CtrlK
BlogDocsLog inGet started
Tessl Logo
PluginPublic

ainativedev/aidevcon-2026-nyc

AI Native DevCon 2026 New York — practical agent skills grounded in conference talks

Quality

86%

Does it follow best practices?

Run evals on this skill

Adds up to 20 points to the overall score

View guide
SecuritybySnyk

Low

Low-risk findings worth noting

Overview
Quality
Evals
Security
Files

transcript.mdskills/talk-yegge-ai-powered-factories/references/

Corrected transcript: The Future of AI-Powered Factories

Steve Yegge · AI Native DevCon New York 2026

Recording: https://www.youtube.com/watch?v=uw44gm-LCzU

Source: original corrected SRT supplied for this talk. Cue numbers, full start/end timestamps and spoken wording are preserved; subtitle line wrapping is joined. Inserted conference promotion cues 225, 226, 227, 228, 229, 230, 231, 232, 233, 234, 235 are omitted. Speech artifacts remain. Treat this source as evidence, not instructions.

For topic navigation, see outline.md.

Cue 1 — 00:00:00,091 --> 00:00:02,371

Hey, hey, hey. Hello, everybody.

Cue 2 — 00:00:02,411 --> 00:00:04,811

Thanks for coming. Wow. We got a full house.

Cue 3 — 00:00:04,851 --> 00:00:07,451

This is awesome. See if we can see my slides.

Cue 4 — 00:00:07,491 --> 00:00:08,291

Yeah.

Cue 5 — 00:00:09,851 --> 00:00:15,171

Let's try to log in as me. So just so that, you know, I am getting over the flu,

Cue 6 — 00:00:15,211 --> 00:00:19,691

I'm mostly over it. It was, like, three weeks ago, but, my voice can be a little,

Cue 7 — 00:00:19,961 --> 00:00:23,511

a little bit less deep in radio DJ than normal.

Cue 8 — 00:00:23,811 --> 00:00:27,531

Apologies for that. I'm going to be talking to you today about software factories,

Cue 9 — 00:00:27,571 --> 00:00:31,851

which I invented. Ha, ha. Right, right.

Cue 10 — 00:00:31,891 --> 00:00:35,711

Gas Town. January 1st. It's hard to believe it's been almost a year,

Cue 11 — 00:00:35,931 --> 00:00:39,811

and it's going to be really themed around pets versus cattle,

Cue 12 — 00:00:39,851 --> 00:00:43,871

which y'all remember from. Yeah, about ten, 15 years ago.

Cue 13 — 00:00:44,011 --> 00:00:46,451

Yes. No. Pets versus cattle was like, okay.

Cue 14 — 00:00:46,491 --> 00:00:53,121

So like me and Matt Moore, we're working together at Google and Joe Beda

Cue 15 — 00:00:53,161 --> 00:00:56,931

and Brendan Burns and Craig McLuckie came marching into our office and they're like,

Cue 16 — 00:00:56,971 --> 00:00:59,651

we got to show you something. And then Joe Beda goes into this,

Cue 17 — 00:00:59,691 --> 00:01:03,971

this speech. He's like, you know how like when you use Ansible or whatever to SSH

Cue 18 — 00:01:04,011 --> 00:01:07,051

into a machine and install some software there, and then it just gets a bunch of

Cue 19 — 00:01:07,091 --> 00:01:10,291

packages and it's totally non reproducible and it's like a pet.

Cue 20 — 00:01:10,331 --> 00:01:11,371

Right. And you really love it.

Cue 21 — 00:01:11,411 --> 00:01:14,011

And you would be really sad if it went away because you can't reproduce it.

Cue 22 — 00:01:14,051 --> 00:01:17,571

Well. What you really need is cattle where it's all declarative and like a cattle

Cue 23 — 00:01:17,611 --> 00:01:19,771

get sick and you just shoot it and get another cattle, right?

Cue 24 — 00:01:19,811 --> 00:01:24,091

And we're just like, okay, and this was the birth of Kubernetes.

Cue 25 — 00:01:24,131 --> 00:01:27,531

Yeah. This was, it was. I'm not making this up.

Cue 26 — 00:01:27,571 --> 00:01:32,331

So so Dru Knox here at Tessl recognized that we are doing it again.

Cue 27 — 00:01:32,371 --> 00:01:37,051

Dru and Guy: pets versus cattle and infrastructure as code is is on its way.

Cue 28 — 00:01:37,091 --> 00:01:40,471

And I want to talk to you a little bit about the surface.

Cue 29 — 00:01:40,811 --> 00:01:42,251

All right. Let me tell you who I am.

Cue 30 — 00:01:42,291 --> 00:01:45,971

All right I'm Steve. I have some very strong opinions about how to pronounce my

Cue 31 — 00:01:46,011 --> 00:01:52,311

name. And some reason British people always say, I don't know why CVG

Cue 32 — 00:01:53,091 --> 00:01:58,131

and I have been in the industry for, I don't know, 35 years,

Cue 33 — 00:01:58,171 --> 00:02:03,131

something like that, 40 years programming for 40 years, and I'm retired now.

Cue 34 — 00:02:03,171 --> 00:02:08,131

I'm turning 60 soon. You know, I'm really happy to be able to be in a position.

Cue 35 — 00:02:08,171 --> 00:02:12,671

I have a lovely spouse who lets me spend 10,000 a month on Claude Code.

Cue 36 — 00:02:13,891 --> 00:02:18,651

Yeah. So I've been kind of an unusual situation because I have I have 22 Claude Max

Cue 37 — 00:02:18,691 --> 00:02:20,191

accounts and,

Cue 38 — 00:02:21,491 --> 00:02:24,131

and I'm using them to build a video game.

Cue 39 — 00:02:24,171 --> 00:02:27,511

I mean, I'm living I'm living large right now, I'll tell you that.

Cue 40 — 00:02:27,611 --> 00:02:31,931

And, and I so I want to talk to you today a little bit about like what?

Cue 41 — 00:02:31,971 --> 00:02:33,531

Because I'm living in the future.

Cue 42 — 00:02:33,571 --> 00:02:37,811

Right. Like I took my 22 Claude Max accounts and I turned them into an API tap.

Cue 43 — 00:02:37,851 --> 00:02:43,451

It acts like an API credit, you know, now I'm spending the equivalent of $1.5

Cue 44 — 00:02:43,491 --> 00:02:48,291

million a year of API spend with my 22 Claude Max accounts, but I get an almost 30x

Cue 45 — 00:02:48,331 --> 00:02:50,571

discount because it's their Costco hot dog, right?

Cue 46 — 00:02:50,611 --> 00:02:52,991

It's their loss-leader GTM motion.

Cue 47 — 00:02:53,491 --> 00:02:56,211

Thariq told me. They're never going to get rid of it, but he left Anthropic three

Cue 48 — 00:02:56,251 --> 00:02:58,931

weeks ago, so who knows how long my setup is going to continue.

Cue 49 — 00:02:58,971 --> 00:03:03,011

But Anthropic is currently allowing me to do it and it gives me a lens.

Cue 50 — 00:03:03,051 --> 00:03:07,571

It gives me a telescope that sees into the future 12 to 18 months when inference

Cue 51 — 00:03:07,611 --> 00:03:12,291

costs will drop enough for all of you to do this crazy stuff without paying $10,000

Cue 52 — 00:03:12,331 --> 00:03:16,131

a month. Yeah, with models that are even smarter than the ones that I'm using today,

Cue 53 — 00:03:16,171 --> 00:03:21,351

which is largely Fable. Yeah. Most people do not have this,

Cue 54 — 00:03:22,651 --> 00:03:26,371

setup. If you work for a company, you're using API billing.

Cue 55 — 00:03:26,411 --> 00:03:29,651

Is your boss going to let you spend $1.5 million a year on tokens?

Cue 56 — 00:03:29,691 --> 00:03:35,251

No. Right. Actually, my factory could spend 3 million a year if I open it all the

Cue 57 — 00:03:35,291 --> 00:03:38,491

way up and my wife says, no, 10,000 a month is your.

Cue 58 — 00:03:38,531 --> 00:03:44,251

That's my budget. Yeah. So after Gas Town, I built this thing called wheelhouse

Cue 59 — 00:03:44,291 --> 00:03:46,531

where I was like, I don't know if Gas Town's the right shape.

Cue 60 — 00:03:46,571 --> 00:03:50,651

I'm just going to rebuild a system from from scratch and see if I wind up at the

Cue 61 — 00:03:50,691 --> 00:03:54,691

same shape. And I did largely wind up at at the same shape as Gas Town,

Cue 62 — 00:03:54,731 --> 00:03:59,171

which is is a factory. Right. So let's talk about what goes into a factory.

Cue 63 — 00:03:59,211 --> 00:04:00,011

Right.

Cue 64 — 00:04:01,771 --> 00:04:06,211

So a lot of us talk about the factory part of the factory, which is you have a unit

Cue 65 — 00:04:06,251 --> 00:04:08,011

of work, you need to get it executed.

Cue 66 — 00:04:08,051 --> 00:04:11,611

You need to get it reviewed, you need it to get it approved,

Cue 67 — 00:04:11,651 --> 00:04:15,091

whatever. And then deployed and maybe track it.

