Limitless: An AI Podcast

David:
Welcome to the AI Rollup brought to you by the Limitless podcast,

David:
where we stay up to speed with the emerging trends and developments in the AI space.

David:
I'm David Hoffman here with my two co-hosts, Ajaz and Josh. Ajaz, how you doing?

Ejaaz:
I'm good, dude. It's been quite a week, like subtle on the model updates,

Ejaaz:
but we're starting to see some of

Ejaaz:
the ripples of some of these strongholds that big companies have on ai right

Ejaaz:
so we'll get into some of the stories later but effectively we're seeing like microsoft owning

Ejaaz:
pretty much the entire stack of ai and we'll get into like what kind of effects

Ejaaz:
that has but also apple suggesting that they might update their search and remove

Ejaaz:
google from the default caused google to lose

Ejaaz:
almost 200 billion dollars in two hours and i watched this happen in real time so

Ejaaz:
i think for the last two years or so we've seen kind of ai on this rocket ship

Ejaaz:
right you open ai pop up out of nowhere wow these are it's a really cool product chat gpt etc

Ejaaz:
but now we're starting to see okay right the big boys have come to play how

Ejaaz:
much like of a stronghold do they have on the entire stack and what can they

Ejaaz:
do with this can they control our entire lives or is there room for disruption

David:
Yeah, the Silicon Valley just tightens. They're just going at it,

David:
realizing exactly where their choke points are and where other people's soft bars are.

David:
Josh, this is the first AI roll up that's fully on the Limitless podcast. How does that feel, man?

Josh:
Oh, man, it's exciting. I'm stoked. I'm glad for, I'm grateful for all the people

Josh:
listening who followed with us. There's a lot of AI content that is coming that

Josh:
is super, super exciting.

Josh:
And I think this week is no different. I would frame this week for the people

Josh:
listening as like the highest stakes version of Game of Thrones that we're watching unfold.

Josh:
And it is the largest titans in the world that are fighting for the most valuable

Josh:
prize, which is basically this new form of intelligence, this new thing that

Josh:
can control a lot of people. So it's really exciting. This is a crazy week.

David:
I think that's why I think I really like these episodes. Is this just real world

David:
Game of Thrones being played out?

David:
One of the first episodes that we did, Josh, we talked about just like what

David:
does an intelligence explosion mean?

David:
And it's kind of a big take that I think throws people for a loop.

David:
But the conclusion of that episode that we did, which I encourage everyone to

David:
go watch, is that we are making God.

David:
God is the end product of this. And so no wonder, you know, Facebook and Apple

David:
and Google and OpenAI and, you know, four more companies in China

David:
are all like turning into these gigantic tech titans, trading blows with each other.

David:
And I think, Ajaz, that is what you are going to open us up with,

David:
the blows that are being traded this week as everyone tries to jostle to position

David:
to be the first person to create God.

Ejaaz:
Yeah, so let's start off with the first right hook. Apple is coming for Google's throne.

Ejaaz:
And they're supposedly, or rumored, to be launching their own AI search product.

Ejaaz:
Now, let me give you a bit of context here.

Ejaaz:
You know when you open up the Safari browser on your Apple laptop or on your

Ejaaz:
mobile, and you go to the dress bar and you just type in your search query, right?

Ejaaz:
That uses Google by default.

Ejaaz:
And Apple actually pays Google the privilege of $19 billion per year just to

Ejaaz:
have Google automatically default in that search bar, right?

Ejaaz:
So it's a big deal, just let alone to license Google search, right?

Ejaaz:
And so this week, Apple execs Eddie Q

Ejaaz:
says that the AI search via tools like, you know, like chat GBT or Claude of

Ejaaz:
perplexity, things that, you know, the three of us have used on this show pretty

Ejaaz:
frequently is way better than Google.

Ejaaz:
And I don't know about you guys, but I've kind of like started transferring

Ejaaz:
my habit to that as well. Like I don't use Google to search anything anymore.

Ejaaz:
Even when you search something on Google, it's funny, it shows you a little

Ejaaz:
AI summary at the top of the page as well. So it's using AI itself.

Ejaaz:
And they're saying that they're exploring revamping search, which means that

Ejaaz:
they might replace Google.

Ejaaz:
So as you can imagine, this had a pretty gnarly effect on Google's market cap.

Ejaaz:
As soon as this announcement went live, it knocked $200 billion,

Ejaaz:
which equated to about 10% of Google's entire market cap.

Josh:
Red candle, like

Ejaaz:
Super aggressive over two hours. And I watched this happen.

Ejaaz:
I saw a tweet and it said like, Google's lost $100 billion in market cap.

Ejaaz:
And I was like, nah, that's not true.

Ejaaz:
And then I just saw it like subsequently just go lower and lower and lower.

Ejaaz:
And I think for context here, Apple has the biggest distribution for Google

Ejaaz:
searches. They have, I think, 2.35 billion active devices and all of them have

Ejaaz:
Safari pre-installed and has Google as the default browser there, right?

Ejaaz:
So let's hypothesize for a second and say, if they decided to switch,

Ejaaz:
then Google loses billions of search queries per day.

Ejaaz:
I think another way to look at this is like Safari accounts for,

Ejaaz:
I think, about 18% of all web browsing activity. So that's page views.

Ejaaz:
And in the USA, it's 30%, right?

Ejaaz:
So if this switches, Google would lose nearly around a third of their browsing activity, right?

Ejaaz:
And the cost to Google is larger, actually. Like putting that together,

Ejaaz:
that is, okay, what? Hundreds of billions of searches per year and around $50

Ejaaz:
billion of like gross ad sales.

Ejaaz:
So Apple can basically do this in a single software upgrade, by the way. guys.

Ejaaz:
And don't forget like ChatGPT itself is re-holding their search function.

Ejaaz:
We mentioned it last week. So I just think this is completely nuts and it just

Ejaaz:
shows that the distribution power of Apple at the device level,

Ejaaz:
and Josh, you spoke about this last week as well, isn't something to be joked

Ejaaz:
around with, right? Maybe the moat is at the device level for now. What do you guys think?

David:
I think we've all been watching Apple struggle with what to do about AI.

David:
They had this like whole AI announcement over a year ago and we've all been waiting for it.

David:
Now people are realizing that their AI execution abilities has not been great.

David:
So I'm a little bit confused because this would imply that they are more ready

David:
to make moves in AI than I understand them to be.

David:
Josh is a huge AI Apple fanboy and hardware fanboy. So I bet he's got some pretty

David:
interesting takes here.

Josh:
I do have some takes. I've been, well, very disappointed with Apple's ability

Josh:
to execute on AI, which leads me partially to believe that the market really

Josh:
overreacted to this news.

Josh:
$200 billion getting wiped out in two hours is kind of crazy,

Josh:
particularly because Apple hasn't yet rolled out any AI. And this is not a new

Josh:
problem for Google. Where we've been talking about this for months,

Josh:
there's a trend of the 10 blue links dying.

Josh:
Google search is clearly on the way out. It is a slow bleeding thing.

Josh:
Sure, this adds some accelerant, but I don't think this is as detrimental to

Josh:
Google's business as we think.

Josh:
So Google and Apple kind of have two polar opposite problems.

Josh:
Apple has amazing products, but they can't actually create the software to use it.

Josh:
And Google has incredible software. The Gemini, the new Gemini models are about

Josh:
as good as they can get, but they're really bad at creating consumer products

Josh:
and experiences for people to use them.

Josh:
So what they're doing is they're kind of harming each other.

Josh:
Like Google has the search, Apple has the users, and Apple's severing that link,

Josh:
which hurts Google's revenue. I think 56% of Google's revenue is search.

Josh:
That percent is going to continue to drop anyway, regardless of whether Apple

Josh:
cuts them or not, because AI is just getting so good.

Josh:
I think one of the underrated aspects of this that people aren't discussing

Josh:
is the cost per query of search versus AI.

Josh:
So the cost to generate a token is an order of magnitude higher than the cost

Josh:
to generate a search query. So if

Josh:
AI actually can't support the volume. We don't have the GPUs.

David:
Sorry, the cost to generate a single token is a order of magnitude more than

David:
it is a single search. Yes, it is.

David:
And so for an average AI query, you need hundreds, thousands of tokens.

Josh:
Yeah, you need a lot more data, a lot more bandwidth, a lot more compute power

Josh:
because Google is really just

David:
Nearly a naive 10X. It is incredibly more expensive to do AI search than it is to do a Google search.

