How to Build a Self-Improving Company with AI

0:00This is based a little bit off a talk

0:01Diana gave. There's a video up over the

0:03weekend, which is super cool. Um Jack

0:04Dorsey was tweeting some stuff like two

0:06or three weeks ago that I thought was

0:08super cool. And I've kind of um stolen a

0:10bunch of those ideas and shoved them in

0:13here. This talk is like pretty

0:14conceptual and high-level about thinking

0:17about how to build companies. So, the

0:19Roman legions were designed to

0:23project power over two continents or

0:27something from Rome at the center to

0:29like these people on Hadrian's Wall up

0:31in Scotland. And the idea was um this

0:35nested hierarchies with consistent spans

0:37of control. And you had like named

0:39individuals with spans of control to

0:41pass orders down and send information

0:44back up the hierarchy. And if you think

0:46about most companies today, they are

0:48organized like a Roman legion, where

0:50human beings are the conduit for

0:52information flowing up and down. And so,

0:54Jack Dorsey's tweet that I thought was

0:55great was this like this underlying

0:57assumption that hierarchically organized

1:00companies are the are the way that we

1:02should be organizing like our economic

1:03units of value. And I think AI basically

1:06breaks that. If you talked to people a

1:08year ago about how AI was useful,

1:11they talked about productivity. Like

1:15co-pilots making engineers 20% more

1:17productive, adding co-pilots to

1:18workflows, shipping more software.

1:21But I think that is actually a

1:22broken way of thinking about AI. That's

1:25like P had a great blog post where

1:27basically just like taking the old way

1:29of working and adding like a more

1:30powerful engine onto it. And instead of

1:32that, I think you can reimagine like

1:34what a company is and how it acts. And

1:37so, as Gary's talking, like he

1:39I genuinely believe can produce more

1:41code than an entire engineering team.

1:44The thing that's really stuck with me is

1:45this idea of like extracting the domain

1:48knowledge from your company and defining

1:50it as a as like context or a set of

1:52skills or whatever you want to call it.

1:54But like this idea that there's domain

1:56knowledge or business knowledge or like

1:58some know-how that's inside the heads of

2:01people and in Slack messages and in

2:04emails and in Notion, all of this like

2:06information together defines how your

2:09company works.

2:11And if you can make that legible,

2:13you suddenly can

2:15can move from this hierarchical

2:16organization to a sort of intelligent

2:19AI-powered organization

2:21with AI-native software. AI isn't the

2:23something It's not something you bolt

2:25onto the side of a company. It's not

2:26like a tool you give to engineers to

2:28make them more productive. But I think

2:30you can reimagine what a company is as a

2:32set of recursive self-improving AI

2:35loops. I think this is really, really,

2:37really important because when it gets

2:39there, I think the company starts to

2:42self-improve

2:43even when you're sleeping.

2:45So let me give you an example. Diane's

2:46talk talks about this as well. This AI

2:48loop, you start with like a sensor layer

2:50which is like That's a fancy word, but

2:52really it might be like

2:54emails from your customers. Might be

2:56support tickets, code changes, people

2:59canceling their subscription, uh product

3:02telemetry. It's like sensor data to get

3:05information from the outside world. And

3:07then a a policy layer, decision layer,

3:09like rules about what you can do, what

3:11it has to ask a human permission for,

3:13what it must log. A tool layer that's

3:15kind of Gary's skills and code, like the

3:18tool layer is Gary's code. It's

3:20basically deterministic APIs, things

3:21like query my database or look at my

3:24calendar. Um a set of tools that the the

3:28AI can call. A quality gate, like that

3:31might be eval's deterministic checks,

3:32safety filters, human review for

3:34high-risk stuff. And then a learning

3:37mechanism. It's like your system

3:39interacts with the real world, picks up

3:41where it doesn't work and loops back

3:42into the top again. And if you can run

3:44every single step of that without human

3:46intervention without with minimal human

3:48intervention, your system gets better

3:50and better and better

3:51while you're are

3:52And I can give you actual examples of

3:54this that are live right now. We started

3:56with an agent that you can ask and if it

3:58has deterministic tools to query our

4:00database. Pretty simple, like when did I

4:02last have office hours with this

4:03company? Then it got a little bit

4:05smarter, which was like, for this

4:07company I'm doing office hours with

4:08right now, they need introductions for

4:10anyone in petrochemicals or something.

4:12And it could query the database in

4:13different ways and use rag and all sorts

4:15of stuff to like come up with five

4:16relevant founders for you to meet. But

4:18again, this is like this is a sidekick,

4:20right? This is an agent. This is like

4:21the old This is last year's version of

4:23how a how AI is making me better as a

4:26group partner. It's making me 20 or 30%

4:27more effective.

4:29The aha moment for me came when we put a

4:32monitoring agent on top of that, which

4:34looked at every single query every

4:36single YC employee was doing and saw

4:39when it worked

4:41and when it did not work.

