Once Chat GBT kind of hit the market

late 2022, I was curious with a simple

question, which is can chat GBT beat the

S&P 500?

>> You need to have someone in leadership

that's AI pilled.

>> Only 17% of stocks outperform the S&P

500 over a rolling 10-year period. So,

probabilistically, that stock you pick

at random is going to be a stock that

underperforms the S&P 500. How do you

look at the entire corpus of every

portfolio you've ever generated?

This conversation is provided

forformational and educational purposes

only. The views expressed are those of

the individual speakers as of the

recording date, may differ from the

views of their organizations, and may

change without notice. Nothing discussed

constitutes investment, legal,

accounting, or tax advice, an offer or

solicitation, or a recommendation or

endorsement of any security, investment

strategy, or financial product. Access

to this conversation does not by itself

create an advisory, fiduciary, or client

relationship. Any companies, securities,

or strategies discussed are presented

solely for illustrative purposes. The

speakers or their affiliates may hold

positions in, provide services to, or

otherwise have financial interests in

the companies or securities mentioned,

and those interests may change without

notice. Any forecasts, targets,

forward-looking statements, model

outputs, or hypothetical results are

based on assumptions, and information

available as of the recording date. They

involve inherent risks and

uncertainties, may not reflect actual

trading or investment results and should

not be relied upon as guarantees of

future outcomes. Information obtained

from thirdparty sources is believed to

be reliable, but its accuracy,

completeness, and timeliness are not

guaranteed. All investing involves risk,

including volatility and the possible

loss of principle. Past performance is

not indicative of future results and no

investment strategy can assure a profit

or protect against loss. Listeners

should conduct their own due diligence

and consult their own qualified

financial, legal, accounting and tax

advisers before making any investment

decision. Artificial intelligence

systems and other analytical models may

produce incomplete, inaccurate or

inconsistent results. Any model

generated analysis should be

independently evaluated and should not

be used as the sole basis for an

investment decision.

Welcome to the next episode of Invest

with AI, where we explore the

intersection of fundamental investing

and artificial intelligence. We have a

great product market fit with the next

guest on our podcast today, Doug Clinton

of Intelligent Alpha, who has been

investing with AI. So excited to have

you on the pod, Doug, today and explore

everything you've learned that's worked

well that hasn't worked well on your

journey of investing uh with AI. Maybe

to start, could you tell us a little bit

more about Intelligent Alpha and what

you're doing there?

>> Absolutely, Brett. Good to be with you.

Good to be with you, K. And Intelligent

Alpha is a company that I started uh two

years ago now officially. And it was

born out of an experiment that started

three years ago where once chat GBT kind

of hit the market late 2022, I was

curious with a simple question which is

can chat GBT beat the S&P 500 uh which a

lot of human managers can't do. And so

we started running some tests to see if

that was possible and the short answer

was uh things look very promising very

early on. Maybe it's a little beginner's

luck, but I was excited enough about

what I was kind of seeing from very

early LLMs and betting on the trajectory

of these LLMs continuing to get smarter

and smarter over time um that we wanted

to launch an investment business that

was really built around the idea of

harnessing the intelligence from

frontier models, allowing them to do

investment analysis and then ultimately

portfolio management. The one other

small piece I'd add about my background

and kind of why I even played in this

space in the first place is um almost a

decade ago, I started a venture firm

called Deep Water. Um still a partner

there. Uh I I run kind of uh uh both

sides of the coin being a human

allocator and also relying on AI to

allocate at Intelligent Alpha. But we

invest in early stage technology

startups. We invest now in late stage

tech startups and public equity at Deep

Water. And so it was kind of this fun

natural testing ground where we've been

in this world, we've been investing in

AI companies and intelligent alpha was

the next logical progression for us from

that uh business at Deep Water.

>> Super super interesting and I have many

many questions uh about that. One of the

first um is how you thought about just

the fundamentals of large language

models which operate on lexical

intensity. Right? LLMs know that Paris

is a capital of France because it's seen

it 14 million times in the training

corpus and you know when I've sort of in

the past gone and asked chat GPT for a

portfolio of 10 stocks it seems to

operate on that same sense of lexical

intensity that the stocks like Nvidia

that are talked about a lot on the on

the open web get cited as individual

ideas in the portfolio. That could be a

a signal or that could not be often sort

of over the cycle of markets is probably

a negative signal and retail driven

markets probably positive signal. How

have you thought about that concept of

lexical intensity prevalence in the

training corpus and building a a market

beating portfolio?

>> Yeah, I'd say very early on uh one of

the things we sort of realized was in 23

the paradigm for using LLM effectively

to invest was prompt engineering. I mean

that was all the rage then and really

understanding what what is the context

that you are providing to the models in

terms of the task you want it to output

and so like you just said Brett if you

just say give me a stock of 10

portfolios you're probably going to get

I would say eight of the 10 uh largest

companies in the S&P 500 would be my my

bet um which is fine that might be an

okay portfolio um but I think when we

think about really using AI to make

intelligent investments that's not what

we want and so giving it a different

framework work where you know maybe you

narrow the field to small caps and maybe

you define that within certain

parameters for market cap maybe you pull

in some external data uh from sources

where maybe it's fundamental data uh

maybe it's your own research about a

handful of companies I think that's

where it starts to get much more

interesting using these language models

and seeing how they sort of interpret

what is all the data that's out there

that humans are are looking at right and

and language models can process way more

qualitative ative data and quantitative

data than any human analyst can. So

trying to figure out how can you get

that right data and the right selection

set into the models is kind of step

number one. And then I think as I'm sure

you guys have seen too, right, the

frontier has evolved very quickly where

it was all about prompt engineering a

couple years ago. I think now it's about

agentic workflows, you know,

orchestration. And I think even beyond

where we're at right now, the thing that

we're getting really into is how do you

sort of manage um organizational

context, you know, instead of just

narrow task context, you know, you're

trying to pick a large cap portfolio

that's one specific vertical that you

might be working on. How do you look at

the entire corpus of every portfolio

you've ever generated? or maybe if you

have multiple team members generating

different portfolios with their agents,

how do you sort of understand the entire

broad uh knowledge base that's in your

company and make it useful to the models

when you're doing investment work?