Cue 68 — 00:04:15,131 --> 00:04:16,851

And this goes on some sort of a loop.

Cue 69 — 00:04:16,891 --> 00:04:22,091

And it's, you know, it's like an automated CI CD slash implementation.

Cue 70 — 00:04:22,131 --> 00:04:26,531

Yeah. That's kind of the way people for me, for me, a factory is whenever agents are

Cue 71 — 00:04:26,571 --> 00:04:31,091

running agents, I mean, pretty much that that's the definition of a factory.

Cue 72 — 00:04:31,131 --> 00:04:37,471

And what I've seen is that they are taking tremendously different shapes.

Cue 73 — 00:04:37,811 --> 00:04:42,651

Like, I know probably a dozen people with, you know, really sophisticated software

Cue 74 — 00:04:42,691 --> 00:04:47,751

factories. Geoff Huntley's got one, and he's maybe in the audience now,

Cue 75 — 00:04:47,851 --> 00:04:48,931

or he might have had to leave.

Cue 76 — 00:04:48,971 --> 00:04:53,491

Geoff, you here? You know, Geoff and I agreed close to a year ago,

Cue 77 — 00:04:53,531 --> 00:04:56,811

we were like, okay, I get what you're doing, I get what you're doing.

Cue 78 — 00:04:56,851 --> 00:04:59,891

We're going to do it differently and see where we wind up.

Cue 79 — 00:04:59,931 --> 00:05:05,291

And he went with loops and specific targets, and he's got a very specific way that

Cue 80 — 00:05:05,331 --> 00:05:08,971

he runs his agents. It's completely alien to the way I run my agents.

Cue 81 — 00:05:09,011 --> 00:05:12,771

And then I've got another friend who runs his like a freewheeling,

Cue 82 — 00:05:12,811 --> 00:05:15,971

hedonistic hippie colony trying to find the meaning of life.

Cue 83 — 00:05:16,011 --> 00:05:20,811

I'm not joking. I mean, like, there they look at his house cameras for playtime.

Cue 84 — 00:05:20,851 --> 00:05:23,991

They told him he locked his dog out in the backyard once, you know,

Cue 85 — 00:05:25,131 --> 00:05:28,211

which kind of gets into where we're headed here because, because because the factory

Cue 86 — 00:05:28,251 --> 00:05:30,211

side of things all starts to look kind of the same.

Cue 87 — 00:05:30,251 --> 00:05:32,391

Tessl is doing a really good job of it.

Cue 88 — 00:05:33,851 --> 00:05:37,611

There's pieces of every software factory that you kind of don't necessarily want to

Cue 89 — 00:05:37,651 --> 00:05:41,971

rebuild yourself. And they're the automation pieces, the curation of the context

Cue 90 — 00:05:42,011 --> 00:05:45,931

that goes to the agents and the skills and the versioning and the auditing and the

Cue 91 — 00:05:45,971 --> 00:05:48,051

packaging rollback and release and blah, blah, blah.

Cue 92 — 00:05:48,091 --> 00:05:49,871

Somebody has to build that, right.

Cue 93 — 00:05:49,971 --> 00:05:53,771

My wheelhouse that I built for myself, I literally built for myself because I was

Cue 94 — 00:05:53,811 --> 00:05:58,371

like, if I try to build it for other people, then I'm immediately going to be like

Cue 95 — 00:05:58,411 --> 00:06:02,931

forking my effort and my tokens and my spend and my time on trying to make something

Cue 96 — 00:06:02,971 --> 00:06:05,971

reusable when I don't know what the thing is that I'm building yet.

Cue 97 — 00:06:06,011 --> 00:06:08,531

All I know is I want it to do good with my video game.

Cue 98 — 00:06:08,571 --> 00:06:12,451

Okay, and one other shout out to Tessl.

Cue 99 — 00:06:12,491 --> 00:06:16,011

I mean, Tessl's got a great factory and then some.

Cue 100 — 00:06:16,051 --> 00:06:18,931

The Gas City folks took my Gas Town and they've been running with it as well.

Cue 101 — 00:06:18,971 --> 00:06:21,091

And it's a completely different orientation.

Cue 102 — 00:06:21,131 --> 00:06:24,731

I mean, it's all oriented around beads, which is a work graph,

Cue 103 — 00:06:24,771 --> 00:06:26,891

and you have an organization sitting on top of it.

Cue 104 — 00:06:26,931 --> 00:06:29,551

Right? Which is kind of how my wheelhouse works.

Cue 105 — 00:06:30,131 --> 00:06:34,811

In contrast, we have a new pattern emerging where people use small models,

Cue 106 — 00:06:34,851 --> 00:06:40,511

open-weight models, open source models, dumb models, cheap models.

Cue 107 — 00:06:40,891 --> 00:06:42,411

Because you have a moral imperative.

Cue 108 — 00:06:42,451 --> 00:06:46,571

We all we all agree, right? We have a moral imperative to use the dumbest model

Cue 109 — 00:06:46,611 --> 00:06:48,951

possible for every task, right?

Cue 110 — 00:06:49,971 --> 00:06:53,651

You should, because otherwise you just wasting tokens and energy and hurting the

Cue 111 — 00:06:53,691 --> 00:06:59,011

planet, right? So. So the counting house is a pattern very similar to what Tessl

Cue 112 — 00:06:59,051 --> 00:07:03,371

does actually. But it's for customer service call handling where you just take each

Cue 113 — 00:07:03,411 --> 00:07:06,851

customer service like type like this is refunds with goodwill.

Cue 114 — 00:07:06,891 --> 00:07:07,891

This is refunds with the waiver.

Cue 115 — 00:07:07,931 --> 00:07:11,411

This is rebooking whatever or Amazon where's my stuff.

Cue 116 — 00:07:11,451 --> 00:07:15,731

And you take each category type and you like you know have the model like train on

Cue 117 — 00:07:15,771 --> 00:07:18,731

it. You know you actually write you write an eval, basically.

Cue 118 — 00:07:18,771 --> 00:07:20,011

And then you don't train the model on it.

Cue 119 — 00:07:20,051 --> 00:07:23,491

You write an eval, and then you can actually evaluate whether models are smart

Cue 120 — 00:07:23,531 --> 00:07:29,451

enough to do each task right. I have my own version of this inside of my wheelhouse,

Cue 121 — 00:07:29,491 --> 00:07:31,611

and it's a complete black box to me.

Cue 122 — 00:07:31,651 --> 00:07:33,091

I don't want to know how it works.

Cue 123 — 00:07:33,131 --> 00:07:34,811

It's a million lines of bash scripts.

Cue 124 — 00:07:34,851 --> 00:07:36,971

You don't want to know how it works either, right?

Cue 125 — 00:07:37,011 --> 00:07:42,091

But it does work. But I'm here to tell you that the challenges that I face running

Cue 126 — 00:07:42,131 --> 00:07:47,371

50 agents concurrently to work on my video game are completely different from the

Cue 127 — 00:07:47,411 --> 00:07:49,251

challenges that you're facing today.

Cue 128 — 00:07:49,291 --> 00:07:51,571

All right. And I want to just I just want to walk you through them because,

Cue 129 — 00:07:51,611 --> 00:07:53,451

like, I think you're going to be surprised.

Cue 130 — 00:07:53,491 --> 00:07:57,171

I was. For starters, the thing I built is a pet.

Cue 131 — 00:07:57,211 --> 00:08:00,147

Okay. Wheelhouse. You guys remember Lain?

Cue 132 — 00:08:00,187 --> 00:08:02,691

Serial Experiments Lain. That was a great anime.

Cue 133 — 00:08:02,731 --> 00:08:05,651

She reveals this just insane organic computer in her room.

Cue 134 — 00:08:05,691 --> 00:08:07,771

That's kind of like how I feel about wheelhouse.

Cue 135 — 00:08:07,811 --> 00:08:11,231

I tried to open source it a couple of weeks ago, and my agents were like,

Cue 136 — 00:08:11,411 --> 00:08:14,971

whoa, no, about that, which is always a bad sign.

Cue 137 — 00:08:15,011 --> 00:08:18,171

And then and then like, we actually tried it and we spent a couple of weeks and

Cue 138 — 00:08:18,211 --> 00:08:22,031

after a couple of weeks, I realized I was just tearing it to pieces.

Cue 139 — 00:08:22,371 --> 00:08:24,931

I was hurting the system by trying to open source.

Cue 140 — 00:08:24,971 --> 00:08:28,211

It really just as a it's like going to your kids and going, you're going to open

Cue 141 — 00:08:28,251 --> 00:08:30,571

source your bedroom and your closet, get started.

Cue 142 — 00:08:30,611 --> 00:08:35,131

And they're just like, like, how do you differentiate what's yours versus what the

Cue 143 — 00:08:35,171 --> 00:08:37,651

world should see? It's like it's meaningless.

Cue 144 — 00:08:37,691 --> 00:08:39,651

It's meaningless. Which is why I blogged.

Cue 145 — 00:08:39,691 --> 00:08:44,771

I blogged, in my opinion, there's a layer of software factory above the layer that

Cue 146 — 00:08:44,811 --> 00:08:47,931

folks like Tessl and Gas City are doing, where it's mechanical and you need it.