Josh:
Yes, perhaps not token, perhaps not order of magnitude per token,

Josh:
but per search at least you can think of it each search is at least 10 times

Josh:
more expensive than a or each ai search is more expensive than a single search query because google's

Josh:
query is really just an index and it's just going through an index and pointing

Josh:
to things uh this is actually generating net new content so what i hope google

Josh:
would do google has this unique opportunity where they have i think 270 million paying subscribers

Josh:
they need to figure out how to get them onboarded onto ai they need to create

Josh:
a product they need to create a service they need to to fight back against this

Josh:
trend of the 10 blue links dying.

Josh:
And we're kind of seeing that, like you just mentioned, where when you do make

Josh:
a search query, it shows you this little Gemini prompt at the top where it kind

Josh:
of summarizes their results.

Josh:
But it's not there yet. And I'm really hoping they could figure out a way to

Josh:
monetize this like huge consumer base that they have to save themselves. But I guess we'll see.

Ejaaz:
Why do we think they suck so much? It's kind of been the recurring story of

Ejaaz:
Google, no? For like the last couple of decades, they tried to do the whole

Ejaaz:
like social network. What was it called again?

Ejaaz:
I forgot. Oh, but Circles, that sucked and died.

Ejaaz:
They're trying to do the whole like smart device situation. In fact,

Ejaaz:
actually today they announced that they are rolling out Gemini to all their

Ejaaz:
smart devices, right, their smart TVs, their watches and stuff.

Ejaaz:
And I was thinking about this and I was like, wow, that's pretty cool.

Ejaaz:
They're trying to like get it on different devices, collect different kinds

Ejaaz:
of data. And then I was like, who uses this stuff aside from like the Android

Ejaaz:
cell phones? No one, right? So maybe it's a skill issue.

Josh:
It's certainly a skill issue. And we saw this with the entire AI trend since

Josh:
inception. Like Google invented the transformer. Their researchers created the

Josh:
transformer white paper.

Josh:
That's why we have all these models now. And they had it for many,

Josh:
many years prior to OpenAI coming around and launching the ChatGPT API and then ChatGPT the product.

Josh:
They just can't seem to figure out how to create good products with their technology.

Josh:
It's like a bunch of really bright, intelligent nerds who can do amazing research

Josh:
and form these amazing creations, but they don't actually know how to sells

Josh:
people and that's that's been a serious problem

David:
Maybe we've branded this week marketed this week as like the old tech giants

David:
titans just going blow for blow but it's also there's some notion of like senality

David:
in these things like apple google they're just,

David:
old companies and so they're like old people trading blows and you're kind of

David:
just cringing as you watch it because,

David:
they are hurting each other, but they're not creating any value,

David:
as much value as they are harm, if that is indeed what's going on.

David:
And then by contrast, you have the much more sharp, newer startups,

David:
the Anthropics, the OpenAIs, and they're untested.

David:
And OpenAI doing hardware is a completely new field, but there's something optimistic

David:
about that, I think, than let's just get some new blood into this game and see what happens.

Ejaaz:
So you think that's going to leave more opportunity for smaller startups to

Ejaaz:
take the throne, potentially, David?

David:
Yeah, I mean, if Apple can't make hardware. Yeah, exactly.

David:
Yeah. If Apple can't make hardware that has AI integrated and if Google can't

David:
figure out how to integrate a software into hardware, well, maybe it just takes

David:
a net new generation of companies to solve that problem.

Ejaaz:
Or maybe it takes a net new medium or device or experience to onboard those

Ejaaz:
people, potentially, right?

Ejaaz:
Right. Like I've again, I keep saying this on episodes, but I have never spoken

Ejaaz:
into my handheld device unless on the rare occasion that I call someone and it's usually my mom.

Ejaaz:
And now I'm doing it like every day. Right. So I feel like we're getting closer

Ejaaz:
and closer to this weird little metaversal scenario where everyone is just kind of

Ejaaz:
got their digital avatars real and talking to each other through those digital

Ejaaz:
avatars potentially. pretty crazy.

David:
We've talked about on the show about how too many of our friends outside of

David:
the tech circles are just not using ChatGPT. They're just not using AI.

David:
And I went to one and I just told them, hey, download ChatGPT,

David:
open up the voice feature in it and start talking to it.

David:
And then their first question was like, what do I talk about?

David:
I'm like, literally the first question that comes to mind, just go, just go for it.

David:
And then call me back after you do that. And they call me back like,

David:
20 minutes later it's like i just talked to chat tpt for 15 minutes we just

David:
talked about whatever i was able to ask it questions and and talk about things

David:
that i was not able to talk to my friends about and i'm like.

Ejaaz:
Yes that's what

David:
So and i think that that is a homework assignment that i want to give listeners

David:
if you have not tried the voice,

David:
voice mode in chat tpt open up chat tpt after you're done listening to this

David:
episode and giving us a five-star review and subscribing on youtube open up

David:
voice mode and just start talking to it and see what happens and yeah just tell

David:
us in the comments about what that experience is like.

Ejaaz:
And i'm kind of curious to hear what you folks are talking to chat gpt about

Ejaaz:
i i hear a lot of my friends talk to them yeah

Ejaaz:
it's like it's like relationship advice it's like therapy for like issues growing

Ejaaz:
up as a kid as a kid and then a lot of people start to use it for like work

Ejaaz:
stuff but not like hey can you design this document for me it's like

Ejaaz:
i got into this situation with paul and i think he's encroaching on my territory

Ejaaz:
at work and i wonder like how can i like politely like put to him that he should

Ejaaz:
just basically fuck off and it's it's just he's like deeply personal and

Ejaaz:
i think very contextual to the individual and their own experience so it's kind of like outside of

Ejaaz:
Yeah, that's an interesting thing, actually. It's outside of the social etiquette,

Ejaaz:
you know? Do you know what I mean?

Ejaaz:
Like posting a picture of your ice cream cone at this new kind of ice cream

Ejaaz:
shop is like cool and accepted, right?

Ejaaz:
But talking to, or like speaking to this like weird social genie in a bottle

Ejaaz:
about like your personal afflictions with Paul is kind of weird. Yeah.

David:
It's useful for opening up about subjects that are hard to open up about.

David:
I think especially for dudes who it's easier to say, harder things to a robot

David:
than it is to another human,

David:
and so it kind of gives you that like safe space to be able to like talk about

David:
hard things that are on the cusp of things that you want to talk about but you don't have,

David:
the optionality you don't have the surface area you don't have the person to do it with.

Ejaaz:
So it's solving the the male loneliness epidemic

David:
Yes yes right yeah all of our all of our best friends yeah yeah yeah and this

David:
is how we give all of our data to it.

Josh:
Seriously that's how it's gonna happen i was listening to an interesting conversation

Josh:
that sam altman had actually it was at a sequoia session in sf this week and

Josh:
he was talking about how people use ai in different age brackets

Josh:
so he said the older people generally use it as an optimized search engine if

Josh:
they've even discovered it and then the like

Josh:
mid 20s to like late 30s group they use it kind of as

Josh:
a leverage tool to maybe help them with work and maybe solve some questions

Josh:
and then the younger people who are in middle school high school who are who

Josh:
are much younger they use it as an entire life operating system and they kind

Josh:
of load all their thoughts into it they load all of their homework into it

Josh:
they load everything into it and they consult it before before doing anything

Josh:
and it is like the single point of truth and the single reference that they

Josh:
can consult for for any sort of decision they make so i find that

Josh:
train interesting and going back to the point david that you made earlier that

Josh:
is there a new company that's that's going to dethrone apple and google and

Josh:
and open ai is going for it he very clearly said that they're going for the

Josh:
the ai subscription os plan for your entire life is is how can we build that

Josh:
that ecosystem for you have to

David:
Subscribe to your own life.

Ejaaz:
Uh that's literally the black mirror episode episode one of the new season you

Ejaaz:
know josh i as you were saying that, I was thinking like, why does that sound familiar?

Ejaaz:
And then I realized my sister, who is Gen Z, does exactly that.

Ejaaz:
So we briefly shared a ChatGBT account.

Ejaaz:
Lord knows how that's influenced my algorithm, actually. But she would literally

Ejaaz:
dominate 75% of conversations per week.

Ejaaz:
And it's just questions of like, random, like little social conversations that

Ejaaz:
she didn't know the answer to or things that she could have just,

Ejaaz:
I don't know, figured out herself.

Ejaaz:
And she was just offloading every single bit and I was like what is OpenAI going

Ejaaz:
to do with this data we should we should probably get some kind of like Gen

Ejaaz:
Z roundtable on this show at some point guys and like

Ejaaz:
Just kind of like interview them and see what they're using.

David:
There's this chart. We're going to shift subjects here. There's this chart that

David:
I saw on my timeline this week that I'm hoping one of you guys can explain to

David:
me. It's Microsoft's relationship with other companies in the space,

David:
OpenAI, Windsurf, Cursor, VS Code.

