which is companies that flip-flop between different ideas tend to fail. Companies

that are very determined that actually say, "I want to go in this direction.

This is the direction I'm going to go and they stick with it. They just work and work and work at it can be very successful." And I would argue that's

successful." And I would argue that's probably a good thing for people as well. Right? If you're really wanting to

well. Right? If you're really wanting to do something, really, really focus on it and do it. Love it and do it. If you

don't love it, change, >> right? I'm not the kind of person who

>> right? I'm not the kind of person who sticks with the same thing. Contrary to

my own advice, um I like exploring because I want to I want to learn lots of different technologies and so that's what my career has been. But you know I have plenty of friends who focus in one area. They become very dedicated and go

area. They become very dedicated and go and and pursuit things. In my case I'm very focused on technology. That's my

dedication right. So, it's a different kind of dedication, but but I am dedicated on being a technologist and being able to go do it, which is why I constantly keep up with everything that's going on. It's really easy for

somebody to get to my age and say, "Well, I don't understand the new stuff." Right. Well, it's not because we

stuff." Right. Well, it's not because we gotten old and gray and our brains don't work. It's because you become lazy,

work. It's because you become lazy, >> right? Yeah. You got a drive,

>> right? Yeah. You got a drive, >> right? I've got a drive to go learn. I

>> right? I've got a drive to go learn. I

read papers. I go talk to people. I talk

to young people. I talk to old people. I

try to synthesize all of this stuff together. And honestly, it keeps me

together. And honestly, it keeps me young. It keeps me excited and keeps me

young. It keeps me excited and keeps me going.

>> No, I think that that is like super important, right? Like the key the key

important, right? Like the key the key learning piece. So, so maybe like this

learning piece. So, so maybe like this is a good segue in like you know like what you're working right now at Pepperdata which uh which uh is your primary kind of uh focus uh beside other interesting

things you're working on. So I was looking at pepper data's website like what was pretty interesting is you it said like hey dynamic kubernetes optimization save 75% and as I don't

know a lot about kubernetes and I was like okay what does this mean most probably like hey there's some cost savings involved maybe kubernetes has some fluctuating workloads and you guys are optimizing them but I just want to

hear from you like you know as as a domain expert like what does this mean if if I'm a non-technical user how would you explain that to me and if I'm a technical user how would explain this to me.

>> So, let me let me tie this back a little bit to my my journey. Um, so I've always been very mathematically inclined, right? Hard math problems are something

right? Hard math problems are something I love in my spare time. You know, I sol I solve um problems. Call me crazy, but I solve oompiad problems. It's fun for me. Um, in particular, there's one area

me. Um, in particular, there's one area of math I really like. I love

optimization, right? You know, different kinds of optimization, but I love convex optimization. This is an area that I

optimization. This is an area that I really understand really, really well.

But anything that's got a optimization thing, I'm like, "Okay, sign me up for this thing, right? Because we're gonna we're gonna we're gonna figure out a way to take something hard, squeeze something out of it, and then get get

something much better." I love that.

That that absolutely motivates me. So,

at Pepperdata, it's the same thing. And by the way, I've seen this this story in computer science in general. So in computer science what happens is whenever somebody's developing something and

they're needing resources whether it's compute whether it's IO or it's uh it's it's uh memory or GPU or whatever it is they have an idea of how much they think they need they kind of size things to

the maximum but what happens like every other resource on the planet you size it for the maximum but what happens is you get fluctuations right so at some point I may use the maximum but a lot of the

time I may not be using the maximum so all of a sudden there's this capacity that is sitting there wasted right and so the idea is can you somehow put a

useful work in the capacity that's wasted so I might be using something where the fluctuation is like this then when I'm dipping is there somebody else that can come in and use that resource

and as a result that resource is now used much more right so if I think about uh at any given point in time I want the CPU to be running at 95% I want memory

to be running very high at very high, right? Because what I've done is it's

right? Because what I've done is it's not sitting there empty. I'm not waiting for the guy to finish to be able to go do that. I'm actually using that

do that. I'm actually using that downtime to be able to schedule something new. That's the fundamental

something new. That's the fundamental thesis behind Pepper Data, right? You've

got resources that that are utilized that actually fluctuate with time. What

we do is we realize, look, there's a gap here. That gap can be utilized

here. That gap can be utilized efficiently and we can actually put more workload on there. And the the the the nice thing about what we do at Pepperdata is we don't touch the rest of

the system. We actually make sure that

the system. We actually make sure that the rest of the system gets information from us that allows it to act the right way automatically. Right? So we're not

way automatically. Right? So we're not changing theuler. We're not changing

changing theuler. We're not changing anything that we're going on. But what

we're doing is we're basically telling different things that orchestrate what's going on in the environment, hey, you could do more. Go ahead and send more.

