Stanford AI Club: Chamath on How to Win in the AI Era

Thank you guys for coming. Uh, today we have a very special guest who I think needs realistically no introduction, but we're going to introduce him regardless.

Um, Shimoth is the founder and CEO of 8090, an AI native company focused on rebuilding and modernizing enterprise solutions. Uh, prior to 8090, Shimoth

solutions. Uh, prior to 8090, Shimoth was a founder and CEO of social capital where he backed companies such as Slack, Sofi, Grock, etc. The list goes on. Uh,

before that, he was an SpaceX. SpaceX. Yeah, it's about it's

SpaceX. SpaceX. Yeah, it's about it's about IPO. Um, [laughter]

about IPO. Um, [laughter] uh, before that, an early executive at Facebook where he scaled the company to the size it is today. Um, and a co-host of the All-In podcast. Shamoth, it's a pleasure to have you here today. Um,

thank you for taking the time to join us.

Thanks.

I went to school with his mom and I just met his mom.

It's all full circle. Small world.

collecting them like infinity.

Yeah. [laughter]

All right. Uh so as a brief kind of overview of how the structure is going to go today, guys, uh we have some questions that we collected that you guys kind of entered. Um and then we'll open up the floor to some audience

questions. Um so just if you have any

questions. Um so just if you have any questions, note them down, be ready and later. Uh so Jason, do you want to kick

later. Uh so Jason, do you want to kick it off with the first question?

Yeah, sure. I guess like the first one, you've had a very unconventional career in tech. I think we'd all say obviously

in tech. I think we'd all say obviously with the beginnings as an immigrant and then through Facebook VC um and then now to founding curious just for the broader audience but like how did your early

life shape the way you think the way you think about risk ownership and building?

Okay. Well, first of all, thank you for saying it this way because um a lot of people sometimes like wonder like what's the motivation? I think like uh Kobe

the motivation? I think like uh Kobe Bryant actually had the best answer for this, which is the the winning and the losing is sort of like a byproduct of

what it feels like to be in the moment.

And in the moment, you have the opportunity to like practice a craft, get better at something, and mostly prove something to yourself. There are

many other ways in which people say that. Some people say that it's how you

that. Some people say that it's how you play an infinite game. Like if you play a finite game, you and I are competing for the scarce resource of a promotion

or for venture capital or for some external validation, some award, some title. Um, and that it's finite, right?

title. Um, and that it's finite, right?

And that's a proposition where if I'm winning, you're losing and vice versa.

The infinite game is more where you're just like it's like this constant trajectory. What I've accidentally

trajectory. What I've accidentally stumbled into is an infinite game. Where

I started was playing finite games where I was competing with everybody around me. So now if you ask me what my

me. So now if you ask me what my motivation is, it's like I think that I've done a couple of really interesting things. When I was an executive, I was

things. When I was an executive, I was mostly asked by somebody else, in that case Zuck, to take on a task and I had to be a trusted steward, but

fundamentally it was his show and I was a supporting actor.

And then I left and then I started an investment business which was a completely orthogonal thing.

And at Facebook I was really successful I think um and probably could have just stopped but when I move when I went to social capital

I had to kind of find a way of betting on other people and managing the capital of other people. So like and when you take money from other people, I don't know about how you guys would react to

it, but it was profoundly stressful for me because the people that gave me money like Mayo Clinic, the pension fund of

XYZ, I felt so bad about this idea that I could lose money for these folks. And

so I really took that job seriously. And

then I wanted to prove what it meant to take risk and be an investor. I did that job well. The great thing about 8090 is

job well. The great thing about 8090 is it's my first real shot. I've started

many projects. I've gotten them off the ground. I'll be an executive chairman.

ground. I'll be an executive chairman.

But it's my first real shot at doing the one thing that I've never done, which is I start something from scratch. The

white paper clean sheet design. This is

what I think we should do. And then

build it. And I need to see if I can help make it successful. And if I do that, I feel like I've I will have done everything professionally.

That is a challenge. How do you build a trillion dollar company? Check. How do

you build a multi-billion dollar investment firm? Check. Can you start

investment firm? Check. Can you start and scale and create something that's many billions? I don't know. But the

many billions? I don't know. But the

reason I can be motivated is that I'm framing at it as a challenge of who I am as a person. Long-winded answer, but it is such a key to your success. I wish I

knew that when I was younger. So many

times I blew myself up. So many times I self-sabotaged myself. So many times I

self-sabotaged myself. So many times I just I [ __ ] up the core essence of what the opportunity was. And the

problem with places like this is it can really trick you. I went to an equivalent place like this, not nearly as prestigious, but still quite prestigious in Canada called the University of Wateroo. You're surrounded

by the smartest kids. In that case, it's Canadian for the most part. And from,

you know, Hong Kong, that's what it was.

Canadians of Hong Kong. Um,

but they're little geniuses and uh when you're surrounded by geniuses um it's easy to get very outwardly focused and competitive with one another

and it is a huge distraction. So you were saying earlier

distraction. So you were saying earlier that you shifted from this finite reward system to this infinite reward system where instead of competing with other people for finite resources, you're now

on your own lane if I'm understanding correctly. What were some of the key

correctly. What were some of the key kind of realizations or practices? Like

how do you how do you get to that stage?

And I feel like a lot of us here grew up in very similar circumstances.

You have to run into the wall really hard at high speed.

Maybe there are people that are more evolved than I am. I think women are more evolved than maybe they can realize is men are stupid. So you're going to you have to you have to just really run

into it face first. And and the reason is because one of the most shocking things that you can do is not have an orc chart. We don't have an orc chart.

orc chart. We don't have an orc chart.

And so you know you drop these people into this soup.

It's already stressful.