Doug, on that point did um you mentioned

kind of the evolution from prompt and it

seems like we went from prompt

engineering to content context

engineering [snorts] to agentic

workflows and now with kind of the fable

like models kind of intent engineering

or out you know out output pro output

based prompting

I'm curious how that last bit I mean

we're still early in the fable five days

but um has that last bit changed your

approach?

Um, I'd say our approach is always

changing slightly and I think it has to

actually and I think about these models

in some ways as a reflection of of how

alpha changes in markets. You know, what

what might generate alpha today probably

won't generate alpha tomorrow. And

whatever you're doing to generate alpha,

and that's a very broad word that

probably needs a real definition, but

let's just say outperformance relative

to some benchmark right now. um whatever

you're doing to generate that

outperformance today with the models

probably won't work tomorrow because

there are other people that are

experimenting uh with these models or

finding out creative ways I think to uh

make investment decisions. It might be

much more of a humanoriented process

than what we do at Intelligent Alpha

where we rely very much on the models to

kind of work end to end. Um, but you

know, I think we're always trying to

change our approach and see what has

sort of maybe stopped working based on

what we were doing in the past and all

the data that we have over the past

three years. And then our intuition, and

I think this is where the human still

comes in, sort of our intuition of where

the market's going, what might work in

the future. Um, and in some ways that

might be a little bit like how some

quants operate where a lot of times

thinking of the next factor or thinking

of the next um, sort of data set that

you might believe has alpha is much more

of kind of an intuition than than just

math.

>> Yeah, this is this is a huge structural

difference versus building code. For

example, one of the stats I've been I

sort of point out to all my students is

only 17% of stocks outperform the S&P

500 over a rolling 10-year period. So,

probabilistically that stock you pick at

random is going to be a stock that

underperforms the S&P 500 sort of, you

know, people sort of broadly discusses

70% of long only managers underperform

the S&P 500. And so, you know, public

markets alpha is truly a power law game.

How do you how do you think about

capturing how do you think about

capturing that power law sort of right

tail alpha essence in the agentic

structure? Because if we're sort of like

taking our investment process and

turning that into into an agent

structure, if I'm a median investor, I'm

actually augmenting a median process

which structurally does not have the

essence of that of that alpha capture

ability.

>> I think um what's interesting to think

in that context is is how much of the

intuition coming from the LLM is

valuable versus the intuition coming

from the human. And as I said with

intelligent alpha, the bet we are making

is this long-term bet that over time the

models do continue to get better. We've

seen that in the benchmark data and that

ultimately probably leaving the models

to their own devices with very little

sort of nudges or interpretations or or

influences from humans is the best

process. But today, I actually think

that's that's a little bit different

where I still think the creativity of a

human to say, you know, I want to be 30%

weighted in um some stock because I have

really high conviction that it's going

to work for reason X. Those types of

intuitions and maybe there's some deeper

research that goes on behind the scenes,

things that an AI can't get access to.

you know, those things I think are still

the core blocker from just turning the

whole thing over to AI. And so what I

would say we've seen work really well

right now is almost this sort of like

hybrid model between what would look

like maybe a traditional quant portfolio

and a traditional fundamental portfolio.

And so generally our portfolios, they

don't have a thousand positions like you

might have in in a quant portfolio where

it's literally right, you're just

looking at numbers, a lot of small bets,

and you're trying to get that 52 54%

batting average. You know, we generally

have something on the order of a couple

hundred, you know, maybe up to 500

depending on what strategy we're looking

at. Um but there are in our portfolios

some conviction bets that we allow the

models to make that may look like you

know several percentage points in terms

of concentration in the portfolio. Um

and AI I think has been pretty good in

what we've seen so far in our portfolios

at knowing where to be convicted and and

maybe where to dial it back a little bit

where it doesn't have great conviction.

And it's gotten better over time to the

point of our thesis. Um, and so, you

know, I think that humans can still play

a really valuable role, which is

bringing in that data that the models

just can't get access to. Pushing the

models to take really big bets, but if

you're not going to have that human

influence sitting somewhere between

probably a traditional fundamental

process and a traditional quantitative

process works really well with the LLMs.

>> How important is that data and what role

does it play? Is it you know um yeah

could you talk a little bit about the

the data the types of data that you

bring into it?

>> Yeah we kind of think about it in three

buckets. I mean one is just publicly

available data. So obviously within the

the training data of all these models

and for all intents and purposes I kind

of think of them as they've all been

trained on essentially the same corpus

of internet data. So they know

everything that's on the internet. Um

that's kind of piece one in their

training data. And then on top of that

we do bring in a lot of data from

thirdparty vendors um which is just

fundamental data you know consensus data

stuff like that which again is just is

publicly available. Um bucket two for us

um which is where we we've been trying

to be sort of creative is how can we use

the models to make estimates about

certain things related to companies. So

that may be KPI related estimates. It

may just be strict earnings related

estimates. And so they might take some

of the publicly available data um and it

could include things from Reddit or X.