There's more space available. There's

more more CPU available. Oh, there's

more. There's a different kind there's GPU available. So that what winds up

GPU available. So that what winds up happening is then then whatever is doing the scheduling says, "Oh, okay. I'll go

ahead and put that on there." The

problem with that happens is that if too many things suddenly get put into that slot and now you get failures, right? So

you have to now watch to see whether you're going to get to a failure state or not and then say no more preempt or slow things down, right? That's

basically what it is. So you have to set guard rails to say how how bad can you wind up doing? So, think about, you know, kind of like the the the day my days at Yahoo, right? Yahoo at that time

had roughly 1 million machines. If we

ran those machines at 80% efficiency, roughly speaking, 200,000 of those machines weren't really doing anything useful at any given point in time. So,

what you really want to do is run them at 95% efficiency. And that's we developed a bunch of techniques to be able to do that. One of which is called overallocation. That's part of what we

overallocation. That's part of what we do at Pepperdata. But it's basically the idea is let's get that resource that is sitting there to basically higher efficiency to be able to go make that

happen. Think about office space, right?

happen. Think about office space, right?

We have office space. People only use it during the day. Realistically, if we were clever, we would actually fill that office at night also, right? And now all of a sudden, you get much more result

out of that, right? So that's the point.

There's a resource that's sitting there.

Why not use it for something else, right? be clever and use it for

right? be clever and use it for something else.

>> I love your uh explanations. So, so one followup question there is right like so in Kubernetes this makes a lot of sense in today's era you obviously have this AI wave and obviously Kubernetes are relevant for that as well but GPUs are

super relevant and people are buying left right center GPUs for like pre-training inference etc >> are you guys working in that space as well because I think optimization there is super critical

>> so so we we absolutely are the the first thing that our customers have told us is you look you ask a developer what GPU they want they'll give you the most expensive one that you they can get

their hand on. Nobody's going to say, "I want the one that's two generations back." Nobody's going to say that,

back." Nobody's going to say that, right? The other thing is they're going

right? The other thing is they're going to say, "I need it at 3:00 in the afternoon." Right? That's the time I

afternoon." Right? That's the time I need it. So, they don't think about

need it. So, they don't think about where everything else sits and how everything else. So, the first thing

everything else. So, the first thing that we do is we kind of make it clear what's the demand and what's the pentup demand that is available for these things. What's pending to be able to go

things. What's pending to be able to go do? So now you can immediately identify

do? So now you can immediately identify that, oh, you know what? Everybody wants

to use Blackwell at 3:00 in the afternoon, right? No, we don't have

afternoon, right? No, we don't have enough Blackwell chips to be able to go do that. So now you can go and have a

do that. So now you can go and have a conversation with every user and say, "Hey, do you really need the Blackwells or can you do with, you know, H100's?"

And instead of at 3:00 in the afternoon, can you run this at 2 or maybe I can give you a black well if you do it at 4 in the morning, right? Those are the conversations you can start having where you can now take your supply which is

finite and essentially move the demand around and size it so it sits in the right spot. So what you're doing is

right spot. So what you're doing is you're basically saying I have finite resources. I'm going to allow those

resources. I'm going to allow those finite resources. I'm going to make

finite resources. I'm going to make everything available. Give you an

everything available. Give you an analogy.

Think about conference rooms. You have conference rooms with you know 20 chairs. You have conference rooms with a

chairs. You have conference rooms with a large desk. You have conference rooms

large desk. You have conference rooms with telecommunications capability, video capabilities, right? you have

conference rooms with, you know, shades that come down or or they don't have any, you know, natural light, whatever.

Now, all of a sudden, everybody wants to schedule these conference rooms. Everybody says, "Oh, I want one that's got, you know, 20 chairs in it. It's got

a large conference table. I want it at 3:00 in the afternoon. I want it in this building, and I want it to be so that it doesn't have any natural light in it so I can have clarity of the the video that I want to be able to get." Well, you're not going to get that. But you have

other conference rooms are available.

The other conference room, you may trade off one of these things or the other.

It's the same kind of problem, right?

It's that resource that everybody says, "I need, but if you actually have the conversation, you can do that." So,

right now, we allow people to have that conversation. In the next generation,

conversation. In the next generation, what we do in a few months, we'll do this automatically. So, think about like

this automatically. So, think about like conference room scheduling happening automatically. I understand what you've

automatically. I understand what you've got. I understand what you need. I'll

got. I understand what you need. I'll

actually move it around in the right way to be able to make that happen. Right?

