Brex co-founder and CEO Pedro Franceschi believes most people still underestimate how much AI will change the way companies are built. AI isn't just another tool, it's a new foundation for building products, teams, and companies.
In this episode of Lightcone, Pedro shares why he thinks we're only months into a platform shift as significant as the invention of electricity, how AI has changed the way he works, and why every founder should be "token maxing" to understand the limits of the technology firsthand.
He explains why the CEO needs to be the chief AI officer, how Brex is rebuilding itself around AI, and why founders should rethink what's possible when intelligence is available on demand.
Chapters: 01:13 – How Pedro Became AI-Pilled 04:08 – The Electricity Analogy 05:21 – Free the Claw 06:56 – Making AI Safe for Enterprise 10:57 – Why Most Companies Are Behind 13:09 – AI Teammates, Not Chatbots 14:22 – The Case for Tokenmaxxing 18:24 – The Company of One 20:54 – The One Thing AI Can't Replace 28:06 – Building Customer World Models 32:58 – Rebuilding Brex Around AI 39:02 – The CEO Must Be the Chief AI Officer 43:50 – Building Company AGI 51:43 – Why We're Still So Early
AI Generated·May contain errors
You wake up, whatever problem you have in your life, why can't you solve it with AI?
And just start there. I think the CEO needs to be the chief AI officer.
It's not an engineering team thing, it's not a product team thing.
You have to understand the bounds of the technology better than anyone.
I think a good proxy for how to spend your time is,
what are things that only you can do and the models cannot do?
You have to sort of refound the very concept of what the company's self-identity is.
Welcome back to another episode of The Light Cone.
Today, we're joined by Pedro Franceschi, co-founder and CEO of Brex.
Pedro started Brex in the YC Winter 17 batch
and built it into one of the most important fintech companies of the last decade.
He's here today because Brex has gone deeper on AI than almost any enterprise company we know.
And Pedro's own AI setup is so compelling that when he came to YC for lunch,
it sent our entire team down a rabbit hole of building on their own.
So Pedro, welcome to The Light Cone.
Thanks for having me. Excited to be here.
Thanks for changing our lives.
Oh god, that lunch.
I'm like, I think the model company should be sponsoring me
for the token consumption increase we supposedly generated on that lunch.
That was a precursor to G-Brain, I guess.
I was still working on G-Stack.
I was still a 2013 Web 2.0 engineer who time-traveled instantly to the AI tools of January 2026.
And I was probably half a million lines of Rails code in,
and I created G-Stack because of that to help me make a software factory.
And then after I met you, I realized everything is about freeing the claw.
Free the claw, I knew you were going to say that.
Yeah, and then...
And give it tokens.
Well, no, I mean...
Let it rip.
The craziest thing was realizing what I had gotten wrong
that I think actually most people in software are still getting it wrong,
is they've been treating the LLM like this very precious thing that's very expensive.
And so as a result, you have to literally put the agent inside a Foxconn factory.
And it's like, can you imagine?
I mean, that's what the half a million lines of Rails code was for me.
It's like, no, no, no, I need to control what the LLM sees
because it's about really, really, I only want the context from here.
And let me write all the if statements to make sure, like a Foxconn engineer,
you're waking up at 6 a.m.
And if you don't, you're going to get electroshocked.
I mean, it's just this terrible thing that you do to agents.
And they want to be at the Esalen Institute, and that's what OpenClaw is.
Exactly.
And it's funny because I feel like every single good AI product you've used
is an agent loop with tools.
That's it.
You try to sort of over-engineer the harness and then do certain things.
But at the end of the day, it's skills, tools, and a model.
There's not really much else.
Maybe we start earlier because one of the things we'd love to kind of,
you know, get down as a part of lore is like,
how did you get so AI-pilled and like all the way to the edge?
Well, I'll tell you my encounter with LLMs, which was,
so I remember in the pandemic,
someone gave me an API access to GPT-3.
And I was playing with it, and I was like,
OK, this is really cool.
There's something here that could be special.
But it was the kind of thing that was like,
yeah, it feels like a research project,
the kind of thing that Google used to release,
and you play with it for 10 minutes and you stop.
ChatGPT came out, and I think everybody was sort of interested in it.
Where I think it got interesting was when you started to see
reasoning models and, of course, tools.
But I think everything else was sort of a blip until December.
And the way I describe it to my team is like,
you know, electricity was invented in December.
And I think electricity was Opus 1.5.
And sure, Opus models and, you know,
OpenAI models got better and better since then.
But to me, that was the tip of the spear where you could say,
yes, like coding harnesses actually work.
And, you know, Cloud Code existed for probably a year before,
but it wasn't that valuable yet.
And I remember, you know, during the holiday break,
I was playing with it and it was pretty shocking,
probably similar reaction that everybody here had.
And I think the question becomes, you know,
if you sort of if you think about, you know,
you're sort of standing, you know,
looking at 200 years of history,
and then you imagine you are, we're now in May,
you're sort of five or six months after electricity was invented.
And most people are still playing with candles and, you know,
questioning, you know, what can you do with candles and fire?
And, you know, like...
Who needs light?
Yeah, exactly.
It's not so bright.
What about these lanterns?
And what can you do with it?
And, you know, the steam engine is like, I don't know,
maybe like 20 years away still.
But, you know, electricity already exists.
That, to me, was the sort of the fundamental light behind it.
And I would say, I think since then,
OpenCloud was kind of an interesting sort of next step,
which is, I think, when we realized that, you know,
the reality is good AI products are agentic loops with tools.
And we started doing that in our own product at Brex,
but then on a personal side,
I started spending a lot of time understanding,
okay, what is at the frontier of using OpenClaw?
And I think the insight was just, yeah,
like markdowns can take you really far,
just like configuring and automating a lot of the things in your life.
It's kind of funny.
I remember I had this experience of like buying a movie ticket
entirely in OpenClaw using like a Brex card.
It was provisioned through an API.
And then I showed it to my team and they were like,
oh, but like, you can go online and like book it in 10 seconds.
