{
  "video_id": "X_JsIHUfUjc",
  "channel_slug": "ycombinator",
  "channel_handle": "Y Combinator",
  "title": "How to Build a Self-Improving Company with AI",
  "duration_seconds": 808,
  "url": "https://www.youtube.com/watch?v=X_JsIHUfUjc",
  "upload_date": "20260521",
  "transcript": "This is based a little bit off a talk\nDiana gave. There's a video up over the\nweekend which is super cool. Um Jack\nDorsey was tweeting some stuff like two\nor three weeks ago that I thought was\nsuper cool and I've kind of um stolen a\nbunch of those ideas and shove them into\nhere. This talk is like pretty\nconceptual and high level about thinking\nabout how to build companies. So the\nRoman legions were designed to\nproject power over two continents or\nsomething from Rome at the center to\nlike these people on Hadron's wall up in\nScotland. And the idea was um this\nnested hierarchies with consistent spans\nof control and you had like named\nindividual with spans of control to pass\norders down and send information back up\nthe hierarchy. And if you think about\nmost companies today, they are organized\nlike a Roman legion where human beings\nare the conduit for information flowing\nup and down. And so Jack Dorsey's tweet\nwhich I thought was great was it's like\nthis underlying assumption that\nhierarchically organized companies are\nthe are the way that we should be\norganizing like our economic units of\nvalue. And I think AI basically breaks\nthat. If you talk to people a year ago\nabout how AI was useful, they talked\nabout productivity, like co-pilots,\nmaking engineers 20% more productive,\nadding co-pilots to workflows, shipping\nmore software. But I think that is\nactually a broken way of thinking about\nAI. That's like Pete had a great blog\npost. We're basically just like taking\nthe old way of working and adding like a\nmore powerful engine onto it. And\ninstead of that, I think you can\nreimagine like what a company is and how\nit acts. And so as Gary's talking like\nhe I genuinely believe can produce more\ncode than an entire engineering team.\nThe thing that's really stuck with me is\nthis idea of like extracting the domain\nknowledge from your company and defining\nit as a as like context or a set of\nskills or whatever you want to call it.\nBut like this idea that there's domain\nknowledge or business knowledge or like\nsome knowhow that's inside the heads of\npeople and in Slack messages and in\nemails and in notion. All of this like\ninformation together defines how your\ncompany works. And if you can make that\nlegible, you suddenly can can move from\nthis hierarchal organization to a sort\nof intelligent AI powered organization\nwith AI native software. AI isn't the\nsome it's not something you bolt onto\nthe side of a company. It's not like a\ntool you give to your engineers to make\nthem more productive. But I think you\ncan reimagine what a company is as a set\nof recursive self-improving AI loops. I\nthink this is really, really, really\nimportant because when it gets there, I\nthink the company starts to self-improve\neven when you're sleeping. So, let me\ngive you an example. Diana's talks about\nthis as well. this AI loop. You start\nwith like a sensor layer, which is like\nthat's a fancy word, but really it might\nbe like emails from your customers.\nMight be support tickets, code changes,\npeople canceling their subscription,\nproduct telemetry. It's like sensor data\nto get information from the outside\nworld. And then a a policy layer,\ndecision layer, like rules about what\nyou can do, what it has to ask a human\npermission for, what it must log. A tool\nlayer, that's kind of Gary's skills and\ncode. Like the tool layer is Gary's\ncode. It's basically deterministic APIs,\nthings like query my database or look at\nmy calendar. Um, a set of tools that the\nthe AI can call a quality gate like that\nmight be evalistic checks, safety\nfilters, human review for high-risk\nstuff. and then a learning mechanism.\nIt's like your system interacts with the\nreal world, picks up where it doesn't\nwork, and loops back into the top again.\nAnd if you can run every single step of\nthat without human intervention, without\nwith minimal human intervention, your\nsystem gets better and better and better\nwhile you're sleeping. And I can give\nyou actual examples of this that are\nlive right now. We started with an agent\nthat you can ask and it it has\ndeterministic tools to query our\ndatabase. Pretty simple, like when did I\nlast have office hours with this\ncompany? Then it got a little bit\nsmarter which was like for this company\nI'm doing offices hours with right now\nthey need introductions for anyone in\npetrochemicals or something and it could\nquery the database in different ways and\nuse rag and all sorts of stuff to like\ncome up with five relevant founders for\nyou to meet. But again this is like this\nis a sidekick right this is an agent\nthis is like the old this is last year's\nversion of how AI is making me better as\na group partner. It's making me 20 or\n30% more effective. The aha moment for\nme came when we put a monitoring agent\non top of that which looked at every\nsingle query every single YC employee\nwas doing and saw when it worked and\nwhen it did not work and when it did not\nwork it's like oh why not what would\nhave made this query work do we need\ndifferent deterministic tools do we need\nto update the skills file do we need a\ndifferent database view do we need a new\nindex and this happen this literally\nhappens overnight now let's write the\ncode put in a merge request to the YC\ncodebase have an agent review it and\nmerge it and deploy it. So when a human\ncomes the next day to ask the same\nquery, it will now succeed. For me, that\nwas like the holy [ __ ] [ __ ] right?