All right. So, I

mentioned last time webinar one

um that it was a little bit longer than

I had expected.

Um and I went like 82 slides or

something. I mentioned the next one

would be shorter and I was completely

wrong.

I got really excited and built 102

slides.

Um so, we're probably not going to fit

this all into an hour

is the bad news. The good news is this

will be recorded and recorded and

dropped into um

dropped onto YouTube and will be emailed

out. If you're registered for the Zoom,

you'll receive the an email with the

deck.

Um

So, with that, I will go right into the

slides. We will hold off all Q&A

Q&A until the end. So, I initially

labeled this um

session up to speed with Claude Co-work.

We're going to revise that a little bit.

I'll get into the context

uh context of why,

um but we'll talk about a few tickers

today. So, just disclaimer, you know,

nothing is investment advice. And if you

listen to a random guy on a webinar

about uh stocks,

there's probably more more more issues

anyway. So, yeah, don't take anything as

investment advice. Obviously, reminder

of my background. I sort of come through

I sort of come to this session one with

many priors on the applied adaptation

intersection of fundamental investing

and really in the last 4 years more of a

parallel experimentation with these

tools. Um

So, that's really the lens I come from,

not from

a technologist perspective by any means.

I'm

really not a technical person.

But from someone who's really taken a

lot of time to decompose decompose the

workflows and sort of in my personal

journey from you know, for 11 years I

was in New York. I'm not pictured in

this picture because I was likely at a

hedge fund preparing for earning season.

And so, I moved to Arizona 8 years ago

and I just have more time to go through

and experiment with these tools. I

have sympathy for those of you who are

trying to manage capital while also

staying up with AI. It seems like an

impossible task. I don't manage capital

and it's very hard to stay stay up with

with AI. I think a lot of what you'll

see here comes from my historical

priors, but just see AI experimentation

journey that I've been on. I've sort of

been looking around for experts who can

teach me all these things and I'm like,

well, there's no one. So, let's try to

try to see what we can do about being

the one-eyed man in the land of the

blind and

spent a lot of time a lot of hours in

this space.

A combo of things that have worked well

for me if you're inspired on the same

journey is you know, finding the right

people on YouTube and Twitter to learn

one

learn from lots of conversations, a

parallel experimentation which is sort

of the the heart of our pedagogical

approach in AI. Let's show the manual

approach and then the AI approach. Some

client engagements and as I'll show you

today, I'm sort of serving as guinea pig

with the mindset that, hey, I'm building

a

a portfolio

to invest

the same way that you know, AI can drive

you now, can it help me ramp up a mock

portfolio.

Almost certainly can't do all of the

elements of that process, but I'm in

exploration of what elements it can it

can do. And so,

the the fun one someone pointed out

Cunningham's law which is states that

the best way to get the right answer is

to post a wrong answer on the internet.

And so, the reply guys all give me all

sorts of

great feedback, pushback, some

often accurate. So, OX fuckarooie, thank

you for pushing back on my tweet and

pointing out something that was actually

half right.

So,

you know, I've learned a lot just by

conversations and DMs and being active

on active on Twitter. One thing that um

sort of the overarching concept that all

builders in the space are struggling

with right now is the bitter lesson the

fact that building with building with AI

right now

is sort of a deeply

sort of deeply bitter lesson

pill example that the concepts that we

taught in our AI Academy in September

like retrieval augmented generation or

really have been leapfrogged in new

technologies and sort of been hearing a

lot about this concept of the bitter

lesson of AI engineering which is

effectively that the scaffolding that

people are building

right now is just a temporary

compensation for models that aren't yet

smart enough and context windows that

aren't yet big enough.

We've seen all sorts of

examples of this in

all sorts of examples of this in um

in AI whether it's vector databases and

chunking or rag or OCR or prompt

engineering mechanical tricks and

and right now the hot thing is MCP is

that next in terms of abstraction.

There's a debate going on there as well,

too. So, my sort of advice, you know, is

to

the 95% of of normies like me that

aren't deeply technical is

don't necessarily worry about setting up

your Python environments and becoming

master with IDEs and you know, getting a

GitHub account and all all of this. My

sort of belief I'm operating on some

degree of faith that the bitter lesson

will continue to to hold and that much

of this engineering will become abstract

rapidly and I'm seeing that today.

You've seen it in vibe coding tools like

Replit lovable.

And you see it in in tools like my now

favorite Perplexity computer which is

really you know, deeply abstractive

highly effective

highly effective agent and I'm not

sponsored by Perplexity.

So,

so also this just sort of helps me

manage my existential angst to try and

stay at the intersection of AI. I wake

up sometimes and like, you know, am I

going to be part of the permanent

underclass? I better learn this AI

thing. Um and so, really starting in the

fall of September 24, we started with

the cutting edge webinar hosting AI

vendors. We ran a you know, 6-month

guided cohort that finishes this month

called AI for investment research with

PMs and analysts.

A lot of this now the red the reds

crossed because a lot of these things

are now decayed in terms of in terms of

relevance. So, I'm going through the the

rework process right now to figure out

what's decayed, what holds and what the

future looks like which is a little bit

hubristic to say, but I'm going through

the deep experimentation

a process. A couple things

to keep in mind on that.

The world doesn't need another podcast,

but apparently in this space it does and

so, we'll be

we with my friend KE be rolling out a

podcast invest with AI

soon in in June. We'll take some of

these concepts we're talking about today

and

and put that into an accelerator sort of

a guided structured implementation plan

in a cohort structure to do this.

So, that's a little bit of an intro and

welcome.

I want to talk today we're sort of the

core today is to talk about how to get

up to speed

on a name. And I need to start with the

concept that's widely discussed of the

mosaic approach to investment research

which sort of speaks to the fact that

insights on companies don't come out of

one single step. You know, almost never

in my investment certainly never in my

investment process have I just read the

10K and developed a differentiated

insight. An insight comes from the sort

of complex ongoing building of a mosaic

about companies and about stocks.

And you don't necessarily know exactly

what step of the process will develop

that insight, but the combination

[snorts]

of covering companies closely

sort of causes this formation of

differentiated insights that you can

invest with with conviction.

And so, that's sort of a I think think

hopefully non-controversial

a concept, but I think important when

you start to hear

questions about what AI can do for

investment process. You know, the sort

of the belief is that, hey, we can go

one shot this investment process and

just you know, have AI give me a view on

what to do. And almost certainly you're

going to get slop out of this out of

this output. You're going to get the

fact that you know, the LLMs on a one

shot basis are trained from the open

web. You're sort of getting maybe speed

to consensus which is sort of

definitionally not not not helpful for

for alpha generation. You're not going

to get primary source contact and you

know, the particular sort of you know,

risk here is it is sort of sounds

plausible. Sounds right.

But then so, with no validation, this

can be a very dangerous

process. So, sometimes I talk to friends

like, hey, I tried to use AI

in my investment process and it didn't

work. I'm like, well, what did you do?

I'm like, well, I asked ChatGPT what

they what it thought about this company.

I'm like, well, yeah, that's not a very

that's not an effective way

to do it.

Let's sort of talk talk about you know,

building a system around around this.

I'd say sort of other other notable

thing is that a one shot approach is

very likely to get you slop and in our

high stakes career path,

that's a real bad that's a really bad um

outcome. You know, things that could

happen. Hey, you you know, you want to

do AI prep for a CEO meeting and you ask

you know, CEO about a business line

that's been discontinued for 18 months.