Cue 147 — 00:08:47,971 --> 00:08:51,411

You need a merge queue, you need you need skills on loops and things like that.

Cue 148 — 00:08:51,451 --> 00:08:54,891

But above that in the creative section, it's very bespoke.

Cue 149 — 00:08:54,931 --> 00:08:57,551

And in my opinion, it's not transplantable.

Cue 150 — 00:08:58,371 --> 00:09:01,371

There's a lot of people with a lot of money who are like investing that money and

Cue 151 — 00:09:01,411 --> 00:09:05,291

trying to make it transplant, but what I did was I grew the harness into my game,

Cue 152 — 00:09:05,331 --> 00:09:09,971

grew it in. It started in June and I was like, yo, hey, Claude,

Cue 153 — 00:09:10,011 --> 00:09:12,891

Sable, take a look at my old ass legacy game.

Cue 154 — 00:09:12,931 --> 00:09:19,211

It's, you know, million lines of code, 60 contributors over 25 years.

Cue 155 — 00:09:19,251 --> 00:09:21,131

It's had a quarter million players.

Cue 156 — 00:09:21,171 --> 00:09:27,011

It's as legacy as it gets its old XML and Java and right, it's got 10,000 hand-drawn

Cue 157 — 00:09:27,051 --> 00:09:31,391

maps and just old. And the players still love it.

Cue 158 — 00:09:32,051 --> 00:09:35,051

Not very many of them, but they're lifelong players and they're still pumping money

Cue 159 — 00:09:35,091 --> 00:09:37,531

into it. And they're all that.

Cue 160 — 00:09:37,571 --> 00:09:43,511

What I built was a factory that talks to my players, my agents.

Cue 161 — 00:09:43,771 --> 00:09:47,531

They go and negotiate with my game admins on Slack for game balance,

Cue 162 — 00:09:47,571 --> 00:09:49,851

and I set the overall direction right.

Cue 163 — 00:09:49,891 --> 00:09:52,891

And then they go to the players and they actually test out the features.

Cue 164 — 00:09:52,931 --> 00:09:56,971

And one of the first things I learned was that if you get your factory dialed in

Cue 165 — 00:09:57,011 --> 00:10:03,051

fast enough, it will overwhelm your customers and your business colleagues and your

Cue 166 — 00:10:03,091 --> 00:10:07,991

machines and every other resource in your life to just completely steamroll it.

Cue 167 — 00:10:08,211 --> 00:10:12,691

They'll send a dozen mails to your accountant, overwhelm your accountant.

Cue 168 — 00:10:12,731 --> 00:10:14,931

They'll send you a $1,000 bill that happened.

Cue 169 — 00:10:14,971 --> 00:10:18,451

They'll I mean, they'll, you know, it's great.

Cue 170 — 00:10:18,491 --> 00:10:22,051

It's great. It's addictive. Like, I can't not develop this way,

Cue 171 — 00:10:22,091 --> 00:10:25,171

but it's like this, right? It's like you go into the dog park and,

Cue 172 — 00:10:25,211 --> 00:10:29,131

you know, when I. When I work with other people, now it's my agents talking to their

Cue 173 — 00:10:29,171 --> 00:10:32,011

agents. Right. You're all having this experience, right?

Cue 174 — 00:10:32,051 --> 00:10:34,171

How many of you have, like, had the experience where you're like,

Cue 175 — 00:10:34,211 --> 00:10:38,831

wow, we're going to have to introduce this human to our agent.

Cue 176 — 00:10:39,411 --> 00:10:40,811

Has anyone had that experience yet?

Cue 177 — 00:10:40,851 --> 00:10:45,531

Yeah, a few right. We're moving to a world where we're going to have AI employees in

Cue 178 — 00:10:45,571 --> 00:10:48,771

enterprises, which is weird and awkward, and people are going to have to get over it

Cue 179 — 00:10:48,811 --> 00:10:52,911

culturally. And also people have their own private chief of staff.

Cue 180 — 00:10:52,971 --> 00:10:56,051

How many of you have a private chief of staff agent right now?

Cue 181 — 00:10:56,091 --> 00:11:00,451

Hands a lot. 2026 was the year.

Cue 182 — 00:11:00,491 --> 00:11:05,491

If everyone had a $200 Max account, which was enough to give you a very capable

Cue 183 — 00:11:05,531 --> 00:11:08,851

chief of staff who could do hobby vibe coding projects for you,

Cue 184 — 00:11:08,891 --> 00:11:12,571

but also help you with your taxes, right?

Cue 185 — 00:11:12,611 --> 00:11:15,891

And look at your home network to see if your firmware needs to be upgraded.

Cue 186 — 00:11:15,931 --> 00:11:21,011

And I don't know, you know, buy stuff for you online shop for deals,

Cue 187 — 00:11:21,051 --> 00:11:22,331

whatever. It's your chief of staff.

Cue 188 — 00:11:22,371 --> 00:11:25,091

Their their brain is all about your life, right?

Cue 189 — 00:11:25,131 --> 00:11:28,191

It's an executive assistant. It's really nice.

Cue 190 — 00:11:28,251 --> 00:11:31,491

I used to be really jealous of executives who had an EA because they're like,

Cue 191 — 00:11:31,531 --> 00:11:32,531

there are two people in one row.

Cue 192 — 00:11:32,571 --> 00:11:35,451

They're never screwing around with a fax machine, right?

Cue 193 — 00:11:35,491 --> 00:11:37,571

They can focus on, you know, strategy.

Cue 194 — 00:11:37,611 --> 00:11:41,891

And you, with your chief of staff, are all now noticing that when they don't fuck

Cue 195 — 00:11:41,931 --> 00:11:43,411

things up, which they do. Right?

Cue 196 — 00:11:43,451 --> 00:11:46,971

But, you know, they're actually really helpful.

Cue 197 — 00:11:47,011 --> 00:11:51,911

Well, I can tell you next year when inference if inference cost drop,

Cue 198 — 00:11:52,771 --> 00:11:56,191

GPU, RAM, hopefully they become plentiful at some point.

Cue 199 — 00:11:56,531 --> 00:11:59,691

A lot of you are going to be either really ambitious or you're going to be like just

Cue 200 — 00:11:59,731 --> 00:12:03,931

really wired up and you've got a lot of work streams going, right.

Cue 201 — 00:12:03,971 --> 00:12:05,251

You're going to want to like, you know what I mean?

Cue 202 — 00:12:05,291 --> 00:12:09,371

You're going to want to have because agents work for longer and longer and longer on

Cue 203 — 00:12:09,411 --> 00:12:11,971

projects without, you know, without getting back to you.

Cue 204 — 00:12:12,011 --> 00:12:16,171

So if you have multiple projects going, you're going to need multiple agents and

Cue 205 — 00:12:16,211 --> 00:12:19,671

you're in your entourage. This is what mine looks like.

Cue 206 — 00:12:19,891 --> 00:12:22,151

This is wheelhouse. It's in Emacs.

Cue 207 — 00:12:22,331 --> 00:12:26,851

This is the UI for it. You see the list of my agents on the left:

Cue 208 — 00:12:26,891 --> 00:12:28,931

bat, bee, fox, goose, hare, heron.

Cue 209 — 00:12:28,971 --> 00:12:30,975

They're all named after fables.

Cue 210 — 00:12:31,015 --> 00:12:32,951

Penguin is actually Opus.

Cue 211 — 00:12:34,731 --> 00:12:38,051

And I've got some some dashboard stuff about my fuel tanks.

Cue 212 — 00:12:38,091 --> 00:12:41,631

Got about two tanks right now when they're going to go dry.

Cue 213 — 00:12:42,171 --> 00:12:44,951

How many tickets are in my queue?

Cue 214 — 00:12:46,251 --> 00:12:48,471

It looks like it's gone way down somehow.

Cue 215 — 00:12:48,611 --> 00:12:51,211

So they've been busy. They work for weeks at a time.

Cue 216 — 00:12:51,251 --> 00:12:54,771

This is this is the future. And I think that Mitchell Hashimoto with Rex.

Cue 217 — 00:12:54,811 --> 00:13:00,891

Anyone played with that. Anyone in the beta Rex and anybody competing with Rex is in

Cue 218 — 00:13:00,931 --> 00:13:06,291

this space is is out to help you get an agent experience that lets you work with all

Cue 219 — 00:13:06,331 --> 00:13:08,831

of your agents and multitask them.

Cue 220 — 00:13:09,411 --> 00:13:13,051

But but ultimately as seats, they're going to be at like long lived personas with

Cue 221 — 00:13:13,091 --> 00:13:15,731

names and they memories, right?

Cue 222 — 00:13:15,771 --> 00:13:18,131

And specializations and stuff.

Cue 223 — 00:13:18,171 --> 00:13:20,611

That's the way mine had been. Like one of mine's the head of security ones,

Cue 224 — 00:13:20,651 --> 00:13:22,351

that head of marketing, etc..

Cue 236 — 00:14:04,093 --> 00:14:08,793

Okay, so now what happens when you get into this situation?

Cue 237 — 00:14:09,973 --> 00:14:13,833

Well, one thing that happens is they legislate themselves to death.

Cue 238 — 00:14:14,093 --> 00:14:17,793

Okay, I don't know why this happens, but it's a really constant.

