Screaming in the Cloud

AI agents are moving fast,  but the infrastructure behind them is still catching up. In this episode of Screaming in the Cloud, Corey Quinn sits down with Paper Compute CEO Brian “B Dougie” Douglas to explore building telemetry for AI agents, open-source infrastructure, token economics, and what it takes to create developer tooling in the AI era. From local-first observability to agent runtimes and the future of AI workflows, this conversation dives into what’s next for AI-powered development.


Show highlights: 
(00:00) Open Source Trust Signal
(00:16) Show Intro and Sponsor
(01:07) What Paper Compute Builds
(01:55) Telemetry for Agents Explained
(04:10) Local First Data and Sharing
(06:18) Second Time Founder Story
(09:06) Token Costs and Pricing Psychology
(14:20) Stereos VM and Safer Runtimes
(20:34) Open Source Strategy and Vibe Coding
(24:54) Whats Next and Wrap Up


About Brian: 
Brian is the founder of the Paper Compute Company, a distributed systems primitives for AI agents.
Brian previously founded Open Sauced, a company dedicated to increasing knowledge and insights of open-source communities. In 2024, Open Sauced joined the Linux Foundation, further solidifying Brian’s commitment to advancing open-source initiatives. With a passion for open source, Brian has consistently supported and mentored new contributors through Open Sauced, empowering developers to excel in the open-source ecosystem.
Previously, Brian also led Developer Advocacy at GitHub, where he fostered a community of early adopters through content creation showcasing the newest GitHub features. His experience spans across notable companies in the tech industry, including Netlify, where he worked as an advocate. Brian’s dedication to open source extends beyond his professional endeavors. He currently hosts two podcasts Open Source Ready and The Secret Sauce: A podcast focusing on developer insights and experiences.
Through these platforms, Brian continues to share valuable knowledge and promote open-source culture within the developer community.


Links: 
LinkedIn: https://linkedin.com/in/brianldouglas
Website: https://b.dougie.dev


Sponsored by:
duckbillhq.com

What is Screaming in the Cloud?

Screaming in the Cloud with Corey Quinn features conversations with domain experts in the world of Cloud Computing. Topics discussed include AWS, GCP, Azure, Oracle Cloud, and the "why" behind how businesses are coming to think about the Cloud.

For what we're doing and who we wanna talk to, like open source is actually a

value add for them to be able to see what's happening and like what controls

and like what policies they could set before they even get to the conversation.

Welcome to Screaming in the Cloud.

I'm Corey Quinn.

I am joined today by CEO of Paper Compute Brian

Douglas or B Dougie as the kids today call him.

Thank you for joining me.

I appreciate it.

Yeah.

Uh, glad to be here.

This episode is sponsored in part by my day job Duck.

Bill, do you have a horrifying AWS bill?

That can mean a lot of things.

Predicting what it's going to be, determining what it

should be, negotiating your next long-term contract with

AWS, or just figuring out why it increasingly resembles of.

Phone number, but nobody seems to quite know why that is.

To learn more, visit duck bill hq.com.

Remember, you can't duck the duck bill.

Bill, which my CEO reliably informs me is absolutely not our slogan.

So I have to say that the, my discovery of what

you do was a, is a tail rife with disappointment.

I was so excited for a brief instant that when I heard paper compute, oh wow.

It's an e ink company and I finally will have

something other than a Kindle that uses the technology.

But no, it's an AI thing instead.

Yeah.

Uh, sorry to rug pull you on that one.

It was a quick one though.

Uh, yeah.

Paper compute.

Yeah.

We were doing, uh, distributed systems for

agents, well, infrastructure for agents.

So we've been at it for about two months at this point and been quite the ride.

Obviously, AI is moving pretty quick.

I am hesitant to say too much about the current state of AI because between

the time we record this and the time we publish this, which is not that long,

there will doubtless be a bunch of things that come out that made me look.

Foolish, but when you say infrastructure for AI agents, what does that mean?

Yeah, so it's, we started with two different open source projects.