And I'm like, that's not the point.
You're missing, you're completely missing the point.
But anyway, and then I went obviously very deep in this rabbit hole
and I started spending a lot of time thinking
how to change the fabric of the company
and the way we pull the products in.
Tell us about your personal OpenClaw journey,
because I, before you came for lunch,
I had it like in store,
but I was like way too scared to do anything with it.
We were all scared.
Yeah, don't get me wrong.
Like we deal with financial services data.
We spend a lot of time figuring out
how to be mindful of security and protection,
and yet I think people are a little bit more risk averse
than the technology probably requires them to be,
given where the technology is.
And when we started using OpenClaw personally,
I started doing it on a lot of my own personal setup.
Basically what I did in the V1 was
I'm going to give it read access to everything
and just create like OAuth tokens to my email,
to Slack and to everything to just literally not write.
And I was kind of shocked how far it got me.
And then the next question that we spent time on breaks was,
okay, like how do we actually get it
to write into our systems?
And everybody including our security team was,
well, we cannot do that for all the reasons that we know.
And then basically where I spent,
I don't know, probably four weeks of my time was,
okay, let's solve the hardest problem, which is security.
And we ended up realizing that the only way
to actually do something about it
was to do something in the network layer.
And if you treat the agent like,
the agent has its own wills, desires,
and they go to the Aslan Institute for agents,
and they have to-
Instead of Foxconn's factory.
Instead of Foxconn's factory,
they will try to do things at a network boundary
that could not be the right ones.
And we decided to actually just focus on that.
So a lot of folks were,
and we saw NVIDIA and others on NimbleClaw,
let's build these like open shell forks
that have controls over,
what tools the model calls.
And the reality is, yeah, you can do all that,
but you can also just make an HTTP request wrong.
So we focus on that layer.
And then we build this thing called CrabTrap,
which we open sourced probably about two months ago,
which is actually the way we use
to secure agents at Braxton production.
And the basic premise is you analyze,
you HTTP proxy the entire network boundary of an agent.
And the idea is when a request goes through,
that becomes auditable.
And you basically can use another agent
to analyze the traffic and create a policy
to let traffic go through or not.
And surprisingly or unsurprisingly,
because these models are trained
on hundreds of billions of web documents,
HTTP traffic is actually,
I would say probably the way the models reason
more so than anything else,
because they just literally learn on the web.
So the ability of the model to watch like a thousand requests
and make sense of what's happening
was way higher than we anticipated.
So we actually build that,
put that in production at Brax.
And after you record a traffic of an agent
operating for a day,
you can build a pretty good policy
that sets things that should be automatically approved.
And for things that the agent isn't really sure,
you can just use an LLM as a judge.
And the LLM determines,
is this request something that should be approved or not
based on the policy
for what that agent should be doing.
So for example,
we have like a recruiting agent at Brax called Jim.
We have a policy for Jim.
And all the traffic goes to that same policy
and 98% requests go through automatically,
2% use an LLM.
So we sort of got that problem solved
to a degree that we got comfortable
experimenting much more aggressively
and sort of freeing the clock on the enterprise,
which is really hard inside Capital One.
So I would say if we found a way
to experiment with these things,
and granted,
we don't do the most aggressive things of this stuff yet.
We don't use it on like customer data
to the degree that we want one day to do.
And there's boundaries to how we do it.
I don't see any reason why a YC company
shouldn't be at the bleeding edge of the stuff.
Yeah. I mean,
I think your intuition around like the proxy
at the network level
ended up being quite prescient.
Like I think a lot of the stuff that I'm seeing
kind of around the OpenClose ecosystem
at the moment,
at least, or just agent ecosystem
is essentially doing that.
Like we're seeing that like with credentials,
credential brokering,
like AgentVault is doing a lot of that.
I think you had mentioned the first version
of CrapTrap included like credentials vault.
Why did you decide not to include that?
I think it was just,
let's just do one thing really well.
And, you know, at the end of the day,
I think there's going to be a lot of solutions
that do that.
You could do credential brokering
and other tools already,
but the LM as a judge was,
for us, the determining capability to say,
do you trust us in production or not?
But our security team at Brex,
very rigorous and very good at what they do
for a long time were,
well, you know, not really.
So getting them to a,
yes, we actually believe this is enough
was a big unlock for us.
And look, I always say this,
we're not in the business of building HTTP proxies.
We are in the business of being the bleeding edge
of what we can do with AI.
And to get to the bleeding edge
required us to build this proxy.
That's why we did it.
Hopefully someone's going to build a YC company.
Hopefully we're just going to build a better version
and we're just going to go use it.
But at the end of the day,
that's the journey that took us to
just sort of being at the bleeding edge in that way.
And how much was you like sort of pushing this forward
and how much resistance did you get internally?
And just how did you get AI build?
I think there was a lot of excitement about it.
But the way I describe AI adoption inside most companies is
I think there's like sort of three tiers.
There's tier number one,
which is your token maxers,
like your engineers that are pushing a bunch of code
and typically living inside coding harnesses.
And those are sort of well-known.
We know who those are.
Then you have the sort of average engineer
that is building a few things,
but, you know, not so token maxer to the same degree.
I don't know, a tenth of the productivity.
And then you have like the entire rest of the company.
And the entire rest of the company
typically is interacting with AI
in what I call like Google search mode way,
which is a chatbot with a few MCPs
or G suite equivalent that you, yeah,
you have a few tools from Google,
but at the end of the day,
it's really just like a search.
And I think that our thesis was,
if you think about the value that AI creates
for like a token maxer, for example,
a lot of the value comes from the harness.
And the thesis was how to actually build
an equivalent harness for other teams
that are non-technical.
And our whole sort of thinking behind it was like,
that's a lot of what OpenCloud created,
which is this ability that you can self-bootstrap
a lot of the capabilities of the agent
by the way you edit your skills and markdowns
and sort of set up the environment around the agent.
And how far can we get this ability to,
for the agent to self-bootstrap a capability
without anyone actually going and coding it by hand.
