The Limitless Podcast

Ejaaz:
Okay, we have an incredibly special episode for you guys today.

Ejaaz:
You're about to hear from one of the most well-researched strategic investors

Ejaaz:
in frontier technologies.

Ejaaz:
If you don't believe me, these guys raised $1 million, just one,

Ejaaz:
back in early 2019 for their first fund and turned that into over 10 figures

Ejaaz:
in value in a year and a half. You want to know what's even crazier?

Ejaaz:
The original $1 million that they raised was on credit card debt and loans.

Ejaaz:
So we know that these guys go all in when they have conviction in something.

Ejaaz:
What I'm really excited about is over the last two years, they've been dialing

Ejaaz:
in to all the stuff going on in AI, and I'm really excited to get into their

Ejaaz:
heads about what trends they think are exciting and what they're excited about investing in.

Ejaaz:
Anil, Jan, Jose, it's great to have you guys on. How are you guys doing today?

Anil:
Yeah, thanks for having us. Amazing intro.

Jose:
Yeah, I appreciate that.

Ejaaz:
Great to be here. Let's go. Okay, so from a lot of our listeners that tune into

Ejaaz:
the show, they've probably never heard of you guys.

Ejaaz:
And so maybe you guys can spend a few minutes painting a picture of who you

Ejaaz:
are. And now maybe you can kick us off.

Anil:
Yeah, for sure. So yeah, we're all co-founders of Delphi. The Delphi that everyone

Anil:
knows of today has like three main companies, right?

Anil:
Delphi Research, Delphi Ventures, Delphi Labs.

Anil:
Jan heads up the ventures, he's managing partner there. Jose heads up labs,

Anil:
and I mostly focus on research and ventures.

Anil:
Essentially, what Delphi does is we're very research focused,

Anil:
right? We started back in 2018 with research embedded in our DNA.

Anil:
Jan and I and a couple of our other co-founders all met at our first job out

Anil:
of college at Bloomberg.

Anil:
We did a lot of like TradFi there with research and did leverage finance at Deutsche Bank.

Anil:
Essentially, it fell down the crypto rabbit hole when we realized that maybe

Anil:
the future that TradFi promised wasn't all that great.

Anil:
And yeah, just kind of like fell in love with kind of the promise that crypto provided.

Anil:
We put our jobs in 2018, started Delphi as a research firm. And like,

Anil:
you know, first few years, basically, no one really paid us for research because

Anil:
like there weren't that many fundamental investors in the space.

Anil:
You know, shout out like Multicorin and Hash. They were kind of like our first

Anil:
two, you know, real paying customers.

Anil:
But where we really bootshopped was actually helping design and kind of consulting

Anil:
a lot of these protocols.

Anil:
A lot of what you saw in DeFi and work with like protocols like Aave,

Anil:
Lido very early on. And then especially with gaming as well,

Anil:
with, you know, projects like Axie and Yield Guild.

Anil:
Over the years, Delphi has kind of morphed into this like, you know,

Anil:
three kind of like pronged layer where we build, we research and we invest.

Anil:
Right. And I think these three different perspectives work really well together

Anil:
because it lets us have, you know, as many like different hands on the elephant as possible.

Anil:
So we can really feel what crypto is and where it's going and,

Anil:
you know, have a really good pulse of it.

Ejaaz:
It sounds like Delphi was extremely focused on the Web3 crypto world,

Ejaaz:
right? And that's been your bread and butter since you guys have been incepted.

Ejaaz:
And then over the last two years, you've been like dialing in very much on AI. I'm curious, like,

Ejaaz:
like, what kind of like parallels run between the two technologies?

Ejaaz:
Like, do you just see the AI stuff happening and thought like,

Ejaaz:
huh, I just want to kind of like peek over the fence?

Ejaaz:
And then did you get kind of like more involved in that? Like,

Ejaaz:
what made you more interested?

Anil:
Yeah, definitely. I'd say that like, you know, when we first fell down the crypto

Anil:
rabbit hole, it was almost it wasn't even just obvious to us.

Anil:
It was just like, you know, what else could we work on or spend our time doing

Anil:
other than this, right? It felt like nothing else mattered.

Anil:
And I think over the past two years, you know, shout out Tom,

Anil:
one of our other venture partners and co-founders.

Anil:
He really was early to the AI trend and everything like that.

Anil:
And a lot of people within the hive mind that dealt by got nerds signed by AI

Anil:
and it felt we had that same feeling

Anil:
where it was like, how can we not be infatuated and obsessed with this?

Anil:
And yeah, I think there are a bunch of parallels. I mean, obviously the speed

Anil:
of innovation, just like when we entered crypto, just like today,

Anil:
even with a team of almost 100, right?

Anil:
We have around 88 people across three companies at Delphi.

Anil:
There's just no way we can kind of, you know, keep up with every single thing happening in crypto.

Anil:
I think that's the same thing that we see in AI and why we think,

Anil:
you know, we'll get into this later, why we think it's really important to have

Anil:
a team focused on it and kind of like separate the signal from the noise.

Yan:
Yeah. No, in terms of parallels, I think just when you see something that,

Yan:
looks so glaringly obvious in terms of, you know, its growth and application,

Yan:
but at the same time, there isn't really any widespread adoption of it or it's

Yan:
still, you know, orders of magnitude away from what it'll eventually be.

Yan:
Your eyes tend to light up because you start to think about all the possibilities

Yan:
on the growth building and investing side.

Yan:
And so I think what we saw that with crypto, you tend to see here with AI.

Yan:
And then there definitely overlaps the two in terms of implementation and where

Yan:
they can be synergistic.

Yan:
But I think, you know, holistically, you tend to, I think that positioning is

Yan:
really what gets you excited at first, because, and you know,

Yan:
for and all the reasons you're bullish on it are,

Yan:
you know, fundamental reasons, and then kind of put that against a backdrop

Yan:
of the fact that it's still barely permeated, and there's still very minimal adoption of it is.

Yan:
And I think that position is what really excited us about it in the first place.

Ejaaz:
Well, I think what something that's really interesting is your focus on kind

Ejaaz:
of crypto and web three for the initial fund.

Ejaaz:
Crypto is an incredibly fast changing technology, right?

Ejaaz:
And the whole point around it is it's meant to rebuild a ton of different sectors,

Ejaaz:
finance, media, you know, you name it.

Ejaaz:
AI is exactly that as well. So I'm not, I can't say I'm exactly surprised that

Ejaaz:
you guys are marrying both technologies together.

Ejaaz:
You're doing a ton of stuff in this space. So you just mentioned a few arms,

Ejaaz:
Anil, You're doing the research side, the investing side, and also the incubating side.

Ejaaz:
I saw that you guys are incubating a bunch of AI companies.

Ejaaz:
Maybe you guys can speak more to that. Jose, maybe I can pass this to you.

Jose:
Yeah, very similar to these guys. I had my crypto-pilled moment in 2017 with

Jose:
Ethereum and pretty much had the same experience last year, actually,

Jose:
a bit later than I think some of the other people at Delphi when I read situational awareness.

Jose:
I'd been playing with MidJourney, obviously, and ChatGPT, but I was just so

Jose:
busy and kind out of deep into crypto that, that, uh,

Jose:
I think I didn't realize just how momentous this thing was.

Jose:
And then, yeah, last year, once I read Situational Awareness,

Jose:
it really clicked into place.

Jose:
And we pretty soon decided with labs that we had to start doing some stuff in AI.

Jose:
So we put together our thesis on crypto AI, spent a lot of time on that,

Jose:
just figuring out where the good places for overlap was, and then ended up partnering up with Nir.

Jose:
Ilya is obviously an OG in AI. He's one of the original authors of Transformers paper.

Jose:
And yeah we we partnered with them to run our

Jose:
first accelerator in ai which was really really great

Jose:
had some insanely uh strong founders that applied just through ilia's network

Jose:
um and then did a second one uh about finished about two months ago with with

Jose:
the cyber fund guys which was also awesome um yeah we always think that the

Jose:
best way to i mean like anil said the the old

Jose:
elephant groping metaphor that Anil likes.

Jose:
We like having a lot of hands on the elephant.

Jose:
And I think researching is awesome and we're all kind of researchers in, is it our core?

Jose:
But building, you get a really unique perspective. And that happened for us in crypto too.

Jose:
Like there was things we learned by building protocols and being really deeply

Jose:
involved that we really couldn't have learned any other way.

Jose:
And it's been the same exact thing with AI. So it's just been really interesting.

Jose:
And also it's similar to crypto. It's an entirely new paradigm.

Jose:
Like crypto building is like very different

Jose:
from web2 building like you have this these smart contracts they're

Jose:
immutable well uh they used to be anyway nowadays protocols

Jose:
take a slightly different approach but still a lot of them are immutable and so

Jose:
it ends up being more like hardware like you have to be really careful

Jose:
you have to spend a lot of time researching um and then

Jose:
writing like uh it's less of an iterative approach

Jose:
and more of uh you know once this is out there it's it's out

Jose:
there for anyone to exploit and ai is like a different paradigm still

Jose:
where these things unlike like most

Jose:
of software before it aren't deterministic they're

Jose:
probabilistic and so it's really hard to ensure

Jose:
like a uniform user experience and like they're not even standards for like

Jose:
unit tests or anything like that um i really think the the metaphor of it being

Jose:
a new kind of computer is great so it's just been really useful diving in and

Jose:
and learning um like with our hands in the yeah with just just just getting stuck in and building so

Ejaaz:
There's a lot going on in AI right now, new frontier models are being released

Ejaaz:
like every week at this point.

