Projects
![]()
No projects found
Investors
New
DAO Treasuries
DATs
Lend / Borrow APRs
Chains
Categories
Spaces
SKALE School: Agents & Finance, What’s Real vs What’s Fluff?
Recorded: March 31, 2026 Duration: 0:54:08Space Recording
Full Transcription
gentlemen we are back with another episode of scale school and it is a special one today
it's a special one always as we have some incredible guests but before we jump over
to him so yeah how you doing man i'm doing good i'm doing good i have not had my coffee yet though
i'm in the process of drinking it so i'm'm still a little bit waking up in a sense.
But we've had a really good 24 hours to start the week.
We've launched doc updates.
There's multiple new SDKs we released last week,
new one coming today for a plugin for Therimeter, most likely.
And we've got a new release coming soon. Oh, cool.
And it may or may not have something to do with confidential privacy actions on chain,
right? Being able to do some cool stuff there, which probably aligns quite well with the most
exciting thing that's happened to me personally this week, which is getting joined by Justin.
Because we don't get to talk that much these days. We're both way too busy. But every time we do,
it's a very illuminating conversation for me. I feel like he teaches me more in like 10 minutes
than I learn in a whole month. So you don't know Justin, we'll let him introduce yourself, but you're in for a
treat today. 100%. I appreciate you blazing. Yeah, Justin, been around for a long time. If you don't
know me, started a company in 2020 in DeFi. We started managing hundreds of millions slash billions in 2021 to 2022.
Got pretty wrecked by the credit crisis in 2022, post-LUNA, 3AC, Genesis, whatever,
and decided, hey, we're just going to reinvest into the business and started incubating what later became the first agentic wallet and framework for crypto.
So in 2022, 2023, we were way too early to the game, frankly, but we were whipped around AI agents that were trading and managing DeFi portfolios, etc.
DeFi portfolios, et cetera, using our stack.
That was called Moon.
And we worked with scale way back in the day on Moon stuff with a couple other early AI agent teams.
And that's how we got to know each other.
And, you know, since that time, you know, our work in DeFi has kind of slowed down to now close to zero.
And our work in AI has been ramping up quite a bit.
And now we're launching our AI lab called UV.
And we're doing a lot of awesome stuff.
We trained a hyperliquid LLM.
We're training some really cool vision models for finance.
We're scaling up a lot of our work.
And all of it is centered around the data that we generate on our trading platform.
So quite a bit. We've been around for quite a while.
You've never heard of me and yet, you know, we've got infinite things to talk about. So I'll let you guys kind of lead the way and I'll try to add some color where I can.
100%, 100%. And, you know, you did my favorite pun, you know, scaling. But pun you know scaling um but you know let's jump into let's jump into everything
here um obviously massive buzz so many people talking about ai agents how to make money with
them trading running financial workflows um i guess the big question here is and i'm not sure
from you guys but i have always whenever i've seen somebody
talk about it two days later they say my my ai agent has just put all my money into a honeypot
right so is this real or are we too early are we over hyping it where are we
Where are we?
Yeah, I mean, it really has everything to do with the harness.
You know, so when you think about an AI agent, there are two parts.
There's the model, which is providing the engine, the intelligence.
And then there's the harness, which is like the middleware, which is taking that intelligence and shaping it into something useful.
You know, and the harness is oftentimes much larger and developing the harness is a much bigger effort in a lot of ways than developing the model.
You know, we've been working on our harness for three, four years now.
And we're working with models that took, you know,
six months to train. And obviously the research leading up to that was enormous. But, you know,
when you see models making unforced errors, so to speak, a lot of the times that's just because
they don't have good policy management at the middleware level. and we'll get into training and how you can kind of uh compile all of that
um middleware into the model itself that's kind of what we're doing for finance at uv
but the long story short is uh you know if you see an agent doing that probably just avoid uh
that agent and that project for a while while they get their stuff in order.
you know that agent and and that project for a while while they get their stuff in order
Love it.
So, you know, part of what we, if you're new to this show,
obviously, Justin, I know you're obviously new here.
I don't know if you've seen any of these before, but if not, it's all good.
But one of the things that we try to do as we go through
is we try to break down some of the more complicated topics
that are,
well, we hear a lot, but most people actually don't know what they mean, right? And it's not a
negative, right? We're here to help people learn about why UV and YScale are functionally leaders
in where this industry is going. And so I want to pull out a very specific word you just said because there's been a lot of fighting over it
over the last four weeks,
which I know is like very short term
compared to a lot of the work you guys have been doing.
