Bitcoin and AI - Building a Company For The Future | Matt Hamilton and Becca Amilee, AnchorWatch

The reason why ChatGPT, 2022, was a really good, better Google, was because it wasn't

just taking your words.

It was trying to discern the context of why you were asking a question.

You now have this context that lives on your computer.

It's being empowered to basically use the computer as its own personal resource.

And the major improvement for these models across time is they're getting larger context

windows.

Two, three years ago, that wouldn't be possible.

I want somebody now who will say, this tool will allow me to get so much more done.

I am going to create these automations.

I'm going to take advantage of this.

If you don't want to get laid off in the next five years, you must tinker because if

you as a leader are not thinking about the second order effects of these things are starting

to, you are not doing yourself a favor.

Becca and Rob from Anchor Watch, welcome, guys.

Thanks for having us.

Thanks for having us, yeah.

You've both been here for cheat code.

What was your experience like at cheat code?

It's great.

Rob's, this is Rob's third.

Third time in Bedford.

Yeah.

First, cheat code in Bedford.

I went to the cheat code in Australia last year, but for cheat code, it was my first

cheat code Bedford, but anytime we are in London for work like visiting the Lloyd's insurance

markets, I always try to come a day early or leave a day later to be able to see the

games.

I love that town.

I'm Becca.

This is your first time.

First time.

First time.

So, yeah, a cute town.

Really enjoyed being up there, getting to meet the local Bitcoin folks that had come

out to learn and listen and have conversation.

That was really good.

And then, yeah, seeing the soccer was really great.

Football.

Football.

That's right.

That's right.

I think we kind of make fun about the appeal of Bedford, but it really is nice.

It's a nice, quite typical British town in English town, shall we say.

But for the Bitcoin conference, we really take it over and make it out on, but specifically

because it's such a small and intimate affair, you get to know a lot of people, especially

if you're not really, you know, networked or connected, which is most people on to, isn't

it?

So, you can get to meet all sorts of people.

Yeah.

The people, since everyone's coming around mostly VUK, maybe Europe, everyone's so polite

and just very excited to be able to talk to you.

And this conference is unique where there's not like six stages and a bunch of things

you have to run around and do, you get a lot of time just to have long, thoughtful conversations

with the locals.

And they're always very appreciative of the time.

You know, people are welcome to me asking me like technical Bitcoin custody questions or

just talking about anything in life in general and they're all very nice and approachable.

It's a great group of people.

I actually like single stage events.

Same.

So, I mean, I think I can understand the opposite.

I can understand companies and authors of books.

You know, everybody's trying to get on stage and have their opportunity to talk about

what they're doing.

But in terms of attending, you know, a single stage is great.

You sit down and you tend to listen as a speaker when we speak at those events.

We experience that on the other side.

People are very engaged and in their seats listening.

I think when there's a million stages and people getting pulled in different directions

and maybe trying to run between stages to catch something, they end up just giving

up on the content and focusing on the network and networking.

The best part of any conference always are the side conversations you have and not

it, not the what's on stage.

And unless it's a small conference that has like a no recording policy where you can't

watch it later, I very seldomly make an effort to like watch every conversation because

I want to be there in person, engage with people.

And then on the flight home, I'll just watch the stream and I'll watch the talks that

I missed, right, to be able to see the content I want.

That's usually my bias out of conference.

And I don't at all.

That's why I like the single stage event is because I find myself actually sitting down.

If I don't have a meeting going on, if I'm not actively in a conversation, I'll sit down

and listen to the content and listen to the speakers.

Because once I'm on the flight home, there's no way I'm watching.

I'm watching recordings.

I'm either getting back to work work or I'm trying to catch up on sleep.

So yeah, I really, I enjoyed the event.

I thought it was very, very well put on, you know, good production value and...

Thank you, Jess.

Well done, yes.

I thought it was excellent.

Yeah, it was my first time on stage and I have to echo your sentiments, actually, because

if there were multiple stages, I probably wouldn't have met as many other speakers as well,

because that's a key point for me.

I know a lot of people think, oh, it's a big account, say, probably knows everybody

already.

I don't know anybody.

I've been operating this account and I'm sleep for five years.

You will.

We talked previously about how several years into doing Anchor Watch now and traveling

around on behalf of Anchor Watch, you know, the speakers have become our friends, you

know, our personal friends and so going to events is actually amazing from that standpoint

as well as that's our opportunity to catch up with people and appreciate each other

on a personal level.

So today, specifically, we brought you in to talk about AI.

Everyone's talking about AI, hardly anyone's actually implementing it using it in their

businesses day to day.

And so we want to talk about that, but before we do, we need some context about your business.

So tell us about what Anchor Watch is, give us the spiel, but also, you know, it's the

operating context as well, how old it is, what you're looking to do.

And then when we start talking about AI, people will have a better understanding of the

context as you are operating.

Yeah.

So Anchor Watch is an old nautical term referring to the crew of sailors who watch the ship

when it's at anchor at port or at anchor, whatever, right?

And that kind of communicates the big theme here, but what that is is, you know, keeping

Bitcoin safe.

And we have built, there's two main pieces of our business.

We have built our own Bitcoin native custody infrastructure.

I feel pretty good in confidence in saying it's the most secure, large, like enterprise,

like custody systems out there today, and we can get into the details as to why.

And we offer insurance to that.

We are Lloyds of London cover holders and we're able to uniquely offer one-to-one insurance

in your name with the policy directly from Lloyds of London to protect your Bitcoin.

And when we launched to the public last year, I'd say that that combined product is what

people know us for.

So if you think of all the capabilities that a super high-tech, super-secure enterprise

grade custody system can do, I mean, obviously, it can custody.

It can keep Bitcoin very safe and cold storage, but there's all sorts of different capabilities

that need a secure custody layer at the base of it.

So Bitcoin back loan, where's the collateral being held?

And so there's all sorts of capabilities that we have because of the tech solution.

We can ensure any Bitcoin that's being held in our tech solution on the custody platform,

but we're Lloyds of London cover holders, Rob mentioned, and we can do custom insurance

policies.

We can craft things to meet businesses where they are.

So think of it as a Venn diagram.

We've got our insurance capabilities.

We've got our custody capabilities.

What people know us for is pretty much the little overlap of those two things.

They know us for insured custody.

But really, we can fill out that entire Venn diagram, if you will, and continue expanding

our business into different capabilities that take advantage of both.

Take advantage of our insurance capabilities and custody.

The one just short thing I would add to that is that for any financial service or any

use of Bitcoin, custody really sits at the foundation of anything you're trying to do,

whether it's alone, escrow, any sort of futures contract, any sort of financial product

really on Bitcoin, custody is the foundation.

If you don't have good custody, then the rest of it doesn't really matter.

And so we've been really laying a foundation, pouring that concrete, letting it set and

be stable for us to be able to look forward into the future for on that custody side

and also bundling with insurance, or maybe not even using our custody platform, but also

just getting insurance.

Rebecca, you're the Chief Operating Officer of Bank of Watch, and you've got a very strong

background in operations, special projects and retail.

Just give us a quick overview so we get some context of your experience.

Sure.

So my career has definitely had some phases.

So the first 10 years of my career, I was an apparel merchant, so doing both the financial

planning forecasting of apparel at kind of the corporate level, and then also actually

choosing the style.

So planning the fashion aspect of it.

When I went over to Starbucks, we've chatted about that, spent 10 years at Starbucks, all

at corporate, doing a variety of different tasks.

But one of those roles that I had at Starbucks for three years was implementing 26 new,

well, 26 systems, different pieces of software, hooking those up together that orchestrated

all of the company's outbound spend with the exception of mainland China.

There was a very operational aspect to the systems themselves, so I think as we get into

the conversation, we'll definitely go back into that and to kind of how a COO or a really

kind of back office team should be or starting to think about these tools.

Rob, you're in effect the technical partner, but also to CEO, so how did you get to that

position?

I mean, a brief background in my career, my first job, besides busing tables, that was

my first job at 14.

I was actually a video game referee.

I got flown around the states working at Major League Gaming.

It was like the dream job for a high schooler.

We're going to pay a little per diem and fly you around the country to watch people play

video games, and that was at Major League Gaming, back when esports was starting to become

really big things, like 2005 to 2015, but even I went to college, got a double major

math and econ, and then I went and started working for them, managing their ad server.

So it was just a job I could do to help them out when you would sell sponsorship contracts

who'd run a live stream.

We were actually kind of the first to pilot video ads in live streams.

We were like piloting that with Google, because we would get 100,000 people back in 2009

watching these streams, and we would want to serve ads.

I did that for a couple of years.

I did another, went to bounce around ad tech as it's called for a while.

And then I went to go work at IBM as a consultant and turned that into being a data scientist

and consultant.

And my Bitcoin journey just briefly started around 2014.

I had a, but as a big Ron Paul and the Fed, Gold Bug guy, and a bunch of my buddies, a couple

of my buddies would tell me to check out Bitcoin, and I was like, too complicated, you know,

I'm not going to get into that.

And then Dogecoin came out.

And all of our friends had gaming computers.

And so we would all mind Dogecoin.

And I realized that like with the Comic Sans font in the rainbow, like, like, like, this

was like, I knew internet culture that this was satire.

So I was like, okay, this is satirizing something.

Let me go learn the real thing.

And I started going to the New York City Bitcoin developer meetup in, in, in, end of 2013,

start 2014.

And that was kind of my opening into Bitcoin and what, going into those bit devs meetings,

it would be maybe 10 to 20 people in a room, and just going very passionately talking

about all of these concepts, like decentralization and security and cryptography, things that maybe

I was a little bit aware of, but this was such a serious application of all these principles

and designing the software that it kind of fell in love immediately.

So I started going, you know, to all of those meetings, starting to learn about Bitcoin.

And that was kind of, I never really intersected my professional career with Bitcoin until starting

the anchor watch at the start of 2022.

So during COVID and all of the lockdowns in the States, there was a big social audio

app called Clubhouse.

And that's where I met Becca, I met, like, all of our initial angel investors.

And the whole story for anchor watch came out of that, but I was always, uh, in those

clubhouse terms, like, the someone who understood the technical nuances of how Bitcoin works.

And through that, um, realized that there was a really good opportunity to apply that

in the domain of not just maybe working at a Bitcoin company, but starting one.

So all these more than two years now, since Chatchee, VT launch, and it was, yeah, really

quite an explosion, um, a lot of attention, massive adoption, it's almost a vertical line

of adoption.

But, you know, it quickly just settled into being an advanced form of search for most people

right?

Better Google.

Yeah.

I'm much better Google.

And you're like, okay, that's fine.

But now, I guess it's just, it's almost like the second stage with the, um, the advent

of, is it open core, with open clause, the, yeah, open cloud, no, open cloud, the agent

model.

The self-hosted agent.

Yeah.

And so we're, I don't think we saw as much.

I think when, when, when Chatchee, VT came out, there was a huge, uh, level of interest.

And then with the development of the visual model, say, like mid journey, there's a lot

of hype around that.

And video, but nothing really quite matched that initial launch.

Mm-hmm.

I think, as, from my point of view, as much as the core, um, launch did, what, what are

the implications of, first of all, just Chatchee, VT style models, but then the agentic models

and what, what can we do with those things right now?

If we talk about how people kind of went through that transformation.

So originally, you're right.

I think people were using us a better Google.

And I think, uh, I would say, maybe throughout this conversation, it's almost important

to have two answers because I feel like there's a lot of things that Rob, as a developer,

as an engineer, is going to have very different answers than I, which, you know, let's just call

me a normie.