Ejaaz:
Billions of dollars are being spent to train these things. There are numerous

Ejaaz:
consumer applications that are out there.

Ejaaz:
And I can't help but think that this is like an incredibly expensive game to play.

Ejaaz:
So I'm kind of curious, what's your unique edge when it comes to investing in AI?

Ejaaz:
How do you view the market right now? And where do you think you guys can make

Ejaaz:
the biggest impact with what you're doing?

Ejaaz:
You obviously have the whole Web3 crypto background, and maybe it's something

Ejaaz:
to do along with those kind of principles of investing that you had with that fund.

Ejaaz:
But I'm curious whether there's anything new you guys are seeing in the market right now.

Jose:
I don't think we have an edge right now. I think we're sort of

Jose:
hoping to build our edge over time. We've definitely made a lot of investments

Jose:
in crypto AI. I think we have edge there.

Jose:
We've made a couple of investments in AI, but I think we all sort of recognize

Jose:
that we're sort of paying tuition right now and getting to know the industry,

Jose:
getting to know as many founders as possible and kind of building our edge over time.

Jose:
That's the goal of intelligence really. Same as when we started in crypto,

Jose:
you guys didn't want to start a fund straight away, wanted to kind of build

Jose:
your edge, build your knowledge, and then go for that.

Jose:
And I think it's similar here, except now we have some capital behind us.

Jose:
So it makes sense to invest and start building that.

Jose:
So yeah, the hope is that we build the brand with Delphi Intelligence,

Jose:
get some really tough researchers on.

Jose:
And then we're also doing a couple of other things. Like we've been investing

Jose:
in young fund managers in AI, sort of looking to, like when we started Delphi

Jose:
Ventures seven years ago, it was really hard to raise.

Jose:
And we know firsthand both how hard it is to be a first-time fund manager raising

Jose:
and also how much edge you can have as a first-time manager.

Jose:
And so we're kind of looking to find those people that were in the same position

Jose:
we were in seven years ago in AI and back them and then kind of benefit from

Jose:
that deal flow and that learning.

Jose:
And I think the way we're thinking about it internally is we would like to aim

Jose:
to have edge and to really start accelerating our investment pace.

Jose:
12 to 24 months from now, something like that. So yeah, this whole thing is

Jose:
sort of us aiming to build that edge.

Anil:
I think like even when we first got started and we were writing reports,

Anil:
you know, if we put out a report on say, synthetics or something like that,

Anil:
people would always message us afterwards and say, damn, you guys really knew

Anil:
synthetics really well.

Anil:
That's why the report came out. So, you know, great or anything like that.

Anil:
It's quite the opposite, right? Like we learn about, you know,

Anil:
whatever we're researching when we're putting together of this report that we

Anil:
know is going to get like, you know, picked apart on places like crypto Twitter

Anil:
or by the team or by competitors. Right.

Anil:
So that's why we really do love having research embedded into our DNA,

Anil:
because like it almost provides like this check and kind of this like,

Anil:
you know, high bar that anything we publish, we know is going to be looked at

Anil:
by, you know, people either building the space or other investors in the space, et cetera.

Anil:
So we want to make sure that the research is not just really good for us to

Anil:
use and build conviction, but also meets this bar where it won't get ripped apart.

Anil:
And that kind of fear or intimidation, I think, is really powerful.

Jose:
Yeah.

Yan:
If I had to pick an edge, just to give you some answer to that question,

Yan:
I'd say it comes from a few areas.

Yan:
One, just from investing for however many years we've been doing it,

Yan:
and granted, that's an edge that's kind of consistent across anyone who's been

Yan:
doing it, So it's not necessarily a big one.

Yan:
I think we do have a decent variety of backgrounds and ways of thinking as well.

Yan:
And that's been an edge for us in crypto and should continue to be one here.

Yan:
And I think just being able to operate as a group is a big edge where we're

Yan:
able to take a variety of learnings that each of us are doing,

Yan:
bring them to the table and get kind of immediate feedback and have just a variety

Yan:
of points of view. I think that that's probably one of the bigger ones.

Yan:
And then patience, I think is another one that we've kind of learned over time

Yan:
in crypto in particular.

Yan:
And so here we realize we don't really have an edge and we're trying to understand

Yan:
is where the best opportunity is, right?

Yan:
Is it early stage or does early stage really take too long to get a proper payback?

Yan:
Is it makes sense to kind of invest in some of these growth-stage higher value or higher.

Yan:
Valuation but lower risk type plays where you have a pretty kind of cemented

Yan:
path to becoming a large company. And so that's still something we're exploring.

Yan:
I don't think we have really have an answer there yet, but I think it's just the patience.

Yan:
And I think what's helped with crypto is that you go through so many cycles so quickly.

Yan:
And I think you can draw parallels to kind of other online experiences versus

Yan:
physical ones. So if you think about like,

Yan:
online poker guys have have seen an insane amount of hands right and so they

Yan:
have a lot more experience than someone who plays live despite you know having

Yan:
a long-term career so i think you know there is some benefit uh in terms of

Yan:
taking that from crypto and understanding those cycles and trying to uh draw parallels there.

Jose:
Yeah i think we all agree i definitely

Jose:
agree with jan i think being a venture investor is like a skill that's sort

Jose:
of generalizable across sectors like a lot of it dating founders

Jose:
understand it but you you kind of need to understand the sector to be

Jose:
able to properly do diligence the founder and not get bamboozled

Jose:
by a high by a charismatic um you know sort of charlatan i guess um and so i

Jose:
think what we all agree with is that uh we all agree that this is going to be

Jose:
i think the biggest bubble that that like humanity has ever seen i think just

Jose:
like all the ingredients are there isn't it like already

Ejaaz:
A bubble this this was being said like last year and it's just been up only

Ejaaz:
i think what nvidia crossed like four trillion in market cap this week.

Ejaaz:
I feel like how big do you think this boat was going to go?

Ejaaz:
Because I agree with you, charismatic founders are super important,

Ejaaz:
but I see a bunch of these VC investors talk about theses for decades,

Ejaaz:
right? The next 30 years is going to look like this.

Ejaaz:
AGI, we're going to achieve it in whatever, 2027, or they're arguing about that.

Ejaaz:
How important is the founder when it comes to all of these kinds of things?

Ejaaz:
I'm guessing quite a lot. Yeah.

Jose:
To me, we have different I think focuses even as investors.

Jose:
To me, the founder is the most important thing, especially at the stage that

Jose:
we invest in, which is normally seed or pre-seed.

Jose:
The idea is going to change a lot.

Jose:
And you're really betting on

Jose:
a founder and you want someone that is just exceptional and has a history.

Jose:
And exceptional people leave breadcrumbs. You can sort of look at their past

Jose:
and be able to see some evidence of exceptional behavior before.

Jose:
And ideally, you're looking for the things that are like, he was insane at a

Jose:
video game or something in their youth, some sporting thing,

Jose:
those things are generally better because they're not as priced in as someone

Jose:
having done a successful startup and exited it or whatever.

Jose:
And you're really looking for these kind of freaks, basically,

Jose:
that are insanely motivated, that are able to...

Jose:
Go through walls to achieve what they want. And so that pattern of like,

Jose:
we've seen a few with ventures over the years, and those have been our big winners.

Jose:
And we're just looking for more kind of an AI.

Jose:
And then on the bubble comment, I don't think so.

Jose:
I mean, I think when you look at where, I look at 2000 as my mainly,

Jose:
like, maybe the biggest comp, like the price to earnings ratios of the Mag7

Jose:
equivalent, we're still like, you know, two to three EX what they are now.

Jose:
And then And I think in the private markets, there's definitely a few bubbly

Jose:
things, but there's also like insane growth and fundamentals,

Jose:
you know, like CatcherPT is the fastest company ever to a hundred billion in

Jose:
revenue, to a billion in revenue, to 10 billion in revenue.

Jose:
Cursor, I think was the fastest actually company ever to half a billion in revenue.

Jose:
And you're seeing multiples of these, right? With DAUs, like actual revenue.

Jose:
I do think there's some bubbly behavior and some stuff that's kind of reminiscent

Jose:
of 2000 with these valuations, but I do think there's just a long way to go

Jose:
just because, first of all,

Jose:
you have the most profitable companies in the history of the world that are

Jose:
stuck in this game-theoretic arms race where they're incentivized to spend every

Jose:
single dollar of free cash flow into training better AI models because otherwise

Jose:
they might miss AGI and have their company destroyed.

Jose:
And that's a dynamic that's just going to be a constant tailwind to making these models better.

Jose:
And every startup in the ecosystem benefits from better models. So there's that.

Jose:
And then I think there's just the fact that this stuff, like the internet was

Jose:
kind of like, people got really excited in 2000, but there was all this infrastructure

Jose:
that still needed to be built for the killer apps that people imagined in 2000 to work, right?

Jose:
You needed people to have mobile phones to build Uber. You needed payment rails.