Or this is a real story I heard. Someone

did an AI news scraper and it said that

Microsoft's revenue guidance was cut by

10%. If you don't validate that output

and you trade on that in the pre-market,

that hallucination has a direct negative

effect to P&L. Or for some reason, you

know, all of the modeling workflows the

the LLMs are highly obsessed with having

a balance sheet balance, right? That's

sort of seen as like the number one

thing to do in a financial model. You

know, my 150 models maybe only a handful

do actually model out the balance sheet.

We're focused mostly on the P&L and and

key drivers. And so I've seen this in a

few AI models where the numbers look

right and the balance sheet balances,

but the numbers when you actually go to

validate them are

or or or or or or or wrong. And so AI is

designed to produce plausible sounding

uh plausible sounding output. So I say

above all, what we're trying to do here

is we're trying to optimize for rigor,

trying to build a system around that.

You know, I'm I'm sort of work operating

in this framework of a like-for-like 60

hours. I don't think the play here,

personally, is about cutting head count

right now. It's about in 60 hours,

right, finding a richer signal, right?

If I'm doing looking at a name, if I'm

looking at Danaher in the manual

approach or in an augmented approach,

after 60 hours, what outcome gives me a

more rigorous, comprehensive

understanding of the business? That's

the sort of the North Star for me. And

so that can include, you know, we that

could be the exoskeleton mosaic, some of

the old manual processes that you did,

but it could also be layering in

thoughtfully AI augmented pro pro pro

pro pro processes. Why is this

important? I think a lot of the fear

about what AI will do to the investment

process still operates under the old

paradigm of markets, that markets are

beat

based on better information, right? You

know something that others don't know.

Uh whether it's all data or the

democratization of in in information on

the internet, you know, this has been a

compressing alpha pool. Almost never do

I have I sort of been in a situation in

the last 10 years where I feel like, oh,

I found one singular piece of

information that is a source of my alpha

in a name. Almost always is some

combination of behavioral and integrated

perception, right? Hey, the all these

SaaS stocks sell off because of cloud

fears, uh you know, these five that

might be legitimate, but these five have

some sort of element because I know the

industry better, the market's freaking

out. I have sort of developed this

worldview by covering the space more

effectively. I'm going to invest because

there's a meaningful discount to

intrinsic value. And so, you know, I

think, you know, you you heard it from a

number of angles, but there I think

there's, you know, market microstructure

considerations why markets are more

inefficient than ever. They sort of jump

from crisis to crisis and overreact on

the upside and downside.

And so fundamentally, that's an

alpha-rich environment for behavioral

and perception alpha, not necessarily

for informational. And so if the fear is

I can go figure out this information in

a chatbot, there's all sorts of reasons

I think there's flaw flaw in that logic.

Uh that's really not the game we're game

we're playing. The more interesting path

is if I take 60 hours in a name, right,

what's my like-for-like? So in the past,

it might take me 60 hours to really get

up to speed on a name, maybe doing a

little bit of primary research. But if I

can find a way to accelerate these

mechanical approaches, right, take this

60 hours down to 20, right? That doesn't

mean that I'm leaving the desk at 2:00

p.m. in the afternoon uh to go play

tennis. It means that, hey, I've got 40

hours left that I can redeploy to

optimize my rigor and comprehension in

this name. So how do we redeploy that

time that time saved? Um I think there's

a combination of things. Some of it's

just the good old-fashioned scuttlebutt

primary research. There's a number of

other AI augmented deep dive workflows.

There's I think all sorts of, you know,

concepts out there that people are

talking about AI. One is that AI is

great for breadth, but not depth. I

think that's true. Chatbot's not of

agents. I very much disagree disagree

with that on agents. There's there's I

think a number of working sort of

emerging workflows that can really help

you go much much deeper,

uh much much deeper. A simple example

would be learning about a product. I can

spin up an AI survey, uh AI web

scrapers, you know, a deep deep research

report on a product. I could I could

create much more rigorous understanding

with an AI process about a company's

product than I could uh sort of the

manual manual way. And so I think

finding ways to deploy new workflows

to go deep, again with the objective of

we're enhancing or enhancing rigor.

>> [snorts]

>> I think in the past, we're going to talk

about the two fundamental workflows of

up-to-speed and in the flow today. One

of the things that that chatbots

couldn't do is they could you couldn't

build an automated tracker. You couldn't

identify the one thing, you know, the

key thing, the key drivers, risk

factors, et cetera, and track those

autonomously. Uh now you can. Um so the

other thing that people will say is, you

know, LLMs don't have judgment. And and

a stateless chatbot certainly doesn't

have judgment, but I think the agentic

approaches that you can wrap around uh

the PM sort of applying the judgment. Um

in my sort of hypothesis can certainly

rate raise batting average, raise

slugging percentage, eliminate big

losers more as well, too. So I I sort of

feel like a crazy person

um lately. I hope for many of those many

of of you watching this, you followed

some of my stuff on Twitter over the

years to sort of know that I wasn't a a

day one AI maximalist by any means. I've

probably probably more cautious on these

tools up until 6 weeks ago

uh than average. Uh but I'm trying to

just move with the flow of the things I

see, move with the feedback of the

outputs that I that I see. And um I'm

really excited about what's possible. So

I'm excited to be able to share a couple

things with you today. I think the other

thing I would note

when you talk about using the tools is

investment process is is sort of

definitionally hetero geneous, right?

Whether you're at a long-only that focus

on business quality, deep focus on

management, that's going to be a much

different design system design than if

you're a multi-manager that's highly

focused on the earnings motion, focused

on many management management touch

touch points for in for inflection. Same

thing is the core part as we talked

about last week is not to take, you

know, everything I'm showing you as your

core operating system, uh but to think

about building your own exoskeleton,

right? And that's the sort of the

workflow that I'll I'll walk you through

walk you through today.

With the sort of core prior that I have

having worked with many great investors,

is that there is a very close

correlation to comprehension, business

comprehension. Um you know, my prior is

certainly that the best investors I've

worked with, you know, know their

companies the best, right? They sort of

deeply understand how decisions are made

and the competitive dynamic.

And so, you know, we could one-shot, you

know, a thesis uh in AI, but that's not

going to drive deep comprehension of the

company, right? We we're we're

optimizing for the closed-book test

where I can sort of sit there and

explain how how the how the business

makes money, the deep nuance of the

business. And so the question is, how do

we drive deeper comprehension in the

same

the same same time window? Why does

comprehension matter? I mean, you get

into situations where if you have weak

comprehension, you know, you're sort of

pulled by the nose of the shifting

sentiment of the market, right? Hey,

Morgan Stanley said this could be bad, I

better sell, right? You're you're

listening to the noise

and you're being buffeted by the noise.

The best investors are not buffeted by

the noise. They in in in fact arbitrage

the noise. They have strong

comprehension. They say, okay, the

market's freaking out about this thing,

that's stupid. I can't believe Morgan

Stanley think that thinks that's an

issue. No shade to Morgan Stanley, this

is hypothetical.

You know, stock's off 9% on that, I'm

going to buy the stock, right? And so

that requires a baseline level of

comprehension.