Uh, I'm a technically savvy normie.

It, like, I think that's there.

Yeah.

Yeah, I'm not technically ideas and words.

I don't know any code.

Yeah.

All right.

So if that's how we're defining technical, as I don't know code, whatsoever.

Uh, and so yes, initially, I would say, normies were using Chatch as a better Google.

And it took some months to figure out how to use the better Google and really how to, to

appreciate how much better it was.

And, and so I think there was a little bit of learning there.

And then there was a lot of flurry of activity that I'm not even sure I would have been

aware of if I didn't work in a tech company, where, you know, Rob was starting to go into,

you know, various aspects, which he can talk about himself.

And, and you could tell he was getting very excited, right?

Like, so he was going down rabbit holes and coming up and being like, this is amazing

more of this, just like so excited, um, and telling us what the future was going to

look like.

This is what's going to happen.

I was just like, slow down there, slow down there.

Uh, and then I think, uh, actually, I got a little bit of a, I would say a preview when

Rob pulled our team together and introduced the non-technical team to cursor, which was

some tooling, uh, develop, yeah, it was developer focused tooling, but it was a, enough of,

uh, we'll say like a chat style wrapper on top of it that for the first time, Rob could

see how non-technical people would, would start entering.

And I was very curious about it, uh, and I started playing on it and then I got very

sick.

And so I kind of stopped just because, uh, of medical issues pulling me away.

When I recovered, uh, from that, I kind of recovered and, and woke back up from a couple

of months, uh, of, of just being unwell, right as co-work, right as clog, uh, the new model

in desktop came out and people will talk about this, this big change over the last couple

of months is being clawed co-work.

I, I actually think that is a misnomer.

I think co-work was released at the same time as, as other things, but I think the real

change was actually just clawed desktop.

And so you download a desktop app suddenly it was again, just a very basic tooling on top

of these capabilities that made it norm, norm, norm me friendly, specifically you put

a desktop app on your computer, but the files that you create, uh, that you upload to train

things live in the cloud.

What that means is I can now use it when I travel on my laptop, I can use it on my desktop

at home.

I can invite my team, uh, to join a project and to contribute to a project just like

you would share a Google drive, uh, document.

And so something as simple as that, just organization and, and a wrapper, like nothing

really advanced from my point of view, uh, just that was enough that suddenly you could

use it as a non-technical worker.

I could start to organize things, I could work with my team and delegate things, uh, and

it really kind of changed from being a novelty or a personal tool to suddenly a collaborative

tool, which really does change everything.

Uh, so that was a, a big aha.

And then I think as we go on, we can talk about how operators, how, how my team is starting

to get more towards the code side.

But like, let me pause and talk about kind of Rob's entry into it, maybe?

Yeah, I would even just take a, like, one big step back and to talk about your question

about chat GPT versus open claw, why even GPT is a better Google and the initial, like,

in 2022, this came out.

The reason why the, in many of these things of the step functions of improvement beyond

just like the quality of the models, but the user experience and kind of the distribution

of these tools that are becoming more commonplace is the word context.

The reason why chat GPT in 2022 was a really good, better Google was because it wasn't

just taking your words.

It was trying to discern the context of why you were asking a question and use that for

how the answers get returned to you at a very, very high level.

These models are what's called stateless, which means it doesn't have any active memory.

You have this large neural network model that you put something in, you get something

out.

And I'm not sure if most people realize when you talk to one of these, if you talk to

it in response to you and you say something again, what you're doing is you're taking the

initial question, the response that it gave you and your follow-up question are all being

put into the model at once.

And that's more context.

If you're having a dialogue.

Repeditively.

So it's building on itself.

It's constantly reprocessing the entire conversation at every new request or engagement.

Exactly.

And why that's important is because that's more context of understanding what the conversation

is about.

And in 2022, it became a better search very quickly because of that.

Because you were building on a conversation.

Then I would say the next big step forward were these tools where you were pulling something

onto your computer.

And Tobacco's point about why Claude Desktop in particular is good when you, the early

phases of these when you would interact with them, maybe you could drag and drop a file.

And it could have that as context.

What ultimately, though, happened is when it lives on your desktop, rather than having

to go to a chat history and find the conversation and pull it in, you're able to retain context

across conversations as well as the friction of adding context to the model went down a lot.

Not just I'm uploading this document or this conversation.

Also, when I get an output and I curate a really good response, that doesn't just live

on a server somewhere now.

I now have a what's called a markdown file, which is a formatted text file, which lives

on my computer.

And this is where you start seeing the recursive loops of abilities like skills and everything

else like it starts building on each other.

On the developer side in 2022, I started toying with these things.

The issue, though, is the domain of language for conversation versus code, whereas code

is very strict and it has to work or not.

And so throughout 22 and through 23, really, it would be able to build up maybe a basic

prototype, but there would be many, many headaches and friction points.

Like this is, and this is where the common thing of like, it's hallucinating.

It's like became very part of conversations.

The thing to pay attention to here, though, is not just this, this is now existing, but

the rate of acceleration of improvement.

It's funny, like you're saying that you were sick and then you came back.

That's when we had another major step function up, like in that time.

So I had a similar dynamic with my wife, where she went on maternity leave, came back,

and she was gone for three months.

I was like, you need to do this.

Like you need to check this out.

And specifically for open claw, that just took it to the next level.

So rather than me having a conversation, whatever, you now have this context that lives

on your computer.

And it's being empowered to basically use the computer as its own personal resource to

go out and do other things.

Whether it's searching web, whether it's writing code, or maybe there's an ask you make

of it, and it says, okay, I will every day at 7 a.m. send you a text about a briefing

of what happened overnight.

To do that, though, it's writing a program to do that, storing that program on the computer,

and then calling out to that 7 a.m. every morning.

And the further autonomy you give these machines to have access to a computer to be able

to run things, the better able it's to manage context and build things and figure out the

most optimal way of executing things.

And that ultimately powers you.

And that's kind of like these continual stuff, the functions.

Let alone just when a new model comes out, the major jump you get in improvement.

And the major one lesson, the major improvement for these models across time, is they're getting

larger context windows.

You're now in a position where you would quickly overwhelm it, and you wouldn't be able

to add more information.

You'd have to drop information.

They're getting more and more context, which is allowing it more memory, and these are

the things where now we're at a point where even large complicated code bases are able

to be in one session, all be persisted at the same time, whereas two, three years ago,

that wouldn't be possible.

It gets, then I would say, further complicated because of the rate of change, because nothing

is final yet, right?

Everything is very fluid, that even as a non-technical person, there's model development.

And so how well the context and the recursive nature of the models is improving, so you

can have decisions of what model is best or what model do I want to use, what model is

smartest.

But then there's tooling and tooling developments and different companies have developed

different tooling.

And so, for example, right now, actively kind of this week, I've been largely using

cloud desktop and the chat and co-work functions of it, so I don't yet use code, though I think

I will.

But perplexity came out with computer, and they have different pros and cons or different

capabilities.

But I'm confident that's a temporary thing, right?

So it's how much do you dive in and learn and optimize and build to learn versus build

to implement when you know there might be something very better, very much better on

the near horizon.

So for example, cloud desktop is a desktop app, but the files live in the cloud.

But when it executes tasks, it creates a virtual machine on your computer, on your physical

computer, and it runs on your computer, which means your computer needs to be awake and

active doing that work.

You have to take that into account.

Perplexity computer, the virtual machine is in their cloud.

And so if you are creating, say, repetitive tasks that you want happening overnight and

you don't have your own server rack, you know, maybe if the capabilities of the model are

similar and perplexity, by the way, computer lets you use other models.

So you're not constrained by that.

If you really need that, that ability that the virtual machine, the cloud-based virtual machine

provides, which again, for normies is more about when it runs, not on your local machine,

then maybe you need to use, use that.

And the other thing is one of them, one of the two models has the ability to actually make

changes on your personal computer, which has pros and cons and security things to consider,

but can also help you create things, right, because I don't code.

So it can just do things to my computer to implement what I'm wanting.

The other one can browse the internet and control the internet.

So then there's plugins, you know, there's a cloud plugin for Chrome or perplexity has

Comet browser, which has those abilities mixed in.

So just right now, it's very, very agile.

So I think as leaders, we're both assessing the new tools.

Our team is starting to use new tools, which is exciting and seeing what they're creating

and, but watching how both of our teams are using these tools is giving us both in

our respective roles, you know, big eyes, right, because there's a lot of implications

for all these things.

And maybe just the very last thing I would go back to your original question, because

I think a lot of people just don't understand yet this term, and what we're even talking

about.

So when people are using chat, they're just having a conversation, right, and if you're

using cloud chat, you're still just having a conversation.

When you think about agents or sub agents or skills, these are all ways to indicate

that you are training a very limited task.

And you are trying to narrow the context window so that it can't mess up.

So it can't hallucinate and make mistakes.

And so OpenClaw, I've never used, so Rob can speak to OpenClaw in particular, but that

was, I would say, the big technical agent builder infrastructure, whereas now I will be

starting to use either cloud or computer to do it in the normal way.

Yeah, I guess what I briefly add to that is when you're having a conversation with either

chat GPT or a cloud AI, you're just having a chat dialogue back and forth conversation,

it just sits there and it doesn't leave that, like it just leaves, it's that conversation

that's it.

An agent is taking the capability of these chat models and saying, I want you to take

the steering wheel and go build and do things, right?

That is the next level of kind of breaking things out.

So maybe in 2022, if you were writing code, you would, what you first had to do initially

was like you had to take your block of code, copy it and ask a question, take the answer,

copy it out, put it back and check and see if it works and kind of like do that as your

loop.

The agent model, like in these, maybe what's also called a harness, if you heard of

cloud code or open code, code access, opening eyes, is that we're saying, okay, we want

you to run these large language models, but here is the code base and maybe even like

you're talking like an open clause saying, here is your computer.

Now go do a bunch of things.

The one thing because we're just started talking about skills, the way I view skills and

what these, like you're training on a specific sub-domainer task, you give these ultimately

all distilled down to what's called a markdown file, which is a fancy formatted text file.

And these markdown files, what you're doing is you're adding context.

So you have this stateless model, it has no memory or things, but you're saying, hey,

you are a leading expert in insurance or you are an expert in Bitcoin custody, right?

You're giving it guardrails to go through the infinite universe of what this model's potential

is and you're immediately gating it in and saying, this is your focus.

You are a QA, like quality analyst for code.

You're just a tester, right?

And what you can do is you can take the same model and give it different skills and it

can then narrow the scope of what it actually does.

And the way I view skills is if you think about like a craftsman has a big tool belt,

maybe a hammer and a wrench and a screwdriver and they're all kind of circling around his

belt, what it is is you go to the model and say, hey, use this tool or you can have a tool

that isn't just like an agent personality can be, I want you to run like some early examples

of what I've done is like you, I want you to use these tools to go through and verify

a Bitcoin dress.

And this is exactly what you do and make sure it follows X, Y and Z spec.

And I give it context around the business.

And then when these kind of questions come up, rather than from first principles thinking,

hmm, how do I solve this problem?

What it does is it'll look through its skills list and see what skills do I have and that's

like a craftsman looking at the tool belt wrapped around his waist and saying, ah, here's

my tool that does this thing, right?

And this is, and because it's language domain, like the instructions of the, these skill

files are not code in these skill files, you can have sections and say run code like

this or run code like this.

But you don't have to give it specific code instructions if you're not using it for specific

code things.

But what you're able to do though is kind of encode these different personalities to

start thinking about things.