Jose:
You needed like GPS working. You needed all these different enabling technologies.

Jose:
And with AI, it really feels like you don't. like everyone

Jose:
has a smartphone everyone has a has a computer fast

Jose:
internet like um it there's nothing

Jose:
in the way of this thing just scaling like

Jose:
it's really limited just by the quality of applications uh for people to use

Jose:
and there's so much talent going into it there's so much compute going in there's

Jose:
so much like spending uh happening that i just think it's it's gonna stay extremely

Jose:
uh it's gonna keep moving extremely fast uh yeah so i don't think this is the bubble the bubble yet.

Yan:
Yeah. And on the bubble point, I think you can kind of think of it in multiple phases, right?

Yan:
So right now you have this kind of scenario where the markets are.

Yan:
Really giving credit for just capex. So, so margins are coming down on some

Yan:
of these bigger players and, and it doesn't matter because they need to spend

Yan:
and, and spend and spend and just get to this point where, um.

Yan:
The, like the next kind of wave is proving out that the spend is actually valuable.

Yan:
And I think you're, you're starting to see elements of that,

Yan:
but the, the market is kind of very forgiving right now.

Yan:
And, and so, um, you, you know, for the first time in a while you have this

Yan:
technology that can improve efficiency by an order of magnitude.

Yan:
And it just gets captured in so many ways, right?

Yan:
You'll have the big guys who leverage their distribution to just improve margins

Yan:
because they need to reduce headcount or just become more efficient.

Yan:
On the startup side, you have these smaller teams that can get to unicorn status

Yan:
without really needing these longer term cash raises.

Yan:
And so I think the fact that it's kind of happening across multiple areas is

Yan:
what'll give it legs for quite some time.

Yan:
But yeah, in the interim, you have basically this massive spend phase and that

Yan:
doesn't seem like it's going to be slowing down anytime soon once we're starting

Yan:
to see that there are actual improvements to be made to the base models, right?

Yan:
There was that concern up front where, okay, it was actually kind of solved.

Yan:
And then when there were these big breakthroughs, then everyone,

Yan:
you know, the CapEx got turned back on again.

Yan:
And so it doesn't seem like that's really going to slow down anytime soon.

Yan:
But at the same time, you're having real efficiency gains at the early stage.

Yan:
And so, yeah, I think that the trickiest part is probably the very late stage

Yan:
investing side in the world where they don't necessarily need to bring on that capital.

Anil:
Yeah. The one thing I'd add here too is like, Bubble has this very like negative

Anil:
connotation to it, right?

Anil:
I think like one reason we're really excited is because we actually do exactly what Jan said.

Anil:
We think they're going to be insane efficiency gains. We think there's going

Anil:
to be this huge period of abundance, right?

Anil:
Obviously with this new innovation. And I think I think like,

Anil:
you know, one thing that we think about and we were talking about just this

Anil:
past week at our founders retreat is like, you know, there's this like the churn

Anil:
rate of Forbes 500, the Fortune 500 company every decade has just been going up and up and up. Right.

Anil:
So even if you use the churn rate from like the last decade,

Anil:
I think, you know, probably half of the companies would be kind of churned in

Anil:
the next like 10 years. Right.

Anil:
We actually think, or this is at least my stance, I think that churn rate is

Anil:
going to increase exponentially because of AI.

Anil:
And I think you may even see 350 to 400 of the top 500 companies get churned

Anil:
out in the next decade, which what does that mean?

Anil:
That just means there's immense value creation happening in other areas of the

Anil:
market and capturing even a little bit of that upside. I think it's just going

Anil:
to be the craziest thing that you could have ever hoped for as an investor, right?

Anil:
So yeah, I think we are excited for some of these big companies that already do exist.

Anil:
Obviously, like the Max 7, Fang, they're obviously fighting very hard to hold

Anil:
on to their spots, and there will be a lot of efficiency gains there.

Anil:
But I think more excitingly and obviously going to be much harder to figure

Anil:
out are the companies that will go from zero to some of these top 500 companies

Anil:
in areas all across the map.

Anil:
So yeah, honestly, we're just super excited. But yeah, I think it's going to

Anil:
be challenging, but that's why we're kind of pumped.

Ejaaz:
Yeah. So one of these words that I keep hearing all three of you mention is the word edge.

Ejaaz:
And it's like looking to find the edge. And what I want to ask,

Ejaaz:
because I think this is what I'm personally interested in, a lot of people who

Ejaaz:
are listening, is what the process looks like in finding an edge and what type

Ejaaz:
of topics you guys are interested in pursuing where you can find that.

Ejaaz:
Because a lot of times there are episodes, we're interested in just exploring

Ejaaz:
different frontiers, but there's a lot of different pillars in the world of

Ejaaz:
AI. There's so many different industries and categories.

Ejaaz:
Is there a particular spot you're excited about?

Ejaaz:
And within that spot, how do you go about finding an edge and getting an advantage?

Jose:
It's honestly like a lot of trial and error and being very honest with yourself about where you sit.

Jose:
I think that's something crypto really gives you like to survive and thrive

Jose:
in crypto. You need to be very honest about whether you have edge or not and where you have edge.

Jose:
And in AI, I think for us, it's just been a process of, I think,

Jose:
first of all, we started looking at, obviously, we did crypto AI where we thought,

Jose:
you know, there's an overlap here with crypto, we have an existing brand.

Jose:
The sector is exciting here. I think it's pretty clear that we can have edge,

Jose:
like we're very early to it.

Jose:
And then we started trying to do more AI direct investments. And I

Jose:
Uh, the, the bigger challenge. We were, we were like, some of the stuff was

Jose:
hard for us to get our head around.

Jose:
Um, but also it was unclear to us, um, like whether we had edge and that's always

Jose:
like a bad sign. Like you should, you should kind of know.

Jose:
Um, I guess we know the feeling of, of having edge to some extent.

Jose:
And I think it's a mixture of, um, there's like some reason,

Jose:
something that other people aren't seeing here, which I definitely think we're,

Jose:
we're like more bullish on AI than the average person, but probably not than the average VC. Right.

Jose:
So then we thought, okay, I think this direct investment, there's some negative

Jose:
selection happening here, like the deals that we're seeing are potentially not the best ones.

Jose:
And so we started to look at, I mean, first of all, we started to look at fund

Jose:
managers, which I think was an interesting one where we saw,

Jose:
okay, there's these fund managers raising small funds,

Jose:
first-time fund managers, they're really struggling too, because no one wants

Jose:
to back a first-time fund manager generally, and the fund of funds are very risk averse.

Jose:
And so, and we started seeing, wow, there's some guys here who are super plugged

Jose:
in, insanely well-networked and

Jose:
hungry, and really remind us of kind of ourselves seven years ago in AI.

Jose:
And this could be a way that we can have some edge, like not only will these

Jose:
guys perform, but also the deal flow that we get through them is going to be

Jose:
like pre-vetted and give us some access that kind of overcomes that negative selection problem.

Jose:
So we've been kind of digging into that now, and we think that there's edge there for us.

Jose:
We're also looking at China, like we've been looking at China for a while,

Jose:
one of our, actually, both our members of the investment team spend a lot of

Jose:
their time, of the intelligence team spend a lot of their time in China.

Jose:
I believe China is producing like over half of AI engineers.

Jose:
And also the, it's much, the rounds are much cheaper there because there's just

Jose:
less capital, like the US investors are

Jose:
really able to invest in China, like institutionals. And there's obviously concerns

Jose:
like geopolitical concerns and stuff like that.

Jose:
So you've kind of been looking there and figuring out whether there's a way

Jose:
for us to have edge there and to add some value in helping kind of these founders go global.

Jose:
So I think for me, I'm curious what the other guys think actually.

Jose:
And then we're also looking at kind of these secondaries of the big names,

Jose:
the Anthropics, the the groks um the

Jose:
the open ais and kind of figuring out you know

Jose:
whether we have edge there because i think there we're more just trying to capture

Jose:
the the beta versus have a lot of edge but um yeah for me it's a trial it's

Jose:
a trial and error process of like thinking through things going in doing some

Jose:
research and then figuring out being very honest with ourselves if we if we think we have edge or not

Yan:
Yeah no i think the the honesty is is the important one um edge comes in many forms, right?

Yan:
It's selection edge, it's timing edge, it's some informational edge.

Yan:
And then there's some that comes with experience in terms of bet sizing and everything else.

Yan:
And so for us, what we're in the process of doing now is basically trying to

Yan:
understand where we can have an edge.

Yan:
And I think even that on its own is very valuable or it could even be considered

Yan:
an edge and now we're like using this in a very nebulous way but so you know

Yan:
timing wise it's it's it's on the early side for sure right so i i think that's

Yan:
certainly one having the luxury to commit a.

Yan:
To look at this without necessarily needing to generate a return immediately,

Yan:
I think is a huge benefit, right?

Yan:
Where to some extent, other managers as part of their job, they're forced to deploy, right?

Yan:
And so that I think comes with a disadvantage where you might be deploying in

Yan:
areas you don't necessarily want to.

Yan:
So I think the patience itself is a huge benefit and should give us the opportunity

Yan:
to find those unique plays.

Yan:
I think one of the biggest things, and and this is another learning in crypto,

Yan:
is so much of it comes down to bet sizing, right?