It'll con it'll include things uh in the

consensus metrics etc. and then create

an interpretation about that in terms of

you know we think Apple will beat this

quarter for X reason. Um, so that's kind

of bucket two for us is think of it as

like LLM sort of augmented data. And

then bucket three is the um, you know,

the proprietary sort of feet on the

street research I would call it where

LLM's um, I think there's a path for

them to be useful there. Um, but I think

it's really hard to do that today. And

that's, you know, your channel checks,

it's calling on customers, it's going to

investment conferences. Um, all the

things that are just limitations to a

digital system, um, are are kind of that

bottleneck. And and to me, that is kind

of the longer term frontier for using

LLMs is can you use them to do channel

checks somehow. Can you imagine a world

where actually agents are talking to

other agents doing channel checks, not

just agents talking to humans, which is

kind of easy to imagine today. Um, and

just what does that look like in the

future? Because I really think that I

mean just like it is today that probably

is the long-term real durable source of

alpha versus kind of just understanding

really big contexts better than humans

right now.

>> Doug, have you thought about the sort of

the MCP pathway? I think the consensus

was, you know, four or five months ago

it was the MCP structure was probably

too brittle for institutional use cases.

I've seen that conversation die down,

but there's still a little bit of debate

on what the right data structure looks

like to sort of pipe in the right data

to to make institutional grade

decisions.

>> Yeah, I I think it's probably still an

open question. Um, for us, I mean, we

use um various MCP uh, you know, sources

that that help us pull in some data. Um

and I think that I think the world

broadly is evolving to this idea that um

getting context to the models at the

right time in the right place is super

valuable, right? And you look at like a

carbon arc, right? Paying paying almost

per piece uh of data and you get to

determine what you think is really

valuable. I think we're going to see a

lot more models like that and and

whether it literally uses, you know, the

MCP, that exact protocol, or if it just

looks like an API or whatever. Um, I

don't know how much that matters

relative to having just the right

structures, the right payment models,

and ultimately the right data sources

and the right ability for the models to

say, I need this piece of data. I think

that's the super valuable piece. Um, and

I would say this, like models today are

I think they're pretty good. If I had to

grade them, I'd give them like a B+ as

an analyst sort of understanding like

this is the thing that's probably going

to move a stock in a given quarter. Um,

their creativity isn't quite there like

a human analyst. I mean, some of the

creative things I've heard for human

analysts doing over the years to try to

get an angle on a quarter are

incredible. But again, think about the

future. Think about a year from now.

think about it two years from now, I

think models will think of those super

creative things and be figuring those

things out for us. Um, and that's the

frontier that I'm really interested in

trying to figure out.

>> Doug, when we uh before we started

recording, you had mentioned the

importance of knowledge graphs as it

relates to MCP. Could you say a little

bit more about that?

>> Yeah. Um

the it's it's funny K like the the terms

as you as you mentioned earlier right

have have sort of evolved and like what

is the hot thing at AI I mean it changes

every six monthsish or so it feels like

um and I would say this this concept of

sort of the knowledge graph feels like

it very much the frontier right now and

just trying to figure out you know how

do you take um all of the knowledge that

you have as an organization or as an

investment team or whatever your your uh

your node structure is and make that

entire graph usable by every agent that

you have running on your platform,

right? And that may be hundreds of

agents or thousands of agents depending

on how evolved you are. Um, but it's a

really complex task to let I mean

imagine thousands of different agents

trying to pull different bits of context

real time creating new context right

that other agents should then have

access to. Um, doing that really well I

think is going to be a source of alpha

for people using language models or just

agents in general I think to invest

probably over the next I'd give it 12 to

18 months. I think that'll be a huge

advantage before eventually it probably

becomes a little bit more commoditized.