But the term harness, an agent harness,
I feel like we've been in like the argumentative phase
of what that actually means in AI over the last, like I said, last month or so, as everyone decides what we actually want to call the part that wraps the LLM and handles all of the tools and the middleware and managing the end-to-end lifecycle.
end-to-end life cycle. And so for those of you that are unfamiliar with this term, the harness
is essentially kind of what runs the agent. Like essentially the harness kind of, is it fair,
Justin, to say the harness essentially is the agent without the harness, there really is no
agent. So they're very tightly coupled. Yeah. Think of the model as the engine in a car and the harness is everything else.
Love it.
That's beautiful.
That's beautiful.
We need to save that analogy for later.
And so what's really interesting, I think, is you guys, you said, you know, you've been working on your harness for four years now, give or take, right?
four years now, give or take. And it feels like to me, the reason that you kind of say that is
because you, like many other teams, we're kind of building something else. AI happened, the chat
chiquity moment happens, the cloud code moment happens. We kind of see these huge exponential
accelerations of AI adoption. And what you previously built is
the core of what you need to be very successful, or what we believe you need to be very successful
within using more of an agentic flow as opposed to maybe a human-operative flow. That being said,
you guys were actually one of the first that I had seen to, maybe the first to experiment with
AI agents. You guys actually were doing agentic flows back in,
yeah, that was what, 20,
end of 22 or really 23 maybe?
In the crypto market,
we were the first to do it by quite some time.
First in production by, I think, over a year, yeah.
Love it. Love it. So I think that's fantastic. I'm curious,
Love it, love it.
So I think that's fantastic.
I'm curious, you know,
kind of going back to Mr. Falkor's original point,
when we start to look at identifying what we want to use versus what we, I mean,
maybe not even what we want to use, what we want out of this, what's kind of the end goal for
agendic finance? Because I think this is maybe, this is
probably a loaded question for someone like you, but everyone I talk to is kind of just like,
hey, I would love, I just want to give it like a hundred bucks and have it go turn it into a
million. But we know that's not realistic per se. So what do we actually, what do you actually
think the end goal or maybe even the short, mid and long term goals are here for agentic finance?
Yeah. So right now, objectively speaking, if you can lump finance into everything besides execution and then execution,
besides execution and then execution. Execution being, hey, I sent money to a broker. Their
trading desk is using HFT techniques to execute my trade optimally. The goal is to replace
everything up to the execution level and then maybe even also eat up part of the execution level.
And I think something really interesting that hit the wire recently was Jane Street led a $500 million investment into a lab specializing in low latency, high throughput LLM chips.
And that's like a very strong signal being like, hey, maybe HFT firms are thinking about LOMs for getting closer to execution.
And so if you think of, and I'm using the same wire metaphor, but Exchange wires or the financial system. We have mountains of
infrastructure sitting on top. And the goal is to get these LLMs as close as humanly possible
and as fast as humanly possible so that they can make these decisions at incredible pace,
process data and information at an incredible pace, and operate
in the financial markets really effectively. So just think of like the whole TAM of finance,
you know, minus, you know, niches where LLMs really aren't going to be able to perform.
We own that. And then on the execution layer, these LLMs can write the algorithms,
And then on the execution layer, these LLMs can write the algorithms, write the code, et cetera, to operate.
And so obviously there's like a hardware element with like FGPA circuits and other things.
But, you know, LLMs, I think, you can't kind of break the laws of physics.
You can extract alpha with these AI models and you can outperform, you know, maybe even generate outsized returns.
I think it's a little too early to tell how that will look
in the real world. But I think certainly just looking at the data that our traders are generating,
it will happen. It's just, you know, financial markets, especially at the institutional level,
move slow. They need to move slow. And so we'll see this stuff kind of pan out over longer timeframes, but, uh, yeah, I mean,
the, the, the whole TAM of finance is like fair game, I would say for LLMs, you know, the, the
advisory market alone is huge, you know, like every, you know, Merrill Lynch investment broker,
investment advisor, you know, every wealth manager under the sun, whatever, an LLM is going to do better than them.
99.999999 times out of 100.
So the goal is to really explore the market.
It's more of the way that you set up your AI agent
and the strategy that you put in place.
I think what I've seen a lot of the times is, and, you know, kind of the question of just go make me
money. That's not really a thing. You know, you need to give more context. You need to put in
your guardrails. You need to be able to say, Hey, you know, we want to do X amount return per trade.
And then you are kind of just going back to strategy at that point.