David:
And there's like lines of like ownership and partnerships between Microsoft

David:
and all of these things. And I kind of don't get it.

David:
I kind of understand that like, okay, Microsoft has their fingers in all of

David:
these different things and these weird partnerships. and deals and like org structures.

David:
But like, so I'm sharing it on the screen. It's like Microsoft has a 49% profit

David:
share with OpenAI, which owns Windsurf, which is forked from VS Code.

David:
And OpenAI is also an investor in Cursor, which is also forked from VS Code.

David:
But Microsoft owns VS Code.

David:
Can someone just like tell me what all of this is?

Ejaaz:
So Microsoft, if you guys recall about, I think about a year ago now,

Ejaaz:
guys, do you remember when there was that big saga with Sam Altman as CEO being

Ejaaz:
kicked out? by his board.

Ejaaz:
Do you guys remember this? When that happened, Satya Nadella,

Ejaaz:
CEO of Microsoft, kind of like rubbed his hands together and thought,

Ejaaz:
you know what, now is an opportune time to show support to OpenAI.

Ejaaz:
And we can do this in the sense of financial investment. So I think they invested

Ejaaz:
somewhere upwards of like $20 billion.

Ejaaz:
I don't know whether that was

Ejaaz:
set over a couple of years, but or just that one smack bang investment.

Ejaaz:
And they would invest a further $19 billion in providing all the GPU and CPU

Ejaaz:
infrastructure for OpenAI to run their inference and training.

Ejaaz:
So for OpenAI, that was a pretty good deal, right? It's like,

Ejaaz:
what, you're going to pay for all our compute hardware and help us train our

Ejaaz:
models and we don't, like, hang on, wait, what do you guys get in return?

Ejaaz:
And Microsoft said, well, we would love a bunch of equity and be like the majority

Ejaaz:
shareholder. And Simon Altman basically said, fuck off.

Ejaaz:
But what we can do is we can give you a 49% profit share of all services that

Ejaaz:
we charge for through OpenAI. And...

Ejaaz:
Any products we potentially acquire and integrate into OpenAI.

Ejaaz:
So if we look at the first top half of this diagram, Microsoft gets a 49% profit share of OpenAI.

Ejaaz:
And as we know, last week, OpenAI announced that they're acquiring Winsurf.

Ejaaz:
Now, a little bit of a kind of brain refresher for everyone.

Ejaaz:
Winsurf is the same as a company like Cursor, which were featured on the show before.

David:
It's the second place to Cursor.

Ejaaz:
Yeah, second place to Cursor. And the best way to think about it is these two

Ejaaz:
companies make it really easy to code up different kind of software apps and

Ejaaz:
these could these apps could be games these apps could be apps that appear on

Ejaaz:
your ios store to help you with

Ejaaz:
therapy personal training whatever you can dream up you can basically type in

Ejaaz:
a prompt like you do in chat gpt and it kind of codes up an app for you and

Ejaaz:
these companies like windsurf and cursor

Ejaaz:
Are now worth north of several single-digit billions because of this superpower.

Ejaaz:
And within this product itself, it's known as an integrated development environment.

Ejaaz:
It's basically carefully packaged software tools, code editors, and stuff like that.

Ejaaz:
So it abstracts away all the complexities of being a software engineer,

Ejaaz:
right? So people like you, me, and Josh can essentially code up stuff,

Ejaaz:
even though we haven't got any formal training or expertise in software engineering, right?

Ejaaz:
Now, there's a little circle on the bottom left, which says VS Code.

Ejaaz:
And people are like, I haven't really heard of VS Code. What is that?

Ejaaz:
It was the forefather to Cursor and Winsurf. Think of it as literally the thing

Ejaaz:
that gave birth to Winsurf and Cursor. In fact, there were rumors that went

Ejaaz:
around back in the day that Cursor had basically just forked VS Code's software.

Ejaaz:
So an interesting fact here, in 2024, a survey was done to see you know,

Ejaaz:
how much of the integrated development environment does, you know,

Ejaaz:
these different companies command?

Ejaaz:
VS Code commanded 74% of the entire market.

Ejaaz:
Now, I'm guessing that market share is down a bit, but that's basically the layout here.

Ejaaz:
So Microsoft has its hands on VS Code, and by proxy of OpenAI,

Ejaaz:
Windsurf and Cursor, because OpenAI acquired Windsurf, and OpenAI also invested

Ejaaz:
a hell of a lot of money into Cursor.

Ejaaz:
So I would say that some people have pitched this as like Microsoft owns the

Ejaaz:
entire AI stack and they kind of do, and I'll get into that in a second,

Ejaaz:
but it's mainly the coding stack that they have a stronghold on, on AI. Yeah.

Ejaaz:
Josh, I know you have a ton to say on this, so I want to throw it to you.

Josh:
Yeah, it's funny. It's like Microsoft funds OpenAI, OpenAI funds Cursor,

Josh:
and now buys the cursor rival Windsurf.

Josh:
And all this is on top of Microsoft's own open source code editor that they've forked.

Josh:
And now Microsoft's same dollar is getting taxed three times in the same ecosystem

Josh:
because they're competing against themselves with Copilot, which is Microsoft's own AI offering.

Josh:
So it's this very messy, conflicting thing in which Microsoft found themselves in.

Josh:
I'm not sure how happy they are about this because now Sam is trying to reduce

Josh:
that 49% profit share so that they can go public is the rumor this year.

Josh:
So it's kind of, it's a messy situation.

David:
Is the takeaway here that value is accruing to OpenAI more than it is to Microsoft

David:
and it's on the backs of Microsoft's work?

Josh:
Yeah. So there's an interesting dynamic that's part of the deal that we probably

Josh:
should mention where the nature of the deal is I think it's 75% of OpenAI's

Josh:
profits go to Microsoft until it recoups $13 billion, the total investment.

Josh:
Then after it's recouped its money, it reduces down to the 49%,

Josh:
which we're at, which goes until OpenAI reaches $92 billion of profit.

Josh:
And then after that, they're free and clear. So the special profit ends.

Josh:
So there is a cap on this upside.

Josh:
If Microsoft does hit that cap, that's still an amazing investment.

Josh:
They turn, what was it, $19 billion into $178 billion. So they've done really,

Josh:
really well. I think Microsoft is in a good spot.

Josh:
OpenAI, though, has this open-ended version of this profit margin.

Josh:
And if you ask Sam Altman, he'll probably say, oh, yeah, it's just a decade

Josh:
until we reach a trillion dollars of profit or something outrageous like that.

Josh:
So in the long term, it seems as if OpenAI stands to win.

Josh:
In the short term, Microsoft is doing well, but they also do have this little

Josh:
leech on them where now their largest investment is kind of leeching away users

Josh:
from Copilot, which is Microsoft's own offering.

Josh:
And it does create this weird dynamic where I'm not sure they're stoked about

Josh:
it. And now Sam has a lot of leverage because now he's trying to get that 49%

Josh:
profit share down even lower so he can make it more approachable to public markets

Josh:
when they try to go public.

David:
I I'm not privy to like the internal operations and like balance sheets and

David:
all this stuff but the idea of Microsoft you know I don't know 10xing their

David:
cash on billions of dollars,

David:
doesn't seem to feel so great when at the end of the day they don't own anything

David:
they get more money but they don't own any value they don't own,

David:
their game here the game of thrones is all these companies are trying to make

David:
God and there's no amount of money that is worth trading.

David:
The not having the opportunity of like owning a slice of God.

David:
And so I kind of see Microsoft as like the loser in this situation,

David:
no matter how many times they can multiply their cash.

Ejaaz:
So let's actually dig into that, David. I kind of went down a little bit of

Ejaaz:
a rabbit hole when one of you posted this in our chat.

Ejaaz:
And I was like, okay, well, hang on a second. What does Microsoft Stronghold

Ejaaz:
actually look like here, right?

Ejaaz:
So let's break it down. And then I have a comment on like where I think the moat is, right?

Ejaaz:
So in terms of compute, Microsoft Azure, which is their main cloud service,

Ejaaz:
right, competes with like AWS and Google Cloud, is OpenAI's exclusive cloud

Ejaaz:
for frontier model training and inference.

Ejaaz:
So Microsoft has dropped $13 billion on custom GPU clusters alone,

Ejaaz:
just to fund this, right? Just for the privilege of training OpenAI's frontier models, right?

Ejaaz:
So as long as Azure remains like this kind of, I don't know,

Ejaaz:
training substrate for open AI models, switching costs will keep Microsoft at

Ejaaz:
the center for now. But as you said, there's a heavy reliance on open AI there.

Ejaaz:
And I think this agreement is in until like 2030 or something, right?

Ejaaz:
They don't have overall board control, right? But Microsoft's hand is basically

Ejaaz:
firmly on Sam's shoulder saying like, hey, let's get first look into these other things, right?