So, that's that's important. The

interesting thing is the way we do that is we have this technology which we call fingerprinting. So we look at you know

fingerprinting. So we look at you know uh in in the Kubernetes language right what namespace this comes from whether there are any metadata associated with it what cues the thing is in how is it

named all and all its past history that gives us a really good understanding of what the best place is to put it right and and so we basically use that information create a fingerprint and

then basically say hey next time this guy's asking this is really what they're going to need and we'll put them in the right spot right that fingerprinting is something that we do It turns out that

that same fingerprinting can also be used for execution. So modern day GPUs can actually be sliced, right? So Nvidia

GPUs can be sliced, I think, in eight ways or nine ways, I can't remember. And

what you can do is you can say, "Oh, I give me a fractional GPU." Now, a fractional GPU may not be enough for what you need to do, but it may be. So

what we do at uh with Pepperdata is we basically say we'll manage this this pool of GPUs for you. But instead of thinking about them as full GPUs, they're going to be some full ones, some

half ones, and some third ones. Right?

We've sliced them in half, then third.

And now we're going to move your workload into these different buckets to be able to see if we assign it to a third and it fails, we'll move it to the second half. If it fails in the second,

second half. If it fails in the second, we'll move it to the full. And now

you've got the full execution that you need, right? But what it does is it

need, right? But what it does is it allows more stuff to actually take place simultaneously because not everybody needs the full GPU. Some can actually use the slice GPUs and as a result you

can actually create more utilization on the GPUs. At the end of the day GPUs are

the GPUs. At the end of the day GPUs are hard to get. They're expensive. If you

can increase utilization that's a lot of money that's being spent 100%. So the

question I have is that like you know so a lot of companies are trying to do these optimization things like what's unique about pepper data because I when I Google I'm like hey like let's see like who else is in this space. So there

seems like there's a there's a list of like you know 10 plus folks popped up.

>> So there there are there are people doing different kinds of optimization um you know people doing for specific training people doing for fine-tuning and all this. Ours is more generic. If

you have any jobs that are basically started or using Ray on Kubernetes, we can basically do that for you without it all automatically. There is no u uh you

all automatically. There is no u uh you know there no additional work that you have to do to be able to go do that. Um

we're we're very very unique in that capability. Very very unique. And this

capability. Very very unique. And this

fingerprinting technology is the key thing for us to be able to go make that happen. My understanding is that

happen. My understanding is that obviously you cannot cannot share some of your customer names, but a lot of Fortune50 companies are utilizing you for like core infrastructure today. So

like this is >> Yeah, I'd like to make that stronger. We

have two of the Fortune five.

>> Wow.

>> That's not so bad.

>> That's crazy.

>> That's that's pretty impressive. Like

for a for a small >> not a small for a tiny little pette little company.

>> Yeah. Midsize startup like you guys have like the Fortune 500.

>> I still think we're a tiny little company and and Yeah. Yeah, I mean we have, you know, most startups go for, you know, as as VCs call it, logo velocity, right? They have tons of small

velocity, right? They have tons of small companies signed up and all that kind of stuff with very little, you know, revenue associated with them. We

actually are in in a in a kind of an interesting position. We have a few very

interesting position. We have a few very large customers and they're all names you would recognize. There's not a single one that you would say, "Who's that?" You know, every single name and

that?" You know, every single name and all those names are household names. on

top of that. So, it's not like you have to be in tech to understand it. You can

pretty much understand what we got. So,

we're very fortunate from that regard.

And by the way, tiny little team doing this, right? We are, I don't know, 33

this, right? We are, I don't know, 33 people or so.

>> How did you guys convince some of these logos to work with you? Like what what was >> it's called? It's called proving that the works.

>> Look, at the end of the day, >> basic basic 101.

>> At the end of the day, if it works, people will buy it, right? And and you know the other thing is um the problems that we solve are most relevant to the people with the mo

biggest workloads and they're the ones who have the most economic pain. If

you're a small startup, you know, and you're spending money on this, yeah, it doesn't really matter. But if you're spending hundreds of millions of dollars on something and I can save you 20% on

that, that's pretty significant. But in

order to be spending hundreds of millions of dollars, you can't be small, right? So by definition the our problem

right? So by definition the our problem and our sweet spot is at the P places that are the large ones and therefore those are the guys that we go sell sell you know it it was the case probably 3

four years ago where we would do a P and we would insist that you do the PC on in production and the customers would say no way then we told them who our customers were and like oh it's good

enough for them okay we'll try it now we do PC's you know we do a PC on a non-production system just to prove we don't break anything. But to actually get data and actually do tuning and stuff like that, we insist that it's on

production and we don't have anybody that says no to us. No one.

>> That's amazing, >> right? That's credibility,

>> right? That's credibility, >> right? And that's reference, right? It's

>> right? And that's reference, right? It's

it's it's really reference. We we solve a real problem for really large businesses. They trust us. We support

businesses. They trust us. We support

them properly. We're there. And you

know, look, at the end of the day, selling is not just about the product.

It's about the product, the service that you provide, the relationship that you do, how you understand what they do. And

uh you know the companies that we deal with are also very very sophisticated.

They want to be able to deal on the other side with people who are equally sophisticated and can understand what kinds of things they got. And we're very fortunate in having a team that actually understands distributed systems

extraordinarily well. We understand how

extraordinarily well. We understand how to do, you know, deep low-level infrastructure stuff extraordinarily well. And there aren't many many

well. And there aren't many many startups that can kind of match that.