That's one way. There's nobody. So, you

know, we have like nice, kind people that will that are also struggling and so there's camaraderie in the struggle and they're willing to help you, but they have just as little of a [ __ ]

clue as you do. But that's there's something beautiful in that because now you're again you have to just dismantle all of this traditional way of thinking.

Two, you underhire massively.

If you have 80 humans worth of work, having 40 is an incredibly interesting observational experiment. What happens?

observational experiment. What happens?

And now you combine it with this other pressure of there is no hierarchy to help you. What happens? And eventually

help you. What happens? And eventually

what you see is all of these traditional things melt away. Traditional answers to questions melt away. if this then that all melts away and all of a sudden people are forced to just think from

first principles that is a beautiful place to be especially early in your career that's where the magic happens that's where the magic happens you become a wholly different person so I tell people if you find one of these

places anthropic is such a place openai is such a place Facebook was such a place google was such a place maybe there's places pockets no but maybe there are pockets of that that are like

that now SpaceX is such a And you get into those places because it takes you into that flow state. There's

so much work, there's so much pressure, but it strips away the traditional thinking. You have no time for

thinking. You have no time for arrogance. Like if you talk to like Mark

arrogance. Like if you talk to like Mark Gianosa at SpaceX, one of the most humble human beings, and you'll be like, "How the [ __ ] is this guy?" And he's like, "What can we do better? What can

we do?" I was with somebody this weekend, and we were just commenting on Mark. It's always the question is, "What

Mark. It's always the question is, "What can I do better? What if SpaceX's not doing well enough? When you're a $2 trillion company and the CTO asks those kinds of questions, it's a living embodiment of like what it means to be

in that flow state. But could you imagine the stuff that he is juggling, the pressure that he must be under and so he has no time to be arrogant.

That's what you want. You want to find places like that and you want to run to those places. They're incredibly um

those places. They're incredibly um challenging in some other ways, but they're incredibly unique.

You just want to keep us on track. Um,

you mentioned 8090. This is obviously like a newer company of yours. What was

kind of the rationale? Like what was the the catalyst of why you started 8090?

What was something that you saw some inefficiency that made you realize that I'm not going to start a SAS app or something?

Look, the way the way that my mind works is I always think in terms of is there a framework and a set of rules that allow me to abstract what is happening and

give it a context. The most important context in AI is all the way back to how the internet initially developed. And

the most important context of that is this thing called the OSI reference model. And if you just Google that very

model. And if you just Google that very quickly on your phone, what you're going to see is this image of a seven layer cake. And at every layer, what is the

cake. And at every layer, what is the most important thing to acknowledge is at every layer of that stack, you have seen hundreds of billions to trillions of dollars of value created. So even

back then when those people were imagining what the layers of abstraction would be, they got it so right. That

thing was so valuable to me because I used to refer to that all the time when I would meet a new company and I was thinking is this an interesting company?

I would go back to this OSI model and I would say does this logically fit in any one of these things? Where does it fit?

What will it do? It wasn't the you know the ultimate question but it was a very important starting point. So I have a framework of that for AI and in that framework I spent the first

number of years purely in the silicon helped this thing called Grock get off the ground and initially how 8090 started was I thought that what

we needed to do was essentially a transpiler where you could take

essentially CUDA and allow it to get to tranium inferentia uh AMD, all of this stuff. That's how it initially started,

stuff. That's how it initially started, but it turned out to be wrong. I had

spent like a bunch of years like kind of working and helping the team at Grock.

Then I spent a handful of years completely pivoting and doing what I call the fulcrum assets of AI. What are

those? It's prismatic LFP and it's actuation. Why? Because if you look

actuation. Why? Because if you look inside of a digital AI stack, the most important thing that it lacks is power.

The most critical thing for power is both the power generation and the power storage. The power storage fundamentally

storage. The power storage fundamentally iron. Where's the open AI guy? I mean,

iron. Where's the open AI guy? I mean,

where's the anthropic guy? He was here.

Uh iron and then you'll see look it'll be prismatic LFP. So I went and I helped start a bunch of things there. If you

look at physical AI, the biggest Achilles heel that people will have is the actual actuation. You know, how do you get the sensors to work? How do you kind of like pass the numbum match test as they call it? How do you do that? The

only answer for that is rare earths. And

so I spend a bunch of time in that. Then

when I came back back to my reference model, I was looking around, started 8090, thought it was a transpilation problem, realized very quickly it was not. And then it kind of occurred. We

not. And then it kind of occurred. We

have done so much and I and I hate to offend anybody here when I say this. We

have done so much in AI. Okay, we have.

But we are training models to understand patterns that we have seen. We do not think from first

have seen. We do not think from first principles. it's token and then the

principles. it's token and then the prediction of the next token is prediction of the next token. And I

think that that's it's it's still an incredible leap that we've created or that you know OpenAI Anthropic XAI that they have created. But what is missing?

What's missing is being able to translate that in two dimensions.

Dimension number one is long horizon tasks are fundamentally still a joke. It

doesn't work. And I don't care what anybody says. Don't show me a stupid

anybody says. Don't show me a stupid eval. Don't tell me about some dumb

eval. Don't tell me about some dumb script you ran for, you know, 48 hours.

That's all [ __ ] Long horizon tasks are not well- handled. They just don't work. And second is complex problems

work. And second is complex problems also don't work. They are not well addressed and they're not well handled.

Why is this important?

If AI looks like any other technology, we're going to go through the rise, right? The initial hype cycle. We're

right? The initial hype cycle. We're

going to see a natural contraction because somehow somewhere something is going to fail. We are all going to see this and then that's what's called the trough of disillusionment. I think the

business the NBA folks will confirm if that's true and then you see the slow gradual uptake of the real final solution happened in the internet happened before it's

happened in many many many cases. The

problem with this is that we though are spending hundreds and hundreds of billions to trillions of dollars to try to figure

out how to cross this chasm. So what do we do? If we don't figure this out,

we do? If we don't figure this out, people will hit the hit this trough of disillusionment and say this was a joke.