4:42And when it did not work, it's like, oh,

4:44why not? What would have made this query

4:46work? Do we need different deterministic

4:48tools? Do we need to update the skills

4:50file? Do we need a different database

4:51for you? Do we need a new index? And

4:53this happened This literally happens

4:54overnight now. Let's write the code, put

4:57in a merge request to the YC code base,

4:59have an agent review it, and merge it,

5:00and deploy it. So, when a human comes

5:02the next day to ask the same query, it

5:05will now succeed. For me, that was like

5:07the

5:08holy [ __ ] [ __ ] But that's not just

5:10AI making you 20 or 30% more valuable,

5:12it is the AI going through this loop to

5:15figure out how to self-improve. And I

5:17think basically, if you can identify

5:19parts of your company that work like

5:21this and eliminate as much of have the

5:24human in kind of a monitoring or

5:25supervisory capacity,

5:28you can just throw tokens at this

5:30problem and your company will get

5:31better.

5:32And so, other examples might be, if you

5:34have product analytics, having an agent

5:36go through your product analytics to to

5:38figure out what part of your sales

5:39funnel is presenting the highest amount

5:42of friction, researching best practices,

5:44putting in place an AB test, running it

5:45for a week, picking the best version,

5:47and deploying it. Then doing that again

5:49and again and again for your product. So

5:50you have a self-optimizing like product

5:52loop. Or you do it with customer service

5:54queries. You have customer

5:56suggestions coming in and in and in. You

5:58triage it with a kind of You have to

6:00have an agent which is like your chief

6:01product officer and your chief

6:02technology officer who make kind of

6:04judgment calls about Okay, this is a

6:06suggestion which we just don't want to

6:07do. We'll discard it. But no, this is a

6:08suggestion which is now in line with our

6:10road map. Um we can do it overnight.

6:12Let's write the code. Let's deploy it.

6:14Let's ship it to the customer without a

6:15human being involved.

6:17So I think if you can think about each

6:19part of your company as a self-improving

6:22like recursive AI loop, it becomes very

6:24very different to this like

6:24hierarchically organized Roman legion of

6:26a company. So what? So like if you want

6:28to do this, what are the implications?

6:29One is like burn tokens, not head count.

6:32We are seeing companies get to demo day

6:34with about 5x more revenue per employee

6:37than they did 18 months ago. And I think

6:40that's going to continue to series A and

6:42series B.

6:43And so I think you're going to be

6:45constrained on token usage, not on head

6:47count, really really soon. The blunt

6:49measure now is just like measuring

6:50everyone's token usage, which is

6:51obviously like dumb and gameable at the

6:55extreme, but directionally I think is

6:58correct. We're in the phase of like what

7:00is possible right now. And so everyone

7:03should be experimenting to the max to

7:05figure out what we can even do with this

7:07crazy new intelligence we have. As soon

7:09as you turn it into a leaderboard and

7:10people get promoted or fired based on

7:11it, obviously it gets gamed. Obviously

7:13that's dumb. But I think directionally

7:15figuring out who in the organization is

7:17token maxing, who is not, is like a good

7:20way to think about which employees you

7:22should be spending your time with. I

7:23think middle management is done. I just

7:25don't think you need middle management

7:26for this coordination problem. I think

7:27AI should be doing it. And for me, there

7:29are two roles. Jack Dorsey has three. I

7:31actually don't like the third one, so I

7:33deleted it. But there are two roles that

7:34really, really matter for me. I think

7:36everyone just has to be an IC now, a

7:37builder, an operator. And I think

7:40crucially having directly responsible

7:42individuals to get anything done, I

7:45think you need a named human. Not a

7:46committee, not a group of people, just a

7:47single person.

7:48And I think you can build companies

7:50based on ICs effectively. I I think just

7:53middle management is is over. So,

7:55building the self-improving company,

7:56that's the dream. And by the way, I

7:58think like people are

7:59at the bleeding edge of this right now.

8:01I'd be interested to see where you all

8:02are, but it feels like people are like

8:04exploring the boundaries here. I'm not

8:06sure anyone has a truly self-improving

8:08company in every function.

8:10I might be wrong. You might prove me

8:12wrong. What would I do? First of all,

8:13this is really, really important. I

8:14would make the entire organization

8:16legible to AI. What does that mean? It

8:19means you've got to record everything.

8:21Um simplistically, all of our um partner

8:25emails, now if you email a YC partner,

8:27that email is in the YC database. Every

8:30Slack message, every DM, every office

8:32hour we've started recording for the

8:33last three or four months. Every single

8:35thing that happens,

8:37if it is recorded, it happened to the

8:39AI. If it did not get recorded, it is it

8:41did not happen to your intelligence. You

8:43know what I mean? And so, I was talking

8:45with some founders over here um just

8:47now, and we're having like really good

8:48conversations about their company. But I

8:50every conversation I had, I was like,

8:52"Fuck, I need to be recording this

8:53conversation." Because some guy wanted

8:56an introduction to I can't even remember

8:58who the introduction was now. Uh who was

9:00that?

9:01I was talking to someone about and I

9:02promised you an introduction. I said,

9:03"Yes." And I said, "Email me afterwards

9:05cuz I would I I'm going to forget this.