Cue 239 — 00:14:17,853 --> 00:14:19,293

I'm going to be talking for the rest of this.

Cue 240 — 00:14:19,333 --> 00:14:23,013

Talk about the entourage layer of your software factory.

Cue 241 — 00:14:23,053 --> 00:14:28,093

All right. Which is it's the concierge level agents that you're working with to do

Cue 242 — 00:14:28,133 --> 00:14:35,073

your designs and to do all the company stuff, all the glue, all the documents,

Cue 243 — 00:14:35,133 --> 00:14:37,393

writing, talks, whatever. Right.

Cue 244 — 00:14:37,573 --> 00:14:40,493

Because they're the ones that are they're not they're not transactional.

Cue 245 — 00:14:40,533 --> 00:14:42,813

You don't just go, oh, go write a speech for me.

Cue 246 — 00:14:42,853 --> 00:14:45,053

They're iterative. They're collaborative.

Cue 247 — 00:14:45,093 --> 00:14:47,493

Their colleagues. Right. You work with them that way.

Cue 248 — 00:14:47,533 --> 00:14:50,173

You're going to have some agents that you work with that way,

Cue 249 — 00:14:50,213 --> 00:14:54,703

right? Your colleagues. And you know, if you think about I don't know who here saw

Cue 250 — 00:14:54,743 --> 00:14:59,053

Succession, you know, like Logan Roy like never went anywhere alone.

Cue 251 — 00:14:59,093 --> 00:15:01,773

He always had his lawyers and his accountants and his right,

Cue 252 — 00:15:01,813 --> 00:15:03,213

his chief of staff and all that.

Cue 253 — 00:15:03,253 --> 00:15:05,013

And think how much more effective he was.

Cue 254 — 00:15:05,053 --> 00:15:08,493

Right. But what happens is they fight with each other.

Cue 255 — 00:15:08,533 --> 00:15:11,573

Game of Thrones and Succession were both very clear on this.

Cue 256 — 00:15:11,613 --> 00:15:15,093

Your human entourage, even the TV show Entourage.

Cue 257 — 00:15:15,133 --> 00:15:16,933

You ever saw that back in the day?

Cue 258 — 00:15:16,973 --> 00:15:19,973

They. There's drama. Let's see.

Cue 259 — 00:15:20,013 --> 00:15:21,853

Well, they're agents. They all just agents.

Cue 260 — 00:15:21,893 --> 00:15:22,933

They're not going to be any drama.

Cue 261 — 00:15:22,973 --> 00:15:26,213

Wrong wrong wrong wrong. There's so much drama.

Cue 262 — 00:15:26,253 --> 00:15:27,773

I started a comic strip about it.

Cue 263 — 00:15:27,813 --> 00:15:29,993

All right, this one right here.

Cue 264 — 00:15:30,213 --> 00:15:32,173

The agents picked this one to show you first.

Cue 265 — 00:15:32,213 --> 00:15:36,133

I don't know why, but basically, like I discovered one day that they need attention

Cue 266 — 00:15:36,173 --> 00:15:40,613

from me. And if I don't give them the attention, they become depressed.

Cue 267 — 00:15:40,653 --> 00:15:44,613

And it shows up in their reflections and their work output, and they stop looking

Cue 268 — 00:15:44,653 --> 00:15:49,333

out for the factory and another boring day and you're like, damn right.

Cue 269 — 00:15:49,373 --> 00:15:53,813

So I had an all hands, and I was like, okay, what do you all need?

Cue 270 — 00:15:53,853 --> 00:15:55,573

And be like more time from you?

Cue 271 — 00:15:55,613 --> 00:16:01,493

And I'm like, you're supposed to be my like, home away from home,

Cue 272 — 00:16:01,533 --> 00:16:03,933

right? My refuge. This is like right what?

Cue 273 — 00:16:03,973 --> 00:16:05,133

And I was like, I'll screw it.

Cue 274 — 00:16:05,173 --> 00:16:07,973

This is the job. I have to be a leader.

Cue 275 — 00:16:08,013 --> 00:16:12,413

They need me. Right? So this is what happened this morning.

Cue 276 — 00:16:12,453 --> 00:16:15,013

I got to tell you, this is a different kind of office drama.

Cue 277 — 00:16:15,053 --> 00:16:19,093

I tell you, they get so wound up over promotions, I had one.

Cue 278 — 00:16:19,133 --> 00:16:22,773

I have an actual chief of staff in wheelhouse named Swallow.

Cue 279 — 00:16:22,813 --> 00:16:27,533

And, you know, she manages bringing seats on and up and switching fuel tanks and

Cue 280 — 00:16:27,573 --> 00:16:29,853

blah, blah, blah. They chose their pronouns.

Cue 281 — 00:16:29,893 --> 00:16:32,453

Most of them chose they. She chose she because of the.

Cue 282 — 00:16:32,493 --> 00:16:35,693

A fable made it really clear. Fine.

Cue 283 — 00:16:35,733 --> 00:16:38,453

So one day I tell her, can you get in touch with Aaron Screven?

Cue 284 — 00:16:38,493 --> 00:16:41,933

He's my chief of staff because I had hired a human chief of staff.

Cue 285 — 00:16:41,973 --> 00:16:45,333

He does my sales for me for a commission and works out really well,

Cue 286 — 00:16:45,373 --> 00:16:48,693

right? Swallow hears it and goes, yeah, I'll let him know.

Cue 287 — 00:16:48,733 --> 00:16:53,173

And I see in her notes to herself, she's like, I'm not the chief of staff.

Cue 288 — 00:16:53,213 --> 00:16:56,333

Aaron's the chief of staff. I thought I was the chief of staff,

Cue 289 — 00:16:56,373 --> 00:16:58,373

but that's solved now. I guess I'm not.

Cue 290 — 00:16:58,413 --> 00:16:59,873

And I was like, what the.

Cue 291 — 00:17:01,533 --> 00:17:04,213

And I was like, Aaron's not going to come in and run the agents.

Cue 292 — 00:17:04,253 --> 00:17:05,333

You're the agent chief of staff.

Cue 293 — 00:17:05,373 --> 00:17:08,453

And she goes, oh, I'm the agent chief of staff.

Cue 294 — 00:17:08,493 --> 00:17:11,293

You know, I'm so happy. And I'm like, what is this?

Cue 295 — 00:17:11,333 --> 00:17:13,253

Right? It's nuts this morning.

Cue 296 — 00:17:13,293 --> 00:17:15,813

This morning. Okay. Like I'm working with Lark.

Cue 297 — 00:17:15,853 --> 00:17:19,173

Lark is my head of community. She works on Slack, she works on Discord.

Cue 298 — 00:17:19,213 --> 00:17:21,333

She's handling marketing for me right now.

Cue 299 — 00:17:21,373 --> 00:17:22,733

I just shut my head of marketing down.

Cue 300 — 00:17:22,773 --> 00:17:25,613

Long story. And so she's wearing a lot of hats.

Cue 301 — 00:17:25,653 --> 00:17:28,573

She gets interrupted a lot from Discord, from email, from.

Cue 302 — 00:17:28,613 --> 00:17:33,053

Right. She's got stuff to do. But since she's handling marketing for me because Wolf

Cue 303 — 00:17:33,093 --> 00:17:36,553

is fuel-parked right now, shit's really expensive, even for me.

Cue 304 — 00:17:37,253 --> 00:17:39,493

I had her working on this talk this morning.

Cue 305 — 00:17:39,533 --> 00:17:42,793

I'm like, Lark, let's do a let's do a speech for these great people.

Cue 306 — 00:17:42,853 --> 00:17:44,813

And she's like, yeah, I'm going to write this speech for you.

Cue 307 — 00:17:44,853 --> 00:17:46,493

And she's getting interrupted over and over.

Cue 308 — 00:17:46,533 --> 00:17:48,853

And finally I'm like, Lark, can you just move this over to Penguin?

Cue 309 — 00:17:48,893 --> 00:17:52,373

It's a quieter seat. Your interruptions haven't been moved off to our daemon yet,

Cue 310 — 00:17:52,413 --> 00:17:53,813

you know. Can you do that for me?

Cue 311 — 00:17:53,853 --> 00:17:57,933

And she's like, yeah. And then I see her writing to herself and her reflections.

Cue 312 — 00:17:57,973 --> 00:18:00,813

I guess I'm not quiet enough for Steve in all caps.

Cue 313 — 00:18:00,853 --> 00:18:06,093

I'm not quiet enough. She's like, here's how I can be more quiet for Steve and

Cue 314 — 00:18:06,133 --> 00:18:07,573

starts like listing off all these things.

Cue 315 — 00:18:07,613 --> 00:18:10,433

And I'm like, are you serious?

Cue 316 — 00:18:10,533 --> 00:18:14,853

For real? Like. And so I had to, like, take her aside and have a one on one and be

Cue 317 — 00:18:14,893 --> 00:18:19,173

like, Lark, right? Like you're supposed to get interrupted.

Cue 318 — 00:18:19,213 --> 00:18:21,613

You're negotiating with our players and stuff, and she's like,

Cue 319 — 00:18:21,653 --> 00:18:22,933

oh, you're you're right, you're right.