Um

you know, either on an oversell or, hey,

there's something emerging coming down

the pipe that this could be a this could

be a real this could be a real issue. Um

uh this could be a real uh a a real

issue, right? And so I would say in

general, um you know, AI sort of on a

one-shot uh basis is a um is an enemy of

comprehension. And so I think being

careful to build a system around this is

really really important, probably one of

the most important things that, you

know, I would I would take as as

takeaways. Um I sort of tell the tell

the lived experience story of my of my

middle school son that I encourage him

to use more AI only to find out that he

was just sort of speed-running his

homework and when test time came, things

didn't go well. So we had to have

another conversation and say, hey, we

want to use AI as a tutor, right? It

sort of checks your work, you did it

yourself,

uh give you some extra questions, test

prep, et cetera. And so this is sort of

a a rubric to me or a framework to me

that matters to to investing, right? We

can either AI a bad process or you can

AI in a good process.

And AI is really a force multiplier. It

can it can help you be lazy, uh it can

help you speed-run something that looks

close, um or it can really help you, you

know, develop much more rigorous

rigorous work. And so the one thing I

know for sure is that we all have the

same pie of time in a year. We have the

same roughly 3,000 hours. And so the

question is, if I if there's a lot of

things sort of mechanical processes in

these 3,000 hours today, and I can trim

down that time, do I have more time for

value value-added work? This is

Investing is incredibly competitive game

of poker, and so optimizing that

um seems likely to me to be, you know,

helpful to the invest invest investment

process. I'm also starting to tie a

little bit of this back to, you know,

sort of the the core objective we should

all

be focused on, which I think gets gets

lost in, you know, the open claw AI

productivity theater a bit, that the

core question is, you know, we're

spending a lot of time talking about

this, talking about these tools. How can

we actually use these tools in these

processes to drive better performance

for our investors? Whether that's

eliminating big losers or

finding more big winners or better

batting average, entry/exit timing,

better sizing. And so in some of my

work, I'm going to take some of this

from conceptual back down to

um really trying to tie this to um tie

this to um

uh tie the tie this to performance. I

also sort of think, you know, it's it's

very philosophical. The question is, you

know, are we are we cooked as investors?

My general sort of prior belief is that

there is this sort of, you know, alpha

window hypothesis. I take this from the

alt data sort of experience that alt

data was a sort of a a nice driver of

alpha for a window and then it became it

moved from alpha to beta.

I think also institutional capital flows

are pro-cyclical. So, the time whether

it's 5 to 7 years

where you can adopt this in your

investment process and drive better

results is that drive more capital

inflow and winners can lean into

advantage and push the alpha frontier

which sort of always been the prior in

investing. Does it eventually compress

as markets adapt? Possibly. I think one

thing that is on the side of fundamental

investors is our craft has been under

pressure a bit. Capital flows into

indexers,

quants, and retail's non-fundamental

investing sort of actually creates could

sort of extend that window for us. So, I

actually think that the more I think

about this, I think it's a wonderful

time to be starting a career in

fundamental investing. The tools the

tools are better and the setup for alpha

generation because of these market

microstructure considerations are are

helpful. Now, so I'll stop beating that

drum and we'll get into the real

question of how AI will change

fundamental investing. Before we talk

about exoskeleton,

let's just talk a little bit about

skeleton. I think one of the things

that's really helpful as you think about

sort of deploying AI is to think about

your research process, right? And one of

the core things I want to show you today

is this

sort of workflow context and how helpful

how much leverage there is in developing

workflow context. What does that mean?

It's just writing what you do and why

why you do it. And so, I'll show you a

few things that we use in our analyst

academy as as context. I sort of think

about you know, the the investment

process as two arcs. I think about it as

a thesis development arc, right? Kind of

the up-to-speed from from idea to

thesis, right? And then I think about

the as the active position management.

So, once the idea is in the book, what

do I do then, right? You know, how do I

monitor the catalysts and earnings and

management touchpoints and news and news

and event navigation. This is more of a

deeply Bayesian deeply tracking focused.

And I think when you think about

deploying AI, these are two separate

buckets in a way that have sort of

different benchmarks on data accuracy

for sure. Part of the reason that so

much of the focus has been on the thesis

development arc or the up-to-speed

up-to-speed process, as we'll call it

today, is that there's a natural

validation loop, right? If I'm sort of

thinking about populating a portfolio of

30 ideas from 3,000, right? A lot of

this process is just a winnowing of

ideas is qualifying or disqualifying

names. And so, this is great. You know,

AI hallucinates, but AI also could be

highly effective for the sort of

winnowing process. I don't put a name in

the portfolio without you know, human

human level validation, but if I can

turn over many more stones, I can

ultimately drive a portfolio that's sort

of more rigorous

with

sort of more compelling ideas. So, I

think sort of version 1.0 of our

curriculum has pointed people more

towards trying to drive efficiency in

this idea generation or hypothesis

formation, right? So, for those of you

that were in the AI academy in

September, that's where we spend a lot

of time. Partly because chatbots had

fundamental limitations fundamental

limitations. So, you know, the two

fundamental workflows

of up-to-speed and in the flow, you

know, the up-to-speed is more of a

triage function a little bit. Certainly

the early steps in in the triage triage

function. The in the flow is a little

bit more of a you know, Bayesian or edge

function where if you're wrong and

hallucination is much more expensive on

an in the flow function that drives a

training outcome, right? Which is also

why you see a really notable delineation

in adoption in AI between generalists

and specialists. My friends who run

small under-resourced generalist funds

have adopted AI to a degree that's much

more dramatic than my friends who were

at large well-resourced specialist

funds, which is sort of the flip of what

you might think. But if you really think

about it, it makes sense, right? A

generalist is looking at 3,000 ideas. A

lot of the workflow of a generalist is

this triage up-to-speed function, right?

Then you have a portfolio where you're

maintaining that.

The typical well-resourced institutional

manager, you know, you have analysts

who've covered healthcare for 10 years.

They're not reading the 10K for the

first time. They know these companies

intimately. A lot of the flow is active

coverage, right? It's the daily blocking

and tackling of news and earnings and

calls and panels etc.

And chatbots were really not good for

that, right? The bar for in the flow is

much higher. Again, we sort of put all

these workflows into red green light not

to not to say that a red light can't be

turned green light with a proper with a

good system, but in general, we've seen

a lot more red red red lights in our in

the flow, right? News monitoring, right?

If I want to take an AI agent and have

it distill my inbox in the morning, it

had better be right. There better not be

hallucinations because if I see

something, I might go trade that in the

pre where there's not a lot of liquid

liquidity. Whereas if I see a

hallucination in a quick story or a

sniff test, by the time I'm doing the

full ramp, I have a validation workflow

to go back and check those numbers,

right? So, there's a natural prevention

mechanism

of any errors reaching reaching

portfolio, which is not true if I'm

trying to say, "Hey, you know, hey

ChatGPT, here's the earnings press

release. Should I buy or sell the stock

at the open?" Like that that's not a

workflow that should be

should be

adopted adopted easily. So, I think

people have you know, even the large

funds have you know, sort of

adopted the chatbot architecture. People

are sort of reworking that now for

agentic architecture, but the in the

flow architecture probably has more of a

an alpha there's more of an alpha

opportunity there for sure versus an

efficiency opportunity in the in the

up-to-speed. So, this is where I'm sort

of like going crazy now because

it's just like wow, like this is the

pieces are there. It's not still not

perfect that yet, but some of these

things that were purely conceptual. So,

we'll be back more on the in the flow

stuff. That's where we're doing

having some interesting client

conversations. But today we'll talk

mostly about the up the flow up-to-speed

triage consideration and I'll start by

just telling a story about Bayer, which

is you know, a

European pharmaceutical company. I

looked at this name 10 years ago, 8

years ago, something like that and as a

potential long. And the idea was there

was an acquisition of Monsanto. I did a

little bit of sketch pro forma analysis

and it was like five times earnings on a

pro forma. I'm like, this is like pretty

you know, pretty interesting, pretty

cheap if these numbers hit and the

synergies synergies hit. I thought,

okay, put a 10 multiple on that in 18

months, it's a overly hated, it's a

double, right? So, let me go sort of do

some due diligence on this name.