I think we need to be able to take I AI into very complicated environments and make it work. So what's

my solution to that? I think at a very basic level, you have to have a symbolic space that guides the embedded space. So

maybe a little less kind of jargony.

What is the secret of a company? Like

what's the secret inside of Google?

What's the secret inside of Uber? What's

the secret inside of Tik Tok? All of

that. But so but what is that? What that

is is a symbolic representation. And

it's when you're in a meeting and somebody says, well, you know, here's how we should do the, you know, page rank this way. Here's how we should build GFS in big. It's like, it's all in the symbolic space of English. It's all

these like secrets that a company has that describes what they should do. So I

was thinking to myself like the most valuable thing we could do is build a massive honeypot for the symbolic space for English like for the requirements

for all the secrets for all of the like if you go when you guys go recognize inside a company at some point if your company does poorly they'll bring in McKenzie

and then and you'll spend tens of millions of dollars and what will you get at the end? You'll get a deck, a nice slide deck.

What do you what do you do with the deck? Most of the time, nothing.

deck? Most of the time, nothing.

And that's the business model. You

the executives feel better. You uh I mean I I don't you you'll buy some off-the-shelf piece of software. It will be sold to you as a

of software. It will be sold to you as a solution. You'll try to implement it. It

solution. You'll try to implement it. It

really won't work. Then you'll go to a consulting firm. They'll hire hundreds

consulting firm. They'll hire hundreds and hundreds of time and materials people that look like me.

We'll try to implement it. It won't

work. I'm And this croft builds up. So my idea is I want to build

builds up. So my idea is I want to build basically a control plane for AI. I want

to be able to really focus on the things that is the symbolic golden source, the true source, the English language understanding of what should be done, what is done, what works, what doesn't

work. PRDs, requirements, those are the

work. PRDs, requirements, those are the hard-fought secrets that allows companies to be successful. Code is a mechanisticly deterministic thing that either works or

doesn't work. So, you will get that to

doesn't work. So, you will get that to perfection well before you get requirements and understanding to perfection. So, back to this original

perfection. So, back to this original problem. When you put trillions of

problem. When you put trillions of dollars in the ground and someone finally says, "What is the ROI?"

You're going to have to point to the economy and say, "This these companies are 50% more productive."

Unless you completely rebuild how those companies operate, it will not pay that off and there will be blood in the streets. The stock market will just

streets. The stock market will just totally turn over on itself. So, we have a responsibility to do that. So we have this thing that we call a software factory. You get all these humans that

factory. You get all these humans that can collaborate with a bunch of very powerful models and you can take all these requirements and you can get them into a very amazing sense of what the scope of what you want to do is and then

we bind it to an engineering plan. You

bind that to you know work orders and now all of a sudden it thinks can work in the forward pass, the reverse pass.

We can take tens of billions of lines of or tens of millions of lines of code bases, dump them in and all of a sudden it just propagates backwards and now all of a sudden you have English language

understanding. This is exactly what this

understanding. This is exactly what this does. And you and these companies their

does. And you and these companies their eyes are like we didn't even know. We

have this one customer who's got such an old legacy code base. They have to go and get retirees to I swear to God, dude, this is a hundred billion dollar a year company

out of retirement to come into the office to explain what this cobalt code does.

And none of you geniuses were able to figure that out. And nobody before you was able to figure it out. And it turns out that these kinds of problems are the

thing that's stopping ROI on trillions of dollars of investment. It's all the hidden context I bet.

Hidden context, tribal knowledge, information, information. So, Peter Teal calls it

information. So, Peter Teal calls it like, you know, like like all great companies are hidden around secrets.

And I think an interesting thought challenge is what happens if all these secrets get documented? What happens if there's like So, you guys live in a world where like, you know, GitHub is

like the ground truth. Well, instead of GitHub, what if there was an actual set of wiki documents in plain language language that the CEO could actually read that were the actual secrets? You

know, what if there was a document that Uber had that actually described search pricing in English, not in code? And

when you manipulate the English language understanding of the rule, the actual search pricing algorithm then changes downstream to hundreds of millions of people. What if that were possible? You

people. What if that were possible? You

start to unlock a level of optimization and improvement because you don't have to be technical anymore. You just have to have good judgment and be smart.

There's way more people that have good judgment and are smart than are technical. So this is the whole point.

technical. So this is the whole point.

Like when technology works, you're opening the aperture. Like when I was when I was in college, I had to learn how to code in C and C++ and you're manipulating memory at this ultra low

level. And then at some point, you know,

level. And then at some point, you know, PHP shows up and then Python shows up and JavaScript shows up and you start to abstract up. And what happened? More of

abstract up. And what happened? More of

you knew how to do it. Many much many more of you like Python raise your hand.

Okay, most of the room. C++ raise your hand. Oh, [ __ ] Come on. That stop

hand. Oh, [ __ ] Come on. That stop

that. Come on. Stop. I'm trying to make a point here, guys.

Unfortunately, stay with me here.

Unfortunately, the Stanford class is C++.

Okay. Well, that's smart because you need to understand memory and memory structures. And so, I I get that. But

structures. And so, I I get that. But

for most people, the point I'm trying to make, you might you might get more people with C or may maybe assembly.

Okay. How many of you know assembly?

There's there's a few classes.

Oh my god. Jesus Christ.

There's a few classes.

But but the point I'm trying to make is as the languages get abstracted, the n the number of people that can participate in that ecosystem goes up.

And so the ultimate hack for AI is when things are in English. That is the unifying language for 8 billion human beings.

It is not C. It's not Python. It's none

of these languages here. It's this

language here and the space of innovation are not algorithms anymore.