Ejaaz:
The other part, which I thought was interesting is, Josh, you mentioned that

Ejaaz:
it's competing directly with Copilot, right? Microsoft Copilot.

Ejaaz:
But Copilot is based off of OpenAI's 4.0 and 4.mini models, right?

Ejaaz:
It's so incestual. Yeah, it's so incestual.

Ejaaz:
And Microsoft, I read this and I want to get kind of verification on whether

Ejaaz:
this is true, gets access to the 4.0 weights specifically.

Josh:
Yes, that is true.

Ejaaz:
So this is an important thing. So the more I think about it,

Ejaaz:
I'm like, huh, I think the moat is actually,

Ejaaz:
it only remains for Microsoft in this stronghold position if they continue to

Ejaaz:
get access to OpenAI's model weights, which I just don't see happening for the

Ejaaz:
longevity. Why would OpenAI do that?

Ejaaz:
And secondly, they're relying on developer stickiness through integrated environments

Ejaaz:
that use VS Code or whatever that might be.

Ejaaz:
Again, it depends on whether a good enough app or a software engineering community

Ejaaz:
decides to use Microsoft's VS Code.

Ejaaz:
And again, I don't see that as being too sticky to start off with, right?

Ejaaz:
And then, so when you look at that, and then on the coding side,

Ejaaz:
we can't forget that Microsoft owns GitHub as well, right? That's like 100 million

Ejaaz:
users and VS Code, which has accounted for about 74% of the IDE market share,

Ejaaz:
which we mentioned, right?

Ejaaz:
And then we have some of the equity spillover. So I'm not entirely convinced,

Ejaaz:
as you said, David, that Microsoft...

Ejaaz:
Their hold on this is going to remain quite permanent when OpenAI has so much power here.

Ejaaz:
If they were to switch to like some other provider in 2030, Microsoft's stronghold goes.

David:
I think this saga will continue. But ultimately, OpenAI has the IP.

David:
They own the model. They have the talent.

David:
And I think Microsoft is like scrambling to like work its way up the long tail of crumbs.

David:
And maybe it can like negotiate into like bigger and bigger crumbs.

David:
But that's kind of how I see this is like ultimately like OpenAI is the big

David:
winner here and there's no way for Microsoft to really change that.

David:
That's my takeaway from this section.

Ejaaz:
Yeah, I agree with you. I think the young blood has it here. Yeah.

David:
Yeah, let's get into the drama inside of OpenAI that happened this week because

David:
OpenAI has a new CEO that's not Sam Altman. So that's kind of crazy.

David:
I think we all understood like OpenAI is Sam Altman, but there's this new blog

David:
post and new news. OpenAI expands leadership with, I'm going to butcher this

David:
name, I'm so sorry, Fidigi Simo.

David:
And this is a message from Sam saying, hi everyone, I have some exciting news to share.

David:
I'm hoping to do this in a few weeks, but a leak accelerated our timeline and

David:
they go on to announce the new ceo,

David:
josh what's your takeaway with this is it real is this a real new ceo because

David:
the idea of cm altman not being at the helm of the ship is kind of crazy to me.

Ejaaz:
You guys ever watch the series the boys from amazon

David:
Prime yes yes never seen okay the superhero movie.

Ejaaz:
Yeah the concept josh seeing as you haven't seen it is it's all about superheroes,

Ejaaz:
but the dark side of superheroes.

Ejaaz:
So they're all gamed by political investments and they're like lobbied and a bunch of these things.

Ejaaz:
And it just shows the evil side of superheroes that it's not all fun and glory.

Ejaaz:
And there's this big corporation that manages these heroes, but of course,

Ejaaz:
managing a super powerful talent agency for superheroes.

Ejaaz:
But of course, when your superhero can laser beam you to death,

Ejaaz:
kind of who's in charge is the question there so they appoint the ceo and the

Ejaaz:
running joke in the entire series is she has absolutely zero power she runs

Ejaaz:
and gets them coffee and all these different kinds of things now i am not

Ejaaz:
suggesting that this is the case but the reason why i bring up that example

Ejaaz:
is she has been appointed as ceo

Ejaaz:
of applications at OpenAI. And it was branded across headlines that she was

Ejaaz:
CEO of OpenAI. That's just not the case.

Ejaaz:
She has some kind of equity ownership, but Sam Altman is very much still the

Ejaaz:
person, the man in charge.

Ejaaz:
And I think he's using this effort, at least he claims in the blog,

Ejaaz:
that he is going to now focus solely on research towards AGI

Ejaaz:
and towards aligning AI models to be to the benefit or betterment of humanity

Ejaaz:
and so we can focus his attention 100 on that whilst this new lady fiji simo

Ejaaz:
Can focus on building groundbreaking applications for the consumer end users

Ejaaz:
of open ai and i found this really interesting because well i had a few reactions

Ejaaz:
to this and i'm curious what you guys thought

Ejaaz:
Number one was like, Sam's been talking about AGI, AGI as the most important

Ejaaz:
mission of OpenAI since its literal birth, right?

Ejaaz:
And now suddenly he's making this big move to have someone who has a huge amount

Ejaaz:
of experience in building consumer applications to focus a large part of the

Ejaaz:
company's effort and resources and money on this.

Ejaaz:
So it made me think, are we actually as close to AGI as we thought?

Ejaaz:
Or maybe there are some other nuances and obstacles that we need to overcome.

Ejaaz:
Maybe it's at the technical hardware level, or maybe it's just the fact that

Ejaaz:
these things don't understand contextual awareness enough. And so they can't

Ejaaz:
be as personalized as we want. And they're only going to be used for niche cases. I don't know.

Ejaaz:
But on the optimistic side, I thought I looked into this Fiji Simo person,

Ejaaz:
and dude, her career and experience is absolutely stacked.

Ejaaz:
So she was just head of product or CEO at Instacart, which as you know,

Ejaaz:
is like one of the the top consumer applications in the western side of the hemisphere at least

Ejaaz:
and before that she was head of facebook whatever that title means but it sounds

Ejaaz:
again like a ceo of applications head of facebook at meta for 10 years which

Ejaaz:
is absolutely insane so she obviously has

Ejaaz:
a crazy amount of experience building consumer applications and i'm actually

Ejaaz:
really excited to see whether she builds out this amazing like chat gpt app

Ejaaz:
store i don't know what's your take josh

Josh:
Yeah, this is exciting for me because it displays like a clear divide in open

Josh:
AI. There are now two races that they're running. It's the consumer application

Josh:
race where we need to get these users. We need to lock them in.

Josh:
We need to create the best experience possible.

Josh:
And then it's the other half, which is we need to create AGI.

Josh:
And we need to create these systems that power this operating system that we want to build.

Josh:
So for me, this reminds me similar to X when Elon bought it and Linda Iaccarino became CEO.

Josh:
She's the one in charge of designing the day-to-day experience.

Josh:
But Sam is very much still the one who is calling the shots,

Josh:
who is providing the guidance to where the company is going.

Josh:
But to me, this signals to me like they're getting very serious about winning

Josh:
this race by having Sam fully committed to the AGI race. I think there is increasing competition.

Josh:
They now have a $500 billion investment with Project Stargate.

Josh:
Elon is trying to figure out how to get a terawatt of power to a data center.

Josh:
This is a huge scale battle that they're fighting, and the stakes are very high.

Josh:
So I think Sam focusing on AGI is good.

Josh:
This new CEO coming in, focusing on applications is good. I think she'll be

Josh:
responsible for building that.

Josh:
Life ecosystem that we talked about in the last segment sam is focused on building

Josh:
this super form of intelligence and i think that feels like a winning strategy

David:
Yeah so the way i'm hearing this is that sam is focused on building out the

David:
raw intelligence capacity of the open ai models like this this crude oil this

David:
oil this refined energy and then this new,

David:
ceo of applications is in charge of what that looks like, basically distribution,

David:
but like turning that energy and putting it into applications that fit in all

David:
different corners of our life.

David:
For whatever industry you're in, whether you're like, you know,

David:
pick an industry, any industry, some sort of application is going to be built

David:
that connects the AGI that Sam is trying to engineer to the end user,

David:
end product in a multitude of different ways that needs to come with like,

David:
you know, unique interfaces, unique experiences that fit into the different corners of the world.

David:
So that's kind of how I see, that's how that's how i summarize what you guys

David:
just kind of like put together there.

Josh:
Yeah sam's building the engine and fiji is building the the sports car on top

Josh:
of it or the yes the like really luxurious town that is powered by this single

Josh:
engine right it's kind of a good way to think about

David:
It yeah i can't looking at this photo she's very goth-like in this one particular photo.