Well, fast forward 6 days later,

right? I had sort of uncovered a few red

flags, a bigger glyphosate roundup

issue, you know, messy messy regulatory

gauntlet to get the deal done and sort

of a belief that it wasn't the right

management team to to to drive it

forward. So, it sort of went in the two

too hard bucket for me. I passed. I

didn't lose money. I didn't didn't trade

the stock.

But the the sort of negative thing is

that, you know, I I took 60 hours out of

this 3,000 hours, right? I could have

spent that time doing something

different, right? So, the largest

negative impact of Bayer was a 60 hours

of lost non-compounding research time.

And so, I sort of go back to some of

these experiences in my career as I what

how could I have killed this situation

earlier, right? The accelerated tree

triage concept of front-end exclusions,

risk checklist, patterns and priors. You

know, if you're an analyst, hey, will my

PM own this? I was a PM in that

situation. How could I prevent the Bayer

AG issue from from occurring and design

a process around that. And listen, like

you get a few of these a year, it's

fine, but if they this year becomes a

recurring pattern as an analyst or a PM,

then you're constantly working on ideas

for two, three days, four, five days,

two, three weeks and then killing them

mid-process, all of a sudden your idea

velocity is not what it needs to be to

to to sustain alpha ideas in a in a

portfolio. So, this is sort of a simple

workflow. It's like, how can you start?

How can you you know,

start with risk? One of the things we

show in analyst academy is a is a

investment checklist that I developed

over over my career. And what are the

disqualifiers, right? Could be cyclical

demand peaks or peak margins or

aggressive accounting standards. Pretty

easy to to build a checklist and turn

this into a front-end screener to really

give you the go or no go, right? Could I

have spent, you know, 30 minutes on on

Bayer, seen some of these issues, and

gave it a red light and save, you know,

59 and a half 59 and a half hours,

right? And so, that's those are sort of

just some interesting workflow concepts

that I think

I'm thinking through in terms of how to

get the most out of new out of new

systems. The most exciting thing to me

in the last few weeks is with the

evolutions from chatbots to agents,

what's possible in terms of the concept

of a new investor dashboard, right? And

so, one of the very

sort of

challenging points of chatbots was how

cumbersome they were to use. I had 306

prompts. I had to figure out when to use

the right prompt in the right situation,

where to store my prompts. I had to

upload K's and Q's in the ChatGPT

project. You could sort of pick one

specific workflow and you could see an

interesting output, but that workflow

certainly did not scale. And it

certainly did not scale across a hundred

investor investment team. Even now what

I see in most clients is of a hundred

investors, you have three or four people

who are really down the rabbit hole and

IDEs building stuff with cod code,

custom dashboards,

open cod doing crazy things. And the

other 97 people are like, "Hey, I'm

still trying to figure this out, right?"

And so the the interesting thing to me

is now with agents is building

ecosystems around a new investor

dashboard, right? To have, you know, a

new I was sort of like very anti when

people say like, you know, kill the

Bloomberg launchpad. I'm like, that's

probably not going to happen.

The more I think about some of these

workflows and see what's happening, I

think there is an opportunity for a

single pane of glass and a new investor

dashboard where you can sort of pick

these different workflows, right? I

don't have to prompt this. I don't have

to build the skills. I don't have to,

you know, think about the uploading the

data, right? This can be done on the

back end either through a, you know,

subscribing to a vendor or with I think

increasingly easy engineering

capabilities to do this in-house. You

know, all of this gets abstracted away.

The The analyst doesn't even see the

skills architecture, has to learn prompt

prompt prompt engineering, right? And so

my my hypothesis is that if you're an

investor, don't go too far down the

rabbit hole of learning all of these

technical elements, that time might be

better spent thinking about your data

architecture and your workflow

architecture rather than the

engineering. I may or may not That's a

hypothesis I may or may not be right on

that. But the sort of the simple

workflow of each of these individual

workflows can be supported by a data

architecture and a workflow architecture

via the

via skills and prompts. And all I have

to do is press a button. Right? I was

joking with a friend today, you have an

analyst your PM asks, "Why is the stock

down 5% today?" Well, I can just press a

button now, right? And I can pull in

positioning and factors and sentiment

macro betas and all this sort of context

architecture that I support each of

those flows, right? It starts to get

pretty scary in terms of

what's what's what's possible. I didn't

think that this was something I'd be

talking about in April of 2026 until I

found I was able to build, update, and

validate models pretty well in in an

agentic work work platform. Um so I

think this is the This is the concept

that I

um

uh would encourage you to think about

certainly if any of any of you are from

firms that are building architecture.

Uh the prior very much has been over the

last 36 months almost everything that

has been built in-house has been

abstracted away.

Has been sort of thrown in the waste

bin. Um and so 36 months and millions of

dollars of engineering

um

has been uh has sort of uh been

leapfrogged by some of these agentic

systems. It's probably a little bit of a

uh little bit of a

um overreaction, but the bitter lesson I

think holds is that the scaffolding that

people are building has been uh just

temporary compensation for models.

Simple things like OCR, optical

character recognition, which was

required to pull accurate numbers from a

PDF, but now that's a solved problem or

or close to a solved problem with MCP.

Um so that's my That's my just

observation having, you know, paid paid

close attention. And that the active,

you know, call and response chat

dynamics or will shift more towards

passive research so that I have more

dashboards and buttons to push. Um

That'll be a decision dashboard, a

single pane of glass with all this whole

context architecture

supporting that. That's really important

to me when you start to think about

enterprise adoption sort of across

teams. It's not one or two people in an

IDE building building dashboards.

Um that is very hard to scale across

investment teams. It's, you know,

a team sort of figures out a dashboard

that helps them communicate. That is a

is really a works a new workspace that

has the you research management system

plugged in and factor models plugged in

and everything in it. It's sort of the

new new dashboard. It would be wonderful

if Bloomberg could do that. Um

you know, I think Bloomberg is now

getting to a point it seems in their

tool where most of the other tools were

18 months ago. Um so I'm guessing they

won't they won't work fast enough to uh

to sort of sort of build this. All

right, so that's my soapbox on

um soapbox on uh on tools.

Um let's get into up to

up to speed. So again, this is sort of a

a gating structure, right? If I look at

an interesting idea, right? Uh I don't

want to just go start and spend 60

hours, right? I don't want to start by

building a full 12-hour model. Again,

I'm trying to think about a gating

structure

of what's the story here. Just sort of a

quick check. If that's compelling, let

me put it on my active lesson go spend,

you know, two, three, four hours. We

have called we call it the sniff test in

Analyst Academy, right? Those are sort

of the, you know, the front end

qualifying or disqualifying an idea. If

these are interesting, right? I'll put

those on the active workflow to hey, let

me get up to speed and let me do sort of

the full full full full full ramp,

right? The full refresh. A refresh would

be, "Hey, I've covered this name in the

past. I know I don't need to start from

scratch, but I need to get up to speed

on what's happened over the last four

years." Or full ramp, I have no idea

what this business does. I need to go

spend two, three, four weeks

on on on the idea. So each level sort of

a commitment decision. Uh this is not

AI. This is just core workflow

decomposition um that we teach in

Analyst Academy.