>> Clearly like getting through the Fortune 5 is uh is no small feat especially if you're just a small startup. That that's

that's crazy. uh a segue I want to make here like which which is relevant to like GPU work and in this AI kind of like capex spend we're seeing massive data centers being

built and you know primarily they are just for GPUs are like the core thing which are being hosted and one of the reason is that like the amount of demand for like inference and like training and

all of these things are just off the charts and I was kind of looking into like how these data centers are kind of built and what are some of like the core things which go into them. The the

biggest bottleneck seems to be energy, right? Like cuz you know we can most

right? Like cuz you know we can most probably solve for TSMC moving fast, some of the fabs moving faster and Nvidia's revenue in 2025

close to $200 billion can fund most of these infrastructure uh spend which is needed just by itself irrespective of like you know other

hyperscalers. But when it comes to

hyperscalers. But when it comes to power, it's different, right? Like it's

not like a power company cannot just start producing turbines in six months notice. Like they need stable revenue.

notice. Like they need stable revenue.

They need like long-term perspectives.

There's regulations involved. How do you think US will solve that? Because for a pretty long time, these data centers have used equipment and infrastructure built for the steel mills or the old

autom motor sector. And now we're getting to the the threshold of that.

The other thing that goes along with power is heat, right? Is cooling, right?

Uh, you know, if you look at a rack, the amount of heat put out by a single rack is going up in order of magnitude with almost every generation. So, it's

really, really hard to figure out how to go do that. Also, the physical >> more power is needed for future generations.

>> I mean, uh, you know, I mean, depends on where you place them, whether you use convection or not and things of that sort. I mean there there are people that

sort. I mean there there are people that say oh I'm going to build a um a data center in space and because the problem is there's no convection in space you

know yeah it's how do you how do you actually transport that heat away is a real problem we need convection to be able to do that so I don't I don't quite believe in all that but uh you know uh

power and heat are a real problem so let's in terms of heat we're doing liquid cooling of different kinds but the physical infrastructure has to keep up with that right um you know the more

heat we generate the more liquid essentially that we have to pass by it right whatever conductive u um element there is whether it's air or you know whatever else it is that you want to put

on there and with liquid you got to then move more liquid per unit time or larger volumes of it then the infrastructure the piping and all that stuff becomes

heavier it becomes heavier so you have to size and build those data centers to think about volumes of water and all these other liquids you're going to move 5 years from now or 10 years from now in order to get the infrastructure to last.

Otherwise, you're going to build something that may work today, four years from now, it you know, you don't have enough density in the data center to keep, you know, keep doing something interesting. So, there's all kinds of

interesting. So, there's all kinds of interesting things. It's not clear to

interesting things. It's not clear to me. We used to build data centers that

me. We used to build data centers that say, "Okay, we're going to last 20 years." Right now, with the way things

years." Right now, with the way things are going from heat and things of that sort, uh, you know, 5 years is kind of the time bound. So, you can't pour a lot of concrete. you can't put a lot of

of concrete. you can't put a lot of money in there because you might have to rip it all out. So, what's the design of the data center that needs to be done in order to be able to go do that is a very interesting problem. It's a very tough

interesting problem. It's a very tough problem to to solve. Um the second thing is just the amount of energy that to look the fact of the matter is our grid uh fundamentally in the US is terrible.

It's old. It needs to be upgraded. It's

not clear that we have the wherewithal as a as a country to be able to go do that, which means that we're probably going to need power generation near the data centers. So whether it's a utility

data centers. So whether it's a utility that's doing it or whatever, we're going to need to go and and and go do that.

Now, what's interesting is if you go to a utility that's got sort of fixed output and you say I'm going to suck out all your output, then you know, question is is it going to raise I mean people are complaining like in Virginia, right?

my utility bills have gone up because all these data centers are sucking up everything I've gotten. My water bills have gone up because of the same thing.

Um, you know, but but it's interesting if you add more capacity to this thing and you can you can essentially say the capacity is going to be there.

>> What happens is over time you can actually build enough capacity to lower everybody's cost. So this spike in in

everybody's cost. So this spike in in cost is a short-lived thing and it should go away provided that the utility builds enough infrastructure to be able to supply the demand that they need they need to be able to have. they can't

continue to maintain that that that price delta over time that has to go away. So, uh you know to me um uh uh

away. So, uh you know to me um uh uh small nuclear is probably a really good solution uh for being able to go do that. Um I think there are a number of

that. Um I think there are a number of people who are working on that uh and we should really really think about that.