Ejaaz:
I knew you were gonna say

David:
Something i can't not comment on how goth she looks and then when you go and

David:
you actually just search her on google she's just a very normal.

Josh:
Looking yeah welcome to the media

David:
This one particular photo just looks incredibly goth.

Ejaaz:
Yeah, why did they pick that? My God.

David:
I don't know, dude. All,

David:
right. All right. Let's get into the new models section.

David:
So every single week, almost every single week, there's some new models to talk

David:
about. This week is no exception. Gemini 2.5 Flash and also some new models

David:
out of China as well. Jaws 2.5 Flash. What do we need to know about it?

Ejaaz:
Okay. So firstly, I'm cheating a little bit here this week, guys,

Ejaaz:
because this model was actually announced, I think, three-ish weeks ago,

Ejaaz:
But we haven't updated Frontier models on this show for a few episodes.

Ejaaz:
And there are actually a few more recent ones.

Ejaaz:
Yeah. New to us. New to us. Okay. So Gemini 2.5 Flash.

Ejaaz:
It's funny. Earlier, we were talking about how Google kind of struggles to innovate,

Ejaaz:
at least at the consumer app level, but also at some other levels,

Ejaaz:
aside from just kind of like boring search queries or whatever,

Ejaaz:
you know, that important function that we rely on a lot.

Ejaaz:
Gemini 2.5 Flash is actually one of the leading models right now,

Ejaaz:
prior to O3 popping out. In fact, it is better than O3 in many ways,

Ejaaz:
but we just don't talk about that because OpenAI is still the darling child right now, right?

Ejaaz:
But it just doesn't get the love and attention that it needs.

Ejaaz:
And part of me thinks, you know, the reason behind this is because you can only

Ejaaz:
access it via API right now and not through their like main consumer interface

Ejaaz:
for Gemini. So their chat GPT interface equivalent.

Ejaaz:
But it is much better at reasoning and it shows huge gains over its previous

Ejaaz:
model, which is 2.0 Flash. So, you know, you can basically turn or toggle off

Ejaaz:
thinking or reasoning depending on what kind of thing that you're looking for.

Ejaaz:
And if you pull up this table, I think you guys just pulled up the comparison

Ejaaz:
of Gemini 2.5 Flash to like OpenAI 04 Mini, Croc 3 and all those kinds of things. It's

Ejaaz:
Comparatively cheaper than the majority of the models. I think only Gemini 2.0

Ejaaz:
Flash was cheaper than it. But of course, when you get a step change function

Ejaaz:
in actual ability for the model, it's way better.

Ejaaz:
And it's way better if you look across all metrics when it comes to things like

Ejaaz:
reasoning, science, coding, maths, etc.

Ejaaz:
I'm just wondering why it isn't talked about as much or maybe it hasn't been used as much.

Ejaaz:
My guess is it's because OpenAI is still so sticky and has all the kind of attention

Ejaaz:
when it comes to, hey, I want to try out this AI thing.

Ejaaz:
Like David, when you suggest to your friends, I'm guessing you don't go to them

Ejaaz:
and say, hey, get Claude or God forbid, Gemini. Yeah, no, you don't do that,

Ejaaz:
right? You're like, use ChatGPT.

Ejaaz:
So again, it just demonstrates no matter how good your model is right now,

Ejaaz:
at least, the stickiness with OpenAI is still there. Josh, I wonder if you have

Ejaaz:
any thoughts on Gemini 2.5 Flash before we move on.

Josh:
Yeah, it gets back to the thing we were talking about earlier,

Josh:
where you just have to create good products for these things.

Josh:
To have a new Frontier model is, it's impressive, but it's no longer like,

Josh:
oh my gosh, I must try this because

Josh:
my models that are on my like chat GPT desktop app are really great.

Josh:
And I'm pretty happy with that. So I think the reason why people aren't excited

Josh:
is because like I'm personally not going to go and engage with the API to test this thing.

Josh:
So therefore, it's not really super relevant to me. And I'm excited for the

Josh:
downstream effects if people want to create cool products on top of it.

Josh:
But as it stands now, I think myself and many other people are limited in their

Josh:
excitement because there's not many use cases for this for me.

Josh:
I think I'm most interested now in how this impacts my life or how this impacts

Josh:
the services that I use to improve my life.

Josh:
And if it's just an API call and there's not a lot of developers building interesting

Josh:
things that I use on a daily basis, then I'm like, oh, all right,

Josh:
well, O3 is still pretty great. I'll just stick with that.

David:
Right now, when OpenAI releases new models, like the new one is the O3 model, right?

David:
And that is going around in my life with my friends, talking to you guys on my Twitter sphere.

David:
It looks like and feels like back in the old days when Apple would release a new iOS update.

David:
And iOS would like make a material upgrade and all of a sudden our lives got

David:
enhanced. And it feels like that, like a new ChatGPT model launch is like this.

David:
Everyone's life is getting enhanced. We're all playing with it.

David:
It can do new things. It's very exciting.

David:
No other model release feels like that it doesn't have any impact on me because

David:
i don't use that app but it also does kind of beg the question why why doesn't

David:
this product exist and by this i mean a,

David:
a app like chat gpt or a front end like chat gpt that is a model aggregator,

David:
that also does the work of encode like remembering me and so right now i'm very

David:
excited that my chat gpt remembers all of my conversations and I'm excited to

David:
grow a relationship with ChatGPT to the point where like

David:
I start to work and go back and forth with ChatGPT on how best I want it to talk to me.

David:
I am like nurturing this relationship that I have with ChatGPT so that when

David:
it spits an output, it's spitting it out in a form factor that I like.

David:
Why is there not a product that aggregates all of the models?

David:
Because if you go to ChatGPT and you like do the little dropdown menu,

David:
there's like 4.0, 0.3, 2.5, 2.5 high mini, like blah, blah, blah.

David:
And I don't know what any of those means. And I don't care.

David:
And I just want to have this model aggregator do the work of choosing the best

David:
model intelligently while also doing the work of remembering me and my preferences.

David:
And I don't understand why that product doesn't exist yet.

Josh:
I'm not sure I have a great answer for that. I know very early on in the AI

Josh:
days, there was a company called Poe, which probably still exists.

Josh:
And they were an aggregator in the sense that you You can select your own model.

Josh:
It didn't have memory it didn't have the smart selection it just required the

Josh:
manual oversight to tap it i think we're kind of seeing this happening on the

Josh:
enterprise level with windsurf which we mentioned last week where windsurf is

Josh:
kind of serving as this orchestration layer

Josh:
they call it where as you submit a problem it will choose the optimal solution

Josh:
for that problem based on the model but in the consumer world i would imagine

Josh:
there's probably some sort of technical or privacy restriction

Josh:
on the model level preventing that from happening because it If you think about

Josh:
Apple and their App Store, which is probably the closest connection to this,

Josh:
is they won't allow any third-party applications on an iPhone.

Josh:
They must be approved. They must go through the centralized service.

Josh:
Open AI's largest moat right now is memory, is the new feature that they rolled out. So I'm not sure.

Josh:
That's why I'm thinking there's probably some sort of technical limitation preventing that from happening.

Josh:
It's because I find it hard to believe they would give up all of that data for

Josh:
free to an aggregator that can become much more powerful than them.

Ejaaz:
Yeah, I think cost is definitely a major function. Have you guys seen Open Router

Ejaaz:
by any chance? They kind of do what we're discussing right now.

Ejaaz:
This is the company set up actually by the ex-CTO and co-founder of OpenSea, Alex.

Ejaaz:
So if you pull up Open Router right now, just shove it into Google,

Ejaaz:
You'll see that it's described as the open interface for LLMs.

Ejaaz:
And basically, you can pull across any kind of model that they integrate with.

Ejaaz:
And it includes all of the frontier models from Western companies,

Ejaaz:
as well as some Chinese or Southeast Asian companies like Alibaba or DeepSeek.

Ejaaz:
And actually they're also famous for privately screening

Ejaaz:
if that's the right term new models before they even officially get announced

Ejaaz:
so open ai's 4.0 and o3 were actually tested privately you just didn't know

Ejaaz:
that you were testing it you was given a private code name like vulcan or something like that

Ejaaz:
and people were like wait this model is so good and it beats this benchmark

Ejaaz:
and that benchmark and everyone started speculating oh this might be open ai's new thing

Ejaaz:
And so they have a direct relationship where you basically type in a prompt on this website or you

Ejaaz:
build an app based off your API and it can select whatever model you want.

Ejaaz:
Now, I think the way that it's configured right now for this particular company

Ejaaz:
is you can kind of select which model gets picked for whatever type of query.

Ejaaz:
So it's not quite autonomous and smart.