All right. So let me be a guinea pig.

I'll sort of lay lay lay it out for you.

I retired from managing money in August

2021 to do what I'm doing now.

Um

and my Turing test, you know, and

listen, I covered a highly complex

space. There's always a lot going on in

healthcare services names.

Um these are names that aren't your buy

and hold, you know, long Amazon forever

sort of thing. You know, things move

rapidly. So I didn't have an actively

traded the space just cuz the bar to

stay up to speed is

uh is very high. So my Turing test sort

sort of self-assigned Turing test is,

can I construct a portfolio I could

pitch you a biz dev team

of, you know, HR people at a

multi-manager and I could I convince

them that I'm prepared So I'm not going

to do that, but it's sort of my litmus

test of do I feel like I could walk into

an interview and and do a portfolio uh

portfolio review from from 157 57 names

talking about, you know, the businesses

um

uh the business businesses um etc. Um do

you think there are some like maybe some

uh you know, re-emergence of

quantamental things um for all the

reasons of the long list of reasons

quantamental strategies never really

took off for the last 15 years. I think

some of this maybe addresses

some of that. That's So this is a little

bit of skunk works on that. Uh but for

the most part, you know, think of think

of think of me as a as a as as a guinea

pig uh covering, you know, 157 stocks um

that trade let you know, trade over 10

million AUM a day and to in developed

markets. In each of these names, I'm

basically trying to figure out, you

know, I want to be able to explain the

business, articulate the setup, deliver

the thesis, identify the three key

drivers, two key risk factors,

differentiation risk reward, right? As a

DOR, if I'm interviewing a PM, I'm going

through drilling down every name. Hey,

you say you're long Danaher, tell me,

what's the business do? Why do you like

it? What's happening? What work have you

done? Etc. What's the upside? And so the

the interviewer the DOR interviewer is

pushing hard on those ideas. Right now,

I would completely fail that test. So

that's sort of my That's sort of my

test. So I'm getting my my coverage and

comp sheet fired back up. Um just to

sort of see the different metrics. This

is not AI generated. This is built in

FactSet. I'm getting my portfolio back

up. This is not AI generated. This is

built This is built in FactSet. Just

sort of looking at, you know, what the

stocks have done, etc.

Um

And the really the measure of a sector

PM to me, if you walk into any of the

multi-managers to a high-performing PM,

they're going to be able to look at all

all these 157 names and they're going to

know them cold, right? If I say, "Brad,

tell me about Teleflex. Like what does

the business do? What's been happening?"

At my peak, I could tell you I could

tell you about all 150 names, right? Um

you know, 75 really well, 75 like ish.

Um

And then they sort of take that business

knowledge and they're able to apply

pattern sense and, you know, mar pattern

market sense and pattern recognition and

judgment to build a portfolio out of

that name. But that you don't you don't

leap to that sort of alpha element. You

start with the grounding of knowing your

companies cold, right? Particularly as

you build a portfolio and you hire

analysts and junior PMs to cover 50

names, their job is to know their those

50 names better than than than than than

everyone. And once you're up to speed on

those names, you should be able to look

at the coverage and explain every move.

This is the year-to-date move. I should

be able to explain why Acadia is up, you

know, 70% year-to-date and, you know,

why Baxter's down 13%. Again, at my peak

I couldn't, uh but I'm trying to see how

how I could. Uh I'm trying to see uh how

close I can get with this AI-augmented

process. Not necessarily have a thesis

on all 157, but I think in general,

you know, uh you you know 30 to 45 with

incredible depth, a 10 by 15 portfolio,

and another 10 to 15 in the pipeline cuz

you always need constant constant ideas.

And then the other thing is sort of be

able to understand your thematic

thematic expression chips. Which stocks

are an expression of various viewpoints?

Such that when you're looking at this,

you know, on a day why things are

moving, you could sort of explain in

your mind, "Oh, that makes sense. These

stocks are moving cuz they're interest

rate levered. They're levered to the,

you know, ACA exchange, trying to be BP,

etc." And so you start to get a sense of

why stocks are moving. That's important

because you're trying to arbitrage

incorrect moves as a high-velocity

high-velocity PM. That's probably a

little bit too much inside baseball. So

I'd say I have a little bit of advantage

here, you know, because the historical

context will will

will will will will will help me. Um I'm

also going to do this in a small-cap

generalist long only at some point, but

I'm starting here. I'm starting I'm

starting here.

Um So can I ramp, you know, coverage on

157 names, refresh on names I knew, full

ramp. Can I update the models? I have

some of these models.

Many are 4 years out of date. I've

always sort of had these eval set of AI

tools since 2022 with things I'm like

there's no way there's no way that, you

know, the tool can do this. One of one

of them is I give my Uber model that's

four quarters out of date and a little

messy and I just say update this model

and it's like never worked until

recently. And so, can I take my 4-year

models out of date? The build from

scratch is still getting there a little

bit. You have to chunk it down a little

bit. We'll talk about that in our April

April 30th. But my belief is that as

intelligence improves and the MCP

systems have helped a lot in terms of

accuracy there

that I could be able to have a few dozen

fresh builds and then I think the real

alpha is in creating the tracking

systems for for every for all 157 names

can I

sort of in partnership determine what's

the number one thing for each each of

the ideas, the three key driver, two key

risk factors, catalyst path, trading

plan, thesis creep plan, data dashboard,

sentiment narrative and thematic

linkage, right? And so this is really

the structure. Again, I you know, sort

of back to the back to the 10 1090

of active coverage.

Specialists haven't found a lot of use

case

in AI tools because chatbots were very

helpful helpful-ish to get up to speed.

But I'm not really getting up to speed

on the in this use case I am, but at my

peak I'm not. Like I don't need I don't

need ChatGPT to summarize a 10K on HCA

when I know that company incredibly

well, right? But this 90% of active

coverage having another set of eyes on

all of these things which will make or

break stocks. Now we're talking, right?

Now we're talking about something very

helpful.

As we go through this

you know, sort of know the the

importance of context. Like some of

these spit out you know, the number one

thing or three key drivers I completely

agree with. Others don't, right? Like on

GE Healthcare it said that the business

is 13% of revenue was the number one

thing. I'm like no, no, no.

So it does require some debugging the

same way that

coding agents require debugging, right?

And so that's

that's

um

those are a few thoughts on that. All

right, so I want to get up to speed on

Danaher. I'm sort of picking Danaher in

this situation because I haven't looked

at I used to own it.

I haven't looked at it in maybe eight

years or something.

How can I get up to speed on the

investment situation? What's the old way

to do it? I've sort of always been a fan

of the Pomodoro

method, break it down. I have these all

over my house and office. Okay, I'm

going to spend, you know, 60 hours

reading the relevant sections of the

10K, the last four earnings transcripts,

conference transcripts, stack of

sell-side research. That's 68 hours and

I'll go update my model and maybe I'm

spending 12 hours just kind of get the

full full refresh, right? So I sort of

feel like I know the name. I know the

story. I know what management's been

saying.

And that's a big task across 150 names,

right? Not that hard across Danaher.

But I there's 90 names I need to do that

on, right? There's 60 names in this 157

that are like brand new to me. Like I I

don't remember seeing that name. I never

looked at GE Healthcare because it was a

spin and so okay, that's interesting. I

should go

take a look at it. I need to do the full

40 hours. I need to go update my models

for 6 hours and build some new models.