And over time, realistically, as you're improving on these things, you are going to inevitably

have pieces of code that get peppered into them saying, this is my, I want you to make

this call specifically to look at my calendar, or I want you to specifically look at this

formatting code for how I write things in general, right, for like code format.

So that's like the general mental model I like to give people is that these skill files

are like tools on a tool belt and the agent as it walks through problems will be able

to understand, like, ah, this is what I need to go look at and do.

And maybe a real life example.

That's what I want to think about.

Yeah.

So here's, let's say I want to use AI to help me develop new policy, insurance policy.

And you know, at the, first of all, AI, there's a whole new thing to think about of, you

know, at what point you do legal reviews, you know, there's all sorts of change, right?

But let's say I'm going to use AI, that's my goal.

I want to use AI to help me write this new insurance policy.

I already know the policy I want to write.

I know what I want to cover.

And so I might have a normal chat that you're very familiar with in to set the context

for that.

But in terms of agents or sub agents or skills, let's say that, and the way I think of

it in my head is the more generalist an agent is, the more tasks, I think of it more human

and probably over time, I'll name them to have a, have a name, right?

I was watching a video by a guy called Alex Finn, he's really prominent online, doing

tutorials and stuff.

And he literally names every single agent.

He has a chief of staff with the name, the chief of, he only speaks to the chief of staff

who then command is all the other agents kind of kind of, I mean, that's one way to do

it for sure.

But yeah, the more, the more of a generalist it is, I think of it more as a human.

And then if I'm getting into sub agents or tasks, I really don't.

I think of it as like a thing.

So in terms of the policy example, maybe I would set up a project in, in cloud desktop.

And that's where I'm uploading my, the previous policies that we've written, maybe some context

from the internet on research that I did are prior conversations as I was planning for

it.

And overall, I have an agent who takes those things and will spit out a insurance policy

for me.

But I start looking at it and, man, it's hallucinated some crazy stuff and it's, oh, man, this means

so much review.

It's a legal document, so it's, you know, and so what you can do, for example, is in insurance

policy, there's a definition list.

There's a prescribed format for these documents, it's regulated.

And so it must have definitions and so we put a lot of attention into those definitions.

They have to be extremely accurate because they drive really the rest of the document.

So what you can do then is, for example, develop a dictionary sub agent.

So you could name the dictionary a name if you want to or you can just think of the sub

agent is the dictionary.

And so then I would train a sub agent to say, you are the dictionary.

Our job is when I create policy, you are to take our approved list of definitions that

I've written before.

And maybe to train this dictionary skill, all upload load, all the current definitions

and then based on my aspirations for what I want to do, it will help me create, identify

new words that I need to define and I'll work on those.

So you'll use AI to make the dictionary better, the actual list of definitions better.

And then once you have it, you say, that's a job, is you are the keeper of the definitions.

And I can either manually, when I get to the point that I think I have a good policy,

I could say, hey, run the definition or run the dictionary skill.

And it would go through my policy and it would tell me anywhere anything in the policy

might contradict our definitions, where I might need to look or it might tell me, you

know, it will go through and it will not allow a word to be misdefined, because it only

had one job.

It's only job was to make sure the definitions and the policy match the approved list.

And then maybe I'll have a second agent whose job is to check the documents for that.

Are you creating, do you think of it as kind of like an agentic or structure and line

of command?

I don't yet.

I don't yet.

Right now I view it as tasks.

I all report to you sort of, all these agents feed back to you.

I'm not there in my head yet.

I know many people do that kind of see themselves as CEO.

The first time I drew, drew it out, it was both work, but my personal life too.

And so yeah, like I drew myself in the middle of CEO and I just moved away from it.

It didn't, that mental model didn't help me.

What got in the way?

I just think it's unnecessary.

It's premature.

We talk a lot about different neuro abnormality, atypical thinking, I'm super ADD, right?

So I've already always described my brain as being a cloud.

And so when I learned something new, I just throw it up there and I just let it flow.

And then my brain takes care of connecting the synapses on itself.

And so that is how my brain works.

And when I got into, yeah, prematurely identifying agents and roles, it just wasn't helpful

because my brain is still in accumulation mode on this topic right now.

I'm still learning.

I'm still like drinking from the fire hose and throwing things up there.

Now maybe later, maybe later when we figured out exactly what we're going to do in a professional

sense, maybe I will get to the point that I've got these X number of human people reporting

to me and then these Y number of AI agents reporting to me, but I'm not there.

I would say in general, like the comment about it being premature is not that you can't

get productive output today.

What I would encourage for people in general is to start iteratively tinkering because this

is an entire new domain and how to get work done in the skills that are reflected in them.

And then we talk about this where non-engineers are inherently going to become more engineering

minded, even if they're not literally writing code.

But since these computers are very literal in understanding walking through these conversations,

you need to be almost like an architect, right?

Or a way maybe from a non-engineering perspective would be like you're like a product manager.

And because you're taking this higher level of abstraction, when thinking about these things,

the underlying tools are changing, not just because the acceleration of this stuff is happening.

What's interesting is not just it's getting better, the rate at which it's accelerating

is increasing, that you don't want to get too invested into this is my exact architecture

because very likely in a couple of months, it's going to be entirely transformed by a new thing,

right?

But the encouragement to tinker comes from a place of this is a whole new domain space of being

able to have a conversation or being able to curate lines of thought, basically what you're

doing, you're creating a skill fire in an agent, you're like this is the world in which you live in.

That is a skill in of itself, and as a general tip for people at home, the most important

skill, Claude, open sources, because they're just markdown files, right?

This isn't code, they have a list of all of these different skills, Claude, Anthropics

specifically has the skill building skill.

So meta skill, this is how you actually build skills, and that's the one I would start encouraging

people to and you can start having convert, you can load up that name again, Anthropics

Claude, which version?

Yeah, so any version, right?

So these markdown files are just blocks of text, so it's not like opus versus sonnet 4.5

or 4.6, just the skill file, and this is like the idea of is it more important to have

an ability or to learn to make a new abilities?

And this is in the Anthropic GitHub page, you can call it like it, so Anthropic is the

company that runs the Claude models.

The skill building skill is what it's called, and you can find it pretty quickly online.

It's on GitHub, but it's just a quick text file, you could just click, pull down, load

it into a skill on your computer, and that's a great place to start because you're going

to have decent guard wells on what it takes to build a skill.

And then the reason why you want to take her to is you want to understand, I was way

too specific here, I was way too general there.

I should have added A, B, and C pieces to it, and that's the part of tinkering I think

should be encouraged, understanding that we're moving, we're changing the domain rather

than having different people, you have different text files that are able to start, and this

all goes back to context.

You're just learning how to manage context, and this is an important thing for you get

to a point in these conversations where there's something called compacting.

So let's say there's a one, so you've may have heard the term token, my two sentence summary

of that is that what you do with words for these models is you actually transcribe them

into numbers, which is called vector, you take a word and you make a vector out of it.

And what you do over sentences, all of these words turn into numbers, they get put into

the model, because it's a mathematical transformation, and on the other side you get numbers, and

they convert those numbers back into words, right?

So that's kind of like the magical sleight of hand here, why are words doing all of these

complicated things?

It's because it's all math under the hood.

And where this goes to for the tinkering of the skills is what you're doing, and when

you get to a compaction point, a million tokens, which is more than a million words, but

let's just for the sake of conversation say it's that the model says, I can't retain,

I can't do more.

You've hit my limit.

Let's make this smaller and it'll go back and make a summary, right?

So this turns into memory management, and it'll make an attempt to summarize, but maybe

you have strong opinions on what should be maintained, right?

So this is like the things you start iterating and thinking about over time, and creating

tasks into small bite size tasks is a great way of managing context.

And so loading your entire code base or every single document you've ever written in your

company into one thing is not going to meaningfully actually do that because you actually, if

you get too much, this is one last thing of it's called like a vector database or retrieval

augmented generation where I have my folder of all of my documents, and I ask a question

and they'll go look at the folder and be like, oh, this one, they'll look at the title

and be like, this, this may be similar and pull the whole thing and go back and forth.

It gets to a point where everything kind of clusters together and it's all noise.

So your own curation of how you think about organizing data, which is funny, because

it's an engineering principle of like how you organize and kind of break things off, that

becomes the important skill.

Your data curation, your organization of how you group data, and then also learning how

to build these skills.

That's those are skills that are going to be across time still important because these

models have fundamental mathematical constraints.

By planning those, it made me really start to understand a new area of management requirements,

I guess, in the future.

But it'd be fair to say, and I know it's, you just said what a token is, right?

So I know it's not literal, but is it fair to say that tokens are more or less energy

or money?

I mean, the more tokens your employee uses, the more money.

Oh, that's true.

Yeah, so just at that level, you say gas pump when you go to top off that you're paying

by the token.

Right.

Currently, let's say a non-technical employee of a company will have very little spending,

right?

They might have a corporate card where they can cover certain expenses, they might travel,

they might pay some vendors for some things.

But they themselves, if you're talking a corporate worker, office worker, is not generating

a bunch of spend.

Suddenly, your team could be generating huge amounts of spend, doing work, but how does

a new manager start to manage their budget?

So I think what's going to happen this year and next year is across the world.

Suddenly, the companies who allow their employees to actually use their computers to do these

things.

Suddenly, companies everywhere, managers everywhere are going to find that they blew their budget,

like blew their budget, right?

You could have a team of 10 people that suddenly has generated a million dollars of spend

over the course of a year.

That's an actual possibility.

Businesses can't sustain an unexpected million dollars of spend.

And then it might not have achieved anything, especially in the next two years, where people

are iteratively learning and trying to build things and it goes nowhere, it doesn't end

up being used and all that spend did not accomplish the business purposes of growth or advancement.

There's just two brief points.

One, at the moment, most people are using either a subscription plan, either it's Clawed

Code, they pay anywhere between 20 and 200 bucks a month, or they're using a version of

that for opening eye and chat GPT.

Those are being subsidized at the moment, 95 cents on the dollar.

So people are already complaining about the cost and or if they reduce the limits, well,

you're only going to give me $4,000 of compute for my $200 buy and people are like, I never

use an anthropic again.

I'm going to go to chat GPT because they let me take my $200 and spend $5,000 of compute

and you guys are only going to spend $4,000 of compute, right?

Someday, the subsidies will go away and then what I'm describing will be even more.

Even in the subsidized world, companies are going to get surprise bills or unforecasted

bills.

When the subsidies reduce and eventually go away and people are actually paying their

way.

Well, in the video I watched last night, it was a peter-domineless, is that his name?

No.

Yeah.

So he had a sort of bit of a panel discussion on his podcast and one of the guys said, yeah,

I could let one of these agents go overnight and you come back in the morning, you don't

know how much it's going to spend unless you put hard limits on it.

And so he could go off on all sorts of sort of wild use chase and you don't know what

it's going to come back.

Well, this is the other thing, too, to back us point about like, there's a, at the moment,

there's, these tools are incredibly productive.

It's so easy to be seduced by the illusion of productivity.

The idea that you can have a cup, you can wake up, answer a couple of emails, have an idea

and then by dinner time, something's built, right?

And you're like, wow, this would have taken weeks to get to this point.

And then I think it's really important to have that kind of freedom to iterate prototype,

think about things.

On the engineering side, for a pure proof of concept, I'll code something up in an afternoon

and then I'll realize, oh, I actually missed like five different things in this product

spec.

And that gives me more information for when it's time to actually build things.

But it's so easy to be seduced and be like, look at all of the work I'm doing.

Look at how many documents I've made, right?