Ejaaz:
And as far as I know, they don't have a memory function. Maybe Alex is going

Ejaaz:
to work on that going forwards. but I don't know this is kind of like the underpinnings

Ejaaz:
of something like that maybe

David:
Imagine being Alex and you leave OpenSea in like 2022, right at the end of the

David:
NFT summer to start an AI model aggregator project in 2022.

Ejaaz:
And then Sam Holman picks your company. Yes.

David:
Wow. Just perfectly selling the top of NFTs.

Josh:
Well done.

Ejaaz:
Well done.

David:
That's not the only model that came out. No. We also have one out of China as well.

Ejaaz:
Yeah, I mean, we haven't spoken about our friends across the ocean, but DeepSeek,

Ejaaz:
the ones that kind of like shocked the entire Western AI world by creating a

Ejaaz:
model that was much cheaper, but more effective or as competent as OpenAI's

Ejaaz:
frontier models, which they spent billions and billions on

Ejaaz:
and DeepSeek just spent a couple hundred million millions on, is supposedly

Ejaaz:
releasing their next major model update.

Ejaaz:
So it's not officially out just yet, but this has kind of been leaked that the

Ejaaz:
model will have 1.2 trillion with a T parameters, and it'll be a hybrid mixture

Ejaaz:
of experts model, which means that

Ejaaz:
whenever someone queries the model the entire set of parameters

Ejaaz:
aren't queried at once which is typically how a lot of these models work instead

Ejaaz:
it's more efficient and it might hit like 37 billion parameters which is needed

Ejaaz:
to answer your particular query or your question right so it's much more efficient

Ejaaz:
and that's how they like mainly drive down the cost

Ejaaz:
it's also going to be almost a well almost 99 but it's 97.3 cheaper than gpt4o

Ejaaz:
which is just again insane and we were kind

Ejaaz:
We were kind of talking about like, you know, earlier, what kind of moat would

Ejaaz:
Microsoft have with like Copilot or Gemini have with 2.5 Flash?

Ejaaz:
And you could argue that being cheaper per query or per token could be a moat there, right?

Ejaaz:
Because it all adds up. And if you're a startup that's consuming a ton of data

Ejaaz:
and you can get kind of like near the same kind of quality that OpenAI results

Ejaaz:
give you, then I'm going to take that cost cut and trade off and just use this model instead, right?

Ejaaz:
It's trained on a hell of a lot of data, 5.2 petabytes.

Ejaaz:
And it's meant to have a much better reasoning model. Okay, blah,

Ejaaz:
blah, blah. We always speak about these different characteristics.

Ejaaz:
But what's something that's actually cool here is that last point on this tweet.

Ejaaz:
82% utilization in Huawei Ascend 910b.

Ejaaz:
What they're referencing there is a Chinese manufactured AI chip.

Ejaaz:
And I just want to emphasize how important this is because to date,

Ejaaz:
NVIDIA has literally, and I'm not making this up, dominated like,

Ejaaz:
Like 95% plus of the chip manufacturing industry.

Ejaaz:
And that includes like the compute side of things, chip design,

Ejaaz:
like which informs how your model is going to get consumed, data trained, etc.

Ejaaz:
It influences a heck of a lot. We just don't speak about it as much, right?

Ejaaz:
And now we have China, which was supposedly meant to be a decade behind the

Ejaaz:
Western world, coming up with a chip that is 75% as good as NVIDIA's flagship

Ejaaz:
or one of their old flagship chips, the H100s.

Ejaaz:
So it's just really interesting to see how much China, and we're going to get

Ejaaz:
into this in the next topic, but DeepSeek, Chinese manufacturers with Huawei,

Ejaaz:
they're building their own kind of empire.

Ejaaz:
So maybe it's not Game of Thrones on the Western world. Maybe it's just Game

Ejaaz:
of Thrones, but then China just kind of like swoops in and takes out the head.

Ejaaz:
Josh, I know you have a ton of experience on the hardware side of things.

Ejaaz:
I know you love nerding out about this. I want to hear your take on this?

Josh:
I'm obsessed with DeepSeek in the way that I kind of love the Grok team as well,

Josh:
because they just have the fastest rate of acceleration.

Josh:
And I think when you're competing on these like long time horizons with really

Josh:
high stakes and strong exponential curves, the rate of acceleration is probably

Josh:
the most important thing to pay attention to.

Josh:
So Gemini and OpenAI are very clearly the leaders in terms of frontier models right now.

Josh:
I think what we're seeing with DeepSeek is this hyper focus on efficiency.

Josh:
And that efficiency has exponential scales that I think will probably quickly

Josh:
exceed that of the top leading models in the US.

Josh:
What's interesting is they're winning on the hardware front and they're winning

Josh:
on the software. Well, they're not winning on the hardware front,

Josh:
but they're rapidly getting closer.

Josh:
You said it had each, what was the percentage of the H100 that they were able to capture? 75%.

Josh:
So they're 75% there. And that number was much lower just a few months ago.

Josh:
What China has is really great manufacturing capabilities. They have really

Josh:
smart developers who are hyper-focused on this resource constraint.

Josh:
And what we're getting is these incredible models. 1.2 trillion parameters, I think.

Josh:
Is that a record for a public model if it does go public?

Ejaaz:
Yeah, Meta's supposedly working on a 2 trillion parameter model.

Ejaaz:
But, you know, we're talking about sizes here.

Josh:
So far, if this were to release today, this would be the largest public model.

Josh:
And the cost of it are a fraction of what it costs to use GPT-40.

Josh:
So the trends that we're seeing is an increase in efficiency,

Josh:
an increased rate of acceleration in progress, and an aggressive decrease in

Josh:
cost per query or cost per token.

Josh:
And those three things are like really powerful effects that are going to affect

Josh:
markets in a really big way.

Josh:
I think as we get more power at a lower cost, it unlocks the amount of productive

Josh:
output of these tokens. I forget the amount of trillions of tokens that Microsoft

Josh:
said they generated in the last quarter, but it was many, many trillions of

Josh:
tokens. And if you can get

Josh:
10 times that token output for a fraction of the cost, that's a lot more intelligence

Josh:
you can use to either distill models like we see in the smaller models like

Josh:
that Quinn is doing, or just to apply to even larger models and to do reinforcement learning on those.

Josh:
So I think the cost per token and the efficiency of these models is something

Josh:
to really pay attention to.

David:
Okay, so if R2 is coming in with a 97% cost reduction, as in the cost of running

David:
this model is 97% cheaper, that means we can just, you know,

David:
do 33 times more compute for the same amount of cost.

David:
So you get 33 times more compute for the same cost.

David:
But does that result in a 33 times better outcome?

David:
Because I don't think that that's true. I think the queries that you get,

David:
the products that you get to come out of that compute is running up on some

David:
sort of like constraint, right? like we don't just get 33 times better outputs right.

Josh:
Not directly so the constraint mostly is in the actual model intelligence and

Josh:
those benchmarks that you see so this would be if it gets released 1.2 trillion

Josh:
dollar or trillion token model or parameter model and that would be the upper

Josh:
bound of intelligence so it's not

Josh:
creating that new intelligence it's not agi but what we do have is we have a lot more use of it so

Josh:
currently the the cost per token here is much more expensive

Josh:
and when you get a cheaper token you unlock the second order effects of that

Josh:
which means consumer applications become a lot more cheap it means that we can

Josh:
get access to just a lot more intelligence a lot cheaper and a lot more accessible

Josh:
and i'm not sure i have great examples of

Josh:
of where that comes but i'm not sure there's anyone in the world that wants

Josh:
that doesn't want more intelligence and doesn't want more queries or shots on

Josh:
goal at doing this so it's not that it is a smarter form of intelligence But

Josh:
the amount of it and the accessibility of it goes up significantly.

David:
Well, that's kind of just on brand with like what China tends to bring to the

David:
table. It seems like the United States, Silicon Valley, brings these net new

David:
innovations, more intelligence, increasing IQ, and then China just makes that cheaper.

Josh:
Yes. Although this time, 1.2 trillion parameters, that is the new top dog.

Josh:
So as of now, assuming Benchworks come out well, that would be highest intelligence

Josh:
and lowest cost, which is like a huge win.

Ejaaz:
I actually had a question. If the cost per query goes down and we kind of continue

Ejaaz:
this trend or this directional trend towards post-training via reinforcement learning, right?

Ejaaz:
Wouldn't this basically enable them to scale that up?

Ejaaz:
So you could end up with a much smarter model at a cheaper cost because it costs

Ejaaz:
less to query in itself, right?

Ejaaz:
So you could just chuck in a bunch of, you know, reasoning traces,

Ejaaz:
which for those of you who don't know is kind of like questions,

Ejaaz:
which will show a question A and then an answer B.