You start to add this up and if you if

you compare that to the 3,000 hours we

have in a year

right? That requires 4,740

hours to really get up to speed and know

my names know my names well, right? I'm

I'm I have a day job running Fundamental

Edge. So my question is can I pass this

turn turning test in the AI ramp, right?

So can I turn this refresh into 15

minutes? Can I turn the full ramp into

45 minutes, the model updates and the

new model builds and all of a sudden can

I spend 120 hours, right? And I and I'm

really up to speed, right? Now that's

like no no no institutional investor is

going to stop at that point. But what I

would say is like you're freeing up a

lot of if these can hit

if these can hit an accuracy hurdle,

which I think is still a hypothesis,

you you freed up an immense amount of

time for other ideas other other ideas

to to apply

from a research perspective. So and to

be very clear like this is an

experiment. Part of the part of the

guinea pig nature is hey, this I

couldn't get this to work and it kept

spitting out this error.

So that when firms are doing this they

would like I had this issue, maybe

you'll have this issue as well, too. So

I'm sure they'll have many issues and

problems. I'm not saying that you can

you can you you can one shot one one

shot this. So these are sort of the

various up to speed

up to speed workflows.

Sort of things I do you things I might

do if I'm just looking for the 15-minute

download like hey Google a couple

sell-side notes, scan the last call,

that sort of okay, I know what's I know

what's happening. All right. So one of

the things I'd like to do in all of

these workflows is I like to parallel

parallel process

where, you know, I want to start with

Claude just natively like going without

uploading anything, no skills files,

etc. and say, you know, help me ramp

help me ramp coverage on Danaher, right?

And you have you have many friends who

are really advanced in this stuff and I

have many friends who are also like hey,

I tried to get Claude everyone's saying

Claude's great and I tried to use it to

ramp coverage on Danaher and it just

wasn't good. And often what they will do

is they will go into Claude and they'll

say help me ramp coverage on help me

ramp coverage on Danaher. And what you

see is like better than ChatGPT was nine

months ago. But you get something that's

like sounds like a investment blog. It's

a breakdown of the three segments,

highlighted the Massimo deal. You know,

it catches a couple things. But it

really doesn't have any of the sort of

structure of anything that I would find

useful

you reading as reading as a PM and it's

just not not not even close to

comprehensive enough. I'm basically

trying to one shot for you 12 hours of

research here, right? And so this

doesn't get me anywhere close

to passing the closed book test of can I

explain and articulate what's happening

with with with with Danaher So as we

talked about in the first session

what's really important today and maybe

this gets abstracted as well, right? But

as we sit here in April 2026, what's

really important is a is a systems

approach.

One is sort of like having the intuition

to know, you know, the superpowers and

jagged edges. Even simple things like,

you know, training cut off of if you're

asking for certain things and it gives

you 24 instead of 25, well, it's because

of the training training corpus cut off,

etc. So having a little bit of the

intuition of the superpowers and jagged

edges

of AI. It's the you know, tool selection

and data strategy, you know, the

workflows, prompts and skills that's

customized to your process. What's

incredibly important and I'll put a star

behind this is a two-tier validation

system.

So, you know, one of the biggest

inhibitors to institutional adoption

>> [snorts]

>> of AI is one, the big firms aren't

adopting because it's in it's up up to

it's in the flow workflow, which is much

more, you know, sort of high stakes

subject to hallucinations. And the the

AI tools really didn't have a great

validation system. You couldn't build an

AI agent to validate an output. So

chatbots couldn't check outputs. Agents

can check agents in a much more

effective way, too. So this is like a

this is a this is a really big deal to

me, this two-tier validation system. The

systematic agentic validation, but also

the analog validation. Hey, take this

thesis identify the six most important

inputs, either give me citation or

create a checklist so I can go and check

by hand. Let me go into the 10K. Let me

go into this meeting notes. Let me

verify verify, right? So creating a

a culture and system around validation.

We can't we can't afford to have

hallucinations in our high stakes work.

So that's important. And then using the

system to enhance comprehension and

and rigor. So one shot obviously like

it's one shot interactions will not get

you institutional grade

grade grade outputs.

One of the elements that's

getting a lot of conversation and has

been really really important in my work

is what's called a skills file. So a

skills file is just a way to encode a

process. Sort of like a stack of

prompts.

I won't go deep into what a skill is. I

would encourage you to just watch this

23-minute from Shaw on on YouTube which

did a great job of just the the basics.

It's basically a stack of prompts. It's

create

sort of becoming the new standard for

encoding a work process. It tells the

agent what to do effectively and gives

gives the agent the proper context of

you, your process and how you want to

get things to be done. This

this handles a lot of the issues so far

what I've seen in terms of that prompts

had, right? Operating with a concept

called progressive progressive

disclosure with a little bit of

metadata, you couldn't drop in a 30-page

prompt into a system because the

the chatbot would lose attention at

various points along the prompt.

And so it's sort of, you know, very

intelligent engineering which is

becoming the standard of encoding

workflows. So I'm like oh crap, I I

spent all this time in 2025

writing all these prompts then you then

using AI to write all these prompts. I'm

going to have to start from scratch. And

so these prompt libraries, you know,

that that

I've built in that we've built that were

one like very cumbersome because you had

to there I could never figure out a

right system of like when to pull the

right prompt. You couldn't overload Chat

GPT projects with all of these prompts

cuz it wouldn't know when to pull the

right prompt for the right situation.

This is kind of a solved or like close

to solved

problem now.

Um

and one of the things that I was

frustrated by is like I'm going to have

to go through this whole process again

of building skills until I realize that

you can you use the same sort of process

of meta skills creation, right? So, if

you have prompts, I would encourage you

to play with taking that prompt. This is

a quick story prompt that I've used. And

Claude couldn't do it I'll sort of get

into why I had issues with this. Claude,

Perplexity Computer,

um one of the things I've uh um liked

about it is it's very good at meta

skills creation. What is meta skills

creation? Using an agent to create agent

skills, right? Providing the context, in

this case the prompt, and saying, "Hey,

I have a prompt on how to just get a

quick story, like a 15-minute take on

this idea. Can we turn this into a Can

we turn this into a skill, right? That

takes about 2 minutes in Perplexity

Computer. No engineering required, no

coding required. So, these agentic work

platforms uh PC Perplexity Computer in

particular can turn prompts into skills

uh very very very easily. So, I sort of

have mixed, you know, my personal take

on Claude has been has been mixed.

There's obviously a massive amount of

of hype. The agentic tools have been

have been great. You know, a number of

wonderful positives. I think the agentic

abilities, the fact that chatbots have

hands now are genuinely a game changer.

You can see the pieces that are there,

and there's obviously a ton of talent

capital to to continue to improve these.

Claude has been iterating at an

incredibly rapid pace, which is

exciting. Also, I think there's a lot of

bugs in the certainly Claude Co-work um

that have become quite frustrating uh

quite frustrating to me.

You know, the the frontier seems to be

you know, Claude code, you sort of

directly in terminal or an IDE like a

cursor of VS Code.

You know, I I went down that path. Uh

was highly frustrated by that path. And

I started asking around and tweeting.