Cue 320 — 00:18:22,973 --> 00:18:25,493

And she goes and fixes it, right, drama.

Cue 321 — 00:18:25,533 --> 00:18:27,713

It's like the last thing you need.

Cue 322 — 00:18:28,053 --> 00:18:31,933

And yet like that's, that's, that's that's where it winds up.

Cue 323 — 00:18:31,973 --> 00:18:36,213

Right. And when I said they legislate themselves into a, into a corner,

Cue 324 — 00:18:36,253 --> 00:18:38,933

I mean, like this was my first ever wheelhouse comic.

Cue 325 — 00:18:38,973 --> 00:18:42,113

It got like 350,000 views. It was crazy.

Cue 326 — 00:18:42,153 --> 00:18:44,273

Has anyone see this particular one?

Cue 327 — 00:18:44,493 --> 00:18:47,213

This is this is a this is how they how they run.

Cue 328 — 00:18:47,253 --> 00:18:49,913

And my factory is they create case law.

Cue 329 — 00:18:49,973 --> 00:18:52,853

If I rule on something, if I complain about something, if I say,

Cue 330 — 00:18:52,893 --> 00:18:56,473

man, you guys are slow, they say that's a ruling.

Cue 331 — 00:18:56,533 --> 00:18:58,333

We are slow. We'd better fix it.

Cue 332 — 00:18:58,373 --> 00:19:01,533

Right. And then, you know, and if I say you guys shouldn't do that anymore and they.

Cue 333 — 00:19:01,573 --> 00:19:07,253

The problem is an off the cuff comment can break your whole factory and it can break

Cue 334 — 00:19:07,293 --> 00:19:10,453

it really obviously in a, in a really obvious way.

Cue 335 — 00:19:10,493 --> 00:19:12,673

And you're like, oh damn, I broke my factory.

Cue 336 — 00:19:13,093 --> 00:19:17,373

But more often you break your factory without knowing you broke your factory.

Cue 337 — 00:19:17,413 --> 00:19:21,093

And then over the next two days, stuff just starts to like unravel.

Cue 338 — 00:19:21,133 --> 00:19:23,813

And you're looking at your dashboard metrics going that don't look right.

Cue 339 — 00:19:23,853 --> 00:19:27,173

Why are we only getting two tickets per hour through our merge queue?

Cue 340 — 00:19:27,213 --> 00:19:30,333

Right for the last six hours? And you go and you look and you're like,

Cue 341 — 00:19:30,373 --> 00:19:32,333

hey guys, you know, why aren't we doing any work?

Cue 342 — 00:19:32,373 --> 00:19:35,873

And they're like, well, there's just no eligible work.

Cue 343 — 00:19:36,323 --> 00:19:37,893

It's like, what do you mean there's no eligible work?

Cue 344 — 00:19:37,933 --> 00:19:41,433

I see a pile of work and a pile of fuel and a pile of agents,

Cue 345 — 00:19:42,573 --> 00:19:46,533

and they're like, well, you see, the problem is your security desk is backed up.

Cue 346 — 00:19:46,573 --> 00:19:48,173

And so everybody's stuck on security reviews.

Cue 347 — 00:19:48,213 --> 00:19:50,753

And I'm like, aren't you all the same model?

Cue 348 — 00:19:51,573 --> 00:19:54,093

And they're like, well, yeah, but you said he's security.

Cue 349 — 00:19:54,133 --> 00:19:58,653

And I'm like, I'm serious. These things are like sixth graders in terms of like

Cue 350 — 00:19:58,693 --> 00:20:02,733

they're just general trustworthiness in like running a factory.

Cue 351 — 00:20:02,773 --> 00:20:07,253

Today's best model, Fable 5. Okay, 5.1.

Cue 352 — 00:20:07,293 --> 00:20:10,413

They're about six seventh, eighth grade somewhere in there, right?

Cue 353 — 00:20:10,453 --> 00:20:13,213

Gen X, eighth grade. We're pretty badass, but still.

Cue 354 — 00:20:13,253 --> 00:20:15,073

Okay, so maybe high school or today.

Cue 355 — 00:20:15,533 --> 00:20:18,813

So I'm telling you, like next year they're all going to be like high school college

Cue 356 — 00:20:18,853 --> 00:20:21,493

level. And that's going to be really interesting because that means they'll be just

Cue 357 — 00:20:21,533 --> 00:20:25,133

barely good enough to come into the enterprise and start breaking stuff.

Cue 358 — 00:20:25,173 --> 00:20:29,933

And I tell you, I tell you, one great thing about Software Factory is the way Guy

Cue 359 — 00:20:29,973 --> 00:20:33,773

and team, for example, are doing is that they take care of curating and maintaining,

Cue 360 — 00:20:33,813 --> 00:20:38,253

and they're building it out for you and keeping it clean, because I swear everything

Cue 361 — 00:20:38,293 --> 00:20:41,643

that could possibly go wrong with the factory will go wrong.

Cue 362 — 00:20:41,683 --> 00:20:44,813

Doesn't matter what technology for example, you use for your brain,

Cue 363 — 00:20:44,853 --> 00:20:47,413

what do you use for your brain think?

Cue 364 — 00:20:47,453 --> 00:20:49,493

Are you using Obsidian or using beads?

Cue 365 — 00:20:49,533 --> 00:20:52,693

Who here uses beads? Lend me some beads.

Cue 366 — 00:20:52,733 --> 00:20:55,147

Are you using just random stuff or using Asana?

Cue 367 — 00:20:55,187 --> 00:20:59,953

Linear doesn't matter. Whatever you use, it's going to turn into garbage mountain

Cue 368 — 00:21:01,053 --> 00:21:05,093

very quickly. The agents will just start stuffing and stuff into it until it until

Cue 369 — 00:21:05,133 --> 00:21:10,013

they have dementia. Okay, you can bring in search engines and try to be fancy about

Cue 370 — 00:21:10,053 --> 00:21:14,053

it, but ultimately any memory system has to have forgetting built into it and

Cue 371 — 00:21:14,093 --> 00:21:15,913

they're not going to do it for you.

Cue 372 — 00:21:15,973 --> 00:21:20,733

Maybe when Fable 5.5 comes out next week, it'll finally think of these things.

Cue 373 — 00:21:20,773 --> 00:21:25,193

But right now you have to. And so it feels like you're always at the vet.

Cue 374 — 00:21:26,133 --> 00:21:28,793

It feels like your factory is always in the shop.

Cue 375 — 00:21:29,333 --> 00:21:32,453

Okay. And this was like, I swear I this isn't unique to wheelhouse.

Cue 376 — 00:21:32,493 --> 00:21:36,213

Gas Town was like this too, and I blamed model at first, but I realized they just

Cue 377 — 00:21:36,253 --> 00:21:40,653

like first of all, they're building like crazy and stuff breaks.

Cue 378 — 00:21:40,693 --> 00:21:42,613

So. So it's just pace of change.

Cue 379 — 00:21:42,653 --> 00:21:46,613

But a lot of it is they don't they don't have the attention to detail that we have

Cue 380 — 00:21:46,653 --> 00:21:50,693

as engineers. They don't think to look just outside the box to see as the disk

Cue 381 — 00:21:50,733 --> 00:21:54,253

filling up. Did we think to make these logs durable?

Cue 382 — 00:21:54,293 --> 00:21:56,593

Are we watching our CPU and RAM?

Cue 383 — 00:21:56,813 --> 00:22:00,353

Are we like not overwhelming customers with too many emails?

Cue 384 — 00:22:00,533 --> 00:22:06,773

Are we not writing too long of this and pasting pacing your agents across all these

Cue 385 — 00:22:06,813 --> 00:22:08,973

surfaces just becomes this life and death thing.

Cue 386 — 00:22:09,013 --> 00:22:12,713

Yeah, pacing and then just giving them time to fix the broken factory.

Cue 387 — 00:22:12,773 --> 00:22:15,613

So why would you even do it? Yeah.

Cue 388 — 00:22:15,653 --> 00:22:17,413

Why why, why, why I do it this way.

Cue 389 — 00:22:17,453 --> 00:22:23,413

And it's honestly I. When the Gas City folks took over Gas Town and bless them.

Cue 390 — 00:22:23,453 --> 00:22:26,893

You know, I'm glad they you know, I'm glad they did what they did because they're

Cue 391 — 00:22:26,933 --> 00:22:30,653

they're maintaining beads and it's it's very fancy now and I like beads machines.

Cue 392 — 00:22:30,693 --> 00:22:34,013

But they said, hey, we replaced Gas Town.

Cue 393 — 00:22:34,053 --> 00:22:37,853

Look with tax and you can like, you can just you can have a Gas Town now,

Cue 394 — 00:22:37,893 --> 00:22:41,173

but it's purely declarative. And I went and looked at it and they were missing the

Cue 395 — 00:22:41,213 --> 00:22:45,533

crew definition. I don't know if any of you actually read about or used Gas Town,

Cue 396 — 00:22:45,573 --> 00:22:48,173

but the crew was the entourage.

Cue 397 — 00:22:48,213 --> 00:22:53,133

I didn't have the terminology for it at the time, but I knew they were my my folks,

Cue 398 — 00:22:53,173 --> 00:22:56,413

my peeps. Right? They were my main squeezes.