Ejaaz:
And the model basically has to figure out how to get from A to B.

Ejaaz:
And that's what makes the model smarter through reinforcement learning.

Ejaaz:
Josh, don't you think this will basically open up

Ejaaz:
an opportunity for them to scale to a smarter model i'm guessing that's how

Ejaaz:
they're doing it right now right alibaba is doing the same thing with quen i

Ejaaz:
mean yeah the new model that they just released is only 235 billion parameters

Ejaaz:
but it's as smart if not better at certain math coding and science benchmarks

Ejaaz:
and maybe not so much the social reasoning side of things but on certainly across

Ejaaz:
those benchmarks than some of the frontier trillion

Ejaaz:
dollar not trillion dollar trillion Yeah,

Josh:
No, that's a really great point. I think you're right. What we probably see

Josh:
is similar to Quinn in the distillation of models and then the hyper specialized

Josh:
version of those models.

Josh:
So as basically the way these distilled models work is there's this giant foundation

Josh:
model and it uses reinforcement learning to create these good prompts,

Josh:
these really high quality prompts that get fed into a smaller model that can

Josh:
run locally on a, let's say a laptop or your own device that you have at home.

Josh:
What happens from this probably is is now that you have this higher form of

Josh:
intelligence that you could distill much more cheaply you can create a lot more

Josh:
of these distilled models that are even even tighter in size even smaller and

Josh:
maybe enough to run on your iphone

Josh:
that are effective enough to actually work and maybe you get a distilled model

Josh:
that's kind of like a sirion steroids type thing that you could run locally

Josh:
on a phone but maybe you also just get a really great

Josh:
hyper focused model on a specific type of math problem or a specific category

Josh:
of science or like a very specific data set that it's trained on

Josh:
that costs basically nothing because it's been distilled down and it costs so

Josh:
cheap but it's really really smart at this one particular thing and if you have

Josh:
a series or a network of these nodes that are hyper trained on a specific category or a specific problem

Josh:
then you could kind of build this mesh network on top to to interact with each

Josh:
one based on your the needs of the query so it creates this interesting modular approach

Josh:
versus the large foundation model where if you can distill these things down

Josh:
cheaply enough and broadly enough into all these hyper-localized little nodes,

Josh:
then you could access a lot of general intelligence very, very cheaply.

Ejaaz:
It reminds me of another example. I mean, like the general theme that we're

Ejaaz:
talking about here is Chinese AI manufacturers can basically make things cheaper and more efficient.

Ejaaz:
But the difference, which is typically what they've done with like mobile phones

Ejaaz:
when Huawei is copying a bunch of iPhone stuff.

Ejaaz:
But the difference here is that they're actually innovating on the design and

Ejaaz:
the execution, which is really important.

Ejaaz:
And it reminded me that Huawei this week launched, I think it was like a 718

Ejaaz:
billion parameter model,

Ejaaz:
but it was trained on 6,000 of their new Ascend chips, which is like their equivalent

Ejaaz:
of like NVIDIA H100s or whatever you want to call it, right?

Ejaaz:
And this was a significant reduction from the 8,000 chips that they used to

Ejaaz:
train their previous model, which was half the parameter count.

Ejaaz:
So basically, chip count down,

Ejaaz:
model intelligence up, which is a scary but pretty exciting trend to see.

Ejaaz:
I guess the benefit of this so far is that us on the Western side can kind of

Ejaaz:
benefit from it because they're open sourcing all these models and they're probably

Ejaaz:
trying to vamp attack some of the bigger companies in America.

Ejaaz:
But at some point, they're going to close the gates because they've achieved

Ejaaz:
parity, if not betterment of these models.

Ejaaz:
So it's an exciting trend to kind of watch.

Ejaaz:
And it's not one that's getting publicized in the Western media as much.

David:
I'm just seeing as all of these numbers come down ever since we started this

David:
podcast, you know, token parameters go up, chips go down, and the rate of this

David:
happening is just not slowing whatsoever, which makes me think that just like,

David:
it just keeps on going it just keeps on going.

Ejaaz:
Yeah so we're going to end up in this like infinite abyss david where we can

Ejaaz:
you you can talk to chat gpt as much as you want you know earlier on josh was explaining how

Ejaaz:
you know doing search queries via ai is so much more expensive than doing a

Ejaaz:
google search query but maybe not for long maybe not for much longer

David:
By the end of this year it's not going to be the same.

Ejaaz:
Yeah yeah

David:
There are some things to talk about on the crypto side of things where we got

David:
this podcast started. So cookie.fun is a website I got pulled up right now,

David:
which shows the total market cap of all AI agent tokens.

David:
Now, maybe to call a spade a spade, these are meme coins associated with AI agents.

David:
I think that's a fair take. But this sector of crypto, the AI sector of crypto,

David:
has gotten a lot of life back in it over the last two weeks.

David:
Ijaz, tell us what happened.

Ejaaz:
Yeah, so in this particular instance, life translates to money.

Ejaaz:
The total market cap of these ai agents if you remember back in the day i think

Ejaaz:
it peaked at like 25 billion and then it got crushed basically over subsequent months to about 3.2

Ejaaz:
billion dollars now over the last three weeks it has almost tripled oh just

Ejaaz:
over tripled actually to 11.4 billion dollars as of this screen recording right

Ejaaz:
now right and as you can see

Ejaaz:
charts are green across the majority of the spectrum of things.

Ejaaz:
And you're probably wondering, hey, guys, what's happened? Like,

Ejaaz:
has some like crazy innovations been going on? Like, give me the summary as

Ejaaz:
to why this is happening.

Ejaaz:
So let's go through some of the takeaways.

Ejaaz:
Number one, virtuals is leading this entire kind of progression,

Ejaaz:
right? So if you remember back in the day, we used to speak about kind of like

Ejaaz:
the top AI agent protocols. Virtuals was one of

Ejaaz:
Over the bear market, all three of them were crushed, but all three teams were

Ejaaz:
kind of heads down focusing on pioneering new kinds of innovations.

Ejaaz:
And virtuals really kind of like came through as the leader here.

Ejaaz:
So they specifically added over $600 million in market cap over the last two

Ejaaz:
weeks. And that's just in the virtuals token. So it's back over a billion dollars.

Ejaaz:
And they've made a number of changes since we last checked in on them.

Ejaaz:
So if you pull up this tweet, which basically kind of like has like this fun

Ejaaz:
little table as to what they've basically done.

Ejaaz:
So in January 2025, they go through things like agent revenue,

Ejaaz:
agent commerce protocol, fee structure, chain.

Ejaaz:
Basically, everything on the left was kind of unclear, a bit of flimsy, kind of Ponzi-like.

Ejaaz:
And on the right side, they have a much more structured product and focus strategically

Ejaaz:
for what they're trying to build and how they're supporting creators or agent

Ejaaz:
developers on their side of things. But let me explain what any of this means

Ejaaz:
and what the highlight is.

Ejaaz:
So they created a new agent launchpad. And for those of you who don't know what

Ejaaz:
an agent launchpad is, it's basically kind of like a

Ejaaz:
Cursor, if you like, where you can go on, you can use a no-code environment

Ejaaz:
to design what your agent might look like or what it might do,

Ejaaz:
what functions it might serve.

Ejaaz:
And then you can launch that agent. And usually you can associate a token with

Ejaaz:
that agent. So it's kind of like you're IPO-ing an agent.

Ejaaz:
And they created this thing called a Genesis launchpad, which is meant to be

Ejaaz:
a fairer way to launch agents.

Ejaaz:
So they've completely replaced the bonding curve model with something they call

Ejaaz:
proof of contribution, which basically means you need to do something helpful

Ejaaz:
towards the virtual ecosystem if you want to get access to buying the tokens

Ejaaz:
for that particular agent.

Ejaaz:
And contributions are measured in something called I think it's like virtual

Ejaaz:
points or they're kind of like loyalty points

Ejaaz:
and there are many ways to earn these points you can stake virtuals you can

Ejaaz:
yap or talk about virtuals on x or you can simply just hold virtuals tokens

Ejaaz:
or certain virtuals tokens

Ejaaz:
then there's kind of like this pre-sale and people can commit and and the

Ejaaz:
Of course, that's not the only thing that people are talking about.

Ejaaz:
They're talking about the returns that they're getting on this thing.

Ejaaz:
So the average return multiple on each of these tokens, David,

Ejaaz:
since they launched this launchpad is about a 12 to 20x, which is insane, right?

Ejaaz:
Until you realize that the FTV that you're investing in is like $212,000,

Ejaaz:
not million, $212,000. So you kind of have like a guaranteed ROI.

Ejaaz:
What I don't like about this is probably what you can already guess.

Ejaaz:
And if you pull up this tweet by Vader is it's mainly just about the money right now.

