Translated from the 2026-08-17 version of the original
Sam Altman x Cory Levy at InternapaloozaFrom three months at YC to not outsourcing judgment · Thematic reading notes
Scope of these notes: this is not a complete summary of the interview; it includes only the parts where this reading session actually paused, confirmed repeatedly and connected to real work
- What AI lowers is the cost of experiments
- Do not rush to scale before validating
- Answers can be sought; judgment cannot be outsourced
- Explore when uncertain, concentrate once convinced
- Enter a high-collision environment
- Scale, emergence and generalization
- Spotting people is a practice game
Entry point｜What a relaxed conversation left behind
The original interview is not a tightly structured startup course, but its scattered answers got us interested in Sam, his two mentors and his YC experience, and from there we distilled a set of judgments that transfer to real work
Sam went from a founder in YC's first batch to the leader of YC, and spent years with two mentors of sharply different styles, Paul Graham and Peter Thiel. These backgrounds are not a biography these notes set out to tell in full; they are an entry point for understanding how he thinks about feedback, non-consensus, talent and the timing of bets
What this reading session really cares about is not which startup answers Sam gave, but how he brings his judgment into contact with more of reality, how he keeps room to explore when information is scarce, and how he takes on the trade-offs once conviction forms
1. From three months at YC to 17 minutes with Codex
Cory asks which pieces of the advice YC used to give founders have changed because of AI. Sam answers with an extreme contrast about the production cycle, but what is really worth pursuing is how the feedback loop changes with it
Sam Altman: The big thing that has changed is what we expected people to do in three months at YC with great effort and five people is now like 17 minutes and Codex.
If you can make a whole startup in 17 minutes, the rate at which you can test ideas and make small incremental progress and get feedback is just very different.
There’s a very different feedback loop people should get going.
What YC once expected a five-person team to produce in three months of intense work can now be done with Codex in a very short time. What has really changed is not just that development is faster: the frequency of testing ideas, making small changes and getting feedback can all rise sharply, and a startup needs to deliberately build a feedback loop suited to the new speed of production
AI has lowered the production threshold for both startups and creative work at once: one person can write code, build pages, generate content and ship versions faster. But the amount produced is not itself learning. Only when a new version touches the real world, returns new information and changes the next action does the feedback loop actually close
2. Find something worth scaling first, then scale it
When an audience member asks how to market a new company today, Sam does not start with a growth system. He starts with the first few hundred users, selling in person and the early work that does not scale
Sam Altman: The easiest way to market something is to make something that’s so good or so new that people spontaneously tell all their friends and it becomes a ChatGPT moment.
I don’t think you can make that a strategy you depend on.
What worked the most often is a very hands-on, almost all sales-driven, very little marketing-driven approach to getting the first few hundred users, and then listening to them about what they like about it, how they talk about it, and building a marketing thing around that.
This is the YC advice: do things that don’t scale.
A product so good or so new that users spontaneously tell their friends is the easiest way to spread, but such a ChatGPT moment is extremely rare and cannot be a strategy you rely on. The more reliable method is to win and serve the first few hundred users yourself, listen to what they like and how they describe the product, and then shape later marketing from that real language and behavior. That is what YC means by doing things that don't scale for now
This part connects directly to your work: as soon as you see a method that might work, you tend to quickly build processes, write SOPs, make templates, automate and expand the volume of content. Those abilities are fine in themselves; the real risk is that they fire before reality has validated anything
3. Answers can be sought; judgment cannot be outsourced
An audience member, Sakina, asks where Sam would spend his time and what problem he would solve if he were 18, 19 or 20 today. Sam offers no industry answer; he first explains why a real startup direction cannot be collected from a public conversation
Sam Altman: Whatever the right answer to that question is, it’s not going to be told to you by anyone at a chat at an event like this.
By the time it is obvious enough to say with some degree of confidence to a room of hundreds of people, you’re not going to have a unique lens on it.
When we were doing OpenAI, no one would have told us to start OpenAI. There’s a 0% chance whoever was speaking would have said you should start an AGI effort.