One is telemetry, so traces, if you think about O Hotel, we started

this back in January and at the time O Hotel did not work with LLMs.

I, it works, but it was a, it was kind of hamfisted.

No, that's just hotel.

That's basically its entire tagline.

It works, but it's kind of hamfisted.

It's right on their

website.

Oh, oh, fair enough.

Yeah.

Yeah.

Uh, not from my mouth, but they have been sitting around and like doing their

sort of weekly meetings and they're gonna like, have better structure for this.

But what we wanna do is like telemetry for agents in

particular, and that is the first project, open source, HEPL.

The other one is stereos, which is the play

you the, the place that you play tapes.

And the idea is having a agent runtime.

So it's a VM built on Nix Os, and that's kind of the TLDR.

I could go into like way more detail on the why and

how, but I, I realize we also only only have 30 minutes.

Well, I do wanna dive into this a little bit because when

you say telemetry for ai, that usually means a few things.

There's the metadata around it.

What was the duration of the call?

An application was this tied into what was the provider?

And notably the one I like, what were the tokens?

But then you can also get into it what was the actual

conversation, and that's where I start to feel a little bit hunted.

Which side of that divide do you come down on?

Yeah, so we're doing a bit more of the telemetry where

you have like the tokens and the, the time duration.

We do have the prompt stored from the agent and from the human.

So like the agent does have a bit of the,

with the thinking we'll have like, hey.

Brian asked this thing, and I'm looking at this thing and I

remember this folder here, uh, which is always fun to look at.

I think a lot of that gets squashed in Quad Code and Codex now, but back in,

back in my day last summer, uh, agents would spit that out when it was thinking

it, it's also dangerous on some level

because it it because this has been the case.

20 some odd years where if I wanna make one of the smartest

people I know look like one of the dumbest people on the

planet, all you have to do is just do a read back of things.

They have typed into search engines, like things that, even things that I

know off the top of my head, but I wanna validate like what does S3 cost?

Yeah, I do know that off the top of my head, but every once

in a while I like to make sure that it hasn't changed on me.

And also I have brain farts, which tends to happen as we age.

And yeah, let's just make that work.

The idea now of, oh, something's watching me and actually.

Seeing the dumb question I'm asking the robot.

That's scary.

I'm gonna have to make sure that, uh, I grease some palms.

Yeah.

Yeah.

So like the way it works today, so we're, we're still open

source first and only right now, and it works SQL light.

It basically writes to your, your home directory.

Hmm.

Oh, so it's local telemetry.

Yeah, so it is local, local first, and that's what we have today.

By the time this comes out, we will have a cloud product, we'll

will be multiplayer, uh, where if you choose, you can send your tape

sessions to the cloud for your, your coworkers to also comment on.

You are funded by Heavybit, as am I, and so is Atin, uh, Ellie Huxtable.

And that, that's always been an interesting problem too.

It's, oh, you can take your Shell history

and share that to your colleagues as well.

It's like, oh God, can I pay you extra to not do that?

Like, even now if I've run at two Unwrapped, which is the,

their whole summary thing on this, where it tells me that.

Or the number of unique commands.

Yeah.

My commands failed 23.1% of the time, which as someone who spent 25

years in the Linux admin space, that that's kind of a high number.

Yeah.

It's, it's high number, but I imagine it's

more like golf where, nah, I have no idea.

But I feel like with like, 'cause I'm, I'm a pretty heavy Vim user, so

like I'm definitely hitting like leader key and like hammering like a

bunch of stuff that doesn't exist because it's just all muscle memory.

I'm a big fan of.

Like seeing the data and like you can interpret

the data in whatever good and bad it is.

Uh, so like I'm not a big fan of like, oh, well

90% of other companies are, are failing only 40%.

Like, your company's failing and now you don't,

you're never gonna make it to unicorn status.

Like, I, I think that's a, yeah, it's a bit of a myth, but I,

I think I could see what a competitive nature and leaderboards

would come in there, which a big fan of that too, as well.