\nThat's not just AI making you 20 or 30%\nmore valuable. It is the AI going\nthrough this loop to figure out how to\nself-improve. And I think basically if\nyou can identify parts of your company\nthat work like this and eliminate as\nhave the human and kind of a monitoring\nof supervisory capacity,\nyou can just throw tokens at this\nproblem and your company will get\nbetter. And so other examples might be\nif you have product analytics, having an\nagent go through your product analytics\nto to figure out what part of your sales\nfunnel is presenting the highest amount\nof friction, researching best practices,\nputting in place an AB test, running it\nfor a week, picking the best version,\nand deploying it. Then doing that again\nand again and again for your product.\nJust have a self-optimizing like product\nloop. Or you do it with customer service\nqueries. You have customer suggestions\ncoming in and in and in. you triage it\nwith a kind of you have to have an agent\nwhich is like your chief product officer\nand your chief technology officer who\nmake kind of judgment calls about okay\nthis is a suggestion we just don't want\nto do we'll discard it but no this is a\nsuggestion which is now in line with our\nroad map um we can do it overnight let's\nwrite the code let's deploy it let's\nship it to the customer without a human\nbeing involved so I think if you can\nthink about each part of your company as\na self-improving like recursive AI loop\nit becomes very very different to this\nlike hierarchically organized Roman\nlegion from a company so what So like if\nyou want to do this, what are the\nimplications? One is like burn tokens,\nnot headcount. We are seeing companies\nget to demo day with about 5x more\nrevenue per employee than they did 18\nmonths ago. And I think that's going to\ncontinue to series A and series B. And\nso I think you're going to be\nconstrained on token usage, not on\nheadcount really, really soon. The blunt\nmeasure now is just like measuring\neveryone's token usage, which is\nobviously like dumb and gameable at the\nextreme, but directionally I think is\ncorrect. We're in the phase of like what\nis possible right now and so everyone\nshould be experimenting to the max to\nfigure out what we can even do with this\ncrazy new intelligence we have. As soon\nas you turn it into a leaderboard and\npeople get promoted or fired based on\nit, obviously it gets gamed, obviously\nthat's dumb. But I think directionally\nfiguring out who in your organization is\ntoken maxing, who is not is like a good\nway to think about which employees you\nshould be spending your time with. I\nthink middle management is done. I just\ndon't think you need middle management\nfor this coordination problem. I think\nAI should be doing it. And for me, there\nare two roles. Jack Dorsey has three. I\nactually don't like the third one, so I\ndeleted it. But there are two roles that\nreally, really matter for me. I think\neveryone just has to be an IC now, a\nbuilder, an operator. And I think\ncrucially having directly responsible\nindividuals to get anything done I think\nyou need a named human not a committee\nnot a group of people just a single\nperson and I think you can build\ncompanies based on IC's effectively I\nthink just middle management is is over\nso building this self-improving company\nthat's a dream and by the way I think\nlike people are at the bleeding edge of\nthis right now I'd be interested to see\nwhere you all are but it feels like\npeople are like exploring the boundaries\nhere I'm not sure anyone has a truly\nself-improving company in every\nfunction. I might be wrong. You might\nprove me wrong. What would I do? First\nof all, this is really, really\nimportant. I would make the entire\norganization legible to AI. What does\nthat mean? It means you've got to record\neverything.\nSimplistically, all of our um partner\nemails. Now, if you email a YC partner,\nthat email is in the YC database. Every\nSlack message, every DM, every office\nhour we've started recording for the\nlast three or four months. every single\nthing that happens, if it is recorded,\nit happened to the AI. If it did not get\nrecorded, it is it did not happen to\nyour intelligence. You know what I mean?\nAnd so, I was talking with some founders\nover here um just now and we're having\nlike really good conversations about\ntheir company, but every conversation I\nhad, I was like, \"Fuck, I need to be\nrecording this conversation.\" Because\nsome guy wanted an introduction to I\ncan't even remember who the introduction\nwas now. Who was that? I was talking to\nsomeone about and I promise you an\nintroduction. said yes. And I said,\n\"Email me afterwards cuz I would I'm\ngoing to forget this. I'm going to talk\nto 20 people.