So the analogy we use internally for,
I would say the sort of the company-wide adoption of AI
is we don't believe in the, yes,
like give people a few MCPs and let them go.
Because I think what people really want is,
in my opinion, is really a way of saying,
okay, this is actually a virtual employee almost
that has, you know, this on Slack,
it has an email, I can actually invite it to a meeting,
you can join a meeting, take notes.
And you're trying to replicate that as much as possible.
So how do you build the infrastructure
to support that kind of use case?
And I think the harnesses will look a little different,
probably more like OpenClaw than a coding model.
Jared and I just did this this week for the first time
where we installed AquaVoice
and then you open Telegram with the Claw,
or actually we have it in Slack now.
And then basically it was like me and Jared
and like three engineers
and someone from the events team.
And we're trying to put together,
how do we put together 60 dinners
with 20 people each of attendees from startup school
with 21 partners and visiting partners at YC.
Sounds like a great problem.
And then we just basically started talking about it.
And then I picked that up and I pressed enter.
And then, you know, our Claw just started doing it.
None of us opened Clawed Code.
Like it just sort of built a bunch of markdown.
It did the analysis.
And yeah, people forget that Clawed Code isn't magic.
It's just literally a harness around the same models
you can use in an API, right?
So I think that's the unlock of...
And by the way, there's a few things
that Clawed Code is doing that I think are really cool.
Oh, they're amazing.
And yes, it's just a harness.
And Claw can use Clawed Code.
Exactly, Claw can use it.
And Codex, right?
It really prefers to use Codex these days.
Exactly.
I don't know why.
Exactly.
But ACP helps on that.
So yeah, ACP is good.
Why do you think the adoption of token maxing
hasn't really taken off?
The thing that we've found very curious
working with a lot of startups early on
is a lot of founders are very shy about burning tokens.
I think you really get to experience this
when you really go all the way.
Gary mentioned this point, which is tokens are expensive.
And I think there are...
I'm in a fortunate position to be able to spend on tokens.
But I would say, I keep trying to picture myself.
Imagine if I was like 14 or 12 when I started coding for real.
And I had the technology we have now.
I would be token maxing in the cheapest way possible.
And there are people doing that.
You look at the Chinese models, for example.
They're pretty decent.
There's a huge hobbyist community
where they build a gaming rig.
But then they try to build local LLM.
And then that actually is a totally reasonable way to do it.
100%.
I have a friend that has the exact same setup.
He has this little GPU farm in his house.
And first time I went there, I was like,
wow, heating is on here.
It's really warm.
And he's like, no, no, it's my GPUs.
And I was like, great, power efficiency all the way through.
It's funny because at Brex,
we should talk about managing token costs
and spend management for tokens,
which is a topic we're spending a bunch of cycles on now.
I think the cost part is one.
But even if you take the cost part aside,
the first symptom is a lot more people
should be complaining about the max plan limits.
And you see what's the percentage of Twitter
that probably complains about it.
0.1%.
So I think we're probably still early.
To me, there's this AI pill test, in my opinion,
is whatever problem shows up in your life,
do you default to AI first or not?
It's like, of course, mechanically, you can do it.
But there's a point that it becomes second nature.
And then your whole brain gets rewired
and you cannot think in a different way.
And there's the whole topic about AI dependency,
human-machine interaction.
There's all these things that we can talk about
and put in a corner.
It still sort of surprises me how many people
you go talk to about a problem.
And I'm like, it's so cheap to intimately understand
the bounds of this problem now.
Why haven't you done that yet?
And come in with a much more digested view on the problem.
And I think the second thing is,
I think if you have the luxury of building a company now,
the fabric of the company from day one
can be built in such a different way
that I think, if I were to start a company today,
I would say, okay, the premise is,
why can't it be just me?
And then you start from there.
And your token consumption is probably going to be
a lot higher than if you said,
well, I'm going to have three people
or five people or seven people.
And I think the fundamental constraint
isn't as much, in my opinion,
AI as a cost savings,
or I'm going to be more efficient.
I think the unlock is,
the fabric of the company just looks very different
when the boundaries become type systems, interfaces,
agents talking to each other versus people.
And I think people still didn't fully grasp by,
okay, what does it mean to build code with new agents?
Like, and the new technologies we have,
I think that's like well understood.
But how to live in a world where intelligence is on a tap
and your default answer is,
let me actually solve this problem with AI first,
even if you feel suboptimal.
And then from there saying,
okay, how do I actually make it optimal?
Because I think for the majority of problems,
there is a way to solve it with AI that is probably better.
And your job is to figure that out,
even if it's going to take you more time,
because that will compound.
YC Startup School is back.
We're hand selecting the most promising builders in the world
and flying them out to San Francisco for July 25th and 26th
to discuss the cutting edge of tech and startups.
Apply now for your spot.
When you started Brex, I mean, like it's well known,
like you're like MVP, like had no web URI,
which is like all in terminal, like super scrappy.
Today would have because-
Yeah, that's what I'm curious about.
No one needs HTML, it's just test anymore.
Like, was it actually still the right approach?
Just have a really simple MVP and test that anyone worked?
Or would you have like a way more fully featured?
So I have this controversial view,
which maybe you all will disagree,
which is like, I actually think
if I look into a pattern of companies that succeed,
I think there's a really interesting pattern,
which is minimal surface area.
And the problem is with AI, I think you see,
like look at Stripe, for example.
Stripe early days was like literally an API.
Brex in the early days, no UI, just like literally a terminal.
You look at Airbnb, it's like the website was a form
and the form was just like literally where you put it,
what you needed.
And then someone somehow went there
and figured out how to actually make the booking happen.
DoorDash in the early days, similar, right?
Like it was just like literally...
So the surface area was so small with the customer
and so much of the sort of the intelligence
and the bandwidth of the founders
were spent nailing this one single interaction pattern.
And I think the risk with AI is that
the agency behind choice goes away.
So you have this lack of discipline
on what matters to solve.