It's judgment. It's really good ideas and that many many many more people have. I think touching on this there is

have. I think touching on this there is an interesting analogy um that said like when the first compiler was invented there was a lot of kind of uncertainty

around with this new tool right like you would have this code you'd write it and then pass it through the compiler but you weren't sure that like the assembly that it generated was actually correct and I think we're in like this kind of

similar stage with AI right like we have this grand vision we want to reason on the symbolic level but in order to do that we have to pass it through kind of this software factory as you're describing it and there are all these

kind of like lossy conversions that can happen. Um, and so you mentioned earlier

happen. Um, and so you mentioned earlier as well like these agents like there's so much hype around like oh like open claw was running for 48 hours and solved my life or something like crazy like

that like it's clearly not true. is

clearly an exaggeration and one of the big reasons of this is as you were saying there's a lot of secrets right there's landmines you have to bring in the retired developer from 15 years ago to explain why something happened and

explain why like chatbt deleted this line and now the whole like codebase is broken so in this path towards this ultimate goal right we want to reason in English what are some of the biggest

challenges that you guys are facing that you think more people need to be thinking about and so one is a technical challenge and the other one is one of business. The tech

the technical challenge is we're struggling with the same things that everybody else struggles with. How what

is the interplay and the interface between where the agent starts and ends and where the humans get involved? How

do you actually create the right eval and the right guard rails for the agents to stay on task? How do you actually create a knowledge graph that these these agents can traverse so that they

can stay at the level of embeddings and actually be able to act over reasonably long horizon tasks. So those are a set of technical challenges and that creates

really interesting questions of business strategy. So you know we had a very

strategy. So you know we had a very important conversation today which is like what is the value of a control plane and won't we just be seeding ourselves to cursor as an example which

is a phenomenal product and what you're having to navigate there is just you know in in my language in my understanding I would rather build MS

DOS windows not necessarily you know Adobe Photoshop and in in this context I think that for AI we don't have that control plane that's what I would like

to build I want to have something that basically says the ground source golden truth for all of these agents downstream will always be this hardware independent, database independent,

language independent symbolic representation of what you want to do.

So go and convince business people of that. They get that. So we can sell the

that. They get that. So we can sell the [ __ ] out of that. But then translating that into a set of agents that can act on target and on task is very very difficult. It's a very complicated

difficult. It's a very complicated problem. The business challenge is then

problem. The business challenge is then to convince them why those innovations should stay with us. What do I mean by that? So now I go back to how do you

that? So now I go back to how do you build a trillion dollar company? What

was the term network effect? What is a network effect that has never been built before? There's never been a network

before? There's never been a network effect built around these kinds of secrets around code. Meaning, if I was a

hospital and I'm trying to diagnose a cancer patient or I'm an airplane company designing a new wing, immediately where your brain goes to is those are two totally different sets of

challenges, two totally different sets of problems. And what I would say is wrong. I challenge you to say the

wrong. I challenge you to say the following. That may be a different

following. That may be a different ontology, but it's probably the same problem. And at the level of code and at

problem. And at the level of code and at the level of assembly, it all looks the same. You agree? Yeah.

same. You agree? Yeah.

So what if it were actually possible for this company to be able to leverage that code because it's just recompiled with the different ontology so that you understand it. Now all of a sudden you

understand it. Now all of a sudden you have a network effect. Now the N plus first company gets to leverage all the secrets of the N companies before it. So

when Elon talks about abundance that's how I translate it. That's how you create an al like just a logarithmic expansion of of abundance. You can do anything quickly cheaply. So we are

navigating this very complicated business path to convince these businesses to let us do that with them to convince them why this is important.

Now that's a unique task but if we accomplish it we are giving all of these technical people that work at 8090 a chance to do something very profound. If

we don't have that it's a wonderful business. We'll do the same thing as

business. We'll do the same thing as everybody else does. We'll raise money.

It'll be a unicorn. Then it'll be a decacorn. Blah blah blah whatever. But

decacorn. Blah blah blah whatever. But

that's not what I care about. I want to prove this other thing. There's a

network effect in the code that businesses can have a shared cooperative approach to this thing that there is no reason why Boeing needs to be afraid of Memorial Sloanketering. In fact, it's

Memorial Sloanketering. In fact, it's the opposite. Boeing and Memorial

the opposite. Boeing and Memorial Stoneketering can sort of cooperate and both can thrive. Their costs go down, their value goes up, their downstream impact on their customers and patients

go up. That is a positive some view of

go up. That is a positive some view of AI. There has to be some sort of initial

AI. There has to be some sort of initial apprehension though with these companies sharing my secrets, right?

Huge. It's enormous.

How do you overcome that?

That's about trust and reputation. The

thing that we do is we methodically first of all, so this is more about business than technology.

Sure.

In business, it's very important to sort of fish where the fish are, which is to say that in every technology adoption curve, you'll always go through the same transition. You have the early adopters,

transition. You have the early adopters, then you have this sort of like mass middle and then you have lagards. And

historically, there was a very porative or negative way in which you viewed those three classes, particularly the last two classes. And I've challenged these assumptions to say something

different, which is everybody has a different risk posture. It's not I I I think like it's very unfair to call folks that wait till the end lagards. If

you're a regulated pharmaceutical company, I was reminded of this this weekend, and you do something wrong, you can go to jail. So, it's not like they're like, I want to poo poo this

thing. It's like they have a

thing. It's like they have a responsibility to all of you that when you land a drug inside of your body that it doesn't, you know, grow an arm out your forehead. Do you know what I mean?

your forehead. Do you know what I mean?

Yeah.

So, that's not being a lagard. That's

being responsible for their customers.

One, two, I said this earlier. Right now

we have a very complicated moment in AI which is that there is dumerism on one side and full throttle capitalism for me on the other. Those are the two choices.