And somewhat like a general principle in data science is like garbage and garbage out.

So like if you're just casually like doing a proof, quick proof of concept, it can be productive

and interesting, but it may not be the actual end piece.

And to your point, like, unless you set hard limits on these things, and that's like,

when you get to this this agent model in general, you're very quickly going to hit your

usage limits.

If you're doing generalized, like go explore overnight and do things, tasks, which means

you have to go to what's per token rather than having this like subscription based thing

where you get a certain amount of usage and whatever, you need to actually use the API.

If you're using the API, that's truly like a gas pump.

And you're paying twice.

You're paying for when I send you data and then when the model responds, those each are

input and output and those each have their own cost to them.

And that's just its own reasonable thing to manage.

And then that turns into a skill of maybe you start discerning these when people talk

about chat GPT or Claude, there's for Claude, there's Sonnet, Opus, and Haiku.

Opus is like the best model.

It's more expensive.

And then Haiku is the cheapest model and it runs faster, right?

And then you start realizing like, do I really need to use Claude Opus to ask about the

best coffee shop in my neighborhood?

Is that really requiring the most use, right?

And then there's also is a way you can sub break out agents as you can actually say that

this is a-

Tell us the selector model?

Well, you can just code saying that whenever I talk to you, I want you specifically

to use the lowest quality mode.

And then on the GPT 5.4 side, like the latest model there, there's different thinking

levels.

And this is interesting, the way the models are increasing their thinking and this was

a really big thing that around the time you were getting back that was really like stabilizing

is thinking mode is not like the model is just a mathematical formula.

If you want to call up that, it's not like it turns on an extra brain.

What it does is if you ask a question, it will have a side dialogue with itself and

recursively think on it.

And they're charging you for that usage, like that's all usage.

But the iterative process, it's actually modeling like an inner dialogue of asking your

question, thinking about it before you go.

And the way that gets emulated in this code is it'll talk to itself for a while to think

about what the best answer is and you can set different levels of thinking.

Because inherently don't cost more or less, but if you're taking the higher think mode

or the extra high think mode, that will just inherently be more costful because it's

taking more time.

And a very simple, just very quick simple example of how I do this is when I'm architecting

a project in an idea, I use the ultra high think mode because that's the time to be really

expansive and think about the opportunities of things.

And then it'll come out with a playbook.

And then I can pull down the, since that high think mode, the output is this very structured

product document.

I can then tone down the level of thinking to then just go execute the specific detail

tasks that are at hand.

I think in the future then, I don't mean in the next two years, I mean the next five

and more years.

I think part of being a good worker is that you will inherently learn how these models work

and you will start to design your own work to be efficient.

That will be partly forced upon you by your employer to manage costs and your own time,

not waste time or waste server time if it's happening locally.

But as a user, this one will not be that hard because users will want to do it because

think of it like, I think of it like using a slow phone or a slow computer that needs

an upgrade is you can't, you can't change the phone, right?

But you will, you will change your own usage of it to meet the internet speed that's

provided to you.

So you will or you won't do certain things based on the amount of time it takes you.

So if you build, let's say, if you don't use an agent model and so every single time

it's doing anything, it's searching this giant context window and taking a long time,

it's going to be so annoying.

We're going to be wanting instantaneous reaction just like we demand from a browser window.

And so I think, I think over time, people will learn to use agents and sub agents and skills.

So it's super efficient.

So that way when they're working or doing their personal life, it's closer to instantaneous.

So I think that kind of stuff, some stuff will happen naturally and some stuff will be

forced upon you by your employer as a cost or efficiency or other HR control related

things.

I want to explore some, I guess, tangible ways that people can use AI now.

And I want to break this down between simplified tasks and perhaps some more complex tasks,

both technical and non-technical because a lot of the discussions I had for the sake of

understanding kind of like a generalized conceptual level.

People watching this at home, I think, and yeah, but how can I actually use this when

I'm seated at my cubicle?

What work can it actually do?

I think that's genuinely an exercise in changing your perspective on how to evaluate problems.

And the reason why I say that, it's entirely, it's entirely, here's a very just, like as

a basic example, I was talking to someone who was asking me how to use these tools and

they're like, well, I know what to do.

I'm like, well, what is it that you work on?

That's a repetitive task, and they said it and I was like, okay, then go do it.

And they're like, what do you mean?

I was like, explain your repetitive task, have a conversation with it, and ask it, how

can you help me automate this?

That's like the ultimate, just like directly, and this is where people are used to with

code and software, like you have to go to school and you have to read a bunch of texts

books, and then you have to like get a code editor out, and then you have to think about

it, you have to design it.

All of that is gone in general principle, especially for someone who's not technical and

is looking to just go from zero to one and start using this stuff, having the conversation

and just doing your own inventory audit and think about where you're spending the hours

of your day, and just start having conversations and it's an iterative process.

And this is why I make the point about tinkering is important.

You want to start flexing that muscle and thinking about how to interface with this new

technology, and that's where you can go in general.

I can tell you where I started.

My first one, my first one is, I hate email, I'm terrible at email.

I had a friend and an exec at Starbucks who said when he came back from vacation, he took

all like 470 emails and he hit delete, and he's like, just because somebody sends me an

email doesn't mean they deserve or like they automatically get my time.

I was gone.

If it's important, don't email me again.

I was like, my mind's blown, right?

I wish I could take that out of it towards email, but that's how much I'm bad at it and

I don't like it.

And so I was just like, I want this problem solved.

And one, some of my problems are there's a certain type of spam that I'm certain you're

familiar with that makes it through the spam filter.

And they do it obviously in certain ways.

It's instantaneously recognizable where they don't have any links.

They spin up a new domain so it's not a crazy domain like old spam used to be.

I wanted those out.

I wanted things that I have to delete manually like a calendar invite that I've already accepted.

I want those out because with my ADD brain, those things they suck my time away.

I open up email to do legitimate work, to look for that email from somebody important

and respond to it.

But right before I get to that, there's these 17 emails that irritate me and so I end up

doing that.

I get distracted.

Maybe I run out of time and I have to go to a meeting I never did the email.

So I described that I wanted to create an email agent that could help me solve those problems.

So I talked about how to do that.

And so earlier, remember I said certain models can do things to your machine.

Certain things, certain models or extensions can do things to your browser.

And so I started experimenting with those.

I ended up using both Claude and perplexity comment browser specifically.

I've used them both and they are going through Google web apps, script app, web scripts.

And so it is code, but it's doing it into Google's kind of platform that they already had

in place.

It's important for compliance that you're running this logic.

You talk to an AI agent and it gave you an output that you're now just running within

our enterprise account.

It's not like anthropics reading every email that's coming in.

Right.

Right.

And so it was just to help me create this email filter.

So I talked to it for a bit and I said, okay, run it.

Let's see how it does.

And it said, okay, I ran it.

I said, well, I need to know how you did.

So okay, now when you run the email, send me a daily email, recapping what folders you

cleaned out.

I had eight or ten different things I wanted it to do in terms of filtering and cleaning

things out that were a little more advanced than the native filtering skills.

And so now then the next thing was I got a daily email and it listed everything that it

had deleted or moved to spam.

And if I see something that did poorly, I could go retrieve it out of the spam folder

and save it because it was maybe a bill or something like that.

And then every, let's say four or five days, my email would irritate me again.

And I could tell that my email assistance was not doing well.

And so then when I found time, I would open up that kind of conversation.

I'd be like, the email filter is still sucks.

It's still not doing the spam.

It is terrible that there's a reason that stuff makes it through the spam filter.

Even though we can perceive it, apparently it's very, very hard.

So every maybe five days, I would put in a few more requests or I would show it a screenshot

of something and be like, you know, this is happening.

Why is this happening?

And it would improve it to the point that I have like a moderately useful email assistant.

And then I would say the most advanced one I've done is really was done for plain, at

least most advanced in terms of like my thought process and getting it going is actually just

one that is a get me out of the door when I'm traveling, packing to go to the airport,

tool, skill, I guess, where I tell it what I need to do before I leave.

And now it knows, now it's learned how long each task type takes and it will more or

less estimate my amount of time based on what I have going on and it will remind me currently

by sending me a text message, but where I what I want to do when I get home is a final

step where I want to actually connect it to the Alexa, the Amazon Alexa verbal thing,

which is the world's stupidest model, by the way.

It is an LLM now, so you can have conversations with it, but it's really bad one.

But I want to next take my get out the door tool, hook it up to the Alexa, so then I have

audible call outs.

So that's it.

Those are the types of things I did just to play to learn just a very concise one for

most people who use like many people of their desk jobs use Excel.

If there's regular Excel tasks that you're doing all the time, have a conversation, drop

it in there, it could work well with chat, GPT, if you want to make us, if there's something

you're always doing to a spreadsheet, you can use cloud code or co-work and it will

make a script that'll automate that or to summarize, right?

If you have very large complicated documents, give me a bit of a quick summary calls, like

that's where it excels really well to just a or if you work with a lot of large documents

being able to summarize a large.

Do you use it to learn much?

Learn with what?

Well, an example was I needed to write an email back to Lloyds of London, our contact

at Lloyds of London, and I needed to make an arguments based on insurance math and

actuarial data that I didn't natively understand, and so the goal is just to help me craft

an email response, but I didn't want to just send an AI response.

I wanted to be able to make this argument, and so I actually took four hours in all with

the ultimate goal of just an email response, and I had it more or less teach me actuary

math.

Yes.

Which was amazing.

I could have never written that email.

I could have never written that.

I had a similar experience.

There was something technical, I wonder if it was math or chemistry or something that in

high school, I had, oh well, I did a lot of things wrong in high school, but I just had

a question, and I asked it, and I'd explained it in a way that I've never heard anyone explain

it to me because I'm like, oh, yeah, I wonder if I can go back and learn co-learn with

one of these models, all the topics that I struggled with in class, because I mean, as

much as I was in a conventional sort of learner, I'd always get these weird questions popping

my mind, and I'd treat them as roadblocks, because unless it was answered, I couldn't just

focus on the rest of the class, because I'd be so fixated on, but what about that?

Right.

And of course, you can't interrupt the class every 15 minutes.

So it's one, absolutely.

That's a perfect example.

Tangential example is, homeschool my kids, and my wife makes homeschool worksheets, whatever

they look on, and whatever they're interested in that day, as a quick conversation prints

out a worksheet.

Boom.

Now, we have custom tailored learning, and for general learning, I use that all the time,

and kind of checking my assumptions, digging in deeper, trying to read up on things, for

sure, and this is where people have different relationships with AI and how they use it.

I'm very productivity and work focused, or knowledge, or just general, like information

gathering, but some people use it for a therapist, right?

Having a conversation, right?

And that's, that's never been something I've personally used it for.

Careful.

Yeah.

Never saying, I've never personally, I've never personally used it as that.

But that's where I think learning and exploration in general, and I think it's also interesting

too, is like even on personal side, having, like if you were to use that, you could have

it look at patterns over time, devils advocate your own positions, being able to value other

people's perspective, and thinking about things, and how to communicate, whether it's

a work email, or someone that's in your life, like these are all things that you can use

as, I use it as like a recursive, just like, I have a thesis, challenge it, like modify

it, iterate, move forward.

If people want to start to understand the concept of cloud code, I really recommend they

start to, well, not just specifically cloud, by the way, like coding and the future.

Rob one said to me back when he was getting really excited, but before I, the tooling had

developed, and I, I saw, he's like, no, everybody is going to code, and I said at the time,

my response is, no, they're not, that's not how, like I've managed too many people and

just to know, no, they're not, no, they're not.