Yeah, which I highly recommend not looking at the AT two and wrapped summary.

For example, in 2025, apparently I typed 17 and half thousand commands.

1700 of them were cd, which is my number one.

Number two is NPM Run, and the third is ls, which is just, hmm.

Alright.

Alright.

Yeah.

Where am I and what else is here with me is

sort of the thing that I'm always wondering.

Yeah.

LS is de, that's definitely top three for me for sure.

I don't think I actually, I have no idea.

Uh, I don't run Atune actively right now, so I just wiped the

Mach Miss machine, so I'm still catching up with all the stuff.

It's pretty great.

So one thing I find interesting is folks who do a painful thing

and get hurt and decide, you know what, let's do that again.

This is of course a terrible segue into the reality.

You are a second time founder.

Tell me about that.

Why would you do this again?

Yeah, so I'll mention my first time founder project, which is

Opens sauce pizza, which now pizza is actually not working anymore.

Dot com is the only way you can see it, unfortunately.

But we built insights in the GitHub repos, so

I'm, I'm just a big fan of looking at data.

I gotta see in statistics in college.

So for whatever reason, it's been a chip on my shoulder

to build more stats and the developer products.

But yeah, so built that, sold it to Lennox Foundation.

It worked out pretty well for us.

I thought I was never gonna do it again because I felt like

I got like a, a sort of boulder lifted off my shoulder.

I went and worked at another company and then immediately saw in the AI space,

infrastructure was just like not even a thing we were thinking about in 25.

So now in 2026, we're like, oh.

Maybe computers should like be more agentic friendly and like,

we should probably lock these things down and like, they'll

give it access to your Gmail or like, time out the, the runners.

So all those questions that I had around like building stuff

in the AI space, it had to be answered with some software.

And like in December when I had the ideas, none of it was built.

Now it feels like everyone's building it right now.

It, it's almost like if you go back in time.

And there's folks of us in a certain generational era where we struggled

mightily with the advent of the internet and social media in some

ways that kids today don't because they grew up knowing that they're

at a fishbowl and everything they do is going to be discoverable.

It feels like.

I have been using my computer in such a way that I assumed

I would be the only person using it, and that AI would not

come along and start trying to pick up the pieces behind me.

And if you give me a wiped machine, and now this is my new computer, and from

day one, the expectation is is that the AI is gonna be looking over my shoulder

and using any of the email that I interact with and the browser and the rest.

I can use it more effectively in ways that aren't

potentially disastrous when a robot gets confused.

Yeah, we're in like a weird world, like about a year and a half ago.

Everyone had like these pendants that you'd like record

conversations with, or you have like the MetaFrame glasses that

you just like recording TikTok videos all nonstop, all the time.

Now we're in a space where like now our shell history is

being recorded, or now our prompts are being recorded.

Now we're basically being judged on how many tokens we're leveraging,

or if we're not spending enough or we're spending too much or

we're doing side work on our main machine and all this other stuff.

Like it is very, very interesting in this

current state of the world we are in today.

And I know like the world's gonna change in like three weeks.

So like today, like.

I, I think we don't know what we don't know yet.

And I think what's interesting in this space is like I, I want to

build something that we can at least understand like what we're

sort of passing through and like what tokens we're leveraging.

I think I've been actually super hyper fixated on

is Quad Code has a Max Max subscription for $200.

I easily will spend $4,000 in tokens every week.

So, Mike, I, I get that in Q4.

They're probably not IPO.

So at least I should start tracking to see how much

I'm spending so I can sort of brace for impact.

Uh, but what I found out also in the last couple weeks is

prompt caching and quad code is saving 93% of those tokens.

So, uh, this is basically all the stuff we're building.

We'll have, we'll have this, um, actually you can see this in the tape stack.

If you record your tape sessions, tape system product, the, the open

source projects tape sessions, tape deck is the command line interface.

You could see a dashboard to see your tokens.

Uh, and what's cool about that is you could see the

prompt caching and you could kind of brace for that.

But then.