Yeah. I mean, um, it, it, like intent solving is a big part of,
you know, the LLM experience, you know,
you can go onto our platform codex cod3x.org, um,
and say, go make me money. and the model will know what to do you know um it'll you
know just probably look at three say cod3x.org if you haven't used it you're leaving you know you
should at least for research purposes but um really that's just intent solving. It's like, okay, we know why users
are on Codex, all right? You know, we can even look into their wallet when they log in and be
like, okay, here's their risk profile. We're connected to the D-Bank API so we can literally
see everything. And that's more than enough context to figure out, oh, this is what they're looking for. Like when you go into Cloud Code and say, build me a web app, if you interact with the raw base model, it's not going to do a great job of that.
But the fact Cloud Code is a harness with 500,000 lines of code designed to help take you from zero to 100 um you know it'll be able
to do that no problem and so like i would i would say as well like the more information you give the
more context the more strategy you can help the better result you'll get um and so it's really
like it's kind of a spectrum it's like like intent solving helps take you from the face roll input to an output that is kind of, you know, similar-ish to what you're looking for.
But if you give it really precise inputs, then it's going to give you really great outputs.
And so our most successful traders on Codex are doing really insane prompt engineering.
And we have a system whereby you can trigger runs with other runs and you can link contexts together across runs.
And everything runs on a prop trading engine, so the automation is really granular and crazy.
and crazy. And so we have people creating these cathedrals of, you know, analysis and execution
logic that, you know, if you were to do that as a software engineer, as a quant fund, it would take
you months or even years to like, you know, get running consistently and you can do it on codex and like you know 30 minutes to an hour and and
a few iterations um so that's the goal of any harness you know whether it be clock code for
programming whether it be you know codex for uh trading we're built on hyperliquid it's just perps
um or you know harvey for law it's like all of it is just a harness directing that energy toward, you know, the use case.
There's so much to unpack there. That was fantastic. I think let's kind of start. Let's go in reverse there.
So I think, you know, it's really interesting because we're in an era now where doing has never
functionally been easier.
If you're willing to pay for a cheap coding plan, you can do things that to the average
individual were functionally impossible about three and a half months ago.
I lied.
We're about to go into April. Four and a half months ago. I lied. We're about to go into April.
Four and a half months ago.
Like literally, I think December, what?
December 26th?
December 22nd.
Somewhere in like the Christmas week.
I feel like there was some like trigger
where like, I don't know,
Anthropic did something
and Claude Co. just went like zero to a hundred
and all of a sudden stuff just started working um at least for me like i i was a very i was i was
i was a non-believer in agenda coding i couldn't get it to work 100 i can i can attest to this
because i actually went back in our messages like 10 minutes ago looking for you bitching about ai
and it's just all of this and i'm like i converted yeah it's just like there's just like this period
i mean again like it worked like the like autocomplete and stuff from Cursor was working.
But the ability to basically just be like,
I need this, go figure it out,
didn't, for me at least,
and for a lot of other people,
I think click until kind of that end of December,
early January moment.
And that seems to have been this launch period,
which kind of as we step back
into what you just walked through,
there's so many... One of the hardest things about AI specifically
is there's too much for one person to keep up with.
And so the things we'll talk about are probably irrelevant to many people
and maybe really relevant for some.
But before we started the live stream, you said something really interesting,
which is that you're...
Can I share the thing about the benchmarks?
Oh, yeah, sure.
Cool, cool.
So you had mentioned that you're working on some benchmarks.
You guys at UV, you've trained and fine-tuned your own hyperliquid LLM,
which is, from what I can tell, it's a smaller model that you guys have taken
about 6,000. If I remember the number correctly from the article, you have about 6,000 examples
of trading in there. And you're finding that gives just the average trader a little bit more accuracy, right? You start to see that compound
and improve. Maybe it doesn't have the breadth of intelligence, but for maybe the specific
trading, you start to see improvements and then tack that on and I'll let you kind of brain dump
on us. You had also mentioned that in your benchmarks, you're seeing in many cases
You had also mentioned that in your benchmarks, you're seeing in many cases, Anthropix Haiku model outperform Opus. And so this is something that I have flip-flopped back and forth on a lot over the last year and a half, ever since you and I talked.
Ever since you and I talked and I said, hey, I'm really interested in fine tuning small models for specific purposes.
And your response to me last time we talked was small models suck.
Just throw bigger models at it.
And so clearly at some point, right, you obviously not necessarily had a change of heart, but you started to see a different opportunity.
not necessarily had a change of heart, but you started to see a different opportunity.
And so I'm just curious what's kind of, you know, where that's come from and how that's impacted
your ability to serve these capabilities out to your users and ultimately what the downstream
impact is for others building Ingenic Finance. Well, I will still preface this by saying
Well, I will still preface this by saying I remain a big model maxi.