Like think of the compact of the internet up until AI. The compact of the internet was we would make products where you would participate and contribute and the quit proquo was you

would get a value that you would assess as being greater than what you are giving me. You use Facebook and

giving me. You use Facebook and Instagram. You take photos. Well, thank

Instagram. You take photos. Well, thank

you very much. It allows us to create a network effect. We're able to build a

network effect. We're able to build a trillion dollar company. You still find value. Now, think of what AI says. AI

value. Now, think of what AI says. AI

says, I'm going to learn. I'm going to tokenize that knowledge and then I'm going to sell subscriptions. Thank you

very much. See you later. That is not a positive sum view. Nor is the one that says I'm going to go and then use that to replace you and fire you.

that is deeply and fundamentally irresponsible.

The positive sum view is we're going to work together. We're going to take our

work together. We're going to take our time and methodically navigate this complexity. When you need help with

complexity. When you need help with regulators or the government, we will go beside you and help you figure that out together. We're going to get to the

together. We're going to get to the finish line in a constructive positive sum view. We're going to actually show

sum view. We're going to actually show how you're hiring more people, you're paying them more. That's how we do it.

So, it's slow. It's methodical, but it's really working. Um, because it turns out

really working. Um, because it turns out humans again are amazing, resilient, positive, not zero sum. You know, many people just want to play an infinite

game. They want to take care of their

game. They want to take care of their family. They don't want to [ __ ] get

family. They don't want to [ __ ] get screwed over. And so, why are we here?

screwed over. And so, why are we here?

Yeah.

It's just terrible leadership. Terrible

[ __ ] leadership. So I go talk to the Fortune 1000 and I tell them my view of the world and they can vote with their feet. They vote with their dollars and

feet. They vote with their dollars and so so far so good.

I don't have a billion dollar company but I would sign up.

I want to ask you your opinion on all the hype around like these local AI agents. You have openclaw, open Jarvis,

agents. You have openclaw, open Jarvis, etc. um as the models keep on improving with like quantization, efficient architectures, better ondevice GPUs, we're going to see a lot of the AI

workload shifting to like local compute and local devices, right? Where do you kind of see this like split between like local and cloud go? Does this kind of affect like business creation or like

the enterprise side of the world? Um do

you think there is actually any merit in these or is just like all hype and you don't really care?

I think it's fantastic. I think the most important thing that is going to happen in AI which is going to be ginormously disruptive is uh fundamentally open source models. We don't first of all we

source models. We don't first of all we don't have open source models. Okay. We

have closed source in America. We have

open weight in China. That's what we have. We do not have open source models.

have. We do not have open source models.

But number one is we need an ensemble of open source models. And then number two is we need a fundamentally completely unregulated totally distributed form of

compute initially training and then inference. The combination of those two

inference. The combination of those two things is profound. How open claw and all of these local agents work in those realms I don't know well enough to say

but man that is a huge huge trend.

Subnet 3 of Bit Tensor, um, Folding at Home uh Plurales um, there's a couple of others. I I have I have no stake in any of these, but I

would recommend you to learn about all of them. It is incredibly important, I

of them. It is incredibly important, I think, what's happening. That is a it's a Cambrian explosion that's just going to go crazy because then you are totally delevered

from government oversight and government infrastructure. There is no kill switch.

infrastructure. There is no kill switch.

And why that's important is that if these things do become able to symbolically reason, you cannot have that gate kept by five or six entities.

Because what will happen if five or six entities control it is you get immediately this distribution of you have the people that are aligned to the model maker, you know, sort of like think of it as like a planet with a

handful of moons rotating around it and then you have everybody else which is forced to become effectively a vassal state of that model maker. And I think that's a very complicated outcome and and not a good one. But there are these

projects out there. I encourage you to learn about them. I am deeply fascinated. I spend a lot of time

fascinated. I spend a lot of time tinkering and learning. It's really

cool. So I I would say like instead of openclaw, open ages, I would think about more, you know, how do you like how do how does local training work when you're participating as part of a training

pool? How will inference work? I that's

pool? How will inference work? I that's

hugely technically complicated scale.

8B, 10B, 50B models are happening right now kind of don't matter. But when you get to 100B, you're like you're in the game like you can really compete commercially.

Okay. Uh Shamad, Mr. Shamat Pala Palia, is it right?

No, but go to your question, bro. Go to

your question.

I've been a big fan of you and you're one of my biggest inspirations. It's it

feels like seeing a mythical character out in real life. Before before before I ask you the question, I'll tell you a bit about me.

I'm okay. I'm

He knows kung fu, bro.

I'm Jas.

Be very careful.

And I was born in Hwood, California, and I grew up in South India in poverty.

Education is very important to me. And I

run a tech startup right now called Educat.ai under JSA LLC. Now I go to George Washington University with $250,000 scholarship uh university aly

award and I'm very grateful for that and that changed my life. Education changed

my life and I'm also a Buddhist and a feminist. I'm very empathetic.

feminist. I'm very empathetic.

If you cry I'll cry.

Almost there. Almost there.

I'm almost there.

The girls are behind you. I mean

what do what do you want from me?

Okay. I'm very empathetic. If you cry, I'll cry. If you're happy, I'll be 10

I'll cry. If you're happy, I'll be 10 times happy for you. Now, the question, [laughter] question. I'm almost there. Question.

question. I'm almost there. Question.

I'm an information system student. Uh

I'm thought to sit right between business logic and technical implementation. If AI is moving toward

implementation. If AI is moving toward an AI first English first control plane, should we focus less on mastering specific tools like PowerBI or Python and more on more on the art of

requirements engineering and first principles thinking?

Yes.

Thank you.

Yes. One over there.

Yes.

Let's go right here.