And he was like, yes, they are.

And the thing is, is we're both entirely correct, which is pleasant for a co-founder relationship.

Well, well, yeah, I mean, Chris's thing, yeah, yeah, that's where I'm headed, but yeah,

we're both absolutely correct, because when I talk about my email, that is building,

that's coding.

And just talk to it and have a conversation, I talk to it and it built something for me.

Not, you know, right?

It actually built something for me.

And so in that way, Rob is entirely correct, everybody is going to code, everybody is going

to build, I would say.

And when I'm saying, no, they won't, what I mean is, no, they won't, they will use a

wrapper and they will interact with it in a normy way and it will do the coding for you.

Now a fun anecdote that like we're excited to talk about is actually kind of a, an in-between

that I think is an opportunity for people who have the interest in, in going this way.

And it is this in-between state, where you're not say to an engineer of today, you're not,

you're not a coder of today, but you're much more advanced than say your average office

worker.

And so Chris on our team is our head of ops, he is very technically oriented and interested

and he, the things I was using Google script for using AI.

So AI was using Google script on my behalf.

He was already the type that was experimenting with out on himself and he was, it was just

very laborious, right, because he was kind of figuring it out as he went.

But somebody with that inclination saw a problem, which is we have one of our backend systems

is, you know, it's a small piece of software, meaning a small company.

And you know, like any kind of back office software, there's parts of it.

Maybe there's a bunch of different modules to serve different clients or different businesses

that license the software.

You use parts of it, you don't use parts of it.

And so what he decided is like this part of my job is very frictiony.

And I know that we pay a lot of money for the software.

So I am going to rebuild it.

And so he and I had had a conversation and I was talking about kind of my ah-haws about

how people are going to use these tools.

And he took that as a kind of go ahead to start experimenting.

And I talked to him on a Thursday, apparently on a Friday, he started.

And by Monday, he showed me that he had rebuilt and we'll tell you what we meant by that.

But rebuilt this piece of software in entirety.

And in terms of looking at it on the screen and how it appeared, it worked.

It pretty much worked.

So if customer info went in here, all the proper things would happen.

It would spit out our desired information.

And so he had created this on the surface and think of it as a prototype effectively.

And I gave him some feedback on it.

I said, well, let's say he's the user himself.

So if you're inputting stuff that's currently on these three pages, wouldn't it be

nicer if they're all on one page?

It would be faster, fewer clicks, right?

He's like, oh, yeah.

So then he just by taking screenshots, right?

So he started with screenshots of existing software that he uses.

And then he told it what he likes about it and doesn't like about it.

He explained how he wants to use it.

He spent a lot of effort into context.

But in terms of cloud desktop, I said, it's chat, co-work, and code.

Chat is chat.

Think of it just like GPT.

Co-work, when I said at the beginning of our conversation, was a bit of a misnomer

that everybody seemed to think it was co-work, was the big unlock.

Co-work is, think of it for really building deep analysis or specific tasks.

I almost never use co-work.

Clawed, or I'm sorry, code is this big step farther where it really will write the code

and can go so much farther.

And he showed it to Rob, and I'll hand this over to Rob in just a moment.

When he showed it to Rob, I think Rob, he can do his own reaction.

But I think he was pretty impressed when we talked about it as executives or managers

over our team.

I was very curious, his point of view.

And he said, look, I mean, on the front end, very impressive, on the back end, very

unsecure, not the right languages.

Everything's Jerry rigged.

And so then I'll pause, but that's ultimately then what starts a conversation here about

how do we want to take advantage of these crazy abilities, this insane ability that this

guy who's not just rebuilt a piece of software that we pay money for, right?

And what could it look like in the future when we can overcome the security and the

jankiness?

So before you learn, so I want to set some context, my personal interest in this question

is immense because there is a challenge at the moment that's in my mind is that there's

never been a solo founder entrepreneur who's built a billion dollar company on his own.

I think the possibility with the tools that we've got emerging, I don't think we're

there yet.

Sounds of it.

But I want to hear your opinion.

No, we're, I think it's possible, like, and this is, I'm not the first to say that's

it's possible that that person's building that company right now right now.

It's very possible.

So for this anecdote, like, of building things, it was funny because I got pulled in when

basically like, because like, oh, I just talked to Chris, he's rebuilding stuff.

And I was like, oh, he's, he's building software, like, you know, like in the company,

I like, I managed the engineers and I'm like, okay, let's see how this goes.

And it was impressive in the sense that tell us what you were impressed by first.

Well, that's what I was getting.

Yeah.

So like, from a strictly, just the front end and looking at it and just building something

you can touch for a prototype.

And I made this comment earlier that I really enjoy using these coding tools and actually

it's funny for your distribution.

I almost entirely only use cloud code, even for analysis.

And only if it's like very superficial, quick question, well, I use chat.

And that's basically, but 95% of my work is with cloud code or codex or like these,

like the coding level tools.

I'll get to that back in a second.

But for what Chris did was so impressive was, this is something we would have hired a

junior engineer a year ago to work on for a week or two and just to get to the prototype.

Just to get to the prototype of what like what Chris had done.

And it looked nice.

It was functional code like the front end was like, it was real framework like there was

a library there.

There are certain things that I'll get to in a second that are easily addressable.

There wasn't a database.

There wasn't any security validation of when you talk to a front end website, when you

put data in somewhere, you don't want to let any arbitrary blob through.

It used an insecure version of TypeScript, like which is just a coding language, right?

And in the post mortem of talking about this was one, like we were very happy, like for

Chris to go off and run and do this, even though it's not really like his day job.

But like this was a good exercise to start figuring out what the future of work looks

like for people that are highly motivated office workers.

So as managers, us kind of looking at this as like a post mortem experience and understanding,

okay, this was a fun little experiment.

Took a couple days.

Chris wasn't burning API token credits, right?

So this was all within his subscription limit.

So as a quick iterative loop of understanding, I want to build something.

I'm a motivated office worker.

How do I kind of go to that next step?

What we did was talking about it, I realized that there were several things we could do

across teams on the first side.

When looking at, you know, building this out, I may have mentioned this earlier thinking

like a product manager.

You need to be very disciplined and specific up front because it's way easier as these

machines are building things to have the good context and the good patterns up front.

Then to say, oh, I made a misdissumption or I forgot something, go bolt this back in.

So what does that mean?

Does that mean work on a full product requirement description or product requirement document

first?

Yeah.

Absolutely.

And say, build me, like if it's a photo editing app, build me a photo editing app.

Exactly.

Exactly.

Exactly.

I think it is important to know going into a test, I guess, or a project knowing how it

is intended to be used.

I think doing that in general is not generally necessary if you're building something for

yourself.

If you're dog fooding something for your own tool, it's not going to go elsewhere.

Then as long as you're using your time and resources wisely, you know, it really doesn't

need to have that level of review.

The question is, if we want to use something as a company, if we want to start building

the tools that we need to be successful and do them in-house, what does that look like

if it's involving non-technical workers?

And that's where the PDD, I think, becomes very important and that's what Rob shared

is.

That's where we need to begin now.

That's what we learned here.

Yeah.

Before anything's written, you're going to get way further better returns.

I made this comment earlier about how maybe this is where you'd use your higher quality

models, extra thinking time, being very slow and methodical, and just very detailed building

out specs.

I think what's important for this is having the engineers aware of it because if Chris

were to build something, he's going to build it and then someone else doesn't have to

maintain it.

Software devs in general, if at all possible, don't maintain other people's code because

if there's issues or problems, they are issues and problems that were created by someone

else.

Now it's their job to be janitor.

Chris gets to have a fun four or five days sabbatical building something and then throws

it over the fence.

Now my team's like, well, this sucks.

Chris is great to work with.

This is now my problem.

Now I have to own all of the technical data associated with this.

I have to own all of the complexities or the misassumptions and I have to kind of

bolt and do these things on later or maybe even possibly rewrite things, which then just

becomes tedious because it's way easier for a developer going from the start, actually

building this in methodically building out in a way that makes sense to them that may

not be exactly the way a model has generated something.

I was going to say, yeah, so this is a very real-world non-AI problem, but it's exacerbated

because you don't have anyone you can go to and say, hey, why did you build them like

this?

I'm talking about this up front with the software developers.

I think it's very like why I say this is a cross-team functional thing is that it would

also be really helpful for our developers to say, these are the languages that we use.

This is how our general system is architected, so it's not like, I'm going to do a database,

so I'm going to build it for no sequel, which is a database type where we use Postgres,

right?

It's like one of these things where you want to immediately up front kind of call these

things out.

It's like as in going through our code and thinking about, okay, whenever we build a front

and website, we do things like form validation.

Let's talk about how we do form validation up front and maybe even show some patterns

of how we use that and give that to someone like Chris, so then when there is an intersection

of work streams, it's something that's already following the basic code patterns of things

that we are exhibited to.

No, that's not the only way a company can or should allow their teams.

That was our learning and our decision for how we're going to use this if it's going

to be deployed.

Think about what else we accomplish there and how else you could choose to use it.

At the beginning of talking about this anecdote, Rob said that he was impressed on the front

end because it would have taken a front end dev weeks to mock up this, whereas it took

a non-technical person a few days.

He didn't mention is actually the time to hand over the request to that engineer to start

building.

The way we used to, like if it's a back office tool or whatever, if I wanted a certain

functionality, generally what I would do is I've got a big digital whiteboard or an iPad

Pro with a pencil, something like that.

Usually I would start by drawing my own wireframes, so I'd have here and maybe a rough button

here and then arrows and then goes here.

I would be using words with the devs to be like, and here's what I want this to do and

here's what I want that to do, and they're already maybe starting to get frustrated because

I'm making requests that they know how difficult or challenging something might be.

I don't know.

I just want the button here.

I just want the button here.

You might get to a point where you have drawn out approximately what you want.

Then you hand it over to that front end dev and then you get back something from them,

which may or may not match what I envisioned in my head when I made the request of the tech

team to start with.

Even if you don't want to have your team go as far as deploying software and you need

to have that security and infrastructure done properly, what if it's just a very fast

way to communicate now because he's now built a working prototype that even if they can't

use a speck of code from it, he has now communicated more completely than any hand drawn conversation

could have ever communicated.

He has built exactly what he wants and so a company could choose to be like, look, this

is already a better PDD than I could have ever received.

The product type is when you've got teams, cross functional teams working with one another.

You have the end user, so I want something like this here and then the product manager

or the engineering team would be like, what do you mean?

It doesn't make sense to us.

If you can actually show them what you're doing, this is a really important point.

This was the strength of any person within an organization as people have cross-disciplined

understanding.

To be the bridge, whether it's information or understanding, whatever it is, to be the

bridge is a really good place to be in an organization.

One of the best strengths I had when I was on a consultant at IBM working with people

and they'd be like, oh, that would take two weeks and I'm like, can you explain how that

takes two weeks?

And they'd be like, oh, because X, Y, Z, and I'm like, well, A, B, and C, right?

Whatever the specific mechanics are, the ability to materially collapse inward the knowledge

gaps and make the bridge shorter to move faster.

But then to also, in running an organization, being able to quickly discern who's working

hard and who's being productive, is immensely valuable now.

That is something that is just very inherently where we have a very small engineering team.

We don't have this problem, but at large organizations, if we're working at Starbucks, I want

to update SAP or whatever it was to do X, Y, and Z, and you have to go, you have to file

a help test ticket.

You have to go do a meeting, right?

Not only would you run into things when you're using external software, I did work on

SAP.

So let's use it SAP Ariba.