Yan:
And it's like, it's really knowing what the opportunity is and whether you're

Yan:
allocating one, five, 10, 50% to a position is really what makes or breaks a

Yan:
lot of these or what really drives, I think, the outperformance.

Ejaaz:
How do you personally figure that out though, Jan? I know you say that and that's

Ejaaz:
what all the investors say, but I want to get inside your head.

Ejaaz:
What's the difference between you being like, you know what,

Ejaaz:
I'm going to give you around $1 to $5 million.

Ejaaz:
And then you're going, you know what, I'm going to pump in $20 million into

Ejaaz:
your thing, which is not something you guys are unknown to, right?

Ejaaz:
So what is that difference?

Anil:
Jan is a great person to ask this question to be honest.

Yan:
The big one is just risk. And so it's understanding, you know,

Yan:
how can this go wrong? And realistically, what is my downside?

Yan:
And then I think sometimes, you know, when things are going well,

Yan:
it's also knowing when to, like on paper, you should be taking position down.

Yan:
But I think there's an edge in understanding the position outside of it relative

Yan:
to the rest of your portfolio, right?

Yan:
And saying, sure, by the book, I should probably be downsizing,

Yan:
but it's more about how is this position relative to the rest of the market?

Yan:
Is everyone else underexposed?

Yan:
Will there be a lot of money coming in. And so I think that ends up really,

Yan:
it's like, it's understanding that your winners are winners and they should

Yan:
remain that way. And so you're either doubling down or leaving them as is.

Yan:
And so it's not often that you get really convicted and it's kind of in those

Yan:
scenarios where a lot of those edges line up, right?

Yan:
I happen to be down this rabbit hole and I found this, it's going to be a lot

Yan:
harder to get access to this in the future.

Yan:
I think it's de-risked more than people actually think.

Yan:
And so it's when the stars align in those scenarios that you really need to just kind of have...

Jose:
You're talking about optronic here?

Yan:
That's one of them. Yeah. And where you just have a lot of...

Yan:
And I think the risk tolerance is a big one too, where thankfully from crypto,

Yan:
you kind of get numb to the volatility.

Yan:
And I think that ends up being a huge edge as well, where you're just able to

Yan:
tolerate swings where if it goes wrong, it goes wrong.

Yan:
But ultimately, more often than night, it will go right. And you really want

Yan:
to be able to capitalize on those opportunities.

Jose:
Yeah, I think that Jan's really good at this. It's probably one of his biggest strengths.

Jose:
And we definitely have a lot of experience just from, in fund one,

Jose:
we started with one position in the fund just by virtue of our size.

Jose:
And the rest of the cycle was us just selling that position to buy others.

Jose:
And so we just, you really, from that, like understand deeply,

Jose:
like the importance of bet sizing.

Jose:
And you also naturally have this like hurdle rate, right? Like,

Jose:
is this thing going to outperform DoorChain?

Jose:
Which was our position at the time. But I think the sizing, that's,

Jose:
yeah, Yeah, one of the biggest things is also one of the biggest things I look

Jose:
for in fund managers, like people who are going to be concentrated and not afraid to take big swings.

Jose:
And it's also one of the biggest mistakes early fund managers make.

Jose:
They want to kind of, and like, concentration just drives all the right behaviors.

Jose:
Like it forces you to think about whether this founder is going to be able to

Jose:
return the fund for you, whether this is someone you want to spend a lot of time with.

Jose:
It forces you to actually add value to the founder. It forces you away from

Jose:
like indexing and just following in to around because Sequoia is in or whatever.

Jose:
So, and then the other thing is just like conviction is, it's like a feeling, right?

Jose:
That you build through research and speaking to someone and thinking about it.

Jose:
But when you have it, it's really important to recognize it because conviction,

Jose:
at least for me, it's not like you can have sort of 10x more conviction in something

Jose:
than you have on anything else.

Jose:
And a lot of people will feel that and size them equally anyway,

Jose:
right? Or like I have to have 10 positions or whatever.

Jose:
But actually if you're conviction, if you have 10X more conviction in something

Jose:
else, you should size it appropriately because those things don't come along that often.

Jose:
You know, there's only probably three to five, if you're lucky,

Jose:
spots a year where you really find that kind of conviction where the stars line up.

Jose:
And when you find it, it's really important to size things correctly.

Jose:
And it's kind of the biggest difference, I think, in performance for people.

Anil:
That's why we wanted to build this research team build this conviction right is like

Anil:
we think we feel confident in our ability to see these opportunities.

Anil:
But if you don't have the conviction, you may not take the swing at the right size.

Anil:
Right. And I think that's going to be really important for us.

Anil:
And then, you know, going back to Josh's question about, you know,

Anil:
obviously, we've been using the word edge a lot.

Anil:
I'll say that, like, you know, EJ started it. So

Anil:
not totally our fault. But the only thing I'd add to what these guys said is

Anil:
like, For me, I think one of the biggest edges that we founded with Delphi is

Anil:
just different perspectives.

Anil:
And I think that's what we're going to seek out with Delphi Intelligence as well.

Anil:
And I think that's not even just within our team, which we really do like building

Anil:
those perspectives and insights within the team.

Anil:
But I think more so just within our trusted network.

Anil:
Within crypto, we lean on our network all the time. And that really helps scale

Anil:
the amount and, you know, the speed at which we learn.

Anil:
That's definitely going to be something we lean on, you know,

Anil:
within other areas that we're trying to explore and learn about.

Ejaaz:
Yeah. So as you guys move into the world of AI, I'm curious if Delphi,

Ejaaz:
as a company, if you individually, you have a framework or a structure of how

Ejaaz:
you think about these opportunities.

Ejaaz:
Because AI is divided into a lot of big categories.

Ejaaz:
I mean, on the show, we like to talk about it as a layer cake almost where you

Ejaaz:
have the chips layer, then you have foundation models and you have dev tools

Ejaaz:
and like infrastructure. and then the top's the application layer.

Ejaaz:
And there's all these different worlds that you could explore,

Ejaaz:
I guess, to get that edge.

Ejaaz:
And I'm curious if any of you or if there's a company-wide kind of tooling or

Ejaaz:
a way that you explore these opportunities and find order in the chaos when

Ejaaz:
you're evaluating everything.

Jose:
I definitely think we have, different people have different perspectives on this.

Jose:
We've looked at things across the layer cake.

Jose:
I think personally, I'm most interested in the top and the bottom.

Jose:
Um i just think that's like

Jose:
um those are the places that tend to be

Jose:
the most defensible so we've looked at a couple of we haven't actually pulled

Jose:
the trigger on any although i actually made a mistake on one of them but we've

Jose:
looked at a bunch of chip startups and and and people doing new new architectures

Jose:
and stuff which have been really interesting um and then for me i'm really bullish

Jose:
on the application layer like i think i think chat gpt rapper is

Jose:
uh people use it as a as sort of

Jose:
you know to throw shade but i think chachapati wrappers

Jose:
are going to be insanely valuable and you're kind of already seeing it with

Jose:
with cursor you know and and others like it and to me ai the capabilities that

Jose:
it has already it could do um probably like 100x more than what people are using

Jose:
it for right now and that gap to me is the product opportunity

Jose:
of creating like verticalized applications

Jose:
with really clean products,

Jose:
with really smooth like context engineering and to solve like particular pain points.

Jose:
And I think you're going to have those across every single vertical and they're

Jose:
going to be, yeah, really, really, really big opportunities.

Jose:
So that's one I'm really excited about.

Jose:
But yeah, we look at stuff all across the stack, I think just to,

Jose:
at this point, just to kind of build knowledge. I mean, actually,

Jose:
in the crypto AI area, we did look at a lot of data stuff, too.

Jose:
We kind of had an intuition that that would be somewhere that crypto would have

Jose:
a particular advantage, like being able to, it's always been kind of a crypto

Jose:
thesis, right? Initially, it was this idea of Web3 Social where everyone would

Jose:
own their own data and you'd get paid for it.

Jose:
But I think the idea of coordinating a bunch of humans to provide valuable data

Jose:
to train AIs always was like an obvious or seemed like an obvious crypto AI idea.

Jose:
So we did make a lot of bets there too.

Jose:
I think we're a little bit more cautious now just given where things are going

Jose:
with synthetic data and just RL and we're being a bit more cautious there.

Jose:
But yeah, those are two that kind of came to mind.

Ejaaz:
So you mentioned that you're bullish ChatGPT wrappers.

Ejaaz:
Can you just give us the bull case for them?

Ejaaz:
Because I, like you, have seen so many people shit on them, basically.

Ejaaz:
Yeah. Why are you so bullish?

Jose:
The sort of precondition for me being bullish on a ChatGPT wrapper is the founders,

Jose:
or the app gets better as the models improve, right?

Jose:
So it actually becomes more useful as the models get better.

Jose:
And there's a lot of examples where that's the case. There was a lot of examples

Jose:
initially to be where you're just building some scaffolding on ChatGPT to do

Jose:
code or therapy or something.

Jose:
And that's not interesting. All that stuff will get picked off by the models.

Jose:
What is interesting is just verticalized applications, which improve as the models get better.

Jose:
And some of them, I think even the more interesting ones or the most interesting

Jose:
ones are the ones which actually don't work right now.