Yeah. So when I think about like early Facebook product and growth leaders like Stan or Naomi, I think one way that maybe you differ from them too is just

how you've been able to connect with people and sell. And I think about like the story of how you taught um like Cheryl Sandberg's kids to play poker after Dave passed away. And I just feel

like you have this like incredible like empathy and just ability to connect. So

I was wondering like what experiences in your journey helped you develop that skill?

Um well look I think it's kind of a little it's it's relatively well known but I felt very deeply

uh less than growing up. Um because I would always think to myself like my dad was a wonderful like to now I've I've totally dispelled all this anger but my dad used to I mean truly beat the [ __ ]

out of me and uh you know I said this on Lex once like it's a crazy thing and I remembered it recently. He would allow me to pick the thing that he would beat me with was the belt or stick. And I

remember like learning like how to like judge the tensile strength of various sticks because I was like, well, this one I can get hit by 10 times and it'll break and the other one will be, you

know what I mean? But when when that is done, you're like, why it this must mean that I am just so worthless, right?

Because why would you do that to somebody else, especially somebody that you gave birth to essentially? So that

in the back of my mind was there for a very long time.

And so I would find ways where I could like not not ex like purposely but like I was like it would make me feel better. And so I did it in

a weird way for myself. And then over time I've just learned to like just let it all go. And it's like actually it just happened. It was a thing. And so

just happened. It was a thing. And so

what's left over though is this desire.

It's fun to be around people to make them laugh, you know, to tell the truth, just to be who you are, and it's all just normal.

Um, and it's been it's just been a process of learning that I am just like everybody else, that I'm not worthless.

That's what got me there. But it started from a very negative place, which I don't wish upon you, but I think there are other ways to get there. I just

don't know what they are because I didn't live that journey.

Hi. Um I want to ask a couple questions that should be related. One is uh now that you see what's happening with AI uh and how it's abstracting away processes

and businesses, what should universities be teaching, especially in computer science? And then secondly, what kind of

science? And then secondly, what kind of people are you looking for at 8090?

Uh I want people that are extremely extremely interested in navigating the journey to their own resilience. That's

all it is. Um, I want them to hear the story and I want them to have the courage to either self- select out or

double down and uh, that's what I want in terms of technology. I do think that we are still at a place where technological capability is still paramount and I think for probably the

next 10 or 20 years, it's going to be increasingly critical that people can really refine their understanding. meaning I actually

their understanding. meaning I actually think you're going to have to get more low level than stay at the high level.

Um but in success there there will be this 50 or 100 million artisans that will be building something for billions of people and for

them they'll be able to live in the symbolic language of English and be able to be productive human beings. Um, so

yeah, there'll be both, but you know, I really appreciate the initial statement that you kind of made coming into the podcast, uh, coming into the talk today. Uh, you talked about

talk today. Uh, you talked about something that like kind of parallels like death of the ego and like really like living in the moment. I think it's like emotional intelligence is something that I've also learned being is is one

of the most important things in building a business. How do you continue to

a business. How do you continue to maintain emotional intelligence over time and like maintain that death of the ego and the lack of need for

external validation because I feel like you you get there and it's great to get there but then keeping that up especially like being someone who like seeks validation or something like that historically speaking

superpower which is I have an incredible partner my wife um she is very orthogonal to me she's not motivated ated this way. My wife runs a pharma company. She makes drugs for rare

company. She makes drugs for rare diseases. You know, she's getting her

diseases. You know, she's getting her head bashed in every day. Everything

fails. Everything fails all the time.

She's always on the knife's edge. And

so, when she has a blockbuster drug, she only has it for six or seven years. So,

all that money has to go back into an R&D program. So, she lives this very

R&D program. So, she lives this very practical day-to-day life. And she's

incredibly humble. She's not motivated by ego. So when I am, she basically

by ego. So when I am, she basically slaps me upside the head. She's also

Italian, so she's not afraid to throw things. And I'm like, "Oh [ __ ] I have

things. And I'm like, "Oh [ __ ] I have friends who see me change. They will

tell me. I have one in particular that's really good." So I have my wife, I have

really good." So I have my wife, I have one friend, I have my poker game. This

may sound crazy, but poker is an incredible insight into your mind and your current mental state. And I am a very good poker player to be totally honest with you. But

like at the highest of highest levels, like the people that come into my game are murderers row of the best professionals. But when I lose four or

professionals. But when I lose four or five weeks in a row, it's because of my ego. It's because something has

ego. It's because something has happened. And then I go and I dissect

happened. And then I go and I dissect and then if all else fails, I fly commercial and I kind of [laughter] Yeah, it's a really amazing

conversation. Thank you very much. And

conversation. Thank you very much. And

uh uh let me do a little bit introduction. So we are the air

introduction. So we are the air foundational model company behind the model GM. So we are the open source

model GM. So we are the open source model. And thank you for clarification.

model. And thank you for clarification.

It's only open weights not open source.

But why we want to do open weight?

Because we want to booming the whole ecosystem and to get everyone benefits.

So my question is uh we actually want to achieve the sort of AGI. So I want to know your definition of AGI and do you

think we can achieve the final intelligence through the language model not the other models like physical models or the models or the others?

I I don't I don't think AGI is practical with what we know today. I think we're just scratching the surface. There will

be some forms of super intelligence and I think that those will be some narrow verticals of capabilities but for for for most part I don't think we're dealing with an AGI on the realm of anything. I think when you strip away

anything. I think when you strip away the um the requirements for fundraising it's mostly bluster. Uh hi Shamat thank you so much for the talk uh around

network effects right you said it about enterprises but what if normal people could also contribute to it and uh so we were building in a uh uh especially

around uh the knowledgebased work that Karapati recently announced and so we are building in a very similar space and what if I told you I could give you access today to something that you are

just talking about 30 minutes ago [laughter] Great. I just spent the last 25 years

Great. I just spent the last 25 years trying to get there. But great. Um, one

thing that I think is really interesting is this idea that, you know, like you know how we have open source projects and people contribute to repos. It would

be really interesting to think about that in the context of business and symbolic language and understanding hospital administrators that can work on the way to build the best hospital. And

what it is is a set of foundational documents that that can guide all the AI and all the agents. I think that idea is so wonderful. It's deeply collaborative.

so wonderful. It's deeply collaborative.