Some people out there are like, oh, I know you're paying.

It might just be a time thing.

So I might come in and say, hey, the users of the software, the 10,000 users of the software.

People paying for it.

Yes.

They all want to button here because this task takes, I counted, 17 clicks.

But if you put a button here, it would solve that problem.

Sometimes when you're using external software, like SAP or an Oracle piece of software, sometimes

the engineers can do that for you.

You have to wait for the Devcycle and wait your turn, but sometimes they can put that

button where you request for your company.

But many, many times they would say, no, we can't because other clients use the SENDAT

is at that level.

And so we can't customize it for you.

And so it was just like, OK, so these 10,000 people are all spending nine minutes six times

a week that they shouldn't need to.

And then you do the math and across an organization like the lack of a button cost, you know, X,

X many minutes and therefore dollars of productivity.

And that will be a thing of a pass.

And then Rob also mentioned kind of workers and knowing if people are working.

That was actually my first AHA when I got into using these tools for work versus just

playing and using chat and learning.

When I actually started using them to make myself more productive and move faster, I immediately

saw the change in the workforce, immediate.

People say like entry level is done, certain thing.

I don't actually agree with that.

I say middle of the bell curve is done.

That could be middle of the bell curve intelligence because they're just not smart enough to use

these tools to their maximal ability.

And even the first to go are going to be mid curve effort.

Because mid curve, mid curve effort has a large element of BS built into it.

Mid curve effort is somebody intelligent, ish, or maybe fully intelligent, who uses

that intelligence to make themselves very efficient.

They do their job very quickly and then they steal the rest of their time.

And then they BS about how long tasks take.

So I want somebody now who will say, wow, this tool will allow me to get so much more

done.

I am going to create these automations.

So I'm going to take advantage of this.

And then when they've collapsed, they're 40 hours of work into eight hours of work.

They're going to look up and look for more opportunities to contribute.

And they're going to continue producing deliverables.

When I said the BS was over, what I mean is in the past when somebody said they're working

on it, I can simply ask to see the iteration.

And with my team, for example, we've kind of changed that daily iterations are new

norm because there's nothing that takes more than a day to at least get it to the next

level.

Right.

So there's nothing of like, well, do you want me to do this or this?

I want you to do both of those.

We were having this chat the other day, but just to give people an understanding of how

much this compresses time, give us an example of something that may have taken one or two

weeks, that suddenly it is very reasonable for you to ask for by this afternoon or tomorrow

morning.

At least in iteration.

That intervention.

Not for perfection.

Just to drive the conversation further.

Yeah.

So maybe in the past, I might have asked somebody on my team to investigate or start doing

research on a particular customer vertical because I sense that there's an opportunity

there.

I don't know much about it.

I want to know more about it.

But then also, I want to figure out where is there like a distro list of this profession.

So let's say a state attorneys and trust attorneys.

So I want them to go deep on kind of this as a profession, who they serve.

And based on that, I want a marketing proposal on how we're going to find and locate these

people and go all the way to some sort of deliverable at the end of it.

And if it's starting from something that they know nothing about, I would say, I would

say in the old world, I would have expected it to be about a 10 hour task.

And I understand with meetings and emails on other things to ask a 10 hour task of somebody,

I'm going to generally ask, give them about a week to fit that 10 hours into their other

work.

Now I expect a version tomorrow.

Just an iteration.

And if somebody does BS you and I'll give it to you then, if somebody does BS you, not

that I've had to do this, but I now know what I would do, especially if I was in a corporate

environment where I couldn't just, you know, I literally could be like, if somebody

kind of explains why something's taking a long time and I don't believe them, literally

I could be like, you know, just so I understand your thought process and how you approach

the problem.

So just send me a mark down file of your actual conversation history with the model, just

so I can see how, if they didn't do it, it's not there to show.

I don't know how you think the model's output is.

Well, yeah, that and the way I would even think about this further is that whenever you're

doing work in general, there is the, the moment of I need to do X, Y and Z, doing those

things, taking a half step back, reflecting on it and just as an iterative like with any,

any task that you're doing, whether it's a spreadsheet, a document, a PowerPoint, getting

ready for a pitch, whatever it is, research, organizing things, all of these things like,

just have this little loop.

And what this does is this instantly collapses the loop and what's so exciting about highly

motivated workers is we're collapsing all of the work of the tedious part of getting the

things on page and going back to using your brain.

It's like, all right, I'm going to use my brain, I'm going to use my brain, I'm going

to use my brain, right?

And there's certain people who are just highly motivated.

And this is not in like, in the belt curve of like intelligence, I think it's just strictly

motivation.

I think I would take a one, I would take a 105 IQ person who was highly motivated over

a 140 IQ lazy person any day of the week when it comes to using these tools.

And there's a, the, the fundamental model just in life in general, so you can view words

as spells.

You're having a conversation and you're either trying to persuade or you're trying to

convey or you're trying to do something, right?

And maybe an old work culture, people who would be able to ascend corporate hierarchy structures

because they were really good spellcasters.

They knew how to posture things and they knew how to, you know, phrase things.

And then when it came to work, you had all of these meticulous details of why your work

was so involved and important.

And that's all just gone.

Like it's just completely obliterated.

Like this is going from a swords to muskets kind of just like jump, right?

And the people that are highly motivated are able to execute better.

And additionally, when it comes to like, I'm like, expecting iterations, it's not just

like, this is something that's been very apparent for code for a long time.

Where if someone's like, oh, I worked on it.

Here's my update and you look at it and you're like, okay, you literally just asked a coded

question.

I'm looking at the code changes.

You did not look at this.

You just said, you know, make an update looking at x, y and z.

And what you're actually looking for in these iterations is not just, I had a conversation

and the computer did my work for me and here's what the computer did for me.

It is the additional loop of, okay, I'm going to use my judgment now.

I have an output.

And now I want to think very critically about what this output is doing and maybe do that

two or three times.

Look at it one more time and then hand it over, right?

And this is where I think most people in the workforce, we're such a small company.

It's not relevant to us because it's very tight knit and we're all kind of like sprinting

along.

But in the larger workforce, there are a lot of people are like, someone wants me to make

a report and I make the report, okay, and here's what, here's my report and they are totally

surrendering their autonomy, their own intellectual brain, like they are going into autopilot.

And those are the people that are very quickly, with the compounding ability of these tools

that they're, it's going to further differentiate.

Everyone gets more productivity, but the people that are going to leverage this on top of

it and compound on top of it, we're going to get way further ahead than someone who's

not.

Well, the ones who don't will get fired, I mean, I'm transparent, right?

They really will because we, we will go ahead, very quickly.

This is a funny thing because like you saw maybe with cash at block, like they'd laid

off a bunch of people.

So 40% in the stock rows, and one, I think in general, there was a massive overhiring during

COVID and salary inflation.

This was a great way to clean it up.

You see these massive layoffs and it just very quickly became like, oh, these are the

people who are just not even trying, right?

Like it very quickly, you can discern out and understand.

And that's a general statement, like I'm not saying that everything on the individual person

and there's always office politics and mechanics and details of things.

And luck.

Yeah.

Like especially at a large work like that, like there are some arbitrary things, like it's

not a conclusive thing, the people are highly motivated are going to be okay.

But when you're thinking about this and these like large layoffs, I think it's just very

quickly going to compound, I think people at the moment are using, because AI is here,

we had to do layoffs because we have an increase in productivity and I just think that it's

less of that.

And more of AI very clearly is starting to show who your highest performers are.

And it's not that, it's another way of saying it, like AI, like things, it's not just

like, because the computer is augmenting anything, they're quickly seeing these are our

high performers.

These are the ones that are really leveraging these tools and that we're able to get a large

amount of output with less.

And this is a general principle like Mark and Jason or someone else with any organization,

the square root of the number of employees is half of the work.

So if you have 100 employees, 10 do half the work.

And if you have 150, 12 are doing half the work.

And I think just AI is very quickly showing that like, it accelerates that principle in

general because they're the ones that are be highly motivated.

I did the math on that last night actually with Starbucks and because I'd always set

os parato, 20% of people do 80% of the work, there will be little bumps along the way

of learning.

I'd like to think I'm a top performer.

I had more or less two embarrassing things like three days apart where I accidentally shipped

out slop externally and had to, you know, in neither case did it create any negative

repercussions, but it was embarrassing because usually I wouldn't typically send out mistakes,

but what I quickly learned and learned with my team as well because in some, in one of

those two cases, also like my team had also checked it.

And I believe they did.

I believe they put in the time in their old way of checking things.

So you think of it like in Excel form if you built it.

So let's say you built a little calculator that does whatever you were trying to do.

If I built that, I'm the one that typed in the formulas when I make a change, or if

I type in a number and I see everything else change, I know what's happening because

I'm the one that built it and I actually understand.

And so when I'm checking my work, if I built it, I can spot check.

So I can see like I'll check cell C11 and F22.

And if those are correct, I'm reasonably sure the document is correct and I can send it

out.

Same thing in words.

Like if I'm the one that wrote something and I'm just kind of spell checking and do grammar

stuff and I do very little time editing and most of the time in creation.

I was taking that way of working into this new world and I did some spot checking on

an Excel and the cells were correct and I sent it out only to figure out after the fact

that like there were totally made up numbers on it.

Like I'd sent out absolutely just incorrect information.

And then I did a similar thing on a written document where I just, it had hallucinated

a different custody model than ours.

And because I didn't go back and read every page, reread every page, you know, it sent

out this garbage.

So now I know that creation is not the time consuming thing, editing and judgment like

Rob mentioned is the time consuming thing.

And that's what I'm paying people for.

If I'm going to pay anybody six figures ever again, it's for the use of their judgments

and their minds to interact with these things.

It's not for tasks ever again.

That's the big evolution change.

And that's bottom of the bell curve.

I might still employ people if we have tasks that somehow are not automatable.

But those mid-effort people are gone gone forever.

And that's just on the coding side, it's turning more to code review.

Not actually writing the code.

It's okay.

You've proposed a change.

Let me now look at a poll request, which is like the changing of the code and let me

look through that and understand it, ask it questions, and try to avoid the orboros

of the snake in its own tail of, well, I have an AI agent that reviews the code before

I deploy.

I mean, for the work we do specifically with Bitcoin custody, I'm one, I'm very happy

that we've built all of this before the big AI boom happened.

I say that strictly because seeing people like vibe code a quick wallet for a fun project

is fine.

But with Bitcoin money, the code is money, like in the money is the code.

So like there really isn't a margin of error for any of that, and I feel very good about

that.

And when these tools started coming out, as I started red teaming, like to red team

is like to try and attack your own infrastructure.

And felt very good getting to them.

I was like, oh, wait, this is great.

I was like, oh, let me check out like the ledger hardware.

Let me check out the code card.

Like I started like looking at things and poking around and felt pretty good from like

this initial assessment of things.

That's some, not sure how much of this is hype and marketing.

The latest models that are becoming out in middle end of April are so good at the moment

actually at cybersecurity where they're finding bugs in the Linux kernel, they're finding

bugs in the Chrome browser and Firefox like very serious bugs.

And they're trying to really like distribute to cybersecurity teams before like it gets

to them.

Do you think, do you think in terms of, so I would say cybersecurity has always been an

arms race, right?

You know, so hackers, you know, coming in and hackers are increasing their abilities and

defenders are increasing their, so this goes up and up, right?

I have to imagine that these tools simply exacerbate that or, you know, does this ever

get so resilient that somehow like hacking is a thing of the past or the opposite?

And it's just like indefensible or is it just the same that we're in today?