Jose:
They're actually just betting on the models improving enough that one day they'll work well.

Jose:
And there was a bunch of examples of that initially, but I think there's There's

Jose:
some interesting ones now too.

Jose:
But to me, the bull case is just, yeah, kind of what I said earlier,

Jose:
what i said before that to get the most out of

Jose:
out of models is actually hard work like you need pretty

Jose:
good system prompts for whatever uh vertical you're

Jose:
using it for right like if you're using a model for for therapy it needs to

Jose:
not be so agreeable uh it needs to actually like tell you hard truths and stuff

Jose:
like this whereas if you're using the model to write you uh i don't know a twitter

Jose:
or shill post or something then maybe you want it to be persuasive and and stuff

Jose:
like this um if you're using a model for investment due diligence you needed

Jose:
to have access to all your investment notes.

Jose:
You needed to know what the thesis is behind your firm.

Jose:
So there's all this, people call it prompt engineering. I like context engineering,

Jose:
which is a combination of meta-prompt and context.

Jose:
And that stuff is actually really hard.

Jose:
It's hard to get the most out of a model. And there's going to be applications

Jose:
that optimize that process for a specific vertical and just give users really

Jose:
refined experiences for it. Cursor is a great example, I think.

Ejaaz:
But also Anthropic just released Claude Code recently, right?

Ejaaz:
And so I'm curious about your thought around how much of the application layer

Ejaaz:
you think the model makers can actually kind of take, right?

Ejaaz:
So I'll give you another example.

Ejaaz:
XAI just launched GROC 4 and they have this huge distribution network, right? Which is X.

Ejaaz:
And granted, Elon is a very unique case because he's just buying everything.

Ejaaz:
He's probably going to be influencing the chip sector at some point as well.

Ejaaz:
He's putting chips into our brain, blah, blah, blah.

Ejaaz:
And he's building up a massive competitor in terms of data centers.

Ejaaz:
What edge do you think application builders that either you're investing in

Ejaaz:
right now or that you're looking for right now have over what model producers

Ejaaz:
can just kind of replicate themselves?

Ejaaz:
Is it in the context engineering that you're talking about, Jose?

Ejaaz:
Is it the fact that these founders can basically and intuitively describe how

Ejaaz:
an app should behave? Because a lot of this is just around social behavior.

Ejaaz:
The thing that makes an app successful is if you go on it and a bunch of people

Ejaaz:
like it and really vibe with it. That's it.

Ejaaz:
OpenAI just launched their agent yesterday.

Ejaaz:
And the number one bit of feedback I've seen was, this is cool,

Ejaaz:
but what am I going to use it for?

Ejaaz:
And if you have your potential target market saying, what am I going to use it for?

Ejaaz:
You haven't nailed the application there. So I'm wondering whether like there

Ejaaz:
is like, you know, maybe just a list of items that you think separates kind

Ejaaz:
of like founders that are building applications in AI versus like model producers

Ejaaz:
that are just going to like steal their stuff eventually.

Jose:
I think it's a great question. It's kind of the golden question if you're investing

Jose:
in AI applications, like is this something that the models can do?

Jose:
I think coding is an interesting one where, like, if, I think if Claude turns

Jose:
out to be the best coding model for everything, it's going to be hard for,

Jose:
for Cursor to, to win, right? Right.

Jose:
If it's just literally a Claude wrapper, although there's still like cool stuff

Jose:
that Cursor's built, like the, the rules, you know, which are,

Jose:
which I think is a really interesting primitive.

Jose:
I don't know if you guys have used cursor much, but it's a very interesting

Jose:
UX framework that they've built and there's other stuff like that.

Jose:
And I think there's definitely advantages to being laser focused on just pretty

Jose:
much user experience and not having to build your own models.

Jose:
It's hard to answer in the abstract and in the general. I think you have to

Jose:
go kind of like application by application.

Yan:
Yeah, user experience is a big one in the sense. I think one parallel is looking at Gemini, right?

Yan:
And how underutilized it is because it's just the UX is tough, right?

Yan:
And so it's kind of clunky. It doesn't really, it's not as widely used as you'd

Yan:
expect it to be considering how many people are using Gmail and all of that.

Yan:
And so I do think, you know, the UX is a big component.

Yan:
And so it depends on how much of the value is just in the raw processing ability

Yan:
of the model versus how much of the value in the product is in building out

Yan:
everything else around it and making the experience fluid.

Jose:
There's a lot, for instance, Harvey's an interesting one where they've just

Jose:
built a lot of scaffolding, as I understand it, a lot of scaffolding to make

Jose:
the document creation for lawyers extremely fast and seamless.

Ejaaz:
So Harvey AI, just for context for the listeners is like ChatGPT for lawyers. Is that right, Jose?

Jose:
Yeah, basically, for creating memos and stuff like this. And you want to be able to have your firm's

Jose:
standard boilerplate stuff and like whatever the style

Jose:
is that your firm writes in the key

Jose:
documents and you want to go document by document because this isn't

Jose:
this is like very high uh fake stuff that

Jose:
you don't want to get wrong and and i think that's going to be the case for

Jose:
like almost every vertical is going to have this uh and because like reliability

Jose:
is also a huge thing kind of talked about that before but these these models

Jose:
are not uh they're getting more and more but they still have hallucinations

Jose:
and not super consistent um that that's another thing that the

Jose:
kind of verticalized applications can can help fix with really good scaffolding

Jose:
and system prompts and stuff um but yeah i think harvey and cursor probably

Jose:
the two biggest examples of ones so far that i think have have built cool stuff

Jose:
on top of um like a basic wrapper nice

Anil:
Yeah i do also think customization is going to be a big key and i'm you know

Anil:
i wanted to jump in after you because I think like, this is something I go back

Anil:
and forth on a lot is a lot of these model creators obviously have a lot of

Anil:
data on, you know, who is paying for compute,

Anil:
how much they're paying and, you know, can very quickly figure out why,

Anil:
you know, if this person is paying, they're obviously building something that

Anil:
is valuable. Let's go copy and paste that.

Anil:
And yeah, to Ejaz's point, obviously a lot of, you know, these guys are all

Anil:
going towards this agent space, towards like creating something that is scalable

Anil:
to, you know, the masses.

Anil:
I think, you know, the last decade was very much about, you know,

Anil:
there's an app for that. and I think

Anil:
upcoming decade will be very much like there's an app for you,

Anil:
right? So very like custom app.

Anil:
Maybe Jose, like, I don't know if you want to leak or share some of the conversations

Anil:
we were having this week about like, something labs is building for Delphi itself.

Anil:
I don't know if you want to go into that. But like, I think that's a great example

Anil:
of like something that, you know, yes, we know a lot of these model creators

Anil:
will have something that will probably accomplish 70 to 80%,

Anil:
if not, maybe even more, you know, in the future for us.

Anil:
But it's something that, you know, So I think Louds wanted to roll up their

Anil:
sleeves, get their hands dirty and build something custom fit for us that would

Anil:
be, you know, fulfill basically 100% of our needs.

Jose:
There's like Delphi, we operate, we like to call it like the hive mind,

Jose:
right? It's also the name of our pod.

Jose:
And it really operates that way where there's a bunch of people in different

Jose:
divisions, some doing research, some building stuff, some investing that are

Jose:
having a bunch of interesting calls.

Jose:
And right now it's the sort of bandwidth between surfacing the interesting conversations

Jose:
for the whole firm to benefit from is really slow.

Jose:
Like we have to schedule these like bi-weekly calls. And then by the time that's

Jose:
happened, people have forgotten about it.

Jose:
And so I think the initial sort of vision is for it to be sort of an organizational

Jose:
knowledge base, or like we call it, you know, Delphi OS or Hivemind OS,

Jose:
which can just, first of all, like have all the conversations that people are

Jose:
having across the firms in a retrievable and like queryable format and then

Jose:
building like intelligence on top of that.

Jose:
So this thing can, for instance, generate IC memos really easily.

Jose:
Like I have a bunch of calls of the project and then it has our IC memo format.

Jose:
Maybe I can put in podcasts that the founder's done and then I can answer some

Jose:
questions to the AI and then it can just generate an IC memo format,

Jose:
you know, something that takes me kind of hours to do.

Jose:
You might have the same with research. Or for instance, if we want to have a

Jose:
kind of CRM of all the companies that we've ever spoken to, we can see all the

Jose:
conversations people have had with people at this company and also all the conversations

Jose:
people have had about this company, right?

Jose:
We can sort of search this and see, oh, this founder actually leaked to Malifaux.

Jose:
Like these guys are not performing well. They ended up using a different service provider or whatever.

Jose:
Like, and we want to have, and I think every company will basically have this in the future.

Jose:
Like it'll all the knowledge of the company will feed into this to this uh central

Jose:
like memory knowledge base whatever you want to call it and then there'll be

Jose:
various kinds of agents you can run on it that both help the company operate

Jose:
better and just automate and augment its people to be able to to be able to do more you

Anil:
Know you could kind of see this getting kind of crazier as time goes on right

Anil:
like recently we just had this big founders retreat and we always like to like

Anil:
kind of like share a book that we all read and stuff like that and this book

Anil:
for this last week was Essentialism by Greg McCohen, right?