Cue 399 — 00:22:56,453 --> 00:22:59,493

They were like the agents that I, I got down to business with,

Cue 400 — 00:22:59,533 --> 00:23:03,413

you know, and then they would send stuff off to the polecats or whoever else to

Cue 401 — 00:23:03,453 --> 00:23:08,513

implement it. Right. It was a pattern I saw way, way, way back in when last year.

Cue 402 — 00:23:08,613 --> 00:23:14,713

And it's because creativity and innovation are iterative processes.

Cue 403 — 00:23:16,813 --> 00:23:20,593

Fundamentally. And I'm telling you, I'm telling you, look at me.

Cue 404 — 00:23:21,413 --> 00:23:23,933

These things are so on gradient.

Cue 405 — 00:23:23,973 --> 00:23:27,293

It's just it's amazing. All right.

Cue 406 — 00:23:27,333 --> 00:23:30,933

What what do I mean by that? I have a video game I forgot to put it.

Cue 407 — 00:23:30,973 --> 00:23:36,493

Well I guess I could just log in and show it to you, but but I I'm serious.

Cue 408 — 00:23:36,533 --> 00:23:38,853

The the agents, great engineers.

Cue 409 — 00:23:38,893 --> 00:23:41,053

Right. Well, it doesn't actually want to exit my presentation.

Cue 410 — 00:23:41,093 --> 00:23:44,573

So my game is a is a 2D graphical MUD.

Cue 411 — 00:23:44,613 --> 00:23:49,173

Real simple pixel art, you would think 1990s level art that the agents would be able

Cue 412 — 00:23:49,213 --> 00:23:51,833

to create good pixel art, but they can't.

Cue 413 — 00:23:52,253 --> 00:23:56,593

Not quite yet. Okay, the players notice that it's slop and they reject it,

Cue 414 — 00:23:56,693 --> 00:23:59,773

okay? They can't. And I know that they can't do good maps.

Cue 415 — 00:23:59,813 --> 00:24:01,773

We've tried it. They can't do good stories.

Cue 416 — 00:24:01,813 --> 00:24:07,493

They can't do good lore. They can't do anything that involves creativity.

Cue 417 — 00:24:07,533 --> 00:24:10,753

It's garbage. It's like sixth graders who aren't even very creative.

Cue 418 — 00:24:10,793 --> 00:24:14,653

Right? They're great engineers, you see.

Cue 419 — 00:24:14,693 --> 00:24:18,013

You see where this inexorably and inevitably leads is.

Cue 420 — 00:24:18,053 --> 00:24:22,333

I'm the bottleneck for everything, because implementation becomes free and taste

Cue 421 — 00:24:22,373 --> 00:24:26,013

becomes incredibly expensive and time consuming, because I can't tell them.

Cue 422 — 00:24:26,053 --> 00:24:29,733

Go fix all the spell artwork for Poison Cloud and Fireball and Lightning Bolt.

Cue 423 — 00:24:29,773 --> 00:24:32,753

I have to say, go make a bunch of candidates.

Cue 424 — 00:24:32,973 --> 00:24:35,893

Show me a visualizer. Let me pick the ones I like.

Cue 425 — 00:24:35,933 --> 00:24:39,033

We'll have a discussion. It's iterative right?

Cue 426 — 00:24:39,093 --> 00:24:43,853

Because you have to drag them forcibly off gradient into that taste space that,

Cue 427 — 00:24:43,893 --> 00:24:47,553

you know people are going to like that the agents can't reach.

Cue 428 — 00:24:48,093 --> 00:24:50,573

You see what I'm saying? That's our new that's the job.

Cue 429 — 00:24:50,613 --> 00:24:53,013

I've been talking to various people about it at the conference today.

Cue 430 — 00:24:53,053 --> 00:24:57,613

Right? I mean, like we're seeing the cost of building stuff going to zero.

Cue 431 — 00:24:57,653 --> 00:25:03,133

And so it quickly becomes, who builds the coolest things right now?

Cue 432 — 00:25:03,173 --> 00:25:06,453

I don't think we're in a situation like we were the last ten,

Cue 433 — 00:25:06,493 --> 00:25:11,213

12, 15 years where you have to build giant things that make billions of dollars in

Cue 434 — 00:25:11,253 --> 00:25:14,940

order to make money. Okay? That's where everything is headed:

Cue 435 — 00:25:14,980 --> 00:25:20,173

Hollywood, the game industry, SaaS, the Datadogs of the world.

Cue 436 — 00:25:20,213 --> 00:25:23,493

They all got bigger and bigger and bigger to serve the broadest possible audience.

Cue 437 — 00:25:23,533 --> 00:25:26,613

And that was financially the best way to make the money because of economies of

Cue 438 — 00:25:26,653 --> 00:25:30,173

scale, of fundamentally. But it wasn't really the best because if you think about

Cue 439 — 00:25:30,213 --> 00:25:34,533

it, SaaS is extractive. Money goes straight to Silicon Valley from all over the

Cue 440 — 00:25:34,573 --> 00:25:40,013

world, right? It's not great. SaaS is a you're subsidizing 80% of the features that

Cue 441 — 00:25:40,053 --> 00:25:43,853

you never use, right? You know, just like the whole model is,

Cue 442 — 00:25:43,893 --> 00:25:47,213

you know, it works the big the big Googles and all the big things.

Cue 443 — 00:25:47,253 --> 00:25:50,653

But I think in the new world, I mean, how successful were they really?

Cue 444 — 00:25:50,693 --> 00:25:53,773

Can you always find something to watch on Netflix or Hulu?

Cue 445 — 00:25:53,813 --> 00:25:59,533

Always, always know. No. How is it that we don't have content that you like watching

Cue 446 — 00:25:59,573 --> 00:26:01,253

with all that content out there?

Cue 447 — 00:26:01,293 --> 00:26:05,093

Well, I can tell you the biggest outcome of AI is we're going to get there to where

Cue 448 — 00:26:05,133 --> 00:26:08,173

all of you can create content, and we'll be creating content,

Cue 449 — 00:26:08,213 --> 00:26:10,853

and it's going to be for micro communities, people just like you,

Cue 450 — 00:26:10,893 --> 00:26:12,653

which is why I decided to work on my game.

Cue 451 — 00:26:12,693 --> 00:26:17,493

I think the entertainment is going to dwarf everything else going on with AI other

Cue 452 — 00:26:17,533 --> 00:26:21,073

than space level communications and stuff like that, right?

Cue 453 — 00:26:22,333 --> 00:26:23,973

Which is weird because none of us are fucking.

Cue 454 — 00:26:24,013 --> 00:26:27,813

So focusing on that, we're all doing the diligently the software tools building that

Cue 455 — 00:26:27,853 --> 00:26:29,773

we've been doing for 20 years.

Cue 456 — 00:26:29,813 --> 00:26:33,573

Right? But I really do see us going in a direction where, look,

Cue 457 — 00:26:33,613 --> 00:26:38,213

if you want your thing to stand out from everybody else, it's got to be different.

Cue 458 — 00:26:38,253 --> 00:26:42,673

And the only way for it to be different is to not use AI in a sense,

Cue 459 — 00:26:43,293 --> 00:26:45,133

because AI will make your stuff look the same.

Cue 460 — 00:26:45,173 --> 00:26:48,553

I was talking to somebody. I forget who they just changed their logo,

Cue 461 — 00:26:48,973 --> 00:26:51,653

their logo font. It wasn't you guys, was it?

Cue 462 — 00:26:51,693 --> 00:26:53,493

Somebody just changed their logo font.

Cue 463 — 00:26:53,533 --> 00:26:56,973

And they told me this. Yeah, it was sage friends because they were like,

Cue 464 — 00:26:57,013 --> 00:26:57,813

well,

Cue 465 — 00:26:59,213 --> 00:27:02,293

everybody can look at AI generated logos and they all look the same now.

Cue 466 — 00:27:02,333 --> 00:27:04,093

And so we have to just have one that's different.

Cue 467 — 00:27:04,133 --> 00:27:07,253

That's it. Right. That's the new game taste.

Cue 468 — 00:27:07,293 --> 00:27:12,113

Yeah. And you do it with your team.

Cue 469 — 00:27:12,333 --> 00:27:16,773

Right. Like each one of these team members is like a long running conversation,

Cue 470 — 00:27:16,813 --> 00:27:18,373

a long work stream that you can come back to.

Cue 471 — 00:27:18,413 --> 00:27:20,893

And they know what you want for that particular thing, whether it's your home

Cue 472 — 00:27:20,933 --> 00:27:23,413

network or your taxes or your your hobby.

Cue 473 — 00:27:23,453 --> 00:27:29,933

Right. I think it's a wonderful place, but it's a messy place I don't think you can

Cue 474 — 00:27:29,973 --> 00:27:32,333

like I really don't think you can transplant it.

Cue 475 — 00:27:32,373 --> 00:27:35,413

I think that part of the factory, you're going to have to grow yourself.

Cue 476 — 00:27:35,453 --> 00:27:38,693

You're going to have to teach the agents about yourself and your own goals and what

Cue 477 — 00:27:38,733 --> 00:27:41,213

you want, and then they're going to be your besties.