Whether that's because some thirdparty

vendor introduces a really easy to use

system which just doesn't exist right

now that I've seen um or right cloud

code and codecs make it so easy to just

say hey implement this and it just

works. It's fascinating too because the

sorry Brett the like schemas and

ontologies again these kind of buzzwords

that no one was talking about six months

ago

>> but it is this this weird idea that you

have this data but you have this

structure that oftentimes is just

intuited by your process by your

analysts by your PMs by the industry and

that might be the gray matter that the

LLM hasn't figured out yet. And just

like watching the the race to

disentangle that, to categorize that, to

label that, to you know, to make it

visible is just uh it's just really

fascinating right now. And and I guess

probably early days to know if it's even

effective,

>> I think. So, and it comes back to this

idea of taste. you know, you hear the

word taste often in AI, which I think is

sort of this nod to longer term human

value, this sort of instinct that we've

talked about. Um, I do think to some

extent like creating these ontologies

and these knowledge graphs, there is an

element of okay fine like codeex pointed

at something, give it a really great

creative loop and it'll probably figure

something out that's passable. Um, but

to make it A+, I think it does need to

have that element of taste from a human

who really understands like this is what

it means to generate alpha. Not just

like what the internet has taught you as

an AI model, but like this is why it's

hard, right? And this is kind of like

maybe even the history of how how alpha

sort of evolves in a in a sense and

being able to take that taste and apply

that to the ontology. That's where I

think you get this unique value from

humans coming in and taking sort of AI's

baseline work which is like B+ and then

getting it up to that A level and that's

where I think you can really get some

additional value from it. Could

>> could we drill into this this concept of

knowledge graph a little bit more? Like

what does it actually look like in

practice? I'd be curious K your

perspective on like how what's the

what's the sort of zero to one like if a

a client comes in wants to build a

knowledge graph what does that mean? Is

that internal file structure? How do I

get my data out of Excel files and

OneNotes and and PDFs and Outlook inbox

and Slack? Like, how do you what's

what's that actual process looking? Oh,

I mean, I'm going to give you the

training wheels version of this, but um

I think that it's a lot of the things

that you just said where you have all

this this data, but even something like

earnings or IVIDA that can mean 15

different things to 15 different people

even within a firm within a subsector

and so on. So if you go in and teach the

model like when I say you know for us

the concept of ibida means you know

these adjustments you don't never do

this you have to teach the model that

and by the way you have to basically

like give it a map that like these are

the ways that we define all of these

things. So once you give them the map

and so this map can be just like a

markdown file right and once you give

them the map then you have this data

right you have let's say 10 10 port 10

companies with you know 10 years of

quarterly IBIDA across 10 different

sectors you have to basically say that

like well the way we look at IBIDA in

software is completely different than

the way that we look at it in

industrials and in doing that you know

when you want to grab that number you

have to make this adjustment you have to

look here, you have to do do that. And

so you're basically kind of putting it

together so that when the model goes in

and you're like, what's the IBITA for

Salesforce this year? It starts to go

down this graph. It's like, well, I know

that this like the way that Brett

defines IBITA is this. And I know that

in SAS, you look at it differently. You

go here, and now I know that there's a

Salesforce data here, but there's this

other data I need to add back in or out.

And so it's basically creating that like

the term is like ontology which is like

again dumb person here but you know the

the definition your definition of what

these words mean so that you can

translate it to the model. So then you

give this you give it all this data

again markdown files CSV files skills so

on um and then you want to query it

right and so there's all these different

ways of quering it right you could do

brute force agentic search which is like

command f ebita but like still go

through the ibita knowledge graph to get

to the data point or you can use like

some of our other guests have discussed

semantic search through rag and

embeddings it's like oh okay you know

you're asking me for uh profitability

ility which can mean a lot of different

things and how does that tie to this

definition of IBIDA and Salesforce and

da da da you can then add like keyword

search and you know different databases

that like have sanitized the data and so

on. So that's conceptually how I'm

seeing folks do it and I work with much

smaller firms where they're just

starting with like a small group of

names and they're just, you know, we're

going to do the Salesforce ontology here

and then we might roll that up to the

SAS ontology over here and just like see

if it works. See if it just like speeds

up the process of getting the right

context at the right moment and with the

lens that I'm used to looking at it

with. And then you're getting dev shops

that are coming in. And like the the

crazy thing about the dev shops coming

in because I've like sat in in a few of

these conversations is that a big chunk

of the time is interviewing a PM. It's

like when you say Ebida, what do you

mean, right? Because that's so baked

into their heads, but you need a

translator. And that's been a a bit of a

challenge with a lot of firms like like

wait, I thought you were going to give

me the answer. It's like, wait, no, you

need to tell me how you understand this

concept and then I'll write code for

you. And then a lot of people like,

well, I don't have time for that. Like,

I thought you were just going to do it

for me.

>> I don't know, Doug.

>> What do you have to add to that?

>> And it's funny like, okay, that that's

what makes it valuable, right? is that

you put your stamp on it. You know, if

you just say, well, just give me your

general definition of how you think

about IBIDA or whatever the metric is,

um, it sort of it defeats the purpose

because then it becomes this sort of

generalization instead of something

specific to the knowledge that should be

inherent to your organization that

reflects how you think about investing,

how you think about the world. Um, and

it does take a lot of work. And I I

think that that's actually, you know,

the most important thing when when you

think about trying to implement, you

know, whether you're trying to do what

we do and really rely on the LLMs to

make investment decisions or if you're

just trying to build better investment

processes that are still built around

humans, it's going to be a big

investment in terms of time, in terms of

building the systems. And u the thing

that I would I would suggest everybody

think about no matter which end you are

on that spectrum just from from having

lived it so far for a couple years is

that things are changing so fast.

There's always this tension I think in

the investment world um for the the

desire for perfection. Like we want to

nail uh 99% accuracy rate on on numbers

or you know whatever it might be. And I

think we're living in a time right now

where you have to accept a little bit

less perfection in the name of speed

because if you go and you you go on this

like sixmonth you know really deep

complex process to build some really

elegant knowledge graph I guarantee you

in six months something will be

different. the whole paradigm will have

changed and you're like, "Oh, shoot. Now

we need to build XYZ thing. That's the

new hot thing, right?" And so, just

don't let yourself get excessively um

you know, hung up on the details. Get

something that works, that's reliable,

and know that it's going to evolve very

quickly over time and evolve with it.

>> Yeah. How do how do you think that sort

of obsolescence risk has been has been

massive like Yeah. And in in in three

months is there a vendor where you just

put a listening device in your office

and it gathers context for three weeks

and then builds a custom knowledge graph

for you, right? Like the sort of crazy

things are in development now. H how do

you think about navigating that

obsolescence risk in in building your

infrastructure?