It's efficient. um you know now they're they're they can be designed in such a way that the the risks of all the you know fallout stuff that everybody imagined uh you know end of day kind of

thing can be dramatically curtailed and they're small they're contained and things of that sort you know there's all obviously a question about nuclear material and terrorism and all that but I think we can we can put our arms

around that but that's probably the best the second one is probably using um is using um renewable resources um but renewable resources that are available all the time are not

necessarily local to you, right? And

they may be far away. So, we need transport and the grid to be able to go do that. If you look at China, China is

do that. If you look at China, China is building some of the largest solar farms on the planet, right? I think

cumulatively they'd be larger than everybody else's solar farms combined.

And they're putting the infrastructure in to get that electricity back to where it's needed. you know, places like

it's needed. you know, places like Shanghai and Beijing and all these other places and where they've got their data centers, they're moving that stuff, right? They're also investing in battery

right? They're also investing in battery technology and stuff because the thing that we need to be able to do is we need to store it, right? To be able to say, "Oh, you know what? The the during the day I generated it, during the night I

can't generate it. I got that." There's

also backup, right? Will we use natural gas? Probably. Um, you know, I'd like to

gas? Probably. Um, you know, I'd like to hope that we're not going to use coal very much, but that may still be the case. But natural gas, oil are two and

case. But natural gas, oil are two and and coal are certainly possibilities to fill in that gap when there's demand and we don't have it. But I think if we could do a good job of generating it,

being able to store it so we can use it offline and generating it in a way that's friendly to the to the environment, uh then I think we can actually solve this problem. So good

news is battery technology is going really really well. We're move we're making massive improvements in battery technology as a result of that. this,

you know, the the the battery power that's generated elsewhere can be used and stored. But I still come back to one

and stored. But I still come back to one fundamental thing, right? We don't have a grid that we can go do this with. We

need to upgrade the national grid to do it. Until then, you're going to get

it. Until then, you're going to get local generation. You'll get local mini

local generation. You'll get local mini grids that people will set up that allow all this stuff to happen, but energy will be generated many different ways. I

think the longest the the best uh answer in the long run is going to be small nuclear. No, I think battery tech

nuclear. No, I think battery tech obviously is is a big bet as well like with long size small nuclear. So that's

that's really helpful. One one kind of thing I I wanted to chat with you about is because you know you work around the globe, you work for like multinational companies uh in very different regions. So we

obviously in this AI wave there are two big countries and to some extent it has become like a wave of AGI and like you know countries are trying to approach but really there are just China and US

who is really competing for this. if you

realistically think about it. And I was looking at a chart about the energy buildup over the last like 10 years and China is just crushing it compared to

everyone else. And then secondly,

everyone else. And then secondly, obviously their model companies are pretty active. They have some of the

pretty active. They have some of the best uh refinery capabilities for rare metals which go into a lot of these ships. Uh and then one interesting thing

ships. Uh and then one interesting thing is that every three years chips gets depreciated and the race starts again from from an infrastructure perspective.

So maybe like China doesn't win in the short term, but long term like they may have a real >> I think I think I I think they've really thought this through, right? So let's go

back to rare earth, right? Rare earth is fundamental for building infrastructure whether you're talking about chips or you're talking about batteries,

>> right? So they have invested heavily in

>> right? So they have invested heavily in all of that stuff. They realize that's the building block for whatever it is that you wind up doing. Um you know allows them to build solar cells

cheaper. Allows us to store the data.

cheaper. Allows us to store the data.

Allows them to get uh by the way pollution out of the environment by having many many uh car uh you know car companies that are electric and going fully electric in that. Uh they've done a really good job of being able to say we're going to lock up that

infrastructure. They've done a very very

infrastructure. They've done a very very good job of that. Um the other thing that I think the US had a lock on intellectual work right through our university systems, our PhD programs and

all these other things. And so what we started seeing was more and more foreigners coming in to work on PhDs, you know, from India and China and staying here to be able to do the work

that there is because their their universities weren't quite on par. They

weren't quite there, right? The IITs are good, but they're not on par, so to speak. China was the same way. Tingua

speak. China was the same way. Tingua

had the same problem. But now you look at those universities. They're amazing.

They're producing very very high quality graduates. And so even the people that

graduates. And so even the people that are coming here, they're actually seriously think do I stay here or do I go somewhere else? That lock that we had on sort of intellectual capital is

becoming unlocked. We're losing that.

becoming unlocked. We're losing that.

And I think the the the more we rob the um educational system of juice to be able to go do that, the worse it's going

to become. That's number one. You look

to become. That's number one. You look

at the number of Chinese authors in all these different papers. It's massive.

The number of Chinese authors or Chinese authored papers, just Chinese authored papers in AI now exceeds those that come out of US.

Right? So now you're you're now starting to see innovation actually come there.

Then you look at Deep Seek. I mean

they're doing some clever things. You

know, you know the old story, mother is ne necessity is the mother of invention, right? Okay, you started them of chips.

right? Okay, you started them of chips.

Well, they're going to do something clever with it, right? And they did.