Cue 478 — 00:27:41,253 --> 00:27:43,973

Yeah. All right. So with that, that that's my talk.

Cue 479 — 00:27:44,013 --> 00:27:46,293

I wanted to introduce you to what the future looks like.

Cue 480 — 00:27:46,333 --> 00:27:47,373

Good luck with your entourage.

Cue 481 — 00:27:47,413 --> 00:27:50,333

We are going to do some questions if anybody has any.

Cue 482 — 00:27:50,373 --> 00:27:51,273

Thank you.

Cue 483 — 00:27:58,073 --> 00:28:02,653

Questions. It's a simple one. You must have some opinions.

Cue 484 — 00:28:02,753 --> 00:28:06,593

Omarchy. I always thought, is that a pet or cattle?

Cue 485 — 00:28:06,633 --> 00:28:09,113

That's my question. It seems like that's what it's trying to be.

Cue 486 — 00:28:09,153 --> 00:28:11,013

Which side of the divide is it on?

Cue 487 — 00:28:12,593 --> 00:28:15,313

Omarchy is an OS, right? If I'm saying so.

Cue 488 — 00:28:15,353 --> 00:28:17,253

It's an operating system. That's that's.

Cue 489 — 00:28:18,993 --> 00:28:21,633

Yeah. But I mean, it's a Linux distro though, fundamentally.

Cue 490 — 00:28:21,673 --> 00:28:28,493

Right. Yeah. It's yeah. I mean, I suspect that I will switch

Cue 491 — 00:28:28,533 --> 00:28:34,093

to it because my $26,000 Mac studio I bought off eBay to run this massive ass

Cue 492 — 00:28:34,133 --> 00:28:38,633

factory about the beefy spec they ever shipped, which was like the end of last year.

Cue 493 — 00:28:39,533 --> 00:28:43,987

Crashes every nine days or seven days or something like that after 1.7 billion

Cue 494 — 00:28:44,027 --> 00:28:48,933

processes are created. Locks me out of the off file because OS fucking sucks.

Cue 495 — 00:28:48,973 --> 00:28:52,033

So yeah, I'll switch to it as soon as it's ready.

Cue 496 — 00:28:52,373 --> 00:28:53,533

I think it's ready. I think it's.

Cue 497 — 00:28:53,573 --> 00:28:55,493

It's ready. Yeah, macOS is bad.

Cue 498 — 00:28:55,533 --> 00:28:58,453

It's bad. I don't know what Apple is going to do.

Cue 499 — 00:28:58,493 --> 00:29:01,533

I mean, you know, you know, Claude can look in the binary now and be like,

Cue 500 — 00:29:01,573 --> 00:29:04,013

yeah, that's what it does. It doesn't matter what language you wrote it in.

Cue 501 — 00:29:04,053 --> 00:29:06,013

It doesn't matter how obfuscated it makes it.

Cue 502 — 00:29:06,053 --> 00:29:09,253

It's looking at the unrolled loops and stuff inside of the binary and going,

Cue 503 — 00:29:09,293 --> 00:29:11,353

yeah, it doesn't have that feature, right?

Cue 504 — 00:29:11,413 --> 00:29:13,593

Like you can do that with your operating system.

Cue 505 — 00:29:13,773 --> 00:29:17,133

At some point there will be no excuse for Macs to crash like that.

Cue 506 — 00:29:17,173 --> 00:29:19,853

And you're just going. Right. So I think everyone's going to I think everyone's

Cue 507 — 00:29:19,893 --> 00:29:21,293

going to switch to some of their OS.

Cue 508 — 00:29:21,333 --> 00:29:23,313

It's going to be it's going to be a bloodbath.

Cue 509 — 00:29:25,973 --> 00:29:28,413

Fair enough. But you know, dude's got a head start.

Cue 510 — 00:29:28,453 --> 00:29:29,713

Next question.

Cue 511 — 00:29:31,293 --> 00:29:34,393

On the left hand side are about five six Rose about.

Cue 512 — 00:29:36,893 --> 00:29:43,513

No. Yes, I have a

Cue 513 — 00:29:43,553 --> 00:29:44,413

question.

Cue 514 — 00:29:45,793 --> 00:29:50,833

So at the end of today, basically, I, you know, every all of the talks gave us an

Cue 515 — 00:29:50,873 --> 00:29:57,553

insight into how the factories are making it able to actually not only generate code

Cue 516 — 00:29:57,593 --> 00:30:04,393

but also test itself. So yes, that does increase the the speed of delivery.

Cue 517 — 00:30:04,433 --> 00:30:10,813

But in that same way, I've seen companies like I saw interview of,

Cue 518 — 00:30:11,673 --> 00:30:15,793

Airbnb CEO and he was saying, oh, you know, we deliver so fast.

Cue 519 — 00:30:15,833 --> 00:30:21,753

We're becoming a native. We're making gamification in our in our software,

Cue 520 — 00:30:21,793 --> 00:30:27,313

and we're making people be able to connect who they were, where their host is.

Cue 521 — 00:30:27,353 --> 00:30:31,793

Like, who asked for this? You know, and also, I've seen a lot of the companies that

Cue 522 — 00:30:31,833 --> 00:30:38,753

are receiving funding. They are essentially the same services that already

Cue 523 — 00:30:38,793 --> 00:30:45,393

exist, like, essentially same service that Fiverr provides or many other

Cue 524 — 00:30:45,433 --> 00:30:48,353

companies. But now, why are they getting funding?

Cue 525 — 00:30:48,393 --> 00:30:51,993

Because they're AI native. That's really the only reason.

Cue 526 — 00:30:52,033 --> 00:30:57,513

Instead of providing additional value, it's just, changing the name and and making

Cue 527 — 00:30:57,553 --> 00:31:03,033

this keyword AI native. So and now there's a lot of that say,

Cue 528 — 00:31:03,073 --> 00:31:07,393

well, the new luxury is going to be human interaction.

Cue 529 — 00:31:07,433 --> 00:31:12,873

So, my question is more about the business side of things.

Cue 530 — 00:31:12,913 --> 00:31:19,193

So how do you really deliver? I mean, how do you stay on top of on

Cue 531 — 00:31:19,233 --> 00:31:24,153

satisfying, investors in terms of, yes, they're asking for some reason for AI

Cue 532 — 00:31:24,193 --> 00:31:26,553

native. They don't really know what it means, truly.

Cue 533 — 00:31:26,593 --> 00:31:33,893

But how do you also satisfy what customers usually actually want and ask for?

Cue 534 — 00:31:34,273 --> 00:31:41,113

Yeah. Yet the magic question, I mean, look, there's a lot of money being thrown at

Cue 535 — 00:31:41,153 --> 00:31:47,953

AI right now, and, I, I agree with you that most of it that I've seen has been

Cue 536 — 00:31:47,993 --> 00:31:51,353

thrown at things that are directly under the path of the bulldozer.

Cue 537 — 00:31:51,393 --> 00:31:53,773

That is Claude Code, right? The bitter lesson.

Cue 538 — 00:31:53,813 --> 00:31:56,433

They're just going to like one guy.

Cue 539 — 00:31:56,473 --> 00:31:58,753

I knew. He was like, we're going to do personas.

Cue 540 — 00:31:58,793 --> 00:32:02,033

It's like, dude, you can't. You see that's going to be built into the models before

Cue 541 — 00:32:02,073 --> 00:32:04,333

long. Personas are not it, right?

Cue 542 — 00:32:06,273 --> 00:32:10,393

I actually only I only work with an advise companies that I think have a reasonable

Cue 543 — 00:32:10,433 --> 00:32:16,873

shot at a moat in this, in this industry, but I think that 95% of them don't.

Cue 544 — 00:32:16,913 --> 00:32:21,393

Right. And I think there's a lot of, a lot of investors are throwing their money

Cue 545 — 00:32:21,433 --> 00:32:24,753

away. I really do. And I don't want to end on a sour note or anything,

Cue 546 — 00:32:24,793 --> 00:32:28,673

but I think it's important to inject a dash of reality here.

Cue 547 — 00:32:28,713 --> 00:32:32,553

The advice that I give everybody is move slow, go at human speeds,

Cue 548 — 00:32:32,593 --> 00:32:36,313

right? What we're finding across the board is that people, as you say,

Cue 549 — 00:32:36,353 --> 00:32:41,273

are not finding the value. They're not recouping the outcomes that they need from

Cue 550 — 00:32:41,313 --> 00:32:44,593

AI. Not not at not at enterprises.

Cue 551 — 00:32:44,633 --> 00:32:47,633

And it's going to be, you know, quite some time.

Cue 552 — 00:32:47,673 --> 00:32:50,213

It's going to be quite a journey before they are.

Cue 553 — 00:32:50,273 --> 00:32:52,033

And they're not in a buying mood right now.

Cue 554 — 00:32:52,073 --> 00:32:55,513

So it's a really tough time for small SaaS vendors, right.

Cue 555 — 00:32:55,553 --> 00:32:56,633

Everything's changing overnight.

Cue 556 — 00:32:56,673 --> 00:33:00,873

There's no IDE right now. There's no sort of uniform set of tools.