And there is this very tiny paper article by Rich Sutton from March 2019 called The Bitter Lesson. And The Bitter Lesson is kind of like the, you know,
it's like Moore's Law for AI, kind of,
but it's like a compliment to Moore's Law
that basically says, you know, over-engineering
or over-architecting solutions for AI or for neural networks is often a waste of time because the ultimate solution to any problem in AI is just more compute.
And that has remained very, very true.
You know, a lot of people get really excited about, you know, super complex, you know, neurobiology inspired architectures that don't scale.
So before I start talking about small models, I just want to get that out there because
the results I'm seeing are pretty unintuitive for the benchmarking. So I've created basically a
normalized market environment that allows these models to just, that allows me to test their raw
decision-making skills with a very minimal harness, very minimal. Like I'm trying intentionally
not to skew their opinions one way or the other. But the second you introduce a harness and you
start doing context engineering, the big models will mog the small models all day long.
But what I saw looking at the raw, you know, intelligence is that, you know, Haiku outperforms Sonnet, which outperforms Opus in just basic financial tasks.
And there could be any, you know, many reasons for that. Overcomplicating portfolio management, Opus could easily be doing that. Same with Sonnet. You know, as well as potentially, you know, theus 4.5 moment where, you know, a lot of people attribute to takeoff beginning,
being like, you know, the route toward, you know, hyperscaling this automated research,
automated growth. You know, the reason Opus 4.5 was so good is because Claude figured out how to
scale their RL in Cloud Code.
You know, and obviously, like, they're not saying that explicitly.
Like, you can see some of the research coming out at the time.
But, you know, basically the most powerful tool right now for training models is reinforcement learning. So given all of this, I believe that our team is going to be able to scale up our hyperliquid LOM to outperform Sonnet, which we use in our harness.
We use Haiku as well, but Sonnet can just process way more context way more effectively.
We believe we can, at a much smaller scale, outperform these really huge models using domain adaptation. And the 6000 data pieces, very few of them were for training.
And so I'm doing LoRa, which is low rank adaptation of a model.
So basically you're adding a layer of addition on top of the transformer that just shifts
stuff over a little bit.
And when you're doing LoRa, there's this concept called rank, which is essentially like how much you're messing with the model every time it processes a piece of data. And when you're doing
some pretty extreme domain adaptation, you want it to change a lot. And so you can make pretty big
changes if your rank is fairly high. And if your rank is too high, you might overfit. And so you
don't want to show it too much data. You want it to develop an intuition. And a lot of that just comes with like masterfully curating a data set.
And what we did with HyperLOM4B was super simple. Like, you know, I didn't spend too much time on
it because this is just a smoke test. You know, what configuration do I need to start pushing
these models toward finance effectively? And as we scale, we're going to start doing RL.
We're going to start manually curating more evals, more training data.
But that's 6,000 pieces of training data was training it across 12 different capabilities.
You can have very sparse data sets depending on your approach.
Definitely. That makes a lot of sense. And I do think, like I said, it's kind of funny
because I think, right, as with any technology in the world, you get people kind of on both
sides of the fence. And I've seen a lot of people very anti-Laura and kind of
manipulating the LM that way.
Why?
That makes no sense.
I feel like
I've come across people who are basically
like, you can't
you're better off
with just improving its direct context
and trying to over-tune it.
You wind up, to your point,
potentially going too far.
Again, it's more so
the interesting part here.
That's the opinion of people
who've probably never done it.
If they did do it,
they didn't really know why
they were doing it.
Yeah, I would concur.
And so I think the... That is true in some cases though like you you know unless you have unless you know exactly what
you're looking for you're just gonna give the model a lobotomy but um domain adaptation it's
a very surgical process yeah it definitely is and i think you know this is something we've started
to talk about and it's a bit broader than finance, but something that we've started to... I've had a few people I've had live conversations with this about now. And I really do believe that as...
ModelMax is you, but I'm most of the way there at this point.
You see the impact pretty significantly when you're doing longer running tasks, especially.
And so one of the things that I've started to talk with some people about is as more
and more models get to this base level of, call it like, good
enough, right? For example,
we now have
multiple Anthropic models
that are good enough. We've got
multiple OpenAI models that are good enough.
We've got Moonshots
Kimi K2.5,
which is kind of a monster.
We've got, what, a series
of models from Quen that are pretty dang monster. We've got a series of models from Quinn
that are pretty dang good.
We've got Minimax has a bunch of models
that are pretty good.
GLM has a series of models that are very good.