It's mutually, everyone just has the same quality of arms now.

So the complexity of the domain in general, just like going from swords to guns to going

to missiles and nukes, just so now you're doing it, you're walking away 10 paces turning

around and then shooting ballistic missile at each other instead of, right?

So it's just constantly going back and forth like I don't think it goes away fundamentally.

Have you tried the Replet platform briefly a little bit?

Give me your thoughts on that because I've used that.

I think it has a really good non-engineering experience for being able to do that.

I think it's a great for prototyping, it'll also help you build out like your architecture

for your database and kind of like go through these things and having conversations.

I saw the latest Replet announcement, which is really cool where they basically have your

code base and what you're building also live with like your marketing documentation and

having them kind of recursively like it kind of memory management and context is really

well handled and it just very cleanly for someone to be able to just click a button

and deploy code to so all of the infrastructure complexities are just totally abstracted

away.

So I think it's a very powerful tool.

That's like cool, like sorry to read with, we don't build on it, like we have, we own

our own infrastructure, we go through all these different pieces like engineering team.

I have an engineering team.

Yeah, that's right.

Yeah, an engineering team and also have crossed the threshold of skill domain where even

if I wanted to have agents write the code and do that, I would have them help me build

my own infrastructure rather than using something that's entirely self-hosted.

It is something that's within my knowledge and context that I personally have that I

could use AI tools.

And that's the funny thing, right?

You and I, this is actually just a comment, a beautiful thing about these AI models is

they're massively democratized.

There isn't like a secret cabal sort of the people who work at these companies.

It's not like the, it's not like, let's pick it like Amazon has access to like the super

duper secret model and you and I have like a pleb model, like everyone has access to

the things.

Everyone has access to the best, right?

That's just an interesting democratizing thing, but I can use that to build cloud infrastructure,

but since that's not your area of domain, I mean, we do that, but there are probably

many things that you have unique domain expertise on that I do not that you'd be able to

way further accelerate and use it for production of a podcast for like a thousand other things,

right?

And that's just like the differentiation.

Everyone's going to be even more specialized.

Rob's expertise is really important.

I think earlier I said I will use code, cloud code.

And the reason for that or the reason is that I can, I have already run into limitations

of letting it think for me that because I don't have the domain expertise, I cannot yet

or at least really quickly anyway, really quickly redirect it to do things better more

efficiently.

For example, I wanted, I had a series of Google Drive folders that had like 50 PDFs on them,

and I wanted to have it go into these folders.

I had a list of documents.

I wanted to go into these like six folders and download each of them onto my desktop into

its own folder.

That was, that was the task.

And it spent like all day doing this and what it was doing was taking like a screenshot

of my list.

And then like based on that screenshot, it was doing this and there was opening each

folder and comparing things to things.

And you know, I would leave my computer and come back in like four hours later, I see

what it's been doing.

And I'm like, no, you idiot, like why would you do it like that?

That's so dumb.

And right now all I kind of know is that Rob, if he was the one doing this, he would have

said something along the lines of, well, I'm just going to write a script.

That's kind of like, and he would write it and because he knows the best way to instruct

a computer to do that, I don't know the best way.

And so all I can do is gradually learn over the next six to 18 months by watching it do

stupid things.

That's why it's important to tinker.

This is exactly it.

It's you must tinker.

I tell my friends, look at how I like it.

This is exactly how I like it.

It's amazing.

That's not efficient.

And this is why I have spent the last six weeks having coffee and lunch with all my

friends and my family and my sister who are all still in corporate America where I am

urging them to tinker.

You must do this.

All these things we've been talking about today, if you don't want to get laid off in

the next five years, you must, I care about you so much, I care about you, you must tinker

because if you haven't, and all my friends at this point in my career are senior managers

through executives, right?

If you as a leader are not thinking about the second-order effects of these things are

starting to or you can't even picture because you've never had these thoughts in your head

yet because you haven't tinkered, you are not doing yourself a favor for sure.

Really funny, just like near image of that exact anecdote is, so my wife works in biochemistry

at a publicly traded company.

I went to their quarterly earnings and I started just poking around and I looked through

it all in there and I was trying to start how important this is as she's getting back

to work.

I just started asking questions about the research and have you thought about X-Wines

and she's looked at me and she's like, okay, I like to do this podcast and I was like,

no, I just did it and she's like, wait, you, and it was just one of these funny inflexions

which was like, wait a second, this general, this democratizing factor of knowledge in general

that's happening right now is something that many people aren't appreciating, unless

they're tinkering.

That's what you want to tinker.

You don't need to be able to tinker and understand, because the way you describe your

problem, I would have used the Google Drive API and I would have made a very simple thing

and just talk about it like a computer system would and that's why everyone's turning

more into writing code, it's that you're a product manager and the product managers understand,

maybe not the lines of code, but they understand the connectors and the tools, just speaking

the language that can go to the engineers to do that and now your engineer is a chat

box, right?

So you and I could do a race of like, how do you want to do this?

And I, since I understand the, you know, when architecture, but like the closing, the

kind of understanding general architecture, like, you easily could have asked me like, how

would you think about this problem?

I would have been like, boom, boom, boom, and then you would have been done.

Totally.

A 90, a 30 to 90 second conversation would have transferred sufficient knowledge for her

for in perpetuity whenever interacting with Google Drive, you would have known how

exactly.

Now and then I'll ask him, but sometimes I'm actually purposely not because it's the

same thing I actually pushed my team to do, which is they're like, well, I don't, how

to do this, or I don't know how to do this, or I don't want it to like, that's intimidating

to me.

And, and we keep being like, ask it, ask it how to do it, let it tell you how to do it.

And so I'm trying to kind of live that to a certain extent, even when I can see that

this is happening in a stupid way and I'm not figuring it out yet.

I am trying to learn how to communicate with it to wrestle from it what I need and do

some of that learning.

I wouldn't, I would recommend actually getting the tips and tricks for most people because

why waste a bunch of effort and time.

I think it's important to learn those skills.

I mean, I've had funny things where I'll have a bug in software and my camera can help

me fix it.

And what it'll do is it'll just remove the code.

Yeah.

Bugs fixed.

Like, I just, I just remove the code, isn't that better?

Like, I think no more, no more, our test suite has no more errors because you just remove

the code or it'll remove the test, you know, like, you can't fail a test, you know, right?

Right.

So, like, you have to go like, yeah, that's the old Soviet saying you have a man, you have

a problem, you have no man, no problem.

That's right.

That's right.

Now, we've spoken a lot about how people can, uh, position themselves as employees, so

that they can get more out of their work and make themselves more valuable in environment

where, you know, a lot of people are going to come under pressure.

Yes.

As an entrepreneur, how do I qualify someone as with, uh, having potential to work well

with AI or not?

I'm so far for me.

It's not very complicated.

It's looking for top performers and attitude and intelligence.

I don't quite, I don't agree as much as Rob's kind of extreme statement, which may have

been for effect that he would take a 105 IQ with good effort.

105 is pretty, pretty dull, um, but I would take one, yeah, definitionally, one, 105 is

pretty dull, but I would take one 20 and good effort over 140 and lazy.

Um, so honestly, I'm just looking for people.

If I were hiring two day, I'm fighting for my life here.

If I were hiring literally today, I would certainly get into this conversation about

using the tools.

How are you using it today?

How do you think about it?

Give me some examples of things that you've done.

Yeah.

And if they're like, if their answer is something along the lines of, well, you know, I've,

um, I've been using chat and I've been using it more and more and I did some, uh, really

good research or, uh, spreadsheet, uh, one, once, uh, I did that.

That was really useful, but it becomes evident that they have not used it beyond a better

Google because it's so, but it's so new, right?

So it's so new that maybe I, it's fine.

That is a non technical person.

Maybe you haven't, but when I press into that and I start talking about this, if they're

getting excited, uh, and you see their eyes opening up and like, oh, wow, like that is

true.

And like you see these things, um, that means a lot versus a lot of people have this initial

reaction that is like, especially non technical people is I don't want to learn to code.

I don't want to gain that knowledge because it's boring to me.

If I had wanted to do that, I would have become an engineer to start with.

And if, if their initial reaction is to say no, they're, they're probably going to have

a very rough five years because all we're going to be is change, change, change, more

and more efficient.

And the only people who are successful who see that and are like, oh, cool, oh my god,

I want to try that.

Like, oh, cool.

Oh my god.

Oh, so interesting.

Like, and those are the people who are going to use their intellectual curiosity, uh, even

if they have not had a chance to yet, that I, that's what I'm looking for is the spark,

I think.

Uh, and, and not much more than that because the tools are going to enable, uh, a, say

an entry level person with intelligence and a great attitude, they're not going to have

the same judgment in that using the tools as a experienced career professional who has

lots of examples and personal anecdotes to, to bring in.

But if they're smart and have a, you know, good judgment with the tools, suddenly they

can be on the same level.

Let's say I'm talking about a 22 year old university hire.

I think a, a good 22 year old university hire will be performing at the same as somebody

with 10 years experience because of the tools.

Hmm.

I think we, we can't finish this podcast without talking about Bitcoin and AI.

I was about to wonder if we're going to the data center conversation.

So, um, where do you, where do you want to pick this up?

Because the obvious one is transactions, right?

And so the, the assumption has been, we think, well, for me, perhaps you have a different

, um, place to start.

But for me, would be the obvious place between the combination of AI and Bitcoin would

be a transaction.

Oh, okay.

Okay.

So like agents being empowered to go off and do got it a couple of random thoughts, just

as a little bit of, a little bit of, like a little bit of background, um, these data

center.

I thought, well, we'll talk about Bitcoin as money first because I, I, my, my immediately

to data centers, um, I think fundamentally, I'm not a big fan of these, like, mall,

like, like, I know, like, I love BPI, but they did this study that like, we surveyed

a bunch of models and they looked using Bitcoin as money and ultimately these things are,

are tools that are pointed in a direction and they are going to use whatever their system

directs them to do.

I do think a natively digital money makes the most sense for these things to be able

to do instantaneous settlement when you use the token and it's gone and the work has

gone, like, the money has to live somewhere, like these things are critically important

and for transaction stuff.

I think the, it removes much of the friction around Bitcoin because you no longer

have to be a Bitcoin developer and deploy a server that running the lightning node and

like all of these things, like, all of those like intermediary complexities start collapsing

inward.

Um, Bitcoin for transactions and like a machine machine context, I think makes a lot of

sense.

I think also most people haven't put thought into what good security looks like for managing

that and, um, you don't want to be in a position where you have some random code that

spins up that's starting running on a server and you're not software to have all of a sudden

is managing large flow, right?

So I think custody comes back to being important, whether it's within a lightning channel

or you're using R or state chain, stable coins for dollars.

I think all of this stuff, just Bitcoin's a medium exchange can very easily start collapsing

because a lot of the technical barrier starts dissolving.

You no longer have to be a dedicated lightning engineer to be able to figure out how

am I going to send and receive payments.

Well, won't we immediately start hearing stories of, for example, agents that more or less

drain another agent's Bitcoin wallet and things like that that you've given it kind of

permission to charge and more or less create an invoice.

Let's say that's happening behind the scenes.

So you've given permission to agents on both sides to both create an invoice and then

another agent effectively pay and a lightning invoice say, uh, but there will, that will

go badly for some people like there will be accidents where, where suddenly, you know,

you're creating too many invoices and the other agents is just paying them no problem.

Yeah.

And Bitcoin is like some people will probably get rich and some people will probably get

poor accidentally because of this dynamic.

So I think we'll hear things about that.

For sure.