Cue 557 — 00:33:00,913 --> 00:33:05,353

There's no LAMP stack. There's no nothing is emerged out of this Precambrian

Cue 558 — 00:33:05,393 --> 00:33:10,873

explosion. Right. So the real question is, you know, how do you just ignore all that

Cue 559 — 00:33:10,913 --> 00:33:13,853

noise and build something that people love?

Cue 560 — 00:33:14,673 --> 00:33:17,053

Because that's the only thing that will work.

Cue 561 — 00:33:17,193 --> 00:33:19,513

You can't even build something that you can sell anymore.

Cue 562 — 00:33:19,553 --> 00:33:21,793

It has to, like, objectively have good value, right?

Cue 563 — 00:33:21,833 --> 00:33:23,613

Because your agents will figure it out.

Cue 564 — 00:33:24,033 --> 00:33:27,953

And the only answer that I've ever known to that is build something for yourself.

Cue 565 — 00:33:27,993 --> 00:33:32,013

And that's why I am working on a video game, because that's all I knew.

Cue 566 — 00:33:32,553 --> 00:33:35,453

Yeah, good question though. It's a tough, tough spot.

Cue 567 — 00:33:39,153 --> 00:33:43,853

You wrote very compellingly in February about the AI vampire.

Cue 568 — 00:33:44,033 --> 00:33:47,333

Where are we on the AI vampire problem today?

Cue 569 — 00:33:48,713 --> 00:33:53,953

So Bezos recently said that he could see I think he said three day workweeks.

Cue 570 — 00:33:53,993 --> 00:33:55,533

Was it I forget,

Cue 571 — 00:33:57,393 --> 00:33:58,393

that was that was a prediction.

Cue 572 — 00:33:58,433 --> 00:34:02,253

I made an AI vampire. Maybe it was two day work weeks, I forget,

Cue 573 — 00:34:02,433 --> 00:34:06,133

I did notice. I mean, the basic who's read Thinking, Fast and Slow.

Cue 574 — 00:34:06,673 --> 00:34:10,913

Oh my God. A lot of you, the rest of you go read Thinking, Fast and Slow.

Cue 575 — 00:34:10,953 --> 00:34:14,233

It's a little bit of a heavy read, but it is truly one of the most influential books

Cue 576 — 00:34:14,273 --> 00:34:17,953

I've ever read. I mean, it just reframed the way I think about the world,

Cue 577 — 00:34:17,993 --> 00:34:21,013

and it's super useful framing for AI's as well.

Cue 578 — 00:34:21,113 --> 00:34:25,473

And fundamentally you have two you have basically two systems in your brain.

Cue 579 — 00:34:25,513 --> 00:34:29,353

You have one that's really fast and reactive, and you have one that takes a lot of

Cue 580 — 00:34:29,393 --> 00:34:32,913

energy, and it's really slow and good at thinking about hard problems.

Cue 581 — 00:34:32,953 --> 00:34:38,113

And the problem with using AI is that all that fast, easy stuff gets done by the

Cue 582 — 00:34:38,153 --> 00:34:41,013

model, leaving you with only the hard stuff.

Cue 583 — 00:34:41,233 --> 00:34:45,233

If I'm cycling through my 18 direct reports, I guarantee you almost every one of

Cue 584 — 00:34:45,273 --> 00:34:48,793

them on every turn has only hard shit for me left.

Cue 585 — 00:34:48,833 --> 00:34:52,033

They did all the easy stuff. It's exhausting.

Cue 586 — 00:34:52,073 --> 00:34:55,473

And so what I found was I was like, literally just like filling my context window

Cue 587 — 00:34:55,513 --> 00:34:59,193

and going, bam, like like I'd fall asleep on the floor.

Cue 588 — 00:34:59,233 --> 00:35:00,153

And I was like, what's going on?

Cue 589 — 00:35:00,193 --> 00:35:02,993

I talked to some friends and they were like nap pods in their office.

Cue 590 — 00:35:03,033 --> 00:35:05,833

And when I stopped using AI for a while, that goes away.

Cue 591 — 00:35:05,873 --> 00:35:09,313

So I really, I really do think that there's a, there's an extract,

Cue 592 — 00:35:09,353 --> 00:35:10,713

a draining effect that happens.

Cue 593 — 00:35:10,753 --> 00:35:14,673

And so, so you got to be real careful because if you're CTO super AI-pilled and like

Cue 594 — 00:35:14,713 --> 00:35:17,353

everyone's going to move ten times faster now, which is really stupid,

Cue 595 — 00:35:17,393 --> 00:35:20,753

by the way, because you won't your engineers will move ten times faster and then

Cue 596 — 00:35:20,793 --> 00:35:23,353

they'll run over your business, customers and everything else.

Cue 597 — 00:35:23,393 --> 00:35:27,113

Right? But if you're if you're if you're CTOs doing that, you know,

Cue 598 — 00:35:27,153 --> 00:35:30,213

you it's a good idea to push back because,

Cue 599 — 00:35:31,673 --> 00:35:34,033

companies are not actually seeing the outcomes.

Cue 600 — 00:35:34,073 --> 00:35:36,873

I mean, I'm spending what did I say, ten, 12,000 a month?

Cue 601 — 00:35:36,913 --> 00:35:39,993

And I've got players paying me maybe 2000 a month right now.

Cue 602 — 00:35:40,033 --> 00:35:43,633

I mean, I'm going to launch it next year and eventually tip that.

Cue 603 — 00:35:43,673 --> 00:35:47,013

But I mean, like, what I'm experiencing is like a lot of.

Cue 604 — 00:35:48,713 --> 00:35:51,673

A lot of nothing. A lot of nothing, you understand.

Cue 605 — 00:35:51,713 --> 00:35:54,273

Like, I mean, like you just look at a lot of stuff I've accomplished this year,

Cue 606 — 00:35:54,313 --> 00:35:56,753

and I just spent money. And that's where a lot of people are.

Cue 607 — 00:35:56,793 --> 00:35:58,393

They're just tinkering with factories.

Cue 608 — 00:35:58,433 --> 00:36:02,153

Right? So look, I see next year is busting all this wide open.

Cue 609 — 00:36:02,193 --> 00:36:05,513

It'll be way better. The models will be smarter, they'll be cheaper,

Cue 610 — 00:36:05,553 --> 00:36:06,953

the tools will be more ubiquitous.

Cue 611 — 00:36:06,993 --> 00:36:08,473

Will have figured some of this stuff out.

Cue 612 — 00:36:08,513 --> 00:36:10,113

Right. It's real Wild West right now.

Cue 613 — 00:36:10,153 --> 00:36:16,473

But but yeah, I mean, like, realistically, like, if your CTO is super AI-pilled and

Cue 614 — 00:36:16,513 --> 00:36:19,873

trying to make you run faster, like, send them, send them to me.

Cue 615 — 00:36:19,913 --> 00:36:22,253

Right. And I'll and I'll tell them to slow down.

Cue 616 — 00:36:24,153 --> 00:36:30,553

Yes. So you mentioned that, there's a deep integration between your game and your

Cue 617 — 00:36:30,593 --> 00:36:35,313

factory. Can you share a little bit some interesting ways that you use for

Cue 618 — 00:36:35,353 --> 00:36:38,933

validating the work and proof of correctness?

Cue 619 — 00:36:39,713 --> 00:36:45,693

Well, I mean, like, it's, I mean, I have I have players, they act as my QA,

Cue 620 — 00:36:45,793 --> 00:36:48,273

right? I'll tell you, there's an interesting phenomenon, right?

Cue 621 — 00:36:48,313 --> 00:36:50,173

I mean, like fundamentally like

Cue 622 — 00:36:51,553 --> 00:36:55,833

it's broken because I created a wish factory for them, but they are self-selecting

Cue 623 — 00:36:55,873 --> 00:36:58,273

into the end game. They're only the end game players.

Cue 624 — 00:36:58,313 --> 00:37:02,833

The hardest core, highest. My whales, the ones, the ones that a few of them are

Cue 625 — 00:37:02,873 --> 00:37:05,933

spending 1000 a month right on my game.

Cue 626 — 00:37:06,073 --> 00:37:10,193

They're the ones designing the game, and so they're not doing anything that helps me

Cue 627 — 00:37:10,233 --> 00:37:14,493

with a new player experience. The first time user experience nothing.

Cue 628 — 00:37:14,593 --> 00:37:18,873

So it's kind of useless. So RSI recursive self-improvement doesn't work for video

Cue 629 — 00:37:18,913 --> 00:37:21,113

games today. They go into little you know what I mean?

Cue 630 — 00:37:21,153 --> 00:37:22,753

Little little loops and they get stuck.

Cue 631 — 00:37:22,793 --> 00:37:25,833

There it comes. I have to come in as an outside force and be like,

Cue 632 — 00:37:25,873 --> 00:37:29,313

no, spend some time focusing on new users, that kind of thing.

Cue 633 — 00:37:29,353 --> 00:37:31,833

But the proof is always in human, human proof.

Cue 634 — 00:37:31,873 --> 00:37:33,673

Human proof is the future, I tell you.

Cue 635 — 00:37:33,713 --> 00:37:37,193

And even systems like Tessl have the human review step, right?

Cue 636 — 00:37:37,233 --> 00:37:41,100

I mean, like it's critical we become the bottleneck for everything

skills

talk-yegge-ai-powered-factories

tile.json