And I think people tend to believe
that I can just experiment a lot of things
and that's absolutely true,
but that doesn't preclude you
from actually choosing what matters.
I always tell people, like, I think if you don't,
if you can't minimize your surface area
and solve the problem with a very clear set of boundaries,
you haven't found the right problem to solve.
And I think that's...
And you can, of course, find how to compress the problem
into a smaller surface area using AI
and that's really valuable.
But I don't think you should use it as an excuse
to not do that, which I think is,
well, I can just build so many other things,
but I always tell this to people,
intelligence is compression.
So when someone comes to pitch me an idea in the company,
I'm like, it has to fit in a napkin.
Great ideas fit in a napkin.
What's your napkin?
And then someone comes with this
and I'm like, I don't know where you buy napkins,
but the ones in my house are not this size.
How about the step before it then, even?
Actually, a lot of the pivot advice
I give founders during the batch
comes from you talking about
how you found the Brexit idea.
And the approximate view I remember
is that you thought about it as two week cycles
and you're either in exploration or exploitation mode
and you're trying a bunch of things,
but then you want to hone down.
Would you still use that pattern now?
100%. I think one of the hardest things
of building a company is talking to customers
and not just having the conversation,
but how to extract the unspoken signal
from these conversations.
And I think to me, can AI solve this lens?
Whatever problem shows up in your life,
can AI go solve that?
And you think about building a successful company,
why can't you prompt your way into that?
And the reason is very simple.
It's because there's signal
that the models weren't trained on.
And the signal is when you go talk to a person
and they tell you about a problem they have,
they're not going to give you the answer
shaped into a prompt that you can put into an LLM
and that LLM is going to go and output the product
that's going to win and be a billion dollar company.
They're going to tell you a very local optimum answer
based on their worldviews and their constraints
and the way they see things.
And I think a lot of the job now
is to have the wisdom to choose what you want.
And because before the wisdom was not just to choose,
was to choose and know how to execute it.
The execution is out, right?
The execution is gone
and the model is going to do that better.
The wisdom to choose is still, I think,
the missing bottleneck.
And to me, that all comes from which signals
are not in the models.
So say like pre-AI, you had personal bandwidth
to explore like three ideas in parallel.
You're saying like now in AI world,
you'd still do three in parallel?
Would you like 30 and let the models try and-
The way I would probably approach it is,
let's pick a broader universe of things
to do sort of an early initial exploration.
But to me, the lens is, okay, why can't AI solve it?
And which signal is not in the model?
And I think the signal is typically the customer.
And then when you go talk to the customer,
I think I wouldn't parallelize that probably.
I would be, okay, let's try to get into headspace
of this person.
And I think there's like, it's so easy.
And we did a lot of exploration
with like synthetic customers
and building customer role models and things like that.
And those are really valuable
once you know a lot about the customer.
But when you don't know enough yet,
I think there's this like very basic thing,
which is even at Brex, for example,
like one of the hardest things for us as a company was,
we initially sold to founders.
We're founders.
We knew about ourselves.
We knew about our problems.
And then as the company got bigger,
we were selling to finance teams.
And finance teams are different.
So building that mental model of like,
what's the value system?
Like, of course you can eventually make the model
represent that and have that worldview.
But there's an intangible that I think
is where a lot of the alpha still comes from.
And I think to me is like,
I think a good proxy for how to spend your time is,
what are things that only you can do?
And even in the company of one,
what are things that only you can do
and the models cannot do?
And that to me is like one of them.
I think that's so on point.
I think a lot of founders like you
that successfully navigated pivot have this loop.
Basically, there's this book,
Others in Mind from psychology
that has to do with people
that have very good emotional connection with people
are able to simulate what the other person is thinking
and what the others and others, theory of mind.
And I think the founders that get that
and have the empathy to figure out
what the customer is not verbalizing
is what is make the,
I think Gary says this,
make the implicit explicit.
A hundred percent.
Of what are all those desires.
A hundred percent.
And they're very subtle signs a lot of time
because they're murmurs
as founders go through them
and figure out the insights that,
oh, is this really a thing?
But how do you know when to poke for it?
Exactly.
And the problem with relying on models
and right now, which is,
I'm still very optimistic
that there's still a lot of jobs for founders.
Definitely.
Is you don't even know what the right incantation
or set of prompts to ask the model
because you don't even know what to ask.
Exactly.
There's like another meta layer.
Yes.
It's the whole like Elon thing of like,
you know,
which question is the university answer for kind of.
And of course these are generalities, right?
But I think what I've seen is
you have to remember that LMs are not magic.
Like LMs are trained
on a very specific corpse of information,
optimizing for a very specific set of benchmarks
and outcomes.
And I think the biggest pitfall of LMs is
you have no sense of how much training data
the model has seen
for the exact thing that you're asking it.
So imagine if like,
every time you asked an LLM a question,
it gave you like,
yeah, like the sampling frequency of this in my dataset
was, I don't know,
X.
And on this other answer was 0.00001X.
You would trust it's very different, right?
The distribution is so different.
Oh, I would pay for that.
That's a great startup idea.
Exactly.
Someone should do that.
We need to do a model that does that.
Yeah, 100%.
I would pay for it.
Yeah.
Well, because it's fascinating
because then like anything that's out of distribution,
you just go and like fill that in.
Exactly.
I mean, actually as an applications engineer
on top of the LLMs,
that's actually a huge blind spot.
And that's what Mercore
and a lot of the other data companies are doing.
Like a lot of the jobs for them is to say,
well, what are the blind spots for LLMs?
And it's funny,
I think a lot of the data labeling companies right now
trying to understand the pitfalls in the models.
But the problem is in order to do that,
you have to be an expert
to know what the gaps are in the answers.
But the problem is as a founder,
when you're looking for an idea
is you know nothing about it.
So there is a curse of knowledge
and a curse of not even knowing
what the bounds of knowledge is,
which I think can make you believe
that you understand something
that you actually,
you in order to model actually understand.