Basically, the ability to train a good model
or at least the base level of it,
these bigger ones,
we're seeing more and more companies be capable of getting to that point. Whether they're using other models to actually
do that, we're not going to get into that discussion. We're seeing more big models
that are not capable of being this like baseline of good enough. It's now, to your point,
that's the engine, right? It's kind of like in, It's like in Formula One, you've got all these race cars
that are really impressive, but a bunch of them actually use their competitor's engine
or their competitor's power supply or whatever. It's like that. You've got a lot of people using
each other's engine to get going, but now we take that stuff up and now we're immediately into this harness.
And there's all these different decisions that
we have to make.
And it feels like to me,
obviously building an effective harness
is an art
more so than a science, I feel like.
Because there's just
kind of like, what can you
craft a little bit?
You think? It's kind of like what can you craft a little bit? It's a science. You think?
It's a science.
It's like a lot of people treat it like an art because the amount of testing required to perform the science is so expensive and so time consuming just because you need to run thousands and thousands of turns, you know, in these things. But I would say certainly if it's not a size to you, then you need to like if you do not
have a North Star when you're building AI technology or when you're building out a feature,
if you if you can't say exactly what you're trying to accomplish in the most granular
fashion and who you're trying to accomplish it for, then it's just going to blow up
in your face because, you know, these you need clean, verifiable rewards at every level. It's
not just about doing RL. It's about, you know, even just directing the model, because when you
add context, you know, you're effectively changing the framing of everything in the same way you might
during training. And so, you know, you need to have the same mindset being, okay, like, you know,
this, you need to be doing, you know, there's this, I'm trying not to say this too much,
because it's what real researchers use. And I'm more of like a cringe punk manager type guy.
But you need to be doing like, ablations, which, you know,
in surgery, it's like doing, you know, minor, like I'm going to cut off a slice of this brain and
open it up and mess around with it, whatever. A lot of people are like, hey, that's for the
researcher. You know, I'm just building an app. I don't need to do that. But the reality is you do. You need to be absolutely surgical because one body of context, like 100 tokens somewhere in your harness, could be decreasing your results.
I don't know what the right word there is, by like 10% or more.
by like 10% or more. And that's just how volatile these models can be, especially like,
you know, there's no like, there's no, nobody can write a guide that tells you exactly what to do.
I mean, so everybody, every researcher on earth, every AI researcher on earth, and if you're
building an app, you need to be a researcher, whether you like it or not. All they're doing is
experimenting and testing and seeing what works and what doesn't
and trying to break things
and trying to improve things, whatever.
It's just a very, like,
it's like we've created this huge alien
that is like, you know, as big as a skyscraper.
And now you just need to hire all these guys
to be, you know, vivisecting it,
trying to figure out how the hell it works,
how to, you know, change its behavior or whatever.
It's just madness.
So, yeah, it's like I think people need to get out of the like I'm an artist.
I'm a I'm a software artist sort of mindset and just realize that machine learning is it's literally like the most chud domain of all time.
It's like, you know, you have these huge libraries that do everything. They're written in the worst programming language of all time it's like you know you have these huge libraries that do everything
they're written in the worst programming language of all time uh they're all brute forcing everything
everything breaks constantly every time you make an update and you just need to understand that
this is like massive scale chud canon uh it's not like an intricate tourbillon watch it's like
you know someone built the pyramidramid of Giza.
It's just like a totally different ballgame.
You're not creating
the most optimized compiler.
You're hitting
a hammer against
the Empire State
building.
That's the mindset I think people need
to take.
Fair enough. I appreciate that mindset I think people need to take. Fair enough. I appreciate that mindset.
I don't necessarily disagree.
I still believe there's a little bit of art to finding the starting point,
at least when you're building your own.
Sure, totally.
Like, I guess what I mean by that is the,
so I've been experimenting the last couple weeks.
I've been building my own harness because I do not want to use Telegram to talk to my agent.
I don't use Telegram unless I have to use Telegram.
It's just like dumb to me.
And to each their own, right?
Like to each their own.
But at the same time, I noticed that, I noticed that a lot of the existing pages today,
they're very much so, you're either very heavy domain specific, or they're basically open claws
and they're just like, there's very few guardrails. It's like, if you actually want to use this thing
to do anything consistently in your day-to-day life,
there are significant risks involved.
Because anything you give it is functionally exposed.
You just can't.
Especially with, I mean, in the last 24 hours,
we've had some of the biggest supply chain attacks ever.
Our pyramid of Giza, our building blocks and building code are just getting kind of wrecked by malicious actors.
If you didn't see that, don't do NPM install for the next like 38 hours.
But that being said, the reason I bring it up is...
I did it yesterday. Am I okay?
I'll tell you what to do in a few minutes.