Just last weekend they shipped this thing called Fast Mode,

where you can get faster l um, inference, uh, but you'll

just pay extra money on top of your Max Max subscription.

It's pretty wild.

Like

yeah, people did some experimentation with it and said the results

were not materially different and it was expensive at the top of it.

So I, I get it.

If you're production is down, we really need to figure this out.

Okay.

Turn on the, hit the turbo button on the old tower, but.

It's a strange thing.

Uh, there's also a psychology part of this where we

saw back in the day when cell phone companies would

offer fixed amounts of data instead of unlimited plans.

A number of folks, I'm one of them, would wind up going for the

unlimited plan every time just because I don't want the mental overhead.

I still find myself thinking like that when things that are charged

on a usage basis, even though the actual amount of money is.

Irrelevant.

It, we've, we've raised millions of dollars.

Terrific.

I don't necessarily need to hyper optimize

around $4 and 68 cents worth of tokens every day.

Maybe that's not the biggest bang for buck of what I should be

focusing on right now, but I don't know how to turn that off.

I, I guess for, for us, I, I could speak for myself, for us who don't

come from money, like sometimes you've gotta hold onto money longer.

I was also an at t like unlimited plan for years

with the iPhone until eventually it scro like.

Up until, actually I was on T-Mobile up until last summer and I had

like this grandfathered like $80 a month plan for the entire family.

So

if they were to keep first trying to beg you into

changing and then trick you into changing over the years,

that was the thing is like when you go through the BART

tunnel, you get no fast internet, you get nothing basically.

And it's because you were on the lower tier plan.

That they can't give you like high throughput.

It's a crowded network in the bar tu tunnel.

And that's the point you really feel.

It is like at the moment you're sitting at GitHub issue

and then GitHub like this freezes on you on your phone.

So I'm like, oh, I guess I'll pay the $225 a month to

like get up my modern plan so that way I can like actually

send a GitHub issue on my, uh, in the BART tunnel.

For the five, 10 minutes that I'm in there.

Yeah, I, I wound up adding Google Fi as a secondary,

uh, sim for a while and then switched over to using it

full time, just 'cause it was half the price of Verizon.

It seems to work reasonably well, but it's still the, it's

unlimited, but they'll slow you down after a hundred gigs

and I think on my busiest month I wound up using 20 of it.

So it's, it's fine.

I feel that way about the Claude Max plans as well, where I have

hit session limits before when I've been working on something.

But I'm not one of those folks that winds up waking up in the

middle of the night to prompt the fleet of agents to do things.

It's when I'm sitting at my desk and working on something.

I'll make the robot go build something.

Okay, fine.

It's not, I'm not the heavy user and in fact, I

could probably cut my plan further than I have.

I'd like not having to think about it.

And I'll be sad when that goes away.

Yeah, and And it likely will go away.

I think they're kind of like preparing us for this moment.

I imagine we still get this unlimited plan, but it's

not gonna be 200 bucks, it might be 2000 a month.

And I think that's the point where we're all like the haves and have nots.

We're probably.

Showing themselves up.

Yeah, I've been playing a lot with a lot of local models as well to see how far

I could take those, which, yeah, there's a lot of stuff you could do to like,

and almost a prompt router where great, go ahead and, uh, do

this basically complicated find, replace in this entire code base

that can be done with a slower local model, whereas architect the

whole system, maybe I want one of the frontier models to do that.

Yeah, and I've seen a lot of tools that, uh, outside of like the, the agent

harnesses that have approached that problem but haven't done it very well.

Uh, I think Open Router has a, a bit of a feature around

this as well if you just go straight, open router proxy.

But yeah, something I'll be experimenting with a ton.

Yeah.

Exciting times.

And I think you're addressing the right part of the problem, which

is first I, I don't know what cloud code is doing under the hood.

I have no idea what the usage looks like.

I know sometimes it's genius.

Other times it feels like it's not that, and I wonder

how that's playing out under the hood or what's going on.