Can I confess something weird about like,
after creating Gbrain,
now I do use AI in a different way
where now that I have a retrieval system
that is actually usable,
if I have a problem or question about anything,
like for instance,
I was trying to work on a really, really,
like the last humanized prompt.
And a lot of that stuff
probably isn't in distribution yet.
There's a whole Wikipedia article
about characteristics of AI writing.
But now I can just go tell it,
like go spend a day,
like deep research,
literally every single paper article,
like read everything,
put it into my Git repo,
and then I'll be able to retrieve it
and summarize it into something
that actually is usable.
And so that's sort of like filling in
the things that are out of distribution,
like I can sort of like pack it
with whatever context.
And it's like, you can do that with anything.
It's like, if you're interested in,
you know, running a restaurant,
literally you could have,
you could go and read like 500 books
about like every,
the top 500 books
about what it's like to run a restaurant.
And you would have like the compendium
of all information about it.
Yeah. And I think a lot of what,
like, for example,
like one of the things that we do at Braggs now
is building this customer world model
is a similar idea
where we're trying to get
every single touch point
that the customer has of us,
like literally like what,
how many times they click a button
in the dashboard,
all the way to where they tell someone
on an email
or what they say on the phone
or they send a call
and ingest that unsolved data.
Okay, what should this customer
need next from us?
What should this customer be thinking about?
Like, what are the issues that
they will face but haven't faced?
And again, it's just a distribution problem.
This is actually an answer
to one of the questions,
which is like,
will there be jobs or whatever?
It's like, as long as there are limits on RAM,
actually like there will be.
So I don't know.
I mean, that is kind of an interesting one, right?
I think so.
Literally, you can't have a model
that has enough parameters
that could like have everything
that you could possibly need in distribution.
Like there aren't enough atoms
in the universe, right?
It's like a modeling problem.
I think we forget
that the world models
in which the models are trained,
like there is something
that the designers of the models
influence the way the model
actually behaves in the end.
So, you know,
one of the things that we spend
a lot of time thinking is like
how to make LLMs work
for people that look very different from us.
How to make LLMs work
for like the average finance person
in the U.S.
that we're talking about an answer
and, you know, the model defaults
to like AI capex as the finance
as a default category for like,
like, for example,
that's a really funny example.
Like I was playing with AI
for accounting categorization
and like the first example
of like an example of it
is just like writing prose
and an example is like AI capex.
And I'm like, oh, why is it AI capex?
The first example comes up with
because the people that are building
the models only think about AI capex, right?
So there are things like that
that I think is like kind of
interesting to think about
that the mental models of the models,
I think are out of the box
are more biased
than we may give them credit for.
I mean, speaking of AI capex,
like earlier you're saying,
you know, we're so early still.
I don't know.
The funniest thing about AI to me
is how often I find myself
thinking crypto maxims.
Yes.
This is the worst
the models will ever be.
Yes.
My favorite now is telling people
who hate AI coding,
like have fun coding at 1x speed.
Exactly.
I was telling a friend about,
you know, how to be,
you know, long inference
that basically the thesis
that there's going to be
a lot more inference than people think.
And we're expecting a lot of inference
if you just look at public markets
and, you know,
semi supply chain, all that.
People are saying like 10,000x.
Yeah.
But the underwriting,
which is kind of funny,
is I think there's one image,
2,500 dots.
Each dot is a 3.2 million people
on the planet.
And basically, you know,
84% of the world never used AI.
16% have used at least once
a free chatbot.
Then 0.3%,
which is, I guess,
six or seven squares,
paid 20 bucks a month for AI.
And one box out of 2,500
actually use agents
in whatever capacity.
So that's the argument
to be long inference.
And I think it's just starting out.
And I think a funny thing on this
is I think the,
you know,
it will be the biggest expense
in a company like easily.
Right.
And yes,
there's a lot of margin
in tokens right now,
but people always want to be
the bleeding edge.
But even token costs
decreased by 10x,
they're going to have 10x more usage.
So it will be still a large cost.
And we're spending a lot of time
thinking how to help companies
actually manage token spend.
On Brex,
we ended up building
our internal version of this.
We call it Magbuy,
where the idea is you can
effectively,
you know,
every dollar of token spend
in the company,
you can attribute to
a product we have to customers,
an internal tool that we use to serve,
or an internal employee,
and understand model usage, etc.
And we're now figuring out
how to do the analytics
on what are we trying to do
with the tokens
to start to get a sense of ROI.
But anyway,
it's a fascinating topic
that I think has a lot of early,
early, early work
compared to what it will be one day.
Can you share any of the data
that you've gotten from Brex
about just like what token spend
is like in the economy?
It's increasing.
No, look,
I think two things are surprising.
One is,
I think to your point earlier,
on how do we,
how do we look at token maxing?
I do think there's such a thing
as how much cost boundaries
you create internally
dictate token consumption, obviously.
But to me,
I think what's the most fascinating
is when you look into
the sort of 10 mile radius we're in now,
and maybe you include New York,
tons of token consumption.
And you could probably argue,
and we see in the data,
also faster revenue growth.
I think what's really interesting
is the gap between
anyone in these two 10 mile radius
and everything else.
And this is like not small companies
that you look into,
like very large companies
with very large budgets
and that could be token maxing.
And the economic thing for them
to do would be to token max.
And they spend like,
I don't know, 10,000 a month.
And you're like,
you should probably be spending
10 times more or 20 times more
or 100 times more.
That's still surprising.
And I think the reason is, again,
sort of similar to the point
in the beginning,
like we did this exercise
two and a half years ago
where I sat down with,
you know,
a lot of the engineering product
leaders in the company.
And we had this question,
which is if we started Brex again in 2024,
the answer would be
even more different now.
What would we do differently?
And turns out like everything.
And we start going down this route
and it's like,
it's kind of maddening
because you're like,
OK, we have this like
completely old way of like
even thinking about
the fabric of the company
and the way we build the product
and the way we build
our processes internally.
The first best answer is,
yes, we wish we had started now.
Second best answer is like,
let's go do something about it
and change the way we do things.