Anil:
And you could see us using all this data that this knowledge base like fills

Anil:
and then in our chat add an agent that is based off Greg McCohen who like kind

Anil:
of follows our calls and then kind of shits on us whenever we're drifting away

Anil:
from, you know, what the thesis of his book is.

Anil:
So it's not like us holding each other accountable, but this agent almost holding

Anil:
us accountable to the decisions we're making at an org level.

Anil:
So yeah, I think we're super excited to play around with it.

Anil:
And I think it will be super useful for other companies.

Anil:
And at the same time, to answer your question, do I think this is something

Anil:
that like the big models like OpenAI, Anthropic, et cetera, you know,

Anil:
Crop are not going to build in?

Anil:
No, of course, they're obviously building it right now, as we've seen with all

Anil:
these recent announcements.

Anil:
But I think the customization is something that's really special.

Anil:
I think will be like, you know, again, what I said earlier, an app for everyone

Anil:
rather than here's an app for, you know, you.

Ejaaz:
Yeah, this is a super interesting point, right? Because you were able to build

Ejaaz:
Delphi OS using AI, and that would previously have been something that you'd

Ejaaz:
have to go to a larger company or use a lot of resources in-house to develop.

Ejaaz:
It's become much easier.

Ejaaz:
And then you mentioned that, well, Grok is probably going to integrate this.

Ejaaz:
ChatGPT will probably see these types of tools in. I'm curious where you see

Ejaaz:
the most forming, because a lot of the new innovations tend to become commoditized fairly quickly.

Ejaaz:
And I think one of the most that we've seen perform the best,

Ejaaz:
at least in the consumer world, which is what a lot of the people who are listening

Ejaaz:
to. are involved in, is ChatGPT's memory function.

Ejaaz:
And memory is amazing because it includes all the context of previous conversations

Ejaaz:
you've had, and it really locks people into that platform.

Ejaaz:
But outside of memory, I haven't really seen many other moats that make me want to use a model.

Ejaaz:
So I'm curious what your takes on moats are, if they're possible to capture

Ejaaz:
a large amount of a user base, or is it just going to be commoditized software

Ejaaz:
all the way up, all the models get better, they all kind of copy everyone's features.

Ejaaz:
Is there any moats that you guys are excited about?

Yan:
One funny one is there's a big moat to the brand and what kind of gets normalized, right?

Yan:
So as we kind of all agreed on earlier, we are using a very,

Yan:
very small fraction of the potential of these, right?

Yan:
And so if you think about the earliest adopters of this tech,

Yan:
which, you know, ChatGPT has an insane amount of users, but the penetration is still pretty low.

Yan:
And that's why it's so valuable. And so the first cohort is going to be kind

Yan:
of the most diligent about figuring out, okay, this one is better for this.

Yan:
This one is better for this.

Yan:
But each incremental onboarder is going to be less particular.

Yan:
And at the same time, all of the models will keep getting better.

Yan:
So what that basically means is each one will continue to use less and less incremental.

Yan:
Of the potential of this thing and and and they're all going

Yan:
to be relatively commoditized for their use case and so what

Yan:
it'll boil down to is what gets normalized you

Yan:
know uh going back to you know use xerox

Yan:
uh like for copying then google everywhere you google it and then now like chad

Yan:
gpt has has won that so far right that's just kind of the one that comes into

Yan:
mind for anyone who's looking to start dabbling in this and i think you know

Yan:
that as an onboarding tool and as a customer acquisition tool can't really be slept on?

Jose:
In general, AI stuff has less of a network effect than the web two giants did, right?

Jose:
Like social media and ad based stuff has way bigger network effect where it's

Jose:
just much harder to disrupt.

Jose:
But I think, I mean, the moat in AI, there's some things that have a data moat, right?

Jose:
Someone like Tesla that has like so many hours of driving data and there's other

Jose:
like robotics companies that we've looked at where that's a moat.

Jose:
I think OpenAI in itself, the amount of chats that they have and the ability

Jose:
to use that for training and things like this is also somewhat of a moat.

Jose:
But I do think in AI that the main moat is just going to be UX and speed,

Jose:
the team that is the best at constantly shifting to where the meta is and building the next thing.

Jose:
Ideally, you don't want your memory to sit with ChatGPT or whoever.

Jose:
And this is, I think, pretty visceral for people when they're sharing.

Jose:
I've shared some pretty personal stuff with ChatGPT.

Jose:
I think we all have. Yeah, like more personal than I ever thought I would have.

Jose:
So I think ideally, remember, we would actually sit and we have a project that

Jose:
we're incubating that's actually building this.

Jose:
Ideally, you would have private memory built on TE or ideally FHE once that works.

Jose:
And then you would give in sort of a cursor like UX, you'd be able to choose

Jose:
which model you want to give permission to access certain parts of that context

Jose:
to answer a query, right?

Jose:
I mean, the ideal ideal would just be you have a model that runs locally,

Jose:
but I think that's going to be

Jose:
super tough um so i think

Jose:
that's like one interesting area but i agree in

Jose:
general like the moats and that's why we've also been looking at

Jose:
deep tech stuff i do think the moats sort of end up also moving to like hardware

Jose:
to ip um to just things that in the past were seen as not sexy you know like

Jose:
uh it's not software it's it's too hard but i think those things will actually

Jose:
have like some of the most persistent moats in an era of of uh of ai and just insanely

Jose:
deflated cost of software.

Anil:
Yeah. I'd say that like on the memory front, I really hope that's not a moat, right?

Anil:
Like I, if memory is a moat, that just means that you're kind of like stuck

Anil:
into the, one of these ecosystems and you're really relying on that one builder

Anil:
to build every, you know, the best app of everything.

Anil:
Whereas like, you know, um, yeah. So, you know, to Jose's point,

Anil:
yeah, we are incubating a project that is, you know, based off this thesis that

Anil:
memory won't be locked in in one place and won't be disemoted.

Ejaaz:
So I feel like this whole memory term is just like another term to describe data, right?

Ejaaz:
And that's what all the top social media technology platforms have nailed so

Ejaaz:
far, right? They just aggregate the most amount of data.

Ejaaz:
I mean, Jose, you just mentioned that you use so much personal stuff or you

Ejaaz:
say so much personal stuff to ChatGPT.

Ejaaz:
I am talking to this thing for hours on end, right?

Ejaaz:
So at this point, I'm just like naturally inclined to use ChatGPT,

Ejaaz:
even though there's like another model that comes out. I really hope the portability

Ejaaz:
gets figured out Anil, to your point I just don't know what the incentives would

Ejaaz:
be for some of the bigger model producers

Jose:
It's sort of different from social media though because in social media it's

Jose:
not just the data it's the fact that all your friends are on there so you don't

Jose:
just have to port over your data you have to get all your friends to sign up

Jose:
to whatever new thing whatever web through social media thing you're using whereas here,

Jose:
portability you actually have access to all your chatchipity chats and it's

Jose:
not that heavy no matter how much you've talked to it It's text data,

Jose:
you know, it fits on any computer.

Jose:
So I do think if someone builds a great user experience here,

Jose:
it's something where it can actually win because it's a better product,

Jose:
like fundamentally. It just has to work really well.

Anil:
I also think, Ejaz, you put out this tweet, I don't even know if it was today

Anil:
or yesterday, it's been a long week,

Anil:
but you talk about the different personalities of these models, right?

Anil:
I think that's an interesting way to think about Emote as well, right?

Anil:
The conversations I have with Rock are way different than the conversations

Anil:
I have with like O3, right?

Anil:
So yeah, I think that's an interesting way to think about it as well.

Ejaaz:
No, that's a good point. For those of you who are wondering what this tweet

Ejaaz:
said, I basically described all the top models as having different personalities.

Ejaaz:
So I said, Grok was kind of like, whatever, naughty and rude and extremely horny, just to be frank.

Ejaaz:
And then ChadGBT was like kind of this incredibly agreeable personality.

Ejaaz:
Claude was kind of like, yeah, there you go. There you go. It's much more human,

Ejaaz:
right? It's much more intuitive.

Ejaaz:
Anyway, they have a bunch of different personalities and it kind of like attracts

Ejaaz:
a certain type of audience or it kind of like secretly molds you into being

Ejaaz:
some one type of a user, right?

Ejaaz:
You end up saying information to one model that you went to another.

Ejaaz:
And it just kind of like Chris's weird kind of sociodynamics that I think are interesting.

Ejaaz:
But kind of moving on, guys, I remember when you first started your fund in

Ejaaz:
2019, the stuff that you guys were investing in, I thought you guys were insane.

Ejaaz:
And this is coming from someone that like worked in the space, right?

Ejaaz:
And then of course, years passed, and it turns out that you guys nailed it.

Ejaaz:
So my natural question now that you're focusing on AI and investing so much

Ejaaz:
in AI is, and I'm going to put each of you in the hot seat, so prep your answer,

Ejaaz:
is what is one emerging contrarian trend in AI right now that you think everyone is missing,

Ejaaz:
but they should 100% focus on because it's going to become a big thing over

Ejaaz:
the next couple of years?

Jose:
I guess one thing I'd say is I don't think you necessarily have to be contrarian in venture, actually.

Jose:
I think you have to be right, but not necessarily contrarian.