It's genuinely, you know, again, positive sum. I I just think it's a

positive sum. I I just think it's a beautiful, beautiful idea and I hope and I hope there's a moment where we can play a small part in that.

Chimoth, you talked a lot about first principles as well as address the problem of creating guard rails and evals for agents to stay on track for complex tasks. Um, how much of this do

complex tasks. Um, how much of this do you think is an issue of mechanistic interpretability and being able to study these models as physical systems?

[laughter] I don't totally know the question, but try it again. Maybe are you asking like why do agents go off track or I guess yeah and how much of it is

related to being able to study these systems as physical systems?

I I I think like as far as I can tell I think the reason that it goes off task is is that it's really what we have is like a very fancy autocomplete and it's it's it's not trained on first principles thinking because it's not

possible. It just kind of says, "Well,

possible. It just kind of says, "Well, I've seen this pattern so many times, so it must be this thing on this distribution." And so sometimes you'll

distribution." And so sometimes you'll get the right token, sometimes I'll get the wrong token, I'll go boo. You know,

there'll be some reinforcement learning and then then both of us will get the right next token. That's what it is right now. It's a it's a deeply powerful

right now. It's a it's a deeply powerful but also deeply primitive software system. It doesn't mean we

software system. It doesn't mean we should stop. In fact, it means we should

should stop. In fact, it means we should double, triple, quadruple down because there's so much to do. It just means we're super early. That's why I don't like like the overselling of what this is and what this moment is cuz we're

almost like pulling forward the reality of this moment, you know. Hi Chabbat. Uh

great talk. Thank you. I'm wondering

given that uh mammals have a fraction of compute today frontiers model have models have why do you think it is time to look at that symbolic stuff instead of looking into more effective learning

algorithms? Okay, so this is a this is a

algorithms? Okay, so this is a this is a great question because then you're leveraging this extremely power efficient thing called the human brain.

And you're right, maybe what we find out is that the way that we translate the power of the human brain through symbolic language into into this embedding space doesn't actually make sense. But that would mean that there's

sense. But that would mean that there's a completely different compute modality that we shouldn't be inventing. But my

point is this is perfected over billions of years. It is power efficient. It's

of years. It is power efficient. It's

extremely information dense and it just works. And because there's eight billion

works. And because there's eight billion of these things, the compute power we have collectively as a species, I think is unrivaled. You can't measure what we

is unrivaled. You can't measure what we are able to do in flops. I just think it's an extremely sad approximation of how the human brain actually works. So

instead of saying that we need this to change, I think it's in fact we should map everything from this forward. Uh

Chamath you talked a lot about flow state during the during the talk and I think it's very important but uh you shared some ingredients like flat hierarchy and more work than people. Do

you have any other secrets or ingredients that you can share that helps get a group in flow state?

Um meaning like what else do we do at 8090 for example or something that you've seen work?

Well I mean our entire company is extremely young.

Um, I'm not a huge fan of the the recency bias that comes with experience.

I think new things should are largely proven and pioneered by new new ways of thinking which essentially comes with no priors and no bias. Um, so we have a

very young company. We have a very inexperienced in the classical sense.

Um, they're very experienced in the sense that they're good at taking risk and taking shots on goal. Um, they don't question the lack of hierarchy. They

don't question the lack of resources.

Um, I don't know, man. It's a very, it's a very very very special place. Um, it's a very special. It's small, you know, it's

very special. It's small, you know, it's 40 people or so, but it's very special.

Um, it's chaos. It's truly I've never done anything like this. Meaning like

just the structural underpinnings, the organization. I have like a cloud chat

organization. I have like a cloud chat that I that I have an ongoing dialogue with that's called 8090 org design and I and I just keep feeding it stuff and just kind of trying to figure out like

why is this not working? Why did this work? It's just chaos and I come in with

work? It's just chaos and I come in with some of the stupidest ideas and sometimes these guys are like this is a bridge too far. We we but I think you're but I think when you look back like said

differently if you if you go inside of a normal company you look at an org chart the reason why that or chart is like that is I suspect somewhere in the 80s and 90s somebody sold them some shitty piece of enterprise software and

convinced them that that title needs to exist. Oh you're a CXO. Well, here's

exist. Oh you're a CXO. Well, here's

this piece of software perfectly suited for the CXO. Oh, you're the C YO. Here's

a perfect piece of software for the C YO. And you do this on and on down the

YO. And you do this on and on down the line and you have a multi-t trillion dollar complex. But then it turns out

dollar complex. But then it turns out that it doesn't work. So then you take that multi- trillion dollar complex, multiply it by four and you create a services complex to make the multi-t trillion dollar complex actually try to

do something. But then you realize that

do something. But then you realize that most companies have a 15 to 20% profit margin. And you're like wait a minute.

margin. And you're like wait a minute.

But then you look at the few companies that have built everything from scratch like Facebook and Google and they have almost triple. Has anybody connected

almost triple. Has anybody connected these dots at any point and said, "Wait a minute. Is Facebook and Google

a minute. Is Facebook and Google literally the employees like three times smarter than the employees that I'm just making it up. Merca

done. This is 20. Is it industry?" I

don't know. Because when you look at all of the ways you can strip a company down to its studs, you end up realizing that wait a minute, there's all this inefficiency that sometimes gets trapped by poor

software. And so we should just rip that

software. And so we should just rip that all out. should rewrite it from scratch.

all out. should rewrite it from scratch.