Right.
And I think a lot of our approach
in terms of like adopting AI
has also been,
you know, how do you,
how do you pause and say,
OK, like there is a discontinuity
in the not just in
how we solve the problem,
but on what the definition
of the problem actually even is
and sort of take a step back
and rethink it.
And, you know, like there's
there's like millions
of examples of that.
But, you know, one example,
which is kind of funny is,
you know, we're redesigning
our KYC process.
Like whenever we onboard a customer,
we have to do all these checks
to KYC the customer.
And KYC historically
is something that you can automate
like 80 percent of it,
20 percent is manual.
And of course,
the original impetus for anyone
is that it's with an agent that does it.
Yes, we can go do that.
But what we decided to do
is actually say let's redesign
the entire process end to end.
And then what we redesigned
is the entire onboarding process.
And when you redesign
the entire onboarding process,
what you realize is
there's a very important thing
that happens in the beginning
of the funnel,
which is deal qualification.
Like, is this customer
even remotely qualified
to be a Brex customer?
But when you have KYC for free,
you can KYC a lead
versus the customer.
So you start to have
risk orientation up in your funnel.
And that changes who you even target
because you know who's going to qualify.
And the same thing is true
for credit to some degree.
So now the bounds of the problem
have changed.
And you can go in and say,
and I think a lot of all,
including a lot of our competitors,
have this approach of saying,
oh, I have this entire old process.
Let me go and latch on AI on top of it
or latch on AI on top of our product.
And I think the biggest discontinuities
in a positive way that we've had
were when we said, hey,
let's keep this old way here.
Keep put it in a corner.
And how would we design it
if we started the company today from scratch?
And then just doing that.
It takes a little bit of founder energy to do that.
But I think it's the only thing
we've seen working to really sort of inflect.
I think that reminds me a lot about this
is sort of way back.
I don't know if you ever tried to compile
Arc distributions in Linux.
The culture within power uses of Arc Linux
versus Ubuntu is very different.
Very different.
I think that Ubuntu people
kind of feel more like people
that try chat GPT.
Stuff kind of just works out of the box
or some stuff that you can get up and running.
There's still not a lot of people
that use Linux, by the way,
which I think it feels where AI is.
But with Arc, you're like super hardcore.
And I think that's what
OpenClaw and Hermes feel like
because you have to really customize it
to your own use case,
maintain your skills,
have all the markdowns.
And if you get it working,
you can build something awesome.
One of the most impressive thing
I've seen people build with Arc
is actually, I don't know if you know,
Valve, the Steam engine.
The operating system that runs,
that makes it feel like a Nintendo Switch
is actually built on top of Arc.
Oh, interesting.
They customize all the drivers,
over-the-air updates.
It works with all consoles.
It works with all sorts of hardware out of the box.
But they super duper customize it.
And I think this is kind of what's happening.
If you get your OpenClaw
to work really well for you,
you can kind of build your own
custom Nintendo Switch for whatever you need to do.
Yeah, I always have this thing that I tell people,
which is funny,
which is, think about your time two years ago.
I feel like you're working a lot more now
than two years ago, right?
And probably the same for everybody here.
So then the argument is,
but what about the productivity?
Where's the productivity?
And I was talking to the CFO
of a very large public company this week.
And she was telling me that
we see all this token consumption
and we're trying to measure
product velocity.
And we're seeing more lines of code pushed.
So yes, maybe that's the way to measure the ROI.
But is it really there?
Because people are spending so much on tokens.
And I think this analysis,
yes, of course,
I think having a sense on ROI on tokens is important.
But I think it misses the point
that you're standing in the timeline of history
and it's six months after electricity was invented.
Imagine someone saying in the 1800s,
like, oh, my electricity bill is so high now.
Like, gosh, let's use a little less.
Let's push the steam engine
to come maybe 20 years later
because the cost savings.
Yes, of course.
Don't bankrupt your company on tokens.
It's actually a perfect analogy
because I don't know if you know this,
but when electricity was first invented,
it didn't work very well.
And the ROI was actually bad.
And so shortly after the invention of electricity,
some accountants had done this analysis.
They would have been like,
this electricity thing is like,
is it never going to be a thing?
The ROI sucks.
Why do people stick to it?
And it wasn't the cost savings.
It was just because people were curious about it.
And I think the point of why I was yesterday
until 2 a.m. playing with slash workflows
and Opus 4.8 and all that
is because I think I'd be doing the exact same thing
if I wasn't making any money
because you just see the possibilities
and you see what it can do to technology.
And that just drives people to behave differently.
And I think that to me is the ultimate litmus test.
And it's a good separator.
And sure, if tokens are so expensive,
they're going to be, I think,
over the fullness of time, probably free.
If you project this, I don't know, 100 years online,
almost compared to what electricity,
now we don't think of electricity costs in our day-to-days,
but unless you're in a data center.
But I think there's something similar for sure.
We talked to a lot of founders of later-stage companies
who wish that their companies
could be as AI-pilled as possible.
And you run this big company now
with all of these employees.
And that's only the Breck side.
There's also the Capital One side.
I'm curious what you've done
to bring the rest of the company
along with you on this journey
and if you have advice for other people,
for other CEOs.
There's a lot to do.
I think the CEO needs to be the chief AI officer.
It's not an engineering team thing.
It's not a product team thing.
It's like you have to understand the bounds
of the technology better than anyone.
I would argue that unless you really experience
the limits of technology every day,
I think it's really hard to even understand
what it can possibly do.
Oh, you know why?
It's because nobody can say no to the CEO,
except the board.
And the board won't be in the weeds, per se.
That is 100% true.
When you go think about the whole example
of KYC that we were saying,
the KYC team would never think
of using the KYC technology to score a lead.
The only people that can think
about the organization of the system itself
is if you have the context of the whole.
And to me, the single most important question
that any CEO needs to answer
is forget about the competitive landscape.
Imagine you could get the state
of the technology today
and transport to the moment
you started your company.