I just know that I need to be careful enough around this when

I'm having it do things that if this is gonna make a mark,

maybe make sure that that test coverage is where I think it is.

This episode is sponsored by my own company, duck Bill.

Having trouble with your AWS bill, perhaps

it's time to renegotiate a contract with them.

Maybe you're just wondering how to predict

what's going on in the wide world of AWS.

Well, that's where Duck Bill comes in to help.

Remember, you can't duck the duck bill.

Bill, which I am reliably informed by my

business partner is absolutely not our motto.

To learn more, visit doc bill hq.com.

Yeah, and the one thing I'll mention as well, so I mentioned

tapes in like the telemetry part, but stereos is the

vm, so the actual computer, like essentially it's N os.

You could run it on your Mac Os, you could

run analytics currently not Windows yet.

So we just need more people to care about Windows and we'll focus on it.

So open the issue please.

That's what I keep telling people.

Open the issue and then we'll track it.

But the idea there is like the same way Quad Code does sub subagents,

it's great because it takes that token usage onto its own channel.

So you don't get that sort of.

Now it's a million tokens context for like a session.

Previously it was like 250, but what I'm getting at is like, it will take

that session off to its own session, uh, and then come back with a hash.

So like with tapes, we're using a Merkel dag, just like get, you could

take a session, you could hash it, you could search it based on the hash.

So you can do lookups, you could generate skills, but then you can also have

separate sessions within subagents, and that's what we're doing with stereos.

So stereos is more likely something you use on your own metal.

Um, there's a lot of these like sandboxes and

computers that people are leveraging in the cloud.

Our bet is that you probably wanna build your own cloud or your own enterprise

that wants to do this, and that's what stereos comes in in the question.

Yeah.

What I've been doing is there are risks to it and I'm aware of that, but I

have a dedicated AWS account called Superfund because it is expensive and toxic

and it has no access to any data in it or so I thought, and then I just run

an EC2 instance there that is, and run Claude and dangerous permissions mode.

It has full root access to the box.

Go ahead.

All all you're gonna blow up is my bill.

Then one day I realized it had access to read

things off of my calendar, which was not expected.

It turns out that at some point in the last few months, Claude's

skills and connections were now tied to your Anthropic account, not

to your actual, uh, not, not to just the machine you installed them

on, which is great most of the time and terrible in this one use case.

So first thing I did is slap the control into the settings

json to make sure it doesn't do that anymore, just for.

Data isolation reasons, but yeah, that is

not something most people are going to do.

They're not going to build a box that does this.

Claude Code gets very greedy with the Rams.

So this is a couple hundred month bucks a month for the EC2

instance to sit there running, uh, four or five of these things and.

There needs to be a better solution.

I'd love to have something like that on the laptop, since

I'm mostly around when I want this stuff doing things.

Yeah.

I, I wrote it on a, on a, I wrote it actually on this machine I'm talking to

you on, just as of last weekend I did open claw on A box is a, uh, a repo.

I put together just to like, if you wanna spin up this

and you wanna have like constraints and you wanna have

like network shut down on certain interactions, so.

The use case I use is on Sundays.

I go through my email, it labels a bunch of stuff

for me, and it only, it times out at two hours.

So if it can't get it done in two hours, then it's

probably just, it's, it's too long or I need less email.

So like, I can figure that out.

I'll go call the government and tell 'em to sit, stop

setting up email or whatever the, uh, I'm thinking of the

Seinfeld, uh, when, when, uh, Kramer said, no more mail.

Not sure if that's possible yet.

Yeah, I have a, uh, in the next room I have a Kubernete

singular running on a cluster of 10 raspberry pies.

And one of the things that runs on there is Billy the Platypus,

who's now my virtual ea, where he'll wind up receiving the

email writing a barely professional sarcastic response.

And now with an additional fun feature, he'll go out to a data enrichment,

API figure out what you look like, and then integrate a cartoon of him

dunking on you in some way into the insulting response, which is just.

Amazing.

It is terrific.

People are trying to sell me things or be obnoxious in email in various ways.