It's probably too late anyways.
But anyways, the reason I bring it up is that I think,
you know, there's,
I was talking with a bunch of people
and this conversation kept coming up of
why do developers keep building their own agents?
We have all these great harnesses.
We have all these great tools.
Why do we keep building our own agents?
And I think it comes down to,
we're all after the perfect thing for ourselves.
And that's one of the things that makes
AI so exciting is we all see it differently.
And so that's where I think that little bit
of the art comes in
is it's creating your own masterpiece.
But then to your point, actually making it
happen is a thousand percent.
There's a
method to that madness. And that
is quite literally the running in circles.
I think like anytime there is a huge growth spurt, like something like OpenClaw or in late 2024, early 2025, it was like, you know, Eliza virtuals in 2023, 2024, it was Tao.
You know, whenever you give,
whenever something pops off, like people get inspired
and they think, hey, I can fork this and make it better.
I can do this, that, the other.
And then, you know, eventually they realize,
you know, they've been given this, you know huge book um and they're just writing these small
sentences and uh you know to actually structurally change the content of the book requires an
enormous amount of time and effort um and to give you an idea you know we've been building our
harness for longer than claude Code has been building theirs.
Claude Code is, it just got leaked, the code base this morning.
It's 500,000 lines of code.
Ours to do secure agentic finance is one and a half million.
You know, and I don't think lines of code is a good measure of anything.
And I don't think lines of code is a good measure of anything.
But I'm saying like this is like a very carefully implemented one and a half million, you know, and, you know, something like Twitter, probably like five to 10 million or more, you know, something like Atlassian, you know, probably beyond the pale.
beyond the pale. But, you know, I think like, you know, these harnesses, this is going to be where
most of the work gets done and the code gets written. And if you want to take something like
OpenClaw, which is like, frankly, a total slop cannon, you know, like the domain specific ones,
you think of it like a rail gun, like the goal is to just penetrate as far and as deep into a domain as humanly possible, given the technology.
And you can keep upgrading the engine as new models come out and drill deeper, deeper, deeper, deeper, and maybe even like expand capabilities.
like OpenClaw, it's like because it's so broad and vast, there is like any given domain, like
pointing it at any given domain, there's probably not a lot of work that has been domain-specific
work that's been done. So you can find success for a little while and then suddenly everything
else starts becoming technical debt and suddenly you realize, oh, I would have been better starting
off on my own. And some people are now starting off on their own and they're realizing, oh, like after a few hundred thousand lines of code, it's kind of unmaintainable.
And they're vibe coding it all. Like we hand wrote a grand majority of our harness and we could not have gotten to where we are today if we hadn't. Like we very lucky, very fortunate that we were forced to build this before, you know, the apex of generative coding.
And so it's just like everybody thinks stuff is easy now, but stuff is not easy.
It's just bigger, badder, better.
And so getting to like what was cool five years ago is easy but getting to what's cool
now is like a tremendous herculean effort um and nothing ever changes it's like the tides are
rising everywhere at once it's not like you have you know you're some special genius who's uh you
know the only one whose tide is riding rising. It's like everybody is building.
And so standards are changing.
You know, stuff that was cool five years ago is table stakes now.
You know, a junior engineer is, you know, writing the kernels.
And, you know, the seniors are, you know, trying to conquer the world.
Yeah.
But, yeah, it's just like I think fundamentally like AI is so hype driven and it's so electric when you're, when you're in the terminal and when you're creating, especially like, you know, I've, I started getting into using Cloud Code and stuff that I wouldn't have had time to do normally, I'm now able to do.
And so, you know, I'm doing research, I'm doing smoke tests, I'm, you know, taking on the role of architect, whatever, for all this stuff.
But I'm the dumbest guy on the team, you know, it's like I'm doing a lot of the um a lot of stuff but it's just uh it's so big like it's so big that
um i don't know it's just not stuff isn't easy and i don't think it's ever going to be easy and i
think it's only going to get harder frankly um which you know it is what it is. I agree. It's the era where, I've been saying this a lot,
there's building like really small demos is easier than ever.
Building SaaS is easier than ever within like to a certain extent,
but actually building AI is harder than building SaaS by hand ever was.
Yeah.
Because it will remain that way.
SaaS has always been relatively easy to build.
And the reason these huge SaaS companies have customers is not because of their technology.
You know, it's because of their distribution.
It's because of their integrations.
It's because they were early.
It's because, you know, their moat is not the tech.
Very rarely is their moat the tech.
You can't patent most of this stuff.
And if you do,
there are plenty of alternative
solutions.
Definitely.