And I think it's interesting too because Bitcoin as money is an abstract concept, right?

Like when people start getting to Bitcoin, like, wait, but it's not physical, like, where

does it live?

How do I look at my Bitcoin?

Like, how do I touch my Bitcoin?

And a lot of the interactions we do day to day, if like you make a, and like if you're

using these tools and you're like doing some research, right, go do some research, you

get an answer, keep on going through your day.

There's a whole chain of abstract things happening in the background of the agent going

out and searching and looking for information and organizing things and we're cursively

thinking to itself where like, that's actually a great way where you can use Bitcoin as this

abstracted money and being able to understand, it's actually probably really good for the payment

of these things that are happening all around on the background and you don't have to

understand the exact value proposition of Bitcoin.

It's just from the properties of being natively digital money, the best option possible for

all of these things.

Stablecoins can be seized, they can be reversed.

You know, they have their own kind of baggage and problems and the reason why stablecoins

have just been so massively successful in general is that everyone just doesn't have to

do worry about 4x risk.

But people are probably willing to take some 4x risk and have sovereignty and the ability

to know that I have my Bitcoin.

I don't have to worry about something going wrong later and I can deploy all of these

machines in architecture and just so natively, as I mentioned earlier, the money is the code

and the code is the money.

It just so nicely plugs into how you manage all of that within these tools.

And I think right now, I'm not sure if you've seen money dev kit, Nick, he used to work

at spiral.

He started money dev kit and he actually has replet plugins that allow you to actually

do lightning interactions back and forth, right?

He bought a couple domains on human coffee and on human domains where you actually can

have your AI agent go and buy some of the bag of coffee and ship it to him.

And it's all, and when you go to the website, it's really cool because there's two ways.

You go to the website and it's like, here, buy some coffee and then there's an agent

tab and it's just a markdown file.

And it explains to an agent who visits to this website, how do you buy coffee using the

tools within it?

And rather than it being like, here's the website and the button that clicks to go to Shopify

and do all the things and you pay in Bitcoin, it just says, okay, here's the API link

to get the invoice and this is the payload and how you send me the address and information

and you just have a quick conversation over telegram or your chatbot, whatever it is,

and it'll just send a bag of coffee to you.

Is that visible?

Like when you visit their website, you can see it.

But would it be correct that in the future or the way this can work?

Is it doesn't need to be visible at all?

You're because the information is there.

I mean, it could be say in like transparent text, it could just be in a JSON file.

Yes.

So that we get to the point that it's not a novelty, right?

Like they're doing as a novelty to like, look what you can do.

Look what you can do.

And in the future, I think it's a novelty today.

Your agent will just go to that website and that will be there and it will read it

and do it if that's what you wanted it to do.

And it'll just be a bunch of data in the back.

They'll basically be probably a parallel web that people don't go to.

It'll be a parallel web.

That's right.

And like in the background, it'll just be JSON, APIs and markdown files.

And to us, it'll be tedious and annoying, but like things that we need to as visual creatures

to look at things they don't actually have sight.

They're just reading the code and the words.

So we'll have a very, so today websites are effectively.

There's a front end on the back end, but like there are two versions of the same.

The front end is going to be a markdown file and the back end is going to be like,

there's servers and here's my API to talk to me.

Well, it's almost like there's a third layer.

There's still the front end, which that's what we see.

That's what we see that.

But the difference is so that in the pre-markdown file,

there was the front end, which is what we see.

And there was the back end that corresponds to it.

And there, I mean, maybe you could have hidden stuff, but like that wasn't the point.

You had a front end and a back end.

Now you have a front end, you have a back end, and then you have a whole parallel back

end that's only speaking to other agents.

And the front end is effectively, go to this domain and I'll give you a markdown file.

Or it might have content that's made for humans.

That has nothing to do.

That's what I'm, so the front end will just be about humans.

Right, yeah, yeah.

And the AI version just is invisible, because it doesn't matter.

And if you want to, you can just go to it and it's just going to give you a bunch of

documents, like text files, and it's not going to really be visually appealing to you, right?

I think that's a big potential in the future.

It will fundamentally change commerce.

You can already start seeing with the ads being added into like GPC that like,

this is becoming the new like Google's massive dominance wasn't, was search that allowed

it to sell ads.

And now we're getting to a point where these things are going to be able to start selling

ads.

And then not only are you going to sell ads to humans, they're going to be selling ads

to agents.

What about product placement where like, let's say we wanted to pay for anchor watch to

be recommended as the best resource for Bitcoin insurance.

And you're willing to pay as a company, you know, where is that going that we're allowed

to influence the models for money?

That's already called like agent optimization.

And the malicious way is called content poisoning where people are able to like actually,

with a couple of pointed articles on the web when these models go do a big training run,

they'll have very, very narrow specifically tailored keywords.

And people are able to like, with very specific combinations, just always be the top

hit because they basically self ref, they created a concentration point while these models

were training.

And the model wasn't aware of it to always, and that's like, you can do as grill marketing

too, right?

So you could pay for advertising or you could just do organically by building like a network

of things to be able to pull into there.

And that just creates a whole new domain of like what it is to advertise now the, what's

really changed is that you're not trying to persuade the human.

You still are in some cases, but you're also trying to get these computers to automatically

engage with you first.

Do we get an entire sort of marketplace like Amazon that's solely for these agents to

go showing?

Yes.

Yeah.

Yeah.

Like the thing with Amazon, it has like this network, Lindy effect of them, the warehouses

the distribution of the machines, very well, Amazon may just be the winner because they

have the physical supply and chain infrastructure that's already in place.

But the marginal cost of friction to go from one to another is going to be way lower, especially

if you're using Bitcoin as money because now you don't know, oh, I pre-save my credit

card information into this website, and it's so convenient.

I hit a button and it arrives.

That also kind of starts collapsing away too, if you just use it or using it.

Have these marketplaces popped up already?

I just know enough about physical supply chain because of my background and apparel and

food that immediately I start thinking about the manufacturing aspect.

So less thinking about kind of consumer products right now, but more thinking about manufacturing

and how, you know, if I order, if I need to make breakfast sandwiches, for example, the

primary supplier provides Starbucks with breakfast sandwiches.

The secondary suppliers provide that supplier with English muffins, with sausage patties,

with pieces of cheese, et cetera, and sometimes their tertiary suppliers as well.

If you then take that and go to a much more, many more parts, so like some sort of manufacturing

thing that when you order a part, it needs to source hundreds of secondary things.

And maybe all of those today come from different suppliers all over the world.

It's very manual.

I'm certain that there will be manufacturing type warehouse kind of secondary.

In the food industry, there's something called broadliners.

And that's more or less the grocery store for the grocery store.

It's giant warehouses that supply grocery store warehouses, which supply stores.

So I think all of that is very, very primed for agentic optimization that that's all

just a recipe.

This part is made from so much aluminum and so much plastic and that's a recipe and anything

you make can be drilled down to a recipe and if you can put into a recipe, it can be

automated.

It will happen.

One question I have for you that is to Bitcoin is you mentioned a few minutes ago that

you could do guerrilla marketing or context poisoning on the other side.

It reminded me of when chat came out a couple of years ago, I was still working for Brad

Mills on the side and we would, you know, he had funded certain projects and other projects

were kind of popping up and several of them, you know, were Bitcoin agents of some variety.

But this is a 18-month ago version of what an agent was.

Really just trying to create Bitcoin expertise.

And there is this momentary big effort by the maxi community to start training and trying

to put info out there to teach it correct information about Bitcoin because answers were

coming back very crypto-oriented.

Did any of that effort by those people go anywhere or was it?

I think it did.

I mean, does that actually do something?

No, it does.

It does.

It's a model answer.

No, you're wrong.

Here's the right answer.

No, very much did.

Where a lot of this data was actually trained on Reddit and that's a funny, just cultural

observation.

Sometimes it's like, this is why I would never use it for relationship advice because it's

using Reddit data.

We both leave.

She's cheating.

Every time we read it, it's like, and it's like, yeah, like, and so it's totally like

slow.

It's terrible.

Yeah, right.

Exactly, right?

Boyfriend?

No.

Not for the straights.

For the straights.

That's right.

That's right.

But so like, this is why I would never talk to these models for anything related to

relationship advice because it trained on Reddit.

And people observe that like, hey, these models have a left-of-center, cute, cute skew.

And like, that's because like, the employees too, and they nudge it in a direction.

It does speak incredibly favorably of Bitcoin compared to everything else.

It does.

And we had our own little, and the thing is like, it's funny, you've talked about us

a quarter.

You talked about us like a coordinated thing, and maybe there was a little bit of that.

One, the lag of these models is like, you know, things are like six, eight months,

and then it starts training.

It doesn't get any new data.

So there's a considerable lag in these.

We just truly believed in Bitcoin and talking about an evangelizing public for 15, 12, 15

years that there was already a strong, corpus base of not only like these are our arguments,

but this is how we respond to said arguments, and this is why we've put so much thought

and thorough into it.

And that got baked in and making Bitcoin as a unique asset, for sure.

Can you directly, let's say, content poisoning, but with good intent.

Can you...

Well, let's say guerrilla crowdsource education, where let's just...

We just like an attack.

Like that's where you...

No, it's sure.

If you wanted to go in, so maybe Bitcoin is already well thought of by AI, but like,

let's say it wasn't, and you had interacted with it a lot, and you got frustrated, and

you started educating the tool and said, you know, I just, I keep getting these answers

from you.

You've clearly trained from somewhere, but here's a bunch of information that I think

is better quality information, and you really should be using this when you talk to other

people.

How directly can you actually...

Well, we can't, because we use the enterprise licenses, and they explicitly do not train

on our data.

Sure.

But I mean models and general models, how does that work?

Putting content out there, putting out a podcast like this.

Public.

Public for come soon.

A lot of the coding examples came from Stack Overflow.

A lot of the code stuff came from open source GitHub repos.

This is also an observation of why sometimes it's not great code, because the average coder

who uploads code to GitHub is not a great developer.

And that's just like...

We were talking earlier about like hiring, like hiring like the proficiency for AI.

My engineers actually, they will use it, but they begrudgingly will use it, and they're

very skeptical, because the things that they're working on with Bitcoin and Custy is so

narrow domain specificity, where it'll ask to solve it, and you're like, here's seven

reasons why that doesn't work.

And I'll have these conversations, right?

So what I like about, from an engineering side, and hiring people, is attention, where

you do not want to be delegating, going full autopilot vibe coding for anything that's

a serious software project.

Hey, if you're going to...

The first software project I coded was I built a bot that scraped my local shelter's

website, because they got a bunch of golden retriever puppies in, and it was first come

first serve.

And that's a perfect thing to vibe code.

Yeah, that is a perfect thing to vibe code.

Find me.

Customer data.

I need a website.

And just send me a text.

Right?

Like the most trivial thing.

That's actually...

If you're looking for kittens, like...

I was like, I think I need to let everybody know I'm looking for two adorable Russian

blue kittens, if they can help me.

Did you want to talk about data senties?

I was just going to just briefly say, as an intersection with Bitcoin, because we're

in town this weekend, London, to talk to insurance, and we got the first conversation we sat

down with.

I was like, well, what happens to the Bitcoin network when all of these minors go over

to AI?

I was like, well, every two weeks, the difficulty adjustment updates.

And I was like, oh, I didn't realize that's exactly how it worked, right?

So, like, it's not an existential threat to Bitcoin.

I think in general to these data centers, the first thing to think about is there's two

types of these data centers.

You have training ones that are like Elon's massive one in a Memphis or in Tennessee.