I feel a little sorry for some of the, uh, entry level,

uh, business development types and SDRs who are reaching

out, trying to sell me something that doesn't make sense.

But other times it's, this stuff is so scattershot and

send to the entire world where I like to buy land in Napa.

No, no.

Thank you for asking.

I would not, so yeah.

Have the insulting POTUS smack back.

I love that, but there is risk to this if

it starts tying into all the other stuff.

I mean, the scary thing that I need to defend against and why there's a

human in the loop is it can now schedule things for me, which is great,

but it'll casually look at my calendar and then say, oh, he can't.

Then he's meeting with insert company here.

It's ah, how about no.

That maybe that's, maybe it's fine.

Always having another meeting to yell at AWS.

Yeah, that's a real rarity.

But maybe I'm meeting with a confidential client.

Maybe that's not something that needs to be disclosed in an email.

My challenge is like, I just started paying $20

a month for chat GBT, which gives it memory.

I think you have to enable it.

I did enable it 'cause I was building a, a, a computer, uh, with A GPU in it.

I was like, oh, I'm gonna just like prep the chat GBT to

like, know everything I know about my computer and then like.

Recommend, like, well not Ram at this point, but like, uh,

this was actually last year before RAM was, yeah, good luck.

Money

bags.

Yeah.

But, uh, I was just like looking for different parts and like

seeing if what would work or what, but then like randomly I'll

ask a question about like something like actual work related

and be like, oh yeah, why don't you run on your GPU machine?

They're like, what do you mean?

Oh, the machine you built last year.

It's like, I, well, completely outta context.

Not as bad as your, your sort of confidential.

Call, but like it feels like, it's like, oh, remember

that one like thing you built for your daughter?

It's like, dude, stop.

Not the time of the place.

Read the room.

Yeah.

That's part of the problem that all these agent things have.

Like there are different personas that I take on when I talk to folks when

I'm in a board meeting versus when I am pitching to a customer, when versus

when I'm writing a sarcastic newsletter versus when I am being a parent,

versus when I'm shit posting on the thing known, formerly known as Twitter.

I act very differently, and if you cross

those streams, it doesn't go super well.

Like there's a time and a place for all of this.

Like I, I do a mean conference talk and I do some great standup jokes.

Maybe don't do that when I'm putting together

a wedding ceremony or something like that.

There's a time and a place, and that's something that gets lost sometimes.

Yeah, but it's great for a LinkedIn post.

I think it, it, it could be at the LinkedIn, whatever.

I dunno what that, that vibe LinkedIn has on posting, but, uh, I know

Claude has a good, does a good job of, uh, self a referent and posting.

Oh, absolutely.

And the problem is, is you can't even say that.

That's not what LinkedIn is designed for.

You.

You go to LinkedIn now and it interrupts you when you're writing stuff.

Hey, you wanna rewrite that with ai, it's, what are you?

I think this network is for Exactly.

And then I read some of the insipid stuff on there

and you wanna hope it's ai, but you're kind of scared.

It's not,

well, I guess we'll, we'll figure it out.

Someone's gonna have to ship a this.

Is this ai not true or false.

So I am curious what would possess you, there might be a less confrontational

way to phrase that, but we're going with possess you to start a company

where effectively everything you're putting out so far is open source.

That feels like it has some definite advantages to it,

but it's also the sort of thing that can be fraught.

Like I come from GitHub, I come from open source.

I, I learn how to code because of open source.

So like, it's in, in my ethos, but specifically for what we're building with

tapes and stereos, like I feel like the telemetry parts, it's table stakes.

If you ask any sort of YC company or any sort of

VC of like, Hey, we're doing telemetry for agents.

Like, cool, so we're a hundred other companies, and then

now we thought we were ahead of the game when it came

to like the VM and like building the actual runtime.

But then, uh, like a bazillion other companies

came out and said, we have a runtime too.

Like, Deno has a runtime.

Deno is doing something completely different

for what we're doing and who we wanna talk to.