That's going to continue. So that lock that we had on intellectual capital isn't going to be there. They still

don't have the fabs to be able to manufacture the chips and do all the other things that are there. But they've

shown that they can. If you look at what they did with batteries, they didn't have anything. They just had the rare.

have anything. They just had the rare.

They've now built battery fabs that are just unbelievable, right? They're going

to do the same thing on the other side.

Will they be able to go get that? I

mean, there's no lock that we have on these things. Yes, we have tools. We

these things. Yes, we have tools. We

have, you know, stuff that we can go do, but at the end of the day, they can acquire it. And if not, they'll build

acquire it. And if not, they'll build it. They might be slower, but eventually

it. They might be slower, but eventually they're going to catch up, right? It's

going to get good enough for them to be able to say, "I don't need anything else from the outside to do it." The third thing that we have to understand is the US is much smaller than China

in terms of its ability to produce talent, right? We have an outsiz

talent, right? We have an outsiz appetite for invention, but we don't have enough people to fuel it, right? So

we are rely we rely very much on you know people coming in from the outside to go do it.

>> Well guess what China can actually rely on itself to do it. It doesn't

necessarily need the rest of the world to go make that happen right. It's got

the horsepower to be able to make all that stuff happen. It can be self-sufficient in that thing. So the

talent pool can be local and it can be a vast talent pool compared to what we can create. We were relying on other people

create. We were relying on other people coming in. If we cut that off, well,

coming in. If we cut that off, well, that's going to that's going to create a limit to what we can actually do here from an intellectual perspective, right?

So, those are the kinds of things that are there. So, they're they've set up

are there. So, they're they've set up the right intellectual thing that they can start competing and they have started competing very much in the infrastructure side. They've competing

infrastructure side. They've competing very well. On the energy side, they're

very well. On the energy side, they're competing very well, both generation and storage, right? Um the chip side,

storage, right? Um the chip side, they're not there yet, but on the software side, I mean, you look at some of the innovations that have been great.

I mean the latest one that I can point to is I think it was last week or the week before DeepSync put out put out a um a paper that basically says hey instead of doing text and you know sort

of OCR kind of thing. If I just do convert text into an image and instead of tokenizing using uh using uh letters and words and things of that sort text I

you tokenize use using vision I can actually get tremendous compression tremendous like 10x >> that's phenomenally interesting because if you can do that you can now start thinking about well I can

potentially take the context window and increase this by a factor of 10 >> because I got tremendous compression on this thing right that's a very very nice result, right? Uh just recent they did

result, right? Uh just recent they did it right now. The very first result that they had um u that they talked about that made their model small and stuff like that, you know, the the accounting

for it only cost a few million dollars is a little bit specious. And the

technique that they use actually was invented at Google back in 2013. Um if I remember right, um um who's the the the

guy at Google who actually talked about it? um he's a senior scientist. I can't

it? um he's a senior scientist. I can't

remember his name now. Uh but they actually built it. They talked about this technology to be able to go do it.

It just never got used. If you think about it, Google sat on a whole bunch of stuff for a long time. I mean, you know, Transformers, right? All of the stuff

Transformers, right? All of the stuff that you go do essentially gets to one paper. Right.

paper. Right.

Right. It's it's the one seinal paper that everybody talks about.

>> Attention. You're talking about attention is all >> the attention paper. Yeah. attention is

all you need I think is what the paper's called. But there's a whole bunch of

called. But there's a whole bunch of stuff there. The the one thing I the one

stuff there. The the one thing I the one thing I think where we do have a slightly better advantage than the Chinese and I don't know for how long is I think our systems thinking is better.

We think about not just the software stuff but we think about the whole stack. We have people that have been

stack. We have people that have been around and built a lot of things before because we've been ahead on these things. We can think all the way up and

things. We can think all the way up and down that stack. And I think there's an advantage there. How long that advantage

advantage there. How long that advantage lasts, I don't know. because I know people in China that have now built systems. They're actually assembled systems. They're working on cooling them. They're building all this stuff.

them. They're building all this stuff.

So, they're learning also and they're learning very fast and they're also reading all the papers and talking to colleagues over here. So, the the the the rate at which they can absorb is

very very high and it took us a while because we invented it. But the deep expertise that we have they still can't replicate. But I think it's only a

replicate. But I think it's only a matter of time. if you had to like think about like how can how can the US stay ahead in this curve like what are some things are top of mind for you you

know maybe like a short 30 secondond 1 minute >> well I I think first is is is just produce more more smart people with with advanced degrees uh and and keep them here I think that is where the thing

starts >> your talent >> that tal talent to me is always the most important thing the second is don't take funding away for doing basic research It doesn't make sense. And by the way, what

I just said applies to biotech as well, right?