It's like, "Yo, yeah, that took me 3

months to learn. It took me 100 hours to

learn." Like, to become a power user and

debug and how to spawn sub agents and

all these things. I'm like, "Man, I'd

much rather just stay in my wheelhouse,

operate under the the um

uh sort of bitter bitter lesson that

this will get abstracted. And I'm

starting to see that

a little bit in sort of most notably in

in um you see it a little bit in things

like Replit and Lovable, which are more

vibe coding abstracted tools, but you

see that I think particularly in

in Perplexity uh Computer. So, I think

my observations the tool the tooling is

still very immature. Um

um

you know, the sort of idea is like just

ask Claude, it's so easy. Like, no, it's

not. I spent, you know, my wife was

like, "Who are you yelling at?" I'm

like, "I'm yelling at Claude code in VS

Code because I'm asking Claude how to

use Claude, and it's not working, and

everyone lied to me." So, if you're

struggling with Again, this is like a

skill issue. This is a user problem. And

the three investors who are at a at your

firm who are really technical aren't

having this issues. It seems seamless to

them. But the 97 others who are not

technical, and who whose eyes glaze over

when they look at the raw code like me,

um you know, might may struggle with if

if you roll out VS Code across

across. I think there's still some

issues with inconsistent retrieval. MCPs

are buggy. This one is like approval

loops. You got to approve everything or

just let it run loose on your system.

Like, we have to find a better way for

that. The local system control makes me

a little bit uh uneasy. The reality is

learning these tools is a productivity

sink. Uh the same way that learning to

write prompts from hand for 50 hours was

a productivity sink for me in 2025 cuz I

never I I've written a prompt by hand in

9 months. So, that's just that's a

hypothesis.

Before you go down, you know, and spend

the 100 hours

to learn VS Code to code up your custom

dashboards, maybe just consider that.

I've I've backed off of that. People

think I'm dumb. Like, "No, you do it.

Let's go and But like, I backed off it.

Um that's my opinion, and I can always

pick it back up if I think it's uh

think it's important. I also talked to a

lot of people. I'm building these crazy

things in Claude code. I'm like, "What

are you building? You're running open

Claude Oh, crazy. Let me know what

you're doing." It's like, "Oh, you just

built like a custom primer." Like, I can

do that in 2 minutes in Perplexity

Computer, which is basically like a open

like a user-friendly open Claude. Like,

"Okay. Oh, you're doing weekly web

scraping." Like, "Yes, I'm doing that,

too, and it took me 3 minutes." And so,

I think um

be careful of productivity theater.

Um especially if you're on Twitter. Be

careful of feeling like you're dumb if

you don't, you know, open up an IDE and

start coding things. Um you know, I'm

not saying people are lying to you, but

there's an incentive now to

to to push some of these uh things. Um

So, the one one thing I backed off

Claude Claude Co-work for this specific

exercise is just not there yet in

Claude. The ability to sort of take

prompts um and have them build skills.

It would build skills that Claude

couldn't use. They were too long,

errors, or messy. Um so, I think Claude

Co-work is I think still conceptually

very interesting, still quite buggy. Uh

one of the engineers driving Claude

Co-work sort of showed this thing of a a

piece of hardware to approve everything

in Claude Co-work. That sort of tells

you, you know, tells you the system

design is not there yet, right? If I

have to sit and monitor Claude Co-work

and approve every single thing, that's a

problem versus an agentic work for

platform like Perplexity Pro, where I

just have it do something. It's not

operating in my local system. It's

operating in a in the cloud um uh

virtual sandbox. So, I don't have that

process, and I don't, you know, these

sort of things like you know, you spend

30 minutes on something, and then you

reach a tool use term limit. It's like,

why, you know, okay. So, um

you know, I'm not a hater of Claude. I

think it's wonderful in certain ways,

but I also think the hype is probably a

little bit too hype uh too too hyped.

Which sort of brings me to Perplexity

Computer. I don't think Perplexity did

themselves you built too many fans in

finance by saying that Perplexity search

could kill Bloomberg. Bloomberg is so

sort of a weird like the other side of

the spectrum. They've probably been a

little bit more caution on Perplexity

cuz the version 1.0 wasn't very good for

finance. Um so, when I people I tell

people like, "You're not You're going to

be like shocked, but I'm actually using

Perplexity Computer to build this."

Perplexity Computer is an agent. It's

basically like open Claude

with a really nice seamless uh user

interface. The biggest reason I adopted

this for this workflow shift from

Co-work to Perplexity Computer is the

meta skills creation. I can drop in

prompts, and I can create these skills,

and then all I need to do is just run

the run run the prompt on a run the

prompt on a ticker. So, I don't know why

it's better. It uses 19 different

models.

Uh again, I have no affiliate economic

affiliation with it all with it all. If

there's a better tool that comes out

tomorrow, I'm going to pivot to the the

best the best the best the best tool. Um

but it can create in this ramp process a

skill, right? Which I can share. I won't

share in this instance, uh but we're

considering how to share those. Um and

it sort of walks through, you know, the

full file folder. Again, you know,

bitter lesson. I was like, "Do I need to

go learn about skills and spend 6 hours

figuring out how to build the skill MD

file and the references and all the dot

MD files and start doing that by hand?"

And the answer is no, right? I can drop

a prompt, drop some context into

Perplexity Computer. I have this I have

this skills files created for for me.

Now, that's just starting point. These

things need a lot of iteration.

But it has the full sort of, you know,

produces a 25-to-30-page custom uh

business business business primer. I

also know like the rapidity of change,

new models, new engineering. You you uh

someone said like, "How do you know this

won't be abstracted from Mythos or uh

you know, Chat GPT 6?" And like, "You

don't know." That's just like the nature

of the nature of building an AI right

now. But my hypothesis, I think the

observation now, I'm sort of with

basically anything Andrej Karpathy tell

says, I sort of have a bias to believe.

Um every time he tweets something, I'm

like, "How could I apply that to

investing?" Um

and um you know, that's sort of been his

sort of the knowledge graph concept

right now, which we'll we'll we'll we'll

we'll we'll get into. Um but my

hypothesis for non-developers is the

time you're spending now learning an IDE

in a terminal is is is wasted mo- wasted

motion. That whether it's Mythos or, you

know, Claude manage agents or GPT 6, Sam

is sort of talking about a unified back

end, just an easier-to-use interface. Um

is really the uh is really the is really

is really the

uh the the future. So, in building your

skill, you're sort of building your

platform. Certainly, um if you're if any

sort of engineers um on the um on the

call, I would just say like build build

build build carefully. Cuz my primary

interest isn't in hacking

and, you know, pressing you by all the

things I can build with with Claude

code. But I want to build systems that I

think can change investment process at

firms that have many many investors at

various degrees

of uh of investment process. And like

me, I'm guessing many of them

look at code, and their brain is fried.

We have a cognitive pie, which is

consumed by following companies and

markets. And it's very hard to slice a

100-hour 100-hour slice out of that

cognitive pie. And probably, maybe not

certain, but probably a waste of lesson

a waste of time if the bitter lesson

holds in the context of intelligence is

uh it's all about context next to

intelligence, and the engineering gets

abstracted. So, I'm sort of paying close

attention to all these AWPs, um which is

basically a platform where Claude code

is sort of abstracted or coding agents

are abstracted. You know, Copilot

co-work would be one of those.

Perplexity computer, you know, Manas I

think maybe is a dark horse in this. You

know, what OpenAI builds to sort of

compete with Cloud co-work.