Like open source is actually a value add for them to be able to see what's

happening and like what controls and like what policies they can set.

Before they even get to the conversation.

I spent a lot of time in my first startup where we spent way more

time trying to convince people and having long conversations.

Today we're just like, Hey, use tape stereos if that

kind of fits within the thing that works for you.

Then we have this cloud, or we have this self-hosted on-prem

thing that works one and the same, like not one and the same,

but obviously one's gonna cost more and be more complicated.

But we feel like open source is a lot easier for people to try

before they buy without setting up a 14 day trial or whatever.

What about the current zeitgeist impression that folks are giving off

where this is the, the bad days of SaaS where companies are gonna start

vibe coding SaaS solutions instead of paying, uh, third parties for it?

Yeah, I think that's the same thing of like, I can write a bunch of c plus

plus code, but also I can't because I've never written c plus plus code.

So yes, I can vibe it, but also I don't know

what bad, like unsafe c plus plus code is.

Uh, so when it comes to like SaaS products, like you could probably

build the thing I built like a. Look at my LinkedIn connections

and give me like a list of companies that I can like go reach out

to and hang out with a bunch of friends and show 'em my new thing.

Like I don't wanna pay LinkedIn premium or sales thing

for 12 KA month when I can just like vibe code a thing.

But there will be a point where I bring on a sales person who

I don't want to like use my vibe code a thing and I wanna share

information properly in like a Salesforce manner or HubSpot manner.

Like there's a place where you kind of grow up

into, like, again, I, I don't want a 14 day trial.

I know what I'm doing.

I'm just like running through the, the system and plugging and playing.

And then on the, the vein of like open source, like

I feel like open source is table stakes as well.

Like where you can close source it and like

keep your, keep everything close to chess.

But also I think what we need is like more people building infrastructure.

So like if anybody clones our repo and like says,

I'm gonna kickstart from this and let's get started.

Maybe we're protected with A GPL.

Maybe we're not.

But at the end of the day, it's like what we need is this stuff

to be out in the system and out in the, um, infrastructure.

And hopefully we're the ones that build it, but

like someone else builds it, then obviously we'll,

we at least not going in the right direction.

Yeah.

Like I've built a bunch of internal SaaS tooling that's,

uh, scratches, certain itches I have that I'm not willing

to pay, uh, professional to get up and running for me.

But when we're building Skyway over at Duck Bill

that it's not, it's not hard software in that.

It is effectively a data platform that manipulates data and normalizes it.

Yeah, we, we know how to do that.

We are not pushing the bounds of computer

science here, but the data must be correct.

It's not the sort of thing that you can vibe,

code, and expect good things to fall out of.

Lord knows we've done experiments around that.

Now you can come out with something that sounds very good.

Looks good at first glance until you start digging into it.

But if you're making sincere business decisions based on that data, and

it turns out that data's wrong, uh, no one wants to be in that position.

It's the same reason that Anthropic vibe codes a lot

of stuff, but they pay, I think it's a DP for payroll.

They, they're not just paying for the software that arranges the paychecks.

They're, they're basically paying for expertise.

You don't want your super base table to be

completely open and like read and write.

Well, obviously the, the thing that happened in the last week is delve like.

Everything was accessible and I imagine a lot of the stuff was

just kind of shoot from the hip and a certain point you kind of

wanna bring in the experts to come clean up and polish the stuff.

Naively, I thought compliance was one of those

spaces where you wanted to have expertise in shows.

What?

I know, that's why I've never been on the 30, under 30,

40 under 40.

Perhaps

not even anymore.

I have to be 50 under 50 at some point where I think we just call that folks.

So what's, what's next?

You mentioned you're doing a cloud offering.

Probably by the time that this winds up shipping, I mean

that will presumably address my biggest problem with your

website right now, which is it doesn't have a pricing page.

Yeah.

Yeah.

So we will have a pricing page hopefully in the next two weeks.

Uh, we're fastly approaching that.

We have the infrastructure that works.

Uh, we're sort of pending a. UI on top of it.