It's the exact reason why I feel
like we're seeing companies
now like Linear and others
who are just being like, yeah, we're going to go build an agent and like our product is, obviously those posts are kind of clickbaity.
Same thing I think with AMP, but like realistically is like they have to go in a direction where it's not something you can just like build in a few hours yourself. Nobody is serious yet. Yeah.
Especially in crypto AI,
it's like we've got,
we've got like nobody,
like nobody even knows what's up,
you know,
like the fact that people are arguing about harness nomenclature
and crypto AI.
It's like we had this argument
like one and a half,
two years ago in regular AI world.
ChatGPT started saying harness and we were like, oh, God damn it.
The harness people won.
And so now you just have to say harness.
It's cringe.
You know, it's high up in ChatGPT's next token policy.
But you just have to say it because that's the word, unfortunately.
And, you know, it's like you're putting a harness on a horse, you know, you're in this
metaphor, the AI agent is the horse.
But I would say the harness at this point, like for frontier teams, is more like, you
know, giant mechanized horse armor, you know, with various attachments and other cool stuff.
But yeah, it's, I think in crypto right now, like, you know, we've got like citadels of
technology we're sitting on and we have teams like coming up to us low key, like, hey, would
you build this insane piece of technology for us in exchange for marketing exposure?
That's happened to us twice in the past couple of weeks.
And I think it belies like a fundamental misunderstanding of what it takes to
make these products actually work.
Definitely.
Definitely.
Okay.
So we've had a fantastic conversation up to this point.
I know we don't have too much time left.
And so I want to make sure we double click into a couple of things briefly, which are the other half of
agentic finance. Half of agentic finance is agents and finance. And the other half,
I think at least for us, because of where we spend our time, is the blockchain side.
I think, at least for us, because of where we spend our time is the blockchain side.
Right?
So obviously you guys are very focused on Hyperliquid.
Scale, for those of you that do not know, and you're listening to this for the first time,
Scale is a blockchain network.
And very specifically, we have a number of native privacy features that allow agents
and humans to be capable of doing things they normally can't on-chain on-chain, such
as trade privately, move money confidently, things like that.
And so what I'm kind of curious from you, Justin, is when we start to think about agentefinance,
is it just because of your history with Moon and with everything you guys kind of did
that you guys are kind of agentic finance in this AI crypto crossover?
Or do you have a belief on your side that because of the openness of how crypto works
and because of just the functionally, like, I kind of say, like,
which I stole this from someone else. I promise I did not make this up. I'm not that smart. Someone else said blockchains are
built for machines or agents, not for humans. Right. Because let's be honest, blockchains are
hard. Do you kind of have that belief of like these tools just make more sense for what you
guys are building? Or do you see this as a no one day there
could be you know you could have you know codex creating off of you know a live brokerage firm or
or i don't know different you know real world exchanges if they tokenize things and stuff like
that yeah i mean on the back end we already have all the big exchanges, whatever integrated, you know, we just don't service surface them on the front end.
Like there will be a time where, you know, when we launch the codec CLI, you can touch all this stuff.
But front ends need to be kind of opinionated if your users are going to be able to use them.
I would say, you know, for years, kind of my justification for blockchain was twofold.
for years kind of my justification for blockchain was twofold um you know one uh the fact that you
don't need to worry about compliance kyc all this other stuff the fact that you know the non-custodial
software takes care of it out of the box just makes it way easier to build uh dynamic systems
um you know your biggest blocker doesn't exist uh So integrating with Hyperliquid for us is a no-brainer,
because it's like, oh, we can get users on board instantly.
And then beyond that, you kind of
have this deterministic layer for your probabilistic agent
slash model, which is helpful in theory.
In practice, you know, we don't use it as much as we necessarily could
just because like, you know, a regular policy agent works fine in most cases.
But, you know, if we really, really wanted to, you know, impress a client
with like guarantees about, you know, assertions or versions, whatever,
that would take us, you know, no time at all to build into the system. So that's also very cool.
And I think, you know, as we've been building throughout the years, now that's evolved into,
you know, it makes sense even from like a serviceable market perspective, because it seems like all of these companies are moving on chain anyway.
And what we've known to be true about DeFi for years, that it's just way more efficient.
You know, you can do way more with fewer people.