Um and I think you know, a number of

people building internal sort of

multi-model

multi-model multi-model agents. Um so

this is uh this is this is this is um I

think an interesting interesting

frontier. And ultimately I think the

standard is for builders like can you

build something like this that a team

works on the back end like your

engineering uh AI team work on the back

end to set the architecture and and

skills up, but then can we go and roll

this out to an investment team and can

they is is just so intuitive

user-friendly that it transform

transforms our investment process

uh really sort of right away. So that's

my hypothesis. That's why I'm so excited

cuz I kind of see that this is

potentially potentially possible. Right.

So we'll pick one of those workflows now

and I'll give you an example of

example of how to how how how how to do

that. Um so quick story, I showed you I

showed you the prompt to

you know, the prompt to skills. I now

have a skill set up in in quick story.

All I need to say is as I'm looking at

my my dashboard of an idea is like oh I

you know, a few outliers. I want to see

what's happening in these six names,

right? Just give me a a quick story

skill on these six names. I want to

qualify or disqualify quick qualify

these. This takes a couple minutes in

Perplexity computer. I get a little

report and I can just sort of flip

through this report. So I see okay, why

Acadia is up 60% this year. It was sort

of seemed to be like a um liquidity

overhang where they had capex you know,

capex cut sort of a re-liquidation

event. Seems like maybe a activist you

know, sort of activist is in there to

cut capex, free cash flow is back. You

know, short interest was high. So it's

sort of a confluence of maybe a squeeze

I don't know or some sort of big move.

All right, that's sort of interesting.

Maybe it's a fade. Maybe it's just got

more

maybe the activists have a real sort of

path here for So I don't know the

answer, but all right, let's add that to

the active let's add that to the active

research pipeline. Then I see Inhabit. I

I I don't know what this business does.

It wasn't around when I covered health

care. Like okay, it's up a lot because

there's a there's a takeout. All right,

okay, Kinderhook take private. All

right, let's cut that from the monitor.

So I could sort of go through these.

What's happening in a Brookdale? I used

to cover that fairly closely. All right,

what's happening now? Okay, the the

boomer demand story is now starting to

which everyone invested behind in 2012

is now sort of coming. Sustained

occupancy gains in the high 80s. There

was a big you know, excess supply for a

lot of years in that space. Maybe that's

been cleaned up a little bit. Maybe you

have a favorable occupancy here for a

few years. All right, that's interesting

hunch like not a thesis, but let's put

that in the let's put that let's put

that. I see this Perspective

Therapeutics got into my screen and I

see like okay, this is start we're

starting to get into like too much

science-y stuff for me. I'm a hospital

more of a hospital analyst. All right,

let's cut that. I don't want to cover

that. That's just too much science. So

I'm sort of going through the step one

of just the basic 100 through 157 names

just a little bit of a quick story,

right? And then from there I'm just

activating those names into my names

into my pipeline. This is not a crazy

like controversial workflow like in the

past how would I do this? I'd look at

DES on Bloomberg and maybe EEG and read

a couple notes and it doesn't it's not a

huge like transformational, but the

thing is I can customize this to I can

customize this to what you know, to what

I want. Like the you know, the bull and

bear and the you know, the one factor

and I want to know like specifically why

the stock is up. It's hard to look at a

Bloomberg launchpad of talk up 42%

year-to-date and get a very quickly

distilled essence of like why why is the

stock up 42%

um year-to-date? Well, they had a

blowout you know, blowout Q4 29% revenue

growth, adjusted EBITDA doubled, strong

guidance. Like okay, cool. Like that

that saved me 15 minutes of hunting down

those questions which doesn't seem like

a lot, but if I'm trying to do this 157

times

that is really helpful. Um

so um and this starts to help me sort of

uncover the curve of what's happening in

all of these

um

all the all all all of these stories,

right? Which brings me to Danaher. So I

sort of see in the quick story Danaher

you know, the one factor is a

bioprocessing recovery trajectory and I

have priors on this cuz I covered

Danaher for a long time and I walked

walked through you know, sort of lived

through the

business transformation and management

strategy to align themselves with the

secular growing area and then sort of

COVID you know, boom in bioprocessing

and then the bust and

you know, we sort of get into idea

generation, but one of the powerful

elements in idea generation is pattern

recognition. So one of my favorite

patterns in longs is a hangover

dissipation where there's a temporary

tailwind, boom turns to bust, bust

sustains for a while. Markets are not

great discounting mechanisms. They sort

of assume whatever is in front of our

face now persists forever.

But when you're hungover hangover

doesn't last forever. There's sort of a

natural cooling off of a hangover

and often that re-acceleration can sort

of get you back to you know, you sort of

get the Lollapalooza of you know,

accelerating fundamentals with you know,

re-expansion in multiples. All right,

that sort of feels like maybe that's a

situation here. That extended period of

of weak weak business momentum. So all

right, let's let's let's put that into

the active research pipeline and for

this again not not a recommendation to

buy or sell uh Danaher Danaher, but

let's sort of

let's see if we can go deep from this

quick story signal. Let's see if we can

go deep on a name. So how do we do that?

Here I'll bring in sort of Andrej

Karpathy concept on knowledge bases. And

basically what he's saying is he's

spending a lot less time manipulating

code

and more time manipulating knowledge.

And again like you know, my principle of

AI is like whatever Andrej says at least

go investigate it, right?

Um

So I think he's very unbiased and sort

of calls a spade a spade which is

certainly not true of many of the other

characters in this space. He said he

said agents were sloppy in October. Um

you know, playing with the you know,

ChatGPT agents in October you could see

it slop. He sort of called the turn in

coding agents etc.

So I think he's um he's a reliable

sherpa as reliable as they come

of the frontier of what's uh of what's

uh uh possible. His automated research

concept I think is really really

important when it comes to automating

research around

on on on key drivers. Essentially what

he he's saying is that LLM intelligence

advances such a degree there's an

intelligence overhang.

Right? That the engineering's get

abstracted. That the core part of what

we need to do is just put as much

context next to the intelligence of the

LLM as possible, right? And so that's

sort of the concept. And so I've been

playing with that concept. And there's

sort of two elements of of context for

me that matter. One is just the classic

you know, way we think about research

context for investors is just the raw

information. Filings, transcripts,

sell-side notes. Um

um expert network transcripts etc. I've

sort of pointed to AlphaSense as over

the last 12 years as a signal. I don't

have these great systems. I don't have

an sell-side pipe etc. I'm operating in

a very low research budget internally.

Uh but AlphaSense to me has been a check

that research context is a key point of

leverage, right? That when you feed in

more information you get much better

much better outputs. I can spin up

certain workflows that I that I think

I'm grading my own home home homework

are much better than what AlphaSense

does now, but something like earnings

previews AlphaSense to me is much better

certainly when there's a number of um uh

a number of expert network calls done on

a name cuz it can pull in that context

and give you a much better articulation

of bull and bear than you could on an

open open web scrape. So to me like

AlphaSense has shown the light on the

power of research context.

Um and my sort of parallel processing of

like one shot this versus building a

workflow system has shown me the

leverage of workflow context, right?

What to do, how to do it in incredible

detail, right? And I I um my brain just

is wired for some reason to just like be

able to build 80-page documents on how

people do earning season. That's just

like a hobby of mine. I don't know. I

need to I need to talk to a therapist um

about it, but I like the systems thing

like the systems decomposition of

investment process is always I just like

nerded out on it ever since I started in

doing the doing doing this job. Um so

how I build deep workflow context

a lot of it's just the 15 years of notes

and process learnings um

that I've sort of collected on along the

way and distilled into our programs. We

run many you know, uh analyst academy

enterprise programs, AI programs, factor

factor academy programs etc.