So next is basically you have tapes, you have sessions, you have agents.

You wanna play those in a multiplayer fashion.

So like you and your team could all look at your weird prompts and all the

tokens you're using, uh, and then you sort of like dictate what what is next.

So something actually very cool I've been working on is like

taking your prompts and then training a model with that.

So if you wanna like.

Do you like some fine tuning off down the road for your local models?

How heavy a lift is it to do something like that?

Because I've toyed with the idea, but every experiment I come up with is like,

step one, you're gonna need about $10,000 worth of GPU time, which, okay,

if you know anybody, Nvidia, some friends that

will, well give you some, some rental space.

But yeah, if you have data, data's one of the hardest parts to get for training.

I, uh, particularly have not done it myself.

I've been in the room when this is happening at my last role at continue.

So I have a good idea of how this works.

But then folks like Unsought, they can help you do the fine tuning part.

Uh, so like there's, there's a bunch of like.

Prosumer, like ification, ways to go about it, but also quad's

my friend as well, so I might just like chat with, with them

and, uh, build out a six week pipeline to, to make this work.

I'm still trying to figure out why, when I prompt for

certain things versus my teammates prompt for certain things.

I get such vastly different results in terms of tone, sarcasm,

the way it builds things, and I, I have no answer for that.

It seems like a telemetry play might be the right answer here.

Yeah, I think it's, it's got your taste down to, uh,

let's, let's get the first thing out to Corey and then.

Co-founder gets the better, the better taste.

Exactly.

Yeah.

Mike is the one that winds up doing all the serious button down stuff.

Like we can make this talk to Microsoft Excel and I'm sitting over

here going, we can make it rank the US Presidents by absorbency.

And it's, it's one of those things where it, some point between the two of

us, we're gonna reach a happy medium where it's both whimsical and useful.

We're not there today.

The litmus stuff I have right now is WW JD, which is what would John do?

And I ask that question every time I go write in one of his

projects, which is like, I don't write a lot of Go code.

I'd like professionally.

I've done some go code, but really not to the level that John has.

So what I wanna do is take John's sessions to help inform.

My prompt sessions.

Uh, so I've been generating skills from his sessions and so far going well

shared team skill of Corey Quinn voice.

I need to go in and sabotage soon if some of it gets uncomfortably close.

And that's my space in the universe.

Thank you very much.

Amazing.

So I wanna thank you for taking the time to speak with me.

If, if people wanna learn more, where's the best place for them to go?

Uh, best place to go is uh, I have a site called be dougie.dev.

That's my, this catchall, well I guess they would call it win trees

back in the day, but LinkedIn, I'm happy to like take a DM there.

No promises on response and then X at BW and then blue sky at pizza.

Pizza because I wasn't taking that platform

seriously and I just picked the random BRL.

Yeah, I am having some challenges myself with the answer of where do I direct

people to for all of these things that I, that I have going on, and every time

it feels like I wind up narrowing it down by collapsing two projects into one.

Three more spring off.

I don't know that there's ever a way to get away from it.

Yeah.

Yeah.

Five man Minecraft.

There we go.

Uh, it's similar to, someone showed me a SSO page screenshot once,

and it showed all the different services you could use to log into it.

And like they had GitHub, they had LinkedIn,

they had Google, and they had the Lego website.

Apparently, which acts as an identity provider, which is just absurd to me.

I want to have that as a enterprise class SSO login, just to see the reaction.

The, the Lego Envoy proxy would be amazing.

Like, let's, let's get that to work.

Thanks again for taking the time to speak with me.

I appreciate it.

Yeah.

Have a good one.

Brian Douglas.

CEO at Paper Compute.

I'm Cloud economist Corey Quinn, and this is Screaming In the Cloud.

If you've enjoyed this podcast, please leave a five

star review on your podcast platform of choice.

Whereas if you hated this podcast, please leave a five star review on

your podcast platform of choice along with an angry, insulting comment.

Be sure to include the tapes replay because there's no

way you were able to be that clever all on your own.