The opportunity was still the same,
but just the possibilities of the way
to build a company are totally different.
How would you do it?
And then diff this versus what you have.
And then first suffer in silence
for a little bit because you will.
I mean, I do every day.
But then the second thing is,
okay, what do you do about it?
And how would you do it
if you were starting from scratch?
You'll be the one figuring out,
okay, how do we design our onboarding process
or how we design our growth engine
and our customer acquisition?
And the way we talk to users
and the way you synthesize the data?
And all of that would be
redesigned from scratch.
So I think it's almost like
you have to sort of refound
the very concept of what
the company self-identity is
and the way the functions
and people's sense of success
get structured.
AI is an umbrella that I think
has like three things,
the way we talk about it internally.
There's product AI,
the product actually ships to customers.
There's operational AI,
which is things that directly affect
our ability to serve customers at scale.
Think of customer success,
risk, onboarding operations, et cetera.
And then there's corporate AI,
which is how people work internally.
The three agendas matter
and they matter in different ways
depending on the timing of the company.
And I think people will sometimes
sort of pigeonhole themselves
in one of the three.
But in reality,
I think you have to take a step back
and be like the same thing
we were talking about earlier.
Why can't you solve everything with AI?
At a limit, that's the question.
And then sort of start from there
and start problem solving around that question.
It's a turnaround almost.
I think you have to assume
that if you're a big, large company
that's not AI native,
you're doing a turnaround to some degree.
I guess we've been making fun
of Foxconn factories for some time.
But on the other hand,
if you look at them,
they're like this paragon
of very extreme efficiency.
But they also are designed to be that,
to create one thing perfectly
back to back to back.
And so you have to build a factory like that.
And most companies are designed that way.
I think processes are designed not to change.
There is a certain amount
of broken glass required.
The question is,
I think it's 10x easier for the CEO
to break glass than an executive.
And 10x easier for an executive
than an employee.
So a lot of times,
someone comes to me and says,
I'm trying to do this with AI,
but someone is saying no
because we haven't tested this
in this use case or in that thing.
And I'm like, okay, what are you trying to do?
Do you understand the risks?
Do you understand the guardrails?
Yes. Okay.
It takes me literally 10 seconds
to solve that problem.
And it would take someone 10 hours
to go in into the meetings
and escalate and understand,
okay, can we do this with AI?
Or maybe never.
And I think the conclusion is probably never
because most people would say,
you know what?
I'm just going to build this product
in the old way
because why wouldn't we?
It just works.
We know it's here.
That guy's going to hate me.
And then I have to look at that person
in the lunch line every day.
And it's like,
I want people to be happy and like me.
So I'm just not going to do that.
And what I tell people is,
I think the escalation paths
need to be desensitized in the system
because the company builds antibodies
against any sort of disturbance
to the social cohesion of the company
typically gets rejected by the antibodies.
And I think making escalations faster
and being like,
hey, we're going to go try this thing.
I understand the risks.
Let's take this risk
because the biggest risk is not taking that.
It is just literally missing
the opportunity to rethink a problem
from what would you do
if you started the company today?
On the corporate AI leg of that stool specifically,
do you buy into the Jack Dorsey view
of every company is essentially
trying to build its own little company AGI?
I do, but maybe in a slightly different way.
I do think domain specificity matters.
So I don't believe in the like,
oh, I'm going to have a single company model
that has every piece of data
in a single, with no judgment
or lens into anything.
So, and the way I think about it more
is like, it's more of the sort of
the virtual employee analogy, so to speak,
which is like, how do I build an agent
or virtual employee that is exceptional
at understanding everything
that matters about this customer?
That is a well-defined problem
with clear boundaries of like clear APIs
of who people, who depends on the data,
who interacted the data that is self-contained.
Then there's another agent that can be,
okay, given all the customers that we have
and the problems they have,
how do I manage my product roadmap?
That can be a separate agent,
but that builds on top of this customer world.
Like a virtual exec team, basically.
Exactly.
Functional and domain knowledge still matter, right?
These things are not going to go away.
And I think the way knowledge is structured,
I think is still true, right?
It doesn't necessarily change that much.
And you should separate the agent
and the systems that are actually emitting code
from the system that is talking to customers
and the system that is reasoning
about the conversations of customers
and translating into product roadmap.
Three separate things.
We're kind of like the Tesla for AI.
We're like, I don't believe in anything
that doesn't have real usage.
So it's like, yeah, I build this great model
and I'm like, okay, how many people are using it?
Is it actually displacing the need
to hire a person inside a company?
Is it actually displacing the need to,
you know, spend literally hours?
Like how many hours is this thing saving?
And I think a lot of times people say,
well, you know, it's a cool model.
And I'm like, yeah, but like,
that's not going to cut it, right?
Once you have that orientation,
I think customer role model.
For example, our client sales team
now runs on our customer role model.
So I know it works.
I'm actually having lunch of a customer tomorrow.
And I don't know the state of that account
as well as I probably should.
Customer role model answered the question for me.
And I now have a report with,
including things that the team didn't know about
that came through support tickets.
And, you know, an executive that was traveling
had an issue at an airport with their car.
Total information awareness, right?
That is a well-defined problem.
That is working.
I can trust this building block
as part of my company model as a whole.
And you can have evals on it.
Like we know, like, you know,
I think a very, we should talk about evals.
There's a bunch of learnings on this
and how to build evals into the fabric of a company.
But anyway, I think it's more of like,
you have to decompose the problem a little bit.
Yeah, my favorite thing about evals
is just running cross-modal evals against each other.
So one of the things that we're doing
is related, but I think it's really fun,
which is how do you have every single human interaction
in the company becoming an eval
when you have an AI agent?
So for example, we have the onboarding agents
doing something.
And then you have a team that actually goes in
and looks at KYC exceptions
that the model can't figure out.
How to make that a breaking change.
And, okay, like this manual interaction
will become an eval case.
You know, we have an expense agent in Brex.
Whenever someone has a conversation with the agents
that flags an issue or a bug
or something that feels like
the conversation didn't go as smoothly,
that creates a bug.