>> Encourage curiosity. Un curiosity, give them money, go ahead and and actually go do this. Uh, you know, it's exactly what

do this. Uh, you know, it's exactly what the National Science Foundation did in the ' 50s, '60s, and 70s, right? the

national lab, all the programs that were there, all the research things that were there, the shot to the moon, all of these other kinds of things. These

created fundamental technology things that actually made stuff happen, right? SpaceX wouldn't

exist today if it weren't for all of the stuff that NASA did to launch Apollo, right? I mean, there's a lot of learning

right? I mean, there's a lot of learning that come along with that, right?

There's a lot of understanding. There's

a lot of physics that came along with that. and and and we need to keep doing

that. and and and we need to keep doing that. We need to keep basically saying

that. We need to keep basically saying we are the world's place where you go to invent and to discover.

If you create that environment where people can invent and discover and you give them the the ability to go do that.

If the invention and discovery happens here, the fundamental ones, the companies will happen here, the companies happen, the rest of the environment kind of picks up to be able to do it. At the same time, we do need

people. So, immigration is important.

people. So, immigration is important.

>> Those are two solid points. So, on on that note, like you know, I would love to wrap up, but I definitely want to get a story from you cuz you you've seen like so many different ways you met

different people at different stage of their career, successful and failures.

What is a story which has stuck with you from a motivation perspective which g gives you even inspiration and motivation today which you may want to share with us. Uh to wrap this up, I

could take a commercial story which is which is uh an interesting one. Um

uh and I I may just sound old and a little bit um you know long in the tooth on this thing.

Today access to capital is so easy that people blow it very very quickly. They

don't value the money that an investor puts into a company. In the old days, every dollar made a difference. And it

wasn't the land of free money the way we have it right now. You had to prove you had expertise. You had to have done a

had expertise. You had to have done a bunch of things. You came back and you looked at it. Maybe innovation was maybe a little bit slower, but at the same time, the value of the dollar made a

difference. And the second thing was the

difference. And the second thing was the fact that somebody gave you a dollar wasn't like, well, if I don't succeed, doesn't matter. They have plenty of

doesn't matter. They have plenty of money, which is kind of the attitude now. Our attitude was if somebody gave

now. Our attitude was if somebody gave you a dollar, it's your job to return it and more. That's really part of your

and more. That's really part of your job. That's a responsibility. That's a

job. That's a responsibility. That's a

moral obligation. That's lost today, right? I see, you know, young startup

right? I see, you know, young startup founders say, "Oh yeah, we, you know, we raised $50 million or $15 million. Oh

it didn't work. Okay, $15 million has gone down the drain." There's no feeling of, "Oh my god, I blew somebody's $15 million." Doesn't happen.

Which is really problematic to me. But

if you value money, spend it properly and spend it wisely, success does happen. I'll give you a perfect example

happen. I'll give you a perfect example of it. There was a company called

of it. There was a company called Veritas run by a man named uh Mark Leslie. Mark uh I think took on the

Leslie. Mark uh I think took on the company when it was like around $10 million or whatever. And for years they barely made payroll years. Then all of a

sudden hit a hockey stick. Suddenly

billions of dollars in revenue and all these other things. then and and in in the valley people were writing about look at this startup in the last three years look how well they've done everything else and Mark would remind

people uh you forgot about the six years before that when we just basically were really really really you know hunkering down to be able to do it was the hunkering down the belief in

what they wanted to do the reason why they wanted to do it and everybody was there to say yep we share that vision we'll go through the hard times when they could have jumped to other companies as well easily at that time you know think about Sun Micros

systemystems and all these other things.

But they stuck with it. They went there.

They went through it and they got it and then they wound up getting Escape Velocity and you know the company was worth billions of dollars. Mark did

really really well personally. Uh which

is really an amazing thing for him because he stuck it out. He stuck it out. He watched every penny and things

out. He watched every penny and things that sort. And he told me two things.

that sort. And he told me two things.

One is the guy at the end with the most money in his pocket always wins. Okay?

And the second is money in a startup is more important than your mother. Right?

These are really, really important things to remember. Long-term companies

survive through hard times. Hard times

are perfectly fine. Just because the guy next to you took a took an exponential path, doesn't mean you're going to do it. But if you believe in what you do,

it. But if you believe in what you do, if you believe that you've got real defensible technology and defensible things, I'm not saying, you know, share go after a fool's errand and keep going

after, but if the market is telling you this is working and you can continue to stay alive, it sometimes takes time to be able to go do that and it takes off.

You know, another example of a company that does this is ARM. ARM has been around for a long long time and look at it now. It's just taken off,

it now. It's just taken off, >> right? because it's like the the the

>> right? because it's like the the the right thing for everybody to start thinking about, right?

>> Actually, Nvidia Nvidia was a graphics chipmaker up until around 2012 or so, >> right? At that time, it was still a

>> right? At that time, it was still a graphics chipmaker. I mean, the

graphics chipmaker. I mean, the valuation was sitting around 1820 billion, right? It was doing a good

billion, right? It was doing a good business. It was all about, you know,

business. It was all about, you know, making games and video graphics and and and special effects and all these other things. And then all this stuff came

things. And then all this stuff came along with uh with machine learning and all of a sudden it became more and more useful and oh my god this this is exactly the right processor to do it. So

if you look at Jensen's journey on this it was very patient. It was very deliberate. Did he have in mind when he

deliberate. Did he have in mind when he founded the company that he was going to be at the for at the forefront of AI?