Jose:
Although it helps for sure. Like it definitely is helpful when you're looking

Jose:
at something and you're really bullish on it and no one else happens to be.

Jose:
But I do think just, yeah.

Jose:
I mean, the area I'd say is the one I already spoke about, which is like GPT wrappers.

Jose:
I think a lot of people are sleeping on them and I think they're going to be

Jose:
absolutely like giant kind of businesses.

Ejaaz:
What's a GPT wrapper that isn't like a coding wrapper that you think people

Ejaaz:
should focus on or pay attention to?

Jose:
Um, I mean, this, this application that, that we're building internally,

Jose:
and there's a couple of teams that we've spoken to that are, that are building it.

Jose:
Uh, one of them is, is Den. It's like a YC company.

Jose:
So it's kind of like, think about it as cursor for, for work,

Jose:
right? It ingests like all your work data, your emails, your memos, your calls.

Jose:
Um, and, and, and then you're able to use any model to like run on that data.

Jose:
They also built a Slack clone, which I think is really interesting because the

Jose:
idea being that you're chatting with these models anyway, and actually in the future.

Jose:
And so you can open these chat groups with a model and your team in them.

Jose:
And you can all chat to the model together in these groups and have different

Jose:
models in the different chats, which I think is really interesting.

Jose:
The idea that you're chatting already, why not have a chat app where you can

Jose:
have group chats with the models and they can be on calls and stuff like this.

Jose:
I think various versions of those, I think you'll have a new Slack.

Jose:
I think all the company CRM stuff that Salesforce does right now is going to be rebuilt around AI.

Jose:
I have to think of more some more examples of good rap. Those are the ones I've mainly been focused on.

Jose:
But I think in hiring, for instance, you're definitely going to have something

Jose:
like that that's just going to know exactly what kind of person you're looking for.

Jose:
It can do the interviews for you, sort candidates for you.

Jose:
In every vertical you can think of, AI is going to have, you're going to want

Jose:
AI to do a huge percentage of the work

Jose:
and there's going to be an app that facilitates that workflow, I think.

Anil:
You're giving more high-level ideas, though. I feel like Ejaz wanted specific,

Ejaaz:
Right? I want specific. A specific company. Yeah, exactly. Yeah, yeah, yeah.

Ejaaz:
And the crazier, the better, honestly. Yeah, the crazier, the better. Just lean in.

Anil:
Mine aren't going to be crazy, and I hope Jan and Jose will make up for that.

Anil:
But going off of what Jose said, which is the contrarian part,

Anil:
I think, is over-indexed.

Anil:
And I think in crypto venture, it definitely worked out really well for us.

Anil:
But also, I think nowadays in crypto, there's not much stuff that is contrarian.

Anil:
Every conversation you have, people are bullish, hype or pump or something like that.

Anil:
But I think, you know, when it comes to, you know, just generally AI,

Anil:
I think for us, we thought the contrarian thing was, we thought even the most

Anil:
bullish people were going to be underexposed, right? So for us,

Anil:
we just want to be underexposed.

Anil:
And, you know, the thing that I go back and forth on, you know,

Anil:
to Jan's point is like, I think finding alpha here is going to be extremely difficult.

Anil:
Obviously, we're up for the challenge, but I think it's going to be extremely difficult. difficult.

Anil:
So for me, what I've been kind of pushing internally, and I think,

Anil:
you know, this is open to kind of like any anyone inside or outside of Delphi

Anil:
is, you know, capture a lot of this beta exposure.

Anil:
I think sometimes like investors and people just like to work very hard to,

Anil:
you know, to feel like they're smart.

Anil:
But I think almost like, you know, you can capture, you know,

Anil:
a nice index of, you know, open AI, Anthropic, like Andurl, Neuralink,

Anil:
all this stuff, and capture a lot of this beta upside in a lot of these like

Anil:
sectors that you think are going to be massive, right?

Anil:
Even in the public equities, I think like companies like Google,

Anil:
you know, maybe Tesla and stuff like that, I think are worth like looking at.

Anil:
You know, I'm super bullish Google, for example, even though people maybe,

Anil:
you know, are dancing on their graves because they're thinking that,

Anil:
you know, their big search is going to be like cannibalized by AI, right?

Anil:
Or open AI is launching this browser, which is going to like kill Chrome or

Anil:
something like that. Um, so yeah, I think, you know, again, definitely not contrarian, right?

Anil:
I'm literally fucking talking about Google and, uh, you know,

Anil:
open AI and stuff, but I do think that people will mid curve it and say,

Anil:
you know, that's too easy or, oh, these things have run away.

Anil:
Like maybe the 10 X is behind me or a hundred X is behind me or something like

Anil:
that. So let me try and find that

Anil:
a hundred X and then probably invest in things that go to zero instead.

Anil:
Right. Um, so that's my kind of answer, but, um, and then, you know,

Anil:
more in Jose's vein of like giving, broad ideas and not specific names.

Anil:
I think one idea that I think will be massive in the next, I don't know,

Anil:
12 to 18 months is I think...

Anil:
If you're using Twitter nowadays, you kind of get really annoyed at all these

Anil:
bots, right? And these agents that are like, in your replies, they're really bad.

Anil:
And so a lot of people are kind of like looking for a social network that is

Anil:
like, you know, people only, right? Maybe you do this world corner, whatever the fuck.

Anil:
I think actually the opposite is even more interesting where it's like a one

Anil:
on, you know, one where it's like you entering a social network where it's all agents, right?

Anil:
And you basically can kind of like get these agents to have a conversation about

Anil:
whatever you want based on personalities that you actually do follow, right?

Anil:
Instead of, you know, people listen to all in podcasts, and you're waiting for,

Anil:
you know, the topics that they're talking about, hoping to talk about a topic

Anil:
that is maybe relevant to you, you can kind of create your own podcast of those

Anil:
personalities, personalities you

Anil:
do want to follow talking about the exact topic you want to talk about.

Anil:
So I think something like that will be really cool. And I think will kind of

Anil:
exist in the next like 12 to 18 months.

Anil:
I don't know if the company exists yeah but um that's something that

Ejaaz:
I i think like meta is that's that part

Ejaaz:
of their strategy is just to kind of create a bunch of ai companions grok

Ejaaz:
is launching them as well and i wonder i wish i

Ejaaz:
could somehow track how much time each human user spends with some of these

Ejaaz:
ai agents and companions as they go live i bet you like it's going to be incredibly

Ejaaz:
sticky and what's really interesting about that anil um is that it's basically

Ejaaz:
going to be a reflection of the person to an extent, right?

Ejaaz:
And it depends on how much you dial up the sycophancy trait or if you dial it

Ejaaz:
down and it becomes kind of like your mentor that kind of like abuses you every

Ejaaz:
now and then and says like, no, you need to work harder or whatever that might

Ejaaz:
be. All right, Jan, you're up next.

Yan:
So one area I've spent a decent amount of time looking into and I'm super excited

Yan:
about is the humanoid space.

Yan:
So I think, you know, us speaking to a bunch of emerging managers and early

Yan:
stage investors, It seems as if most of them are kind of fading it to some degree,

Yan:
or they think it'll be more of a application-specific form factor that makes

Yan:
more sense from a cost perspective, from a utility perspective.

Yan:
Part of it is them talking their book, naturally, because building out the humanoid

Yan:
component is very difficult and expensive.

Yan:
And if you're doing early stage investing, it makes more sense to do these targeted

Yan:
use cases that can get to market a lot more quickly and start to generate revenue.

Yan:
And so I think there's a massive world where those make a lot of sense, right?

Yan:
The unit economics can be very predictable because most of the tech already exists.

Yan:
And I agree, there's a huge market for those.

Yan:
But I think fading the humanoid side doesn't make much sense.

Yan:
And the way to think about it is the market for the humanoid form factor is insanely huge.

Yan:
I'm very aligned with the idea that there will be billions of these in probably

Yan:
two decades just because of the amount of time it takes to build them.

Yan:
But I think there will be a massive just supply crunch for them within the next

Yan:
three to five years, realistically.

Yan:
Um the the the the human

Yan:
form factor makes a lot of sense because it can easily slot into everyday life

Yan:
now i think uh the cost component is

Yan:
starting to really get close to achievable so the the human form factor uh business

Yan:
model usually fell off in the transition from uh prototype to scalable model

Yan:
and and that makes a lot of sense right you have these insanely expensive robots that can breakdance,

Yan:
but that's not really valuable from a business perspective.

Yan:
Ultimately, what you want is reliability. So you're paying for hours worked, right?

Yan:
That's kind of what really drives the value prop here.

Yan:
And so I don't think there's a winner take all in this market because the demand,

Yan:
I think, is nearly infinite, right?

Yan:
And as they get better, the surface area for deployment and implementation only grows.

Yan:
They all kind of gather within, you know, they all learn together,

Yan:
which is, I think, something that isn't really appreciated enough where whatever

Yan:
it's learning in one factory, it gets to apply everywhere else.

Yan:
And so, and then, and you also, I think one of the things that gets faded on

Yan:
the humanoid side is the fact that people think there will be kind of a societal

Yan:
uprising, right? They're taking our jobs.

Yan:
But for the foreseeable future, it just kind of amplifies productivity, right?