9:07I'm going to talk to 20 people." Yeah,

9:08so it needs to be on my phone or a or or

9:10smart glasses. Or we deck out every room

9:12with like microphones. But basically,

9:15everything needs to be recorded so that

9:16it can be legible to the AI. And then as

9:17Garry talked about like diarization, you

9:20cannot pump in

9:21100,000 hours worth of recordings into

9:23context window. So, you have to diarize

9:26it. You have to basically aggregate it

9:27down, synthesize it into the important

9:29parts, and then give the AI breadcrumbs.

9:31So, like okay, so here's an example.

9:33Who's read the user manual, the YC user

9:35manual? Hopefully, everyone in this room

9:37has at least opened the user manual at

9:38one point in time, right? Like

9:40it's fine. It was written 5 to 10 years

9:42ago, most of it. It's kind of out of

9:44date.

9:45So, Harj thought uh last weekend, since

9:47now we've got about 2,000 hours of

9:49recorded office hours from the last 3

9:50months, why don't we regenerate the user

9:52manual?

9:53And so, you can click like you give it a

9:55set of instructions, you basically

9:56diarize it down, synthesis like

9:59categorize it into certain areas like

10:00fundraising, hiring, co-founder

10:02disputes, whatever,

10:04and then write me a new user manual.

10:06And by the end of the weekend, he had a

10:07150 page user manual, which is

10:09dramatically better than the existing

10:11user manual. And now we can also update

10:13it every single month. So, our user

10:15manual becomes self-improving. Every new

10:17piece of advice we give, it's compared

10:20with the existing user manual and either

10:21incorporated or thrown away. So, the

10:23user manual becomes this up-to-date

10:25living brain of the advice we give to

10:26founders.

10:28And obviously, it doesn't stop as a user

10:29manual, you then pump it in as context

10:31to an AI agent, and suddenly you can ask

10:33a superintelligent AI

10:34and get the combined wisdom of 16 YC

10:36partners in one.

10:39But only if it's legible. So, you have

10:41to record everything. The second point

10:43is kind of the same, right? Like if it

10:44creates an artifact that can

10:45self-improve, it's legible. If it

10:47doesn't, you throw it away. The third

10:49point then is that every function can

10:52generate This used to say dashboards.

10:54It's not just dashboards, it's on-demand

10:55software. Codex 55 is now good enough

10:57you can one-shot most simple like

11:00most internal software dashboards you

11:02can one-shot to a pretty high level of

11:04quality. I tried it over the weekend on

11:06a bunch of our stuff. It's just unreal.

11:09So, all of your internal operations

11:10teams should be sitting on this layer of

11:13like kind of intelligence understanding,

11:15and then creating their own dashboards

11:17and their own workflows. And I would see

11:20that those as

11:21entirely disposable. I would very

11:24preciously store all the data. So, as

11:26Garry said, he puts it all all of his

11:28emails in markdown. Never throw anything

11:29away,

11:30but then treat these the software as

11:33ephemeral. You can you can generate it.

11:35You can regenerate it. The valuable part

11:37is like the comprehension inside

11:39people's heads of like this is how the

11:40function works. This is how we run a YC

11:43event, whatever. The software to

11:44actually run the event you can generate

11:46for the event. You can throw it away.

11:47The the models get smarter in a month or

11:49two. Throw the software away. Give it

11:52your original set of instructions and

11:53regenerate the software.

11:54So I think the business context and and

11:57skills are the valuable part. I think

11:59the software on top of it is ephemeral.

12:01So what what are humans for in this

12:03world? I think basically we're talking

12:06about a company brain. And I know a

12:07bunch of people in this room are

12:08building this. But the bit in the

12:10middle, like all of your data, all of

12:12your emails, your DMs, the skills, the

12:14know-how, that is like the company

12:17brain. And I think the humans sit around

12:19the edge of this interfacing with the

12:20real world.

12:22So it's where this intelligence makes

12:24contact with reality.

12:26Human beings reach into places the

12:27models can't go yet. That might be like

12:31a conference. It might be a I'm trying

12:33to think of examples. I I would say a

12:34phone call, but I think the AI can reach

12:35into phone calls pretty easily now.

12:37Um I think it's like novel situations,

12:39ethical considerations, high-stakes

12:41moments. You know, it's like it's where

12:42the founder comes to us

12:45and is like thinking about breaking up

12:47with their co-founder. Right? It's like

12:49those real high-stakes, high-emotion

12:51moments where you really want a human

12:52being. I think that's where the human

12:55fits. For all of you, like sales

12:57conversations. I think that's a human

12:58being in the room for the next 20 years.

13:01So the humans live I think around the

13:02edge.

13:03And I'm over time and Kulveer should

13:05bullhorn me. I will leave you this one

13:07question. If you were building your

13:09company today,

13:11would you start it in this shape?

13:14For most of you, you're small enough to

13:16build it right. And so I don't think you

13:17have any excuse. And I know there are a

13:19few of you who are in the process of

13:22ripping up and rebuilding your company.

13:23So with that I will stop and will hand

13:26over to Pete. Thank you for listening.