Combine that with AI, where you can do way more with fewer people. Combine that with AI where you can do way more with fewer people. Combine that
with the hard determinism, the ability to kind of interact across substrates and primitives
relatively easily. I think it just kind of becomes a no-brainer. And AI people are pretty
bullish on crypto for the same reason. It's like, you know, you have, you can program directly,
going back to the wire analogy, you know, you can program directly on top of the wire. And
on blockchain, like the fight against Jane Street, depending on the chain you're on,
but the fight against the HFT firms is actually much different because, you know, if you can spin up a node
or you're willing to bid up an auction and you can do it as effectively and as well optimized
as one of these firms, then, you know, you have just as high a chance for placing a transaction
top of block as they do. And so I think like in a world that is globally distributed um in a world
where you know uh every every you know point of friction is is a potential bounce and and you
know uh customer lost uh blockchain just makes sense and then you know obviously like it comes
with these security guarantees that are in some
ways much more of a headache and much more complicated, but in other ways, much simpler.
You know, the threat model is pretty straightforward. And so, yeah, I think trade offs skew heavily
toward crypto for agentic finance. And, you know, right now, right now if you you know cut a deal with uh
you know a big broker or whatever if you cut a deal with fidelity and they let you noodle around
in their api and do whatever which isn't going to happen but if you can then maybe right now that'll
that would produce a better product uh than necessarily just programming straight to the blockchain.
But like as time progresses, like in a year or two, that will probably not be true at face value.
And it doesn't require the same, you know, BD.
It doesn't require the same song and dance that working with like a big broker would or uh you know a big platform like
that so yeah it's it's like um i don't there are a lot of reasons why i do it i i don't buy into
like you know um i try to stay pragmatic about it you know i guess is what i'm trying to say and
it's just like blockchain just makes sense for us.
And that is because we already know how to build on the blockchain.
And yeah, so very biased.
You know, it's very valid.
I mean, I think, you know, one of the, again, kind of coming back to this,
this AI crypto crossover, something we're definitely seeing, or at least I feel like I'm seeing a lot of,, again, kind of coming back to this AI crypto crossover, something we're
definitely seeing, or at least I feel like I'm seeing a lot of because I'm very much a like,
I build developer tools, right? Like that's, I enjoy doing it. And so I spent a lot of time
building SDKs and CLIs and things like that. And one of the things that I'm noticing as I'm building more agents and agentic processes is using legacy platforms, using just existing software kind of just
sucks. Right? Probably best example right now, I think is probably GitHub or just like
Git, right? Not just GitHub, but Git in general is not very efficient for agentic coding.
It's now, you know, we've seen two companies
in the past three months raise $85 million
to solve different ways of interacting with their,
you know, agents interacting with code.
We've got half a dozen companies,
a dozen companies building specifically new ways
to use the browser
because spinning up a Chromium instance
and trying to peruse that is
really freaking inefficient, right? Now you've got sandbox primitives because normal serverless
in nine out of ten cases doesn't give you enough functionality, right? You've got all these layers
and the end result is we're rebuilding. We're functionally rebuilding all these things. And so
I think kind of putting
blockchain in that perspective, to me, it feels like you're saying very clearly traditional
trading surfaces and platforms just don't necessarily have the freedom or capabilities
to actually do that. Compliance aside, I think that's a...compliance is a bigger topic, right? And I do think for a lot
of these people coming on chain, they ultimately will at a point need some level of probably KYC,
KYB, stuff like that, just because I don't think they can just turn around and bypass all that
because they're on chain. I do think the programmability of the blockchain, to your point,
getting close to that execution engine makes a lot of sense.
But I know you've got a hard stop here in just a couple of minutes.
So I will hand this over to Mr. Falkor to kind of wrap this up
if there's any final questions.
No, no, no final questions.
I mean, Justin, anything you want to,
any final words you want to give to the people?
Yeah, I mean, I guess like, you know,
keep reading and try to understand this stuff better.
It's like, you know, critically important
that you gain these skills over the next few years.
Otherwise, you know, you will be less valuable in the job market, which is never fun.
So just bite the bullet, sit down, learn how to use Cloud Code or Cloud Cowork,
depending on, you know, how you work and whether you're on the engineering floor or the business floor.
And then I would say, like, if you're interested, try out Codex, cod3x.org.
You can follow it on Twitter, cod3x.org.
And then if you're interested in our research, just x.com slash uv.
Two letters, super simple.
And, you know, we're going to be publishing there all of
all of our wins all of our mistakes uh whatever um so if you're interested in financial uh ai then
probably no better place to to follow 100 and amazing justin thank you so much for joining us
dropping the wisdom you know it's been super fun.
Learned a lot. Yeah, thanks a lot, fellas.
I'm going to run to this meeting, and I've been chugging energy drinks,
so I have something to do beforehand.
But, yeah, I'm sure I'll see you all soon.
Bye-bye.
Awesome.
Thanks, Shelton.
Thank you, everyone, so much for tuning in and joining us.
Make sure to come with us again next week.