Yan:
If you zoom out and think about demographics in terms of the population that

Yan:
wants to do some of these roles, that's only going to decrease.

Yan:
So cost of labor will increase. On the other hand, you have electricity costs

Yan:
will come down, production costs will come down, reliability,

Yan:
these things will come down.

Yan:
And these businesses become pretty profitable pretty quickly,

Yan:
especially when you think about their creative kind of forms of financing so i

Yan:
think that space isn't really um as

Yan:
as as appreciated and so realistically in the

Yan:
u.s there are basically three major players for it

Yan:
right you have tesla as the leader with optimus uh figure

Yan:
is second in line they just did a val they did a raise at 40 billion that's

Yan:
kind of getting wrapped up and then i think eptronic is the clear third and

Yan:
um that that's they're trying to do another race soon and that's the one uh

Yan:
we're really excited about internally because we see a lot of value there we um.

Yan:
We think what they excel in is the actuator side, which is basically the joint of the robot.

Yan:
And that's something they've been building for quite some time.

Yan:
And I think there is a moat in that because of how that contributes to the dollar

Yan:
spend per hour's worked formula and in terms of what it does for reliability.

Yan:
And then on the other hand, they're partnering with Google and plugging in Gemini, right?

Yan:
So you have the physical humanoid and then the model and the two needs to work

Yan:
in tandem. And so you can try and build the model from scratch,

Yan:
which is what Figur is doing after their kind of separation from open AI.

Yan:
But I think partnering with someone and focusing on your strength makes a lot of sense.

Yan:
And so, yeah, it turned into an electronic shell.

Ejaaz:
That point around the actuator, Jan, is such a crazy thing to think about.

Ejaaz:
Can you imagine in the Industrial Revolution when humans were just working at

Ejaaz:
factories, that they were each graded by their ability to move their elbow or

Ejaaz:
whatever at a 90-degree angle? That's just insane.

Ejaaz:
The fact that you can program economics into these things is crazy.

Ejaaz:
And I think you're right.

Ejaaz:
Being able to picture and visualize these robots as actual, not some otherworldly

Ejaaz:
creature, but just functioning humans and then monetizing that is just,

Ejaaz:
it's just a new model to kind of like wrap yourself around.

Ejaaz:
It's just insane.

Jose:
I think humanoid is a really good one because you can kind of like,

Jose:
I think being in crypto so long, you can kind of identify what things cause a bubble.

Jose:
And I think obviously the thing has to have very strong narrative potential, right?

Jose:
Like humanoid robots replacing all physical labor has that. And then you also

Jose:
have to have a lot of hate.

Jose:
Like you kind of need, because it both forces people to talk about it and also

Jose:
creates like these really hated rallies.

Jose:
And I think humanoid robots actually has a decent amount of hate from like smart

Jose:
people who just think that specialized robots are gonna win out.

Jose:
So it's a very, I think, good contestant for that. I'd give you two names that

Jose:
I think are interesting, maybe contrarian.

Jose:
I think Anthropic is really valuable.

Jose:
It's like the least valuable of the model companies. I think you could get it

Jose:
at like 60 bill when I last looked a month or two ago versus three to 400 billion

Jose:
for OpenAI and 150 billion or so for Grok or for XAI now.

Jose:
And they're clearly the winners in coding. Like they have been over and over again.

Jose:
I think they have a lot of market share in coding, like every dev and any dev

Jose:
you speak to is using CodeCode.

Jose:
I think that's insanely valuable. If you think software has eaten the world,

Jose:
is going to continue to eat the world, and you are literally the world's software

Jose:
factory, where everyone is going to produce software, I think it's insanely valuable.

Jose:
It's also one of the things that's easiest to train on because you have these

Jose:
easy kind of RL loops that you can do. It's formally verifiable and stuff.

Jose:
So I think they're actually in a really strong position. And

Jose:
it's tough because they don't have their own users i think

Jose:
a lot of people use it via api and that's generally

Jose:
not a not a great place to be but i think if they win coding that's

Jose:
like i think tens of trillions of of

Jose:
dollars like use case like i think it's only going

Jose:
to get get bigger um and then the other one the one we're speaking about at

Jose:
a dinner is just it's in a hated sector it's not to do with ai but it's it's

Jose:
epic games um so those guys they're doing like six billion in revenue and um

Jose:
i haven't found supply for it yet,

Jose:
but it trades at something like 15 billion,

Jose:
which, you know, it's a very depressed multiple and just because gaming is not hard at all right now.

Jose:
Gaming is in kind of a secular decline for the last two years.

Jose:
Sort of the time people have spent, not just crypto gaming, but time people

Jose:
have spent gaming has gone down for two years straight, which no one really thought was possible.

Jose:
No one knows the reason either. A lot of people speculate it's literally just

Jose:
TikTok eating your leisure time that people used to be spending gaming.

Jose:
And people talked a lot about the metaverse in crypto.

Jose:
Fortnite has actually built the metaverse. It's not VR like most people expected,

Jose:
but they have the closest thing to a metaverse in terms of

Jose:
Just different worlds that are player created, all the different maps that are

Jose:
player created, like 500 million users.

Jose:
They're having concurrent players maps with like thousands of players and just

Jose:
a really thoughtful CO and I think like everything is going to be leveraged

Jose:
by AI and I think they will be too just in the speed of what they can do.

Jose:
I think it's an interesting one that like it's always interesting to look at

Jose:
sectors that people aren't excited about at all and I think gaming is one of them right now.

Ejaaz:
Awesome. Before we round up guys you made a big announcement this week around

Ejaaz:
something called Delphi Intelligence.

Ejaaz:
And you gave Josh and I access to the platform beforehand. And we have to say,

Ejaaz:
like, we were super impressed.

Ejaaz:
Maybe you could tell us a little more about what this is and why it's important

Ejaaz:
towards what you guys are doing.

Anil:
Yeah, definitely. Yeah, so obviously, we've talked about this a lot on the pod

Anil:
already. But like, research is just at the heart of everything we do.

Anil:
And to be honest, like, any decision we make, we kind of want to go in with

Anil:
conviction and as much like insight and knowledge as possible.

Anil:
So we know we're not only making the right decision, but when we are making

Anil:
that decision, can size it properly, right? And I think for us,

Anil:
you know right basically you know jose right after he

Anil:
he kind of like passed around the situational witness paper um

Anil:
which he actually read on a you know week off which is like

Anil:
probably when we get the most work done it's like our weeks off um

Anil:
to actually like read and think about you know the future of delphi and everything

Anil:
like that i think that's when we really you know probably nine ten months ago

Anil:
at this point realized that um you know this was like a no not an option for

Anil:
us right we think to be the best uh investors builders researchers in crypto

Anil:
and honestly any area, you kind of need to start building expertise in AI.

Anil:
So that's when we really started, you know, rolling up recipes and doing the

Anil:
hard work of building out a team and building out kind of like an MO,

Anil:
which is just publish a lot of like great work on in areas that we're interested about.

Anil:
So we can kind of build conviction and build expertise in this area to help

Anil:
us make these decisions. So that's what Delphi Intelligence is.

Anil:
It's a research platform, free to access for all. So you can,

Anil:
you know, go on delphiintelligence.io right now, put your email in and you'll

Anil:
get all of our research, you know, basically bi-weekly free.

Anil:
We already have two reports out, you know, one on just like AI in the era of

Anil:
entertainment, and then one on video generation models.

Anil:
Both are like great. We have another one coming out next week on AI powered

Anil:
browsers, which I think is going to be like really top of mind for a lot of people.

Anil:
And essentially like, you know, it's us open sourcing our learning to the world.

Anil:
And what's cool about it too, is it's not just going to be our team.

Anil:
We're going to be curating a lot of great reads from within our network and

Anil:
people we respect, including some of the fund managers that Jose brought up.

Anil:
So yeah, I mean, if you're interested, please subscribe, follow us on Twitter

Anil:
and everything like that. But we're really excited about it.

Ejaaz:
Awesome. Well, thank you all for spending time with Josh and I and kind of going

Ejaaz:
through your thoughts on the AI market.

Ejaaz:
As you can imagine, there's just so much going on and our Twitter feeds or rather

Ejaaz:
our X feeds are off the hook. We are talking to like five different AI models

Ejaaz:
for various different things a day.

Ejaaz:
And it's just not easy to think strategically and long term and have conviction

Ejaaz:
around investments, right? Investments are such a hard thing to kind of nail.

Ejaaz:
So, you know, hearing your perspectives has been hugely informative for us and

Ejaaz:
I'm sure for our audience as well.

Ejaaz:
For the Limitless listeners, thank you so much for joining us for another episode.

Ejaaz:
As you know, Josh and I are trying out something new, which is just put out

Ejaaz:
loads of content as and when it comes live, as and when the topic is trending.

Ejaaz:
So we appreciate you and your feedback.

Ejaaz:
The main bit of feedback that we've got so far is that you love the guest episodes

Ejaaz:
and we want to get more interesting guests on.

Ejaaz:
We hope you see this as one of those pushes towards that.

Ejaaz:
And again, if you have any friends or colleagues or whatever that might be interested

Ejaaz:
in this thing, we appreciate you sharing, liking and subscribing.

Ejaaz:
Thanks, folks, and we'll see you on the next one. See you guys. Thanks.