\" Yeah. So, it needs to be\non my phone or a clip or or smart\nglasses or we deck out every room with\nlike microphones. But basically,\neverything needs to be recorded so that\nit can be legible to the AI. And then,\nas Gary talked about like diorization,\nyou cannot pump in 100,000 hours worth\nof recordings into a context window. So,\nyou have to diorize it. You have to\nbasically aggregate it down, synthesize\nit into the important parts, and then\ngive the AI breadcrumbs. It's like,\nokay, so here's an example. Who's read\nthe user manual? The YC user manual.\nHopefully, everyone in this room has at\nleast opened the user manual at one\npoint in time, right? Like, it's fine.\nIt was written 5 to 10 years ago, most\nof it. It's kind of out of date. So, Haj\nthought uh last weekend, since now we've\ngot about 2,000 hours of recorded office\nhours in the last 3 months, why don't we\nregenerate the user manual? And so you\ncan click like you give it a set of\ninstructions. You basically diorize it\ndown, synthes like categorize it into\ncertain areas like fundraising, hiring,\nco-founder disputes, whatever. And then\nwrite me a new user manual. And by the\nend of the weekend, he had 150 page user\nmanual, which is dramatically better\nthan the existing user manual. And now\nwe can also update it every single\nmonth. So our user manual becomes\nself-improving. Every new piece of\nadvice we give, it's compared with the\nexisting user manual and either\nincorporated or thrown away. So the user\nmanual becomes this up-to-date living\nbrain of the advice we give to founders.\nAnd obviously it doesn't stop as a user\nmanual. You then pump it in as context\nto an AI agent and suddenly you can ask\na super intelligent AI and get the\ncombined wisdom of 16 YC partners in\none,\nbut only if it's legible. So you have to\nrecord everything. The second point is\nkind of the same, right? Like if it\ncreates an artifact that can\nself-improve, it's legible. If it\ndoesn't, you throw it away. The third\npoint then is that every function can\ngenerate this used to say dashboards.\nIt's not just dashboards. It's on demand\nsoftware. Codeex 55 is now good enough.\nYou can oneshot most simple inter like\nmost internal software dashboards you\ncan oneshot to a pretty high level of\nquality. I tried it over the weekend on\na bunch of our stuff. It's just unreal.\nSo all of your internal operations teams\nshould be sitting on this layer of like\nkind of intelligence understanding and\nthen creating their own dashboards and\ntheir own workflows. And I would see\nthat those as entirely disposable. I\nwould very preciously store all the\ndata. So as Gary said, he puts it all\nall of his emails in markdown. Never\nthrow anything away, but then treat the\nthe software as ephemeral. You can you\ncan generate it, you can regenerate it.\nThe valuable part is like the\ncomprehension inside people's heads of\nlike this is how the function works.\nThis is how we run a YC event. Whatever\nthe software to actually run the event,\nyou can generate for the event. You can\nthrow it away. The mo the models get\nsmarter in a month or two. Throw the\nsoftware away. Give it your original set\nof instructions and regenerate the\nsoftware. So I think the business\ncontext and and skills are the valuable\npart. I think the software on top of it\nis ephemeral. So what what are humans\nfor in this world? I think basically\nwe're talking about a company brain and\nI know a bunch of people in this room\nare building this but the bit in the\nmiddle like all of your data, all of\nyour emails, your DMs, the skills, the\nknowhow that is like the company brain\nand I think the humans sit around the\nedge of this interfacing with the real\nworld. So it's where this intelligence\nmakes contact with reality. Human beings\nreach into places the models can't go\nyet. That might be like a conference. It\nmight be a I'm trying to think of\nexamples. I would say a phone call, but\nI think the AI can reach into phone\ncalls pretty easily now. Um I think it's\nlike novel situations, ethical\nconsiderations, high stakes moments, you\nknow, it's like it's where the founder\ncomes to us and is like thinking about\nbreaking up with their co-founder,\nright? It's like those real high stakes,\nhigh emotion moments where you really\nwant a human being. I think that's where\nthe human fits for all of you like sales\nconversations. I think that's a human\nbeing in the room for the next 20 years.\nSo the humans live I think around the\nedge and I'm over time and cool vision\nshould bullhorn me. I will leave you\nthis one question. If you were building\nyour company today would you start it in\nthis shape for most of you you're small\nenough to build it right and so I don't\nthink you have any excuse and I know\nthere are a few of you who are in the\nprocess of ripping up and rebuilding\nyour company. So with that I will stop\num and we'll hand over to Pete. Thank\nyou for listening.",
  "transcript_chars": 14937,
  "ingested_at": "2026-05-21T19:00:09.848225+00:00",
  "source": "retry-no-transcript",
  "yt_meta": {
    "view_count": 5640,
    "like_count": 302,
    "channel_id": "UCcefcZRL2oaA_uBNeo5UOWg"
  }
}