Hey Shamad on your point on the knowledge sharing uh for scaling the the knowledge and AI between companies. So

in 2017 element AI from Canada raised like 150 million to like tackle this issue between making a promise that we can take the data from each companies like banks for example and use that

knowledge to help other banks grow. But

that idea quickly uh proved like it's not going to scale.

So what do you think is different right now and especially with 8090 that companies are okay with sharing their secrets?

Look, I knew it it didn't fail the minute you told me you led with the funding round.

Banks don't care how much money you raise. Banks care that what you said is

raise. Banks care that what you said is actually true and it works. They're not

stupid. They run banks. So, the reason it failed was because whoever was able to convince an investor to give them $150 is not the same person that can sit with the CEO of a half a trillion dollar bank and convince them why it's the

right thing to do. Okay. Well, well, one last question. Like a hard one. No,

last question. Like a hard one. No,

hard. It must be who don't raise your hand unless it's hard, guys. No. All of

you guys think it's hard.

You're all graded on a curve, so you're not used to hard. Shamath, you need to pick on the person.

You need to pick on the hard person.

Why didn't you Citizen judgment.

Yeah, you. It's your judgment here.

It looks like they put your [ __ ] hand down. You You're

occupied enough.

Is it a hard question? Is it a hard question?

Okay. Okay.

How do you what?

Kick yourself up.

Too complicated, bro. Because Okay, bro. There is no

bro. Because Okay, bro. There is no failure. That is what other people think

failure. That is what other people think of you to keep you down.

There is no failure. If you were left on a on an island and it happened, you would just brush it off and move on.

It's everybody else that you think is judging you. But then here's the secret.

judging you. But then here's the secret.

They don't give a [ __ ] about you. They

are living their own lives. So it's your perception of what they think of you.

There is no failure. There is do and learn. Do and learn. Do and learn.

learn. Do and learn. Do and learn.

[applause] Okay, last question. Go. Last question.

We got to go.

Uh, thank you, Shamat. My name is Carl.

I'm from Dartmouth College. Uh, I wanted to ask that you have a very special engineering uh screening pipeline where it's very rigorous and in the end you're trying to not, you know, convince them

to join. So far this is not going to a

to join. So far this is not going to a hard question but but and you want to uh take out people who can survive not four minutes swimming

no not survive I want people who want to see if they can survive it's totally different and that leads to my question how do you actually measure it for example for an

engineers is it the lines of code is it the number of parallel agents running what is it how do you measure it you see it you see it in their character It's not the person that codes the

fastest. It's not the person that codes

fastest. It's not the person that codes the best. It's this ongoing ability to

the best. It's this ongoing ability to process your moment in time, internalize the pressure, learn from it, thrive in

it, see that you are growing and stretching as a human being, you help the colleagues around you, and you go to the next challenge. The best example of this, I I would just leave you with a

parting thought. So if you want to learn

parting thought. So if you want to learn business, go look at the OSI stack. Take

a few seconds to try to build the AI stack for yourself. That's lesson one.

Lesson two is go and Google Raptor 1, 2, and 3. And when you look at that,

and 3. And when you look at that, there's a beautiful picture where Elon took a picture of Raptor 1 beside Raptor 2 beside Raptor 3. And what the first

one was is a gnarly mess.

And Raptor 2 was good. And it was an incredible engineering feat that made it much more refined but still very beastly and rough. And then you look at Raptor

and rough. And then you look at Raptor 3, it is so beautiful. And when I look at that, what I see is not a technical

innovation. I see a process of just

innovation. I see a process of just constantly figuring out try do learn do learn do learn and the from getting from

there to here. That's what I want. the

person that is like, I'm just going to keep chiseling and shaping and trying and I'm going to try things and oh, I [ __ ] this up and oh, I was really bad yesterday and but you did the best of what you could and you don't beat

yourself up and you're like all right and then you have people around you that are not trying to pin it on you, right?

And if you have a a lot fewer people, there's a lot less politics because there's there's you're not fighting over scarce resources. There's too much [ __ ]

scarce resources. There's too much [ __ ] to do. And so you go from Raptor 1 to

to do. And so you go from Raptor 1 to Raptor 2 almost accidentally and you take a step back. You're like, "Wow, we did that. We're in the middle of this

did that. We're in the middle of this thing. I can't really say for an entity

thing. I can't really say for an entity that I can't really talk about."

And I guarantee you that if we get this right, you or your parents will immediately and directly see a benefit from it. And that is something that the

from it. And that is something that the guys that did it, I I shouldn't even say this. It's two [ __ ] people

this. It's two [ __ ] people and it is it's an they sit beside will actually it's an it's an incred it's an incredible thing. No,

incredible thing. No, there's a correlation.

Don't say what it is. Yeah, it's an incredible incredible incredible thing.

So my point is find a place where you can go from raptor 1 to raptor 2 to raptor 3 of yourself.

And it is not about the superficial things of PRs. It's not that's not what it is. It's hard to describe what it is.

it is. It's hard to describe what it is.

It's just about you. Look, we have huge attrition, don't get me wrong. We

[laughter] a lot of people don't fit. Uh because

the problem is we have very orthogonal ways of hiring. Like you go through the screens, you go through the coding reviews, but then when you're in it, it's very clear and it's it's a it's a

mindset. It's a desire about figuring

mindset. It's a desire about figuring out for yourself what you're all about.

No, you're not listening to me.

No, I This is my point. If you want a binary deterministic answer, I can't give that to you. Life will never give that to you.

There is no outcome.

There's only do and sorry.

What?

Do not let this guy back into Stanford ever. Guys, thank you very much.

ever. Guys, thank you very much.

[applause] [cheering]