No. Didn't even exist at that time. What

it was was I'm going to deliberately run a business. I'm going to own a market.

a business. I'm going to own a market.

I'm going to build good solid technology and I'm going to be smart enough to constantly be sniffing and saying where's the next opportunity for me. You

know what he did was he did graphics and he went to self-driving. Right? That was

kind of the next thing that he wound up doing. And then all this AI wave hit and

doing. And then all this AI wave hit and that boom the whole thing took off.

Right? So the last six years has been where all the explosive growth has happened. Before that you know what last

happened. Before that you know what last six years is what 2018 2019 the company was formed in the late 1990s.

>> Right.

some of the early employees of Nvidia like who may have left before 2015 2016 like they miss like the the mass >> so that that's the one thing I I got to

tell you right you can never time these things you never know >> right be happy with what you've got you know and and and uh you know don't envy the guy with a little bit more money

because I can tell you one thing money is not going to create the happiness that you have you know your definition of happiness changes as you get older >> 100% Right? And as you get older, just

understand what your definition of happiness is. Happiness is not about

happiness is. Happiness is not about necessarily about having the most toys.

Maybe at some point it is. We all go through these kinds of silly phases, right? But at the end of the day, I'm

right? But at the end of the day, I'm going to come back to what I said earlier, right? It's about the memories.

earlier, right? It's about the memories.

It's about the people. It's about the set of things that you leave behind.

Because I got to tell you one thing, right? As I watch my elders going away,

right? As I watch my elders going away, they ain't taking anything with them.

>> Yeah. 100%. It's like the memories and the experiences. I'm sure like the

the experiences. I'm sure like the people who who built Nvidia in the early days and who've seen it success love it as well.

>> They love the fact that they were associated with regardless of whether they made $100 million or not, right? I

was part of the team that did it. We

built this thing. It was great and and even the guys who left during the during what I would call the game uh era of uh of Nvidia, they did some pretty cool and they were very happy with they

were on the top of the world. They were

motivated. They did that. They made some money on it. That's awesome. If they

held on to the stock, even better. But

the reality of it is it's the satisfaction of having done something substantial and the experience and the set of people you worked with and the things that you were able to influence.

Yes, it wasn't as big an influence as as the AI thing is right now, but fundamentally change in the gaming experience wasn't a bad thing to do either, right? Because you suddenly

either, right? Because you suddenly changed a lot of cool stuff, right? That

was also very cool, right? Suddenly

being able to have a render farm that was actually pretty reasonable and inexpensive and you didn't have to get these specialized pieces of equipment to do that. That was a big deal for film,

do that. That was a big deal for film, right? So, you know, these are all

right? So, you know, these are all things that you could look back on and say, "Wow, that was great." Like, I have friends who started to talk about to anymore, but to was fundamental in kind of creating the whole DVR thing and

being able to do a bunch of things around that. So be satisfied with what

around that. So be satisfied with what you can do because what happens is, you know, 10 years from now, maybe 20 years from now, maybe the AI wave won't look as interesting anymore. Won't be the big thing. Maybe it'll be the next thing,

thing. Maybe it'll be the next thing, right? Because we saw the internet. Oh

right? Because we saw the internet. Oh

my god. Oh yeah. Oh, I missed that opportunity. Then the mobile, oh, I

opportunity. Then the mobile, oh, I missed that opportunity. Oh, now you have the AI. Who knows what the next one's going to be? Maybe quantum is the next big thing that's going to happen, right?

>> There's always opportunity. Yeah. I

think like that's one thing about like tech or generally like financial markets is like there's always interesting things happening as long as you're focused on core things.

>> Correct. Correct. You know, you look at you look at historically in the past, right? You look at companies that were

right? You look at companies that were built. You look at people's who people

built. You look at people's who people were titans at that time. How many

people actually remember them? I mean,

you know, the the guys from the 60s that were the billionaires in that time.

Practically no one. The reality of it is it's the the thing that you invent that's good for now. Take stock in it, enjoy it, create happiness out of it and and and spread the happiness to other

people. How by first of all your

people. How by first of all your attitude and second by teaching and mentoring. Pass it on. Pass it on.

mentoring. Pass it on. Pass it on.

Right? That's the most important thing you can do.

>> I think this is a great point to end the conversation. Ash, again, thank you so

conversation. Ash, again, thank you so much for taking the time. This is a terrific conversation. I'm sure our

terrific conversation. I'm sure our listeners will enjoy as well. Thank you

so much again for taking the time and uh >> thanks and I really appreciate the invitation.