That bug triggers an agent
that's going to go and modify the code base
and the prompts and everything
to make that eval pass.
And if that doesn't break,
that engineer is going to go in
and figure out how to make that pass.
Because the goal at the end, I think,
is to make the whole thing a self-learning system, right?
And I think a lot of what I see with companies
is they spend a lot of time getting an agent working,
but never thinking
how to make the agent improve every day.
And I think that's like always the biggest unlock.
You need a dream cycle.
You need a dream cycle, exactly.
The dream cycle sees everything every night.
Exactly.
And then it's like, oh, what's going on there?
I need to put this over here.
Exactly.
What actually happened?
Is there a pattern?
How do I cause this?
So how to bake the dream cycle
into the products and into the agents
and into the things that ship.
My favorite thing right now
is I'm building three or four agents for my friends.
Oh, interesting.
And some of it is like,
this is a user research for me for Gbrain,
because I have one.
It's working really well.
I have 350,000 markdown pages in there now.
What a crazy...
I thought it was this wild pie-in-the-sky thing.
And it's going to happen in our lifetimes.
You know, I remember when Neuralink came out
and I used to think about it.
I was like, I don't get it.
I was like, yeah, of course I get it conceptually,
but why is it a thing?
And then now I use AI and you're like, yeah.
Yeah, makes sense.
I'm the ball rack.
Typing is so slow.
I don't know if you use a lot of dictation.
I use a lot.
My most used developer UI right now
is like voice memos to OpenClaw.
I've said this before, like it was maybe accidental,
but I actually just really love the fact that
like Telegram works so well with OpenClaw
because it's forced me to just put more stuff,
like make the agent more intelligent
so that you can do more stuff via voice memos.
Because you have to sort of fight the natural instinct
as like a traditional developer where you're like,
oh, like I can't quite do this or doesn't do this.
So I need to go like build more client or more UI
or take more functionality for it.
Foxconn.
Yeah, exactly.
Just let it do what it wants to do.
Give it some context and it'll just think about,
you know, oh, actually, what about this?
I think a lot of the work to your point is
how do you organize the context for the model?
And you can use the model to help,
but that is the bottleneck for most things.
Once you have the context in there,
it's actually, you can do some pretty crazy stuff.
Like my favorite new feature.
I saw you doing LSD.
Yeah, LSD, yeah.
Lateral Synaptic Drift.
So you just bump the temperature on the search?
It's not just that.
So you have the vectors, right?
And so, you know, if you think about
what conventional ideas are,
like most people give you like,
oh, well, this idea with this idea,
and it's like kind of like in this cone,
LSD mode actually says you cannot combine concepts
that are within this cone.
They actually must be orthogonal
or just like seemingly random.
And then it'll try like, you know,
randomly hundreds of these combinations.
And then it'll rank order them into the ones
that are actually the most coherent.
And then if you do like a hundred of them,
actually like the top five tend to be banger tweets.
You know, what's crazy is like,
I didn't tell Alfred explicitly to be dry.
I went, I actually had like chat GPT generate
like the soul file.
And it was like, based on like everything,
you know, about me, all the interactions here,
like generate like a soul MD for like my OpenClaw agent.
And it was so unerringly like accurate
about like kind of what I would want from like an agent.
And I was like, oh damn, these models know a lot about us.
And my OpenClaw got really interesting.
Once I just ingested my 60 gig Google takeout.
I mean, you have to write a bunch of haiku code
to like only get the emails that are actually real.
But you know, there's like,
it extracted like 4,000 emails out of 60 gigs
that actually matter.
But like, those are like,
oh, actually like a lot of your thinking
and you know, the consequential moments of your life.
So Pedro, thank you so much for being with us.
I mean, you're by far one of the most AI-pilled,
farthest out on the edge,
but also very practical CEOs who is,
you know, playing with this stuff
and actually building it yourself.
What would you say to people watching who are founders,
who want to be founders?
You know, I think that you are sort of the model
for the way people should start companies
and run them with AI as your SLN buddy.
Yeah, I really can't stop thinking
about the electricity analogy,
which is, you know, you're standing,
there's a 200-year timeline of human history.
There's a point in time where electricity was invented.
It sucked in the beginning
or six months after that point.
What do you do differently?
Knowing everything that will be true about electricity,
knowing the data centers
when they will consume electricity and even AI, right?
Well, you do a lot of things differently, I think.
So I think that's one of just marveling
at the possibility of the exact moment in time we're now.
I think the second is like,
have a post-it on your computer,
which is, you wake up,
whatever problem you have in your life,
why can't you solve it with AI?
And just like start there.
And 80%, yeah, you can use a chatbot,
but the 20% that you can't,
figure out why and go build something
that makes you solve that problem.
Less so because of the immediate usefulness
that solving that thing at scale will have,
because it gives you a texture and a feel
for the possibilities of the technology,
which are really hard
if you're not playing with it every day.
And maybe the third thing is like,
I think it's like just measure your total consumption
and how much you're just pushing the limits of the company
and starting at the premise of like,
okay, why can't it just be one person?
Like, why can't it just be me that builds the whole thing?
And you're going to probably face a wall of the elements
of what models can and cannot do.
But at a limit, I think the question is,
how do you spend your time on the things
that only you can do as a founder?
And these things to me are,
number one, which problems are worth solving?
And two, and sort of the choice thing we talked about.
And the second thing is,
okay, given these choices,
what are the limitations of an LLM
that they still cannot do?
And I have to go in and do those things myself.
But almost to some degree,
you're working for the LLM to some point.
And if you're in a bigger company,
you're in a turnaround to put the LLM
as almost a founder and a CEO.
And you're almost architecting the entire company
around that idea.
But I think early on,
so much of it is choosing what matters,
talking to customers,
injecting the signal that the models don't have,
and just rebuilding it the way you would do it in 2026
with electricity being six months old.
Thanks, Pedro.
That's awesome.
Yeah.
Thanks for having me.
Appreciate it.