>> Yeah, we've we've built so much stuff

that we no longer use. Um and literally

some of it has lasted a month, two

months. Um and I I think like with the

emergence of the coding tools that has

accelerated it you know because it is so

easy right now to to write you not even

just like basic code I mean to write

pretty decent code like very functional

programs even for people who aren't like

highle engineers um is is pretty easy if

you know how to use the the systems

well. Um now writing something that

scales to an organization of maybe a

hundred people a thousand people that is

a different sort of challenge but um I I

think that we should assume that that

reality continues to be persistent which

is technology will evolve really fast.

So build your systems as kind of quick

as you can to work really well for the

paradigm today. take advantage of

everything you can today, but be ready

to be nimble and don't have a lot of,

you know, sunk cost in something where

you say, "Well, we can't change that

because we just spent $2 million

building the system, but there is this

better thing." I mean, it's almost like

Nvidia chips, right? Like you buy your

H100s and then the B100 or B200, right,

is way more efficient. You save more

money, you know, per token that you

create by just investing in the better

chips. So you just have to be prepared

to do that just like the data centers

are doing with their with their

infrastructure.

>> I had a a funny story on this. There was

an AI vendor that was very hot on it's

kind of enterprise search uh in 2025

before the MCP takeoff moment which I

would say is like December of last year.

Um and everyone was into this vendor. I

don't want to name it. Uh but was into

this vendor. I did tons of calls with

them last year. And then when the

connectors hit, it went dark. No one

wanted to talk to this vendor at all.

Last week, a friend texted me. He's

like, "Hey, I actually stayed on that

vendor and they have incorporated some

of these elements of knowledge graph,

different semantic search. Permissioning

is a big deal in the knowledge graph.

Uh, and it's really good." That's what

the front what the text told me. So even

in the vendor cycle, you see it. It's

wild. like I would not want to be

building and selling on that side of the

fence.

>> Yeah, I agree. And I think that's

actually a great point too, Kay, because

you think about um in the investment

space, the ability to control your own

destiny too. Um this is just a

philosophy that we have at intelligent

alpha but to the extent that we can sort

of build something ourselves or use open

source and modify ourselves very

quickly. Um, we've tried to very much do

that and not rely too much on thirdparty

vendors because of what you just

described. Like, you know, the vendor

could be awesome when you install them

day one and something new could come out

and and by no fault of their own, just

the fact that they bet on a certain idea

or certain technology, maybe they fall

behind a little bit and six months later

maybe they're awesome again. Like, it's

just really hard to not be able to sort

of control your own destiny given how

fast things are moving. Um, so that's

the other piece for us is just

philosophically kind of how do you think

about how quickly and and what kinds of

investments you want to make in building

your technology infrastructure and is it

something that you want to have a lot of

control over or or are you okay with

relying on good thirdparty vendors for

that?

>> Doug, have you thought about model

routing? You know, sort of feels like

every few weeks you get a new model and

you know the open AI models were

ascendant. You know, this sort of H126

has been all about claude. Now the sort

of conversation shifting back to Kimmy

and Deepseek and the open source models

and to token efficiency argument etc.

How have you thought about evaluating

that systematically and when do you make

the decision to plug in a new model sort

of pivot pivot your routing structure?

>> Yeah, we do a ton of benchmarking

internally. We've actually started to

publish some of this uh work as well

externally um literally today. Um, so

it'll be a few days uh I guess maybe

prior to when this gets uh launched, but

we launched something called the IIA

500, which is a nod to the S&P 500

obviously, but it's a it's a

stockpicking benchmark where we have a

quarterly rebalance structure just like

most uh indexes. And we test 10 of the

top uh models. Top meaning, you know,

most influential. It would be all the

names that you would expect to kind of

be in that group. and we look at their

ability to pick a portfolio of about 200

stocks and then we actually ensemble all

of the picks from the 10 models into a

single portfolio of about 500 stocks and

uh it's performed quite well versus the

S&P 500 over the past year since we've

been kind of putting this data together.

Um so you can check it out at i500.com.

But the the kind of answer to your

question though, Brett, is we're always

doing these tests on an ongoing basis to

see what models sort of stand out where.

And what I would tell you is this. There

um in our view, there is a separation

that is notable between the closed

source models and the open weight models

uh that we see on most of our

benchmarks. In general, I would say when

we're talking about investment tasks,

not just stockpicking, uh the the closed

source models and GPT has consistently

been the best one that we test in this

uh section. Um they always have a little

bit of outperformance as a group

relative to uh the open weight models.

Um so we do see some superiority there.

And I know like you said Claude has been

uh probably the hottest one this year

and uh this is a contrarian take but but

I do think that GPT is probably a better

model overall for the stock related work

that we do. It generally is our top

performing model in most of the things

we look at. Um so that's piece number

one on the open source stuff. I actually

think that to your point is is really

interesting because Kimmy's had a

moment. Uh, Neimatron is is actually, I

think, quite good as well on the

American side. Um, GLM I know was was

hot before K came out with K3. Um, I

think we're going to see kind of the

same thing there in in this open source

space, which is uh just like the clos or

the the closed source models very

rapidly sort of pushing the boundary up

um but maybe not quite hitting the close

source. That's sort of my bet is that

we'll see a persistent gap between

closed and open. Uh mainly because I

think a lot of it is, you know, based on

distillation, how fast some of the the

uh the open source stuff is moving,

which is totally fine, but if you really

want to be on the frontier, and I think

you do want that in the investment space

where 1% edge is important, I think you

really do have to spend more time

looking at the closed source models.