Ejaaz:
And it's not really about much innovation. And I know the virtuals team personally

Ejaaz:
are working very hard on creating value.

Ejaaz:
And you could also argue that on the traditional AI space that they haven't

Ejaaz:
kind of created anything of definitely more marginal value on the enterprise

Ejaaz:
side of things, but not as valuable as we'd expect to see. So it's kind of like,

Ejaaz:
you know, they're kind of fighting for the same kind of thing.

Ejaaz:
But it's good to see that they're kind of like focused on this thing,

Ejaaz:
right? The second thing is virtuals updated their fee.

Ejaaz:
Flywheel, which means that 70% of fees get given back to the creators.

Ejaaz:
Now, previously, that wasn't the case. And the effect that this has had is basically

Ejaaz:
any kind of agent developer that launches a coin can basically get continual

Ejaaz:
money to continue developing their product.

Ejaaz:
And that's been a really good way to kind of like sustain the ecosystem.

David:
So for the AI listeners out there, it's probably worth knowing that there's

David:
this meta happening in the crypto industry around these token launch pads.

David:
The meta that's going on this week is this like Believe app thing,

David:
which allows you to launch a token that is paired with, but other than that

David:
has no material association with kind of a vibe coded app.

David:
So you can vibe code an app and then you launch the app and then you launch

David:
a token with it. And it uses the token to gain virality for the app.

David:
You know, you attract some degens. If they play with your app and then they

David:
like your app, they might buy your token, which again has no material association

David:
with the app other than it's sharing the same brand.

David:
And so on one side of things, we have the AI agent part of this like crypto

David:
industry, which is downstream of virtuals, which is a token launchpad.

David:
And then there's the token launchpad side of things, which is the token engineering,

David:
the crypto degens, the speculative gambling that really turns people off to crypto.

David:
And on the spectrum of like, is this AI innovation or is this token launchpad

David:
financial engineering?

David:
Like, is this for the AI people or is this for the crypto degens?

David:
I'm placing this more on the crypto degens side of thing than any sort of like

David:
AI fundamentals. That's kind of my take about this.

Ejaaz:
Yeah, I think you're right. And I think it's important to just say that,

Ejaaz:
again, this is mostly about the money.

Ejaaz:
Most kind of media coverage on this or like updates that people get excited

Ejaaz:
about is the return multiple that they get, which is, of course,

Ejaaz:
one of the main reasons why people get in crypto. They want to make money,

Ejaaz:
but it's also kind of like it gets kind of like

Ejaaz:
boring or repetitive over time. And we want to kind of see some real pioneering

Ejaaz:
innovation that isn't just some kind of meme coin relationship.

Ejaaz:
Otherwise, you know, what are we doing here? We might as well just like talk

Ejaaz:
about gambling picks every week, right?

Ejaaz:
Moving on, talking about actually on the note of talking about some real fundamental

Ejaaz:
use, decentralized training or distributed training is actually an area where

Ejaaz:
Web3 or crypto innovators are actually

Ejaaz:
Bettering ai or pioneering ai so we've spoken about these companies that are

Ejaaz:
listed on this tweet here a bunch of times before but to give some context

Ejaaz:
In order to build these ai models you need a heck of a lot of compute and money

Ejaaz:
and data and it's very costly so some folks had a bright idea of thinking well

Ejaaz:
what if we do this in a decentralized distributed way

Ejaaz:
where we can aggregate compute from people's idle laptops, computers, cell phones,

Ejaaz:
and we can kind of like combine them into some kind of decentralized network

Ejaaz:
And use that compute to train models.

Ejaaz:
And so a bunch of teams were like, okay, cool, that sounds like a good idea.

Ejaaz:
But it was much easier said than done, because the way you need to network and

Ejaaz:
connect these different GPU servers and laptops and stuff is not an easy task at all.

Ejaaz:
In fact, by the end of 2022, the best minds at Google could only put together

Ejaaz:
a 400 million parameter model.

Ejaaz:
Fast forward to today and Jensen announced that they're using RL or reinforcement

Ejaaz:
learning to train a 72 billion parameter model, which I know isn't as big as 1.2 trillion, Josh.

Ejaaz:
I know it's not as big as like the DeepSeek team, but it shows massive improvement.

Ejaaz:
And this is being led by Web3 folks.

Ejaaz:
So kind of just not really going down the list, but just giving you an overlay here.

Ejaaz:
Teams like Jensen, teams like Prime Intellect, teams like News Research are

Ejaaz:
really pioneering open source distributed training.

Ejaaz:
And I don't think it's something that should be overlooked because if executed

Ejaaz:
well, these blockchain networks could be worth hundreds of billions one day.

Ejaaz:
And I'm not exaggerating because what cost or value would you give to networks

Ejaaz:
that train some of the best local or frontier models of our generation,

Ejaaz:
right? That can be co-owned by multiple people so i thought that was that was really cool

David:
Yeah, the idea of having a decentralized alternative to the very centralized

David:
powerhouses, open AIs, et cetera, is like attractive to me, just like kind of

David:
like on a redundancy perspective.

David:
Humanity at least has like a plan B, but I'm not sure if they can go toe to

David:
toe with the big boys. Josh, what's your take?

Josh:
Yeah, I agree. I'm still trying to wrap my head around it.

Josh:
I'm trying to get excited about it. I think it's an exciting development that

Josh:
is early. And again, like the rate of acceleration seems like it's going fairly well.

Josh:
The the initial take is that all of the largest companies in the world are trying

Josh:
to just gather all the information in the world and distill it into a model

Josh:
and we're seeing that closed source but we're also seeing that in open source

Josh:
with with meta and now meta has up to two trillion parameters on the next model

Josh:
and that's going to be open source

Josh:
open weight and i would imagine if people want to build something interesting

Josh:
from that they could take it and then distill that massive knowledge base down

Josh:
however they would like to kind of customize it into distilled models

Josh:
So I'm trying to just understand the difference between

Josh:
that, like taking an open source model and distilling it into whatever you want

Josh:
based on all this knowledge they've done the hard work to collecting versus

Josh:
going to collect your own

Josh:
data and use it in a decentralized way to build your own ground up model instead

Josh:
of kind of taking a top down approach.

Ejaaz:
Yeah, I completely agree. I think what this is ultimately going to result in

Ejaaz:
is niche use case, kind of bleeding edge or frontier edge models that serve

Ejaaz:
a small percentage of people, right? Maybe it's people that care about privacy

Ejaaz:
of data, or maybe it's people that care about using a certain amount of encrypted,

Ejaaz:
personalized information that they don't want to share with anyone, again,

Ejaaz:
from a safety aspect, to train a model that serves a particular function.

Ejaaz:
It comes back to what we were saying earlier. Like, if you don't have a good

Ejaaz:
enough product that comes from this AI thing, you're not going to make it.

Ejaaz:
And this isn't just unique to decentralized Web3 folk. This is also very prevalent

Ejaaz:
to the top dogs at the top. Like, we're talking about Microsoft, right?

Ejaaz:
What are their unique models beyond just getting direct access to OpenAI model weights, right?

Ejaaz:
If you don't have access to that, then you're kind of like nothing unless you

Ejaaz:
build an amazing product. So it's going to come down to the product.

Ejaaz:
It's going to come down to the founders.

Ejaaz:
I am excited that open source is still pushing really hard at this.

Ejaaz:
And I'm more than confident that some good kind of app program or software will

Ejaaz:
eventually come from this.

Ejaaz:
But you're in a very competitive mode and competitive industry that is being

Ejaaz:
funded equally so on the centralized side.

Ejaaz:
And at this point, you can't even argue that monopolies aren't open sourcing

Ejaaz:
because the folks at Meta, the folks at DeepSeek are all doing it anyway.

Ejaaz:
So let's see how this pans out and let's keep our heads in the game.

David:
Josh and Jaws, it's another week, a lot of news to cover. I'm sure in seven

David:
days, there'll be a lot more news to cover. And I look forward to talking to you guys then.

Ejaaz:
Yes, sir.

Josh:
Awesome. Another great week.

David:
Limitless. This is the Limitless podcast. Brand new feed. If you have not subscribed

David:
to the podcast or subscribed to the YouTube, please do so.

David:
If you are listening to this on podcast, please make sure to give us a five-star

David:
review so we can get this podcast up to the top of the charts.

David:
We think that this can be.

David:
And, you know, already is approaching in some of the best AI news-related content

David:
that you can find. If there are other AI podcasts out there that you guys listen

David:
to, let us know because we want to beat them.

David:
So yeah, give us a five-star review if you're on YouTube. Subscribe to us,

David:
like the video so we can grow this channel. We really appreciate you guys all

David:
here with us. And once again, we will see you guys in seven days.