>> Yeah. K, how have you thought about

this? I mean you you've spent a lot of

time with claude specific training cloud

co-working in particular which has been

sort of the default UI for you know mid

to small size funds super powerful

particularly when you sort of set it up

with the right connectors and the right

skills. If funds have adopted cloud

desktop cla code etc and you see this

model evolution what what's next? What's

the counter punch to that to that

evolution?

>> It's I folks are scratching their heads

right now. Um I think that the the

smaller funds that I work with are are

they're not ready to go to the open

source route. I mean a lot of these

funds use managed you know managed IT

services like they don't even have a

dedicated IT person in house let alone a

software engineer in [snorts] house. So

I think the open source question is is

not for the kind of the long tale of of

smaller funds without the technical

expertise. Um I think that uh there

there was a huge pull into claude. I

think some folks are still using chat

GBT uh web because they haven't like

Codex was very scary you know just the

name intimidated people but CEX is

having its co-work

and it's very very good the new codeex

or chatbt work whatever I don't know I

don't even know what to call it anymore

>> they need to fix that they've got a

branding problem you're right

>> they've got a big branding problem I

still call it CEX just so people know

what I'm talking about um and and I

think some folks are starting the folks

that kept their chatbt subscription open

are starting to dabble. I found that in

in the folks I talked to the in the

intelligence so much of the way LM are

being used is is like the endtoend

workflow or as much of the workflow that

you can capture. So yes, obviously you

want the model that is 2% 3% better, but

if the harness isn't there, then that

raw intelligence doesn't really matter

to my clients. And so they are kind of

poking around on codecs. I think that

until they fix the branding problem

though, no one's really pulled the

trigger because they no one understands

what

codeex is, but it's really co-work. Uh

and I think that's that's the problem

that people haven't uh had. And I think

that just the the issue of cost I think

one of the things that's really

interesting about the chat GB the codeex

versus co-work conversation is that as

Doug said that the 5.6 model series is

very good. It's very good at agentic

work. I can't speak at the the nuances

of financial analysis, the edges of

financial analysis, but it's also

significantly cheaper. And uh I just as

a as a tangent, I've got a $200 Claude

Max plan that I I I personally that I

get close to using up um regularly, and

I've got a $20 GPT plan, and they're

generous on their resets, but I it takes

a long time for me to hit the $20 limit

on the GPT uh plan. So, I think that

folks are trying to figure it out. I

think eventually they'll get to a dual

place, but that's really like an IT and

a mind share question until there's like

a 50 a 20% delta in performance that's

not there at the moment.

>> Yeah. question for you Doug and curious

your perspective on this 2K like in

other areas we've seen the sort of uh

rapper businesses become ascendant again

like Harvey and Lora sort of solves a

lot of these like you know model router

and data wholesale aggregation you know

commercial questions um you know we've

seen that in certain areas of finance

like investment banking with rogo really

sort of ascendant we haven't seen that

yet in investing yet. Is that due to

just the heterogeneity of our craft or

do you think that's just a timing issue

that we'll start to see that moment

where you see some leaders uh you know

Alpha Sense is sort of most adjacent to

that fact that's building out an AI

workspace etc. Um Bloomberg's sort of

perpetually 18 months behind the

frontier, but I'm curious if you've, you

know, heard any interesting vendor

stories you think we'll start to see,

you know, sort of the right workspace

being this finance specific workspace.

>> I haven't heard any um sort of under the

radar companies or workspaces yet. U but

I wouldn't be surprised if somebody does

try something here and maybe it is an

Alpha Sense or or somebody like that. it

would sort of make sense they might try.

Um, we've talked to and I know K you've

talked to a ton of people in the space

too. Um, we've talked to a handful of

pretty big asset managers uh over the

past year just about you know are there

ways we could help them solve problems

and and one of the things that always is

an issue or a sticking point with us

working with them in any capacity is how

do you get through uh compliance related

issues? How do you get through um sort

of data sovereignty related issues and

making sure that you know if they're

sharing data with us in some way, our

systems aren't using that data, that

data is not available to any of their

competitors maybe that are using the

same system. And so um these are all I

think solvable problems, but I think

they're um they're easier to solve

technically than they are to to solve

organizationally, if that makes sense.

And I wonder if that level of headache

is the thing that's maybe holding us

back more than, you know, just the

reality that the technolog is probably

ready. It's, you know, are the

institutions ready? And and our um very

large financial institutions, which

which I would argue, and I hope you guys

would agree, are are maybe not the most

aggressive as it pertains to being

innovative, are just really slow to try

to figure it out, too.

Yeah,

I I would say that I have seen a small

resurgence of kind of like the wrapper

vertical tools like this first pass tool

or something that just gives me all the

investor transcripts and so on like

cleans it up, you know, the three steps

that you had to do in Clawude to get to

sanitize and get that data into the

working format. and people are more open

to those. But it's a little bit aside of

your uh outside of your question, Brad,

in terms of the platform. Um I think

that

people want to people kind of want the

iPhone of all this. It's like you push a

button.

>> Yeah.

>> And there's a recognition that the

minute it becomes the iPhone, then

you've flattened out the edge into the

product,

right? And so you need to have an iPhone

that's more like an Android where you

can customize this widget and this

widget and this widget and this data

source, but then you're kind of back at

the same starting problem. And so I

think that people are um it's a big

change management question. And I think

one of them, one of the questions that I

think larger firms will ask is like what

will get people to use the thing, right?

And if it's giving him an iPhone, then

give them a damn iPhone. But I think

that there are always going to be firms

that are going to say like, I want you

to have the iPhone and I want you to

have Chrome because I want you to take

that widget because we never do the

widget that way.

>> Yeah.

>> Yeah.

>> Yeah. And then the sort that that

intersects the sort of obsolescence or

the bitter lesson dynamic. Like I look

back to our fall 25

>> curriculum where we were teaching people

about chat GPT projects and downloading

documents, uploading them, creating

system prompts, all of this which is

like completely obsolete now. So, how do

you deal with a change management

problem while the sort of like the

baseline of what needs to be changed is

constantly evolving and and and and

continually up for for debate? It's a

it's a hairy

>> lots of work for people like us.

>> It's a hairy it I think, you know, and I

think a lot of um a lot of finance

organizations, larger ones, even

mid-size ones, I think they want to try

to build this themselves in certain ways

and for certain reasons. compliance

related, right? Control related, all

these things. But working with somebody

who knows and has a vision for it, I

think is super important. K, I know you

work with a lot of companies that that

are trying to fix figure this problem

out. Um, but the but the thing I think

every firm needs to think about is they

all need to operate a little bit like a

startup, you know, and and to the extent

that you're not willing to maybe break a

few rules and and really try to figure

things out um in a unique way because

things are changing so fast, I think it

will put you behind the firms that are

willing to take the risk, invest in the

technology, invest in, you know, maybe

even some of the governance risk and and

whatever else comes with that. um and

understands what this new world looks

like with LLMs being a much bigger part

of all investment related processes.

Yeah, the tricky part is if you look at

like the top 100 hedge funds, you know,

certainly near the top of that stack, I

think most funds will build that

inhouse, but even quite large funds, you

know, multi-8 figure, you know, aum

billion AUM funds like there many of

these teams, the investment teams are

still quite small, you know, 10 to 15

people and, you know, relatively small

internal IT teams. This is a big lift to

build these not just to build the first

version but sort of navigate this

obsolescence risk and eval structures

and um it's a it's not a it's not an

immaterial expense. this is not an

immaterial

um you know corporate strategy shift

>> and and to to just piggyback off what

Doug said the need for a vision right

you you you outside of the citadels

where like you know there's hundreds of

people that are paid a lot of money to

solve this problem but once you get to

that tier below you need to have someone

in leadership that's AI pill because

they will not break the thing that needs

to be broken

they will not take the risk that needed.

I mean there are I still talk to very

successful firms like we have web search

turned off on claude like what like I

didn't even know that that was a feature

uh let alone in 2026

right like you need someone with that

vision because from that vision you're

going to nudge compliance you're going

to take a governance risk you're going

to pay some money to give some analyst

who's you know also AI pill a bunch of

fable tokens to go go figure some stuff

out and that that is a hard those are a

lot of hard pieces you know it's like

playing chess like with all these pieces

to move around and someone added and

then the chess table board is vibrating

at the same time so like the piece that

you thought was in the corner is

actually have moved over one cell by you

know within you know a quarter.

>> Yeah. Yeah.

All right, Doug, I asked I asked I asked

Fable, "Build me a stock portfolio of 10

stocks." All right.

>> Okay.

>> Uh, Compute Substrate is 35% of the

portfolio. Nvidia is a 12% position. TSM

10, AVGO 8%, VRT 5%. I didn't I didn't

tell them make an AI portfolio. I just

said build me a portfolio of 10 stocks.

The distribution abstraction layers 35%.

Microsoft, Meta, Amazon, and now another

35%.

the token beneficiaries, Palunteer,

Spotify, another 20% and then it's

giving me a 10% cash or hedge. So

without any prompting it built like a

super concentrated high octane AI

portfolio, no biotech, no consumer

staples, no REITs, no banks. Um, so

yeah, over the last 36 months that

probably shows like massive alpha, but

how do you like, you know, sort of when

the tide goes out, this portfolio is

getting absolutely smacked, right?

>> Yep. Yeah. I think that goes right back

to thinking about context and and what

can you give the system to help it

understand the current regime. And if

you think about how Fable, right, if you

just use Fable for a general task today,

the knowledge cutoff, I believe on Fable

is about April of 26.

Um, so think about the window, right?

The last kind of information the model

had was Q1 ostensibly of this year. AI

trade was ripping then before we had the

the pullback kind of uh March and then

ripped right back after that. Um, so the

model is probably thinking, okay, well

that's what the world looks like and so

I want to have this really super AI

aggressive portfolio. Um, knowing that

if you put anything into a model,

finance related or otherwise, that the

model is going to have some sort of bias

related to its most up-to-date context.

I think is is piece number one is

understanding that you're probably going

to get something that looked really

smart based on something that happened

in the past and your job is to figure

out how can you get it current

information so it understands the real

world um and hopefully can update its

priors. Um and that's what we try to do

at Intelligent Alpha. you know, we would

uh in that example, right, if we were

just going to do a really simple thing

with Fable, uh we'd probably give it

some current information about recent

financial reports, earnings reports, uh

what consensus expectations look like,

um recent transcripts, and so then

hopefully the model would say, okay,

look, the AI trade has been playing out.

Maybe it would even know we had this

July pullback and uh hopefully it would

adapt in the right way to go forward. I

mean, me as a human, I'm still pretty

bullish on the AI trade. I think after

the July wash out, um, if you look back

at the dotcom era, it looks a lot like

1998, kind of LTCM. There's there's kind

of some vibes there. And, uh, the world

got pretty crazy after that in terms of

the dot trade. And so, I'm not saying

that's exactly what's going to happen,

but I think the AI trade is is kind of

far from over. And and uh, I'd be

curious to see if the machine would

agree with me. Yeah, I'd say at like a

high level like I've been pretty

skeptical of this sort of new emerging

AI hedge fund, AI asset management

business where you let the machines pick

the stock. I mean, the last few weeks

I've been a little bit more open-minded

to it. I sort of agree with this sort of

context being a critical piece and this

decomposition of the investment process

into actual signal. Um, Parvis just like

the recent model like you know Fable's

not perfect but the Fable soul ser you

know Kimmy K3 series of models like are

pretty surprising me to the upside on a

few of these pretty complicated tasks in

terms of adherence to a to a complicated

skills architecture. Um so it's going to

be an interesting 9 to 18 months to see

what uh see see what emerges. uh what

what advice would you give for for those

you watching this pod who are thinking

about maybe uh you building an asset

management business with with AI?

>> Yeah, I think just like in investing,

right, you want to look 6 12 18 months

out what you think the world is going to

look like then. Um because ultimately if

you're right about your prediction about

the world, hopefully you're you're right

about what stocks you buy ahead of that.

and um thinking about building with

these models um we're always trying to

figure out okay we know they're going to

get better I think that's a given if

anybody is following the AI space in 12

months these models will be more capable

um but like what are the bottlenecks

today that we think will be uh unleashed

with additional capabilities how can we

prepare ourselves to be ready to sort of

take advantage of those uh things that

get unlocked as soon as possible um and

stay at the forefront front because I

think uh to the extent you can stay at

the forefront I mean there's this great

quote that I I say all the time

internally from Paul Book height um who

was at Google and then Y cominator he

said if you're in the lead if you just

keep running faster than everybody else

no one could ever catch you I think

that's true no matter what you're trying

to do with these models if you can

really stay on the edge and keep

experimenting don't worry about

perfection even though that's really

hard to say in the investment world but

just keep iterating I think that that

will put you in a really good place to

find ways to generate alpha with these

models all the time.

>> Yeah. Yeah. I've been a hater on a lot

of these things or at least trying to

provide some skeptical, reasoned,

empirical push push back. And one of my

friends says, Brett, evidence is a

lagging indicator in AI, right? You sort

of have to operate and build the system

on faith that the thing you can't do

today will be possible in three months

or six months. Um, so I'm trying to take

that to heart as I think about you where

we could be in the spring of 27. It's

also becoming easier because the

evidence of where we're at today versus

six or nine months ago. If I sort of

pull up my outputs of what I could do in

summer of 25 versus summer of 26, it's a

fundamentally different different human

uh fundamentally different, you know,

sort of intelligent source today. Um, so

where we're at next summer kind of hurts

my brain a little bit to to think about.

Any thoughts? Uh, any sort of closing

thoughts or comments? Uh, K?

>> Yeah, I think um that that the ability

to be uncomfortable with uncertainty,

right? And kind of like leaning into it

versus kind of cra uh holding on to the

determinism. But you know what I got

from from our talk with Doug today is

just It's it's messy, right? And I think

a lot of the conversation it's an

industry that doesn't like to be messy

for the right reasons. Like there's a

there's significant pen penalties for

being messy in this industry. But to the

extent that you can kind of bring in

that that experimentation, which is like

what the three of us on this episode and

many of our clients are doing, you know,

again, those pockets will lead you to

kind of stay, you know, skate to where

the puck is going.

Yeah. Yeah. One of the hard hardest

questions I get is almost a simple one.

It's like, "Hey, Brad, I want to set

this up. What do I do?" You know, what

vendor should I use? I'm like, I don't

know. It's kind of like still three

three years in or four year almost four

years into the chat GPT moment. It's

still kind of a hard question. There's

not yet an obvious answer. It's a little

bit more of a process, a journey of

experimentation and learning, which is a

little bit unsatisfying. But the comfort

with uncertainty is probably the exact

right. I think you nailed it. Nailed it.

Nailed it. Nailed it. K, what what what

say you, Doug?

>> Yeah, I think on that on that point,

Brett, the the best way I always try to

tell people to get involved with these

models for any purpose, whether it's

finance or otherwise, is you have to

find something that you're really

curious about and you just get lost in

the models doing. And that could be

designing uh a new clothing line. It

could be writing some piece of content.

could be trying to pick, you know, some

super aggressive AI related stock

portfolio, whatever it is. Um, the more

that you can just get lost using the

models and then understanding like the

intricacies of how they work,

understanding where you can push the

limits and the edges and what you can

have it actually build for you. That's

the best thing you can do is just kind

of learn it by feel by doing something

that's really fun for you. And then you

can take that curiosity, all those

learnings that you have and apply them

to something that might be a little bit

more structured and intentional. U but

until you kind of get through that and

you just experience the models, you

know, unfettered, right, and just let

them go wild, I think it's really hard

to kind of think about the structured

thing because you just don't even really

know what they're capable of yet.

>> Yeah.

Well, thank you so much for uh for for

joining us today. This has been really

fascinating and I'm super excited to

see, you know, where where you're at on

this in six months or nine months or 12

months. Maybe we'll come back in. It's

like, you know, we got the keys. We

figured it out now. So, uh, thank you so

much for for being with us, Doug. This

was really fun fun and an interesting

time in an interesting space.

>> Absolutely. Thank you guys, too.

>> Thanks, Doug.