Don’t take startup ideas from other people.
Once a direction is clear enough that a speaker can confidently tell hundreds of people about it at a public event, the audience can hardly still have a unique lens on it. Back then, nobody at an event would have advised Sam to start OpenAI or launch an AGI effort, so a founder cannot take a direction someone else gives out publicly as their own startup answer
Sam Altman: You can have a descendant of ChatGPT watch your computer screen, watch every meeting, record every call, and have perfect context of your whole life.
You choose what information you want it to have, but you can connect it to your texts or email or docs or Slack or whatever.
And then you kind of have this thing that is not making decisions for you.
It’ll say, “Maybe here’s another idea,” or “I think you’re making a mistake here; you should consider this,” or “I can do this thing for you to help.”
The next-generation ChatGPT Sam imagines can, with the user's permission, understand context over time from the screen, meetings, calls, email, documents and Slack, and while a person writes a sales email or a strategy document it proactively adds relevant information, points out possible mistakes or takes over tasks, but it does not make the final decision for the person
4. Race several horses when uncertain, bet big once conviction forms
Cory asks when Sam put all his time into OpenAI. The answer is the end of 2018, not the launch in 2016, which leads the conversation into a counterintuitive question: must young people do only one thing from the start
Sam Altman: A lot of young people try to work on 17 projects at the same time, and then other people say, “You can’t do that. You’ve got to focus. You’ve got to go all in on one thing.”
I think that’s bad advice. I think it’s fine to work on multiple projects for a little while because you don’t really know what’s going to work.
You don’t really know what you’re going to be interested in. You don’t really know what’s going to be a good fit.
As soon as you figure out what your highest conviction thing is, that is when you’re supposed to do the painful work of getting free from the other stuff and going all in.
While you still do not know what will work, what you will stay interested in and what really fits you, young people can explore several projects in parallel for a while. But once one direction becomes the thing you have the highest conviction in, you have to do the painful trade-off of pulling out of the other projects and going all in
5. Enter an environment that keeps producing unexpected opportunities
Cory turns "what should I do" into "how do I get exposure". Sam is not sure moving to the Bay Area is still the best answer today, but he keeps a more stable principle about environments
Sam Altman: For the last decade and a half, the highest confidence piece of advice for that question is: move to the Bay Area. I’m not sure if that’s still the case.
I always have thought the internet should decentralize that somehow, but it has definitely been the case that you just want to put yourself in a high collision environment and meet a lot of people.
In theory the internet should stop talent, information and opportunity from concentrating in a few places, but reality so far still shows that putting yourself in an environment with dense talent, frequent encounters and the chance to meet many different people and do things together significantly raises the odds of unexpected connections
6. The value of scale may not be more, but different
Explaining why early OpenAI could attract talent, Sam first talks about grounded non-consensus, then about the emergent properties that appear in different systems as they scale. Together these make up his reasons for being willing to bet at the time
Sam Altman: Doing things that are unpopular, but you have some reason to understand that they’re correct.
It’s not just dumb belief. These are the highest leverage moments, and you have to take them when they come.
We understood in 2016 that deep learning was working and getting better with scale.
The truly high-leverage opportunity is something temporarily unpopular that you have specific reasons to believe may be correct. It is not stubbornly holding on because others object: in 2016 OpenAI had already observed that deep learning actually worked and kept getting better with scale
Sam Altman: Surprisingly often across many different kinds of systems in the world, there are unusually interesting emergent properties with scale.
YC is an example of this. You scale up the YC batch and there’s a strong network effect. Weird stuff happens, like good stuff.
Companies themselves have very strange returns to scale that people often underestimate. And it turns out that neural networks also have surprising properties at scale.
When many different kinds of systems grow past a certain scale, existing capabilities do not just increase proportionally; new properties appear that did not exist at small scale. A larger YC batch forms a strong network effect, companies can get unusual returns to scale, and neural networks also show surprising capabilities as they scale
During the reading we brought in generalization to understand scale. This is not a branch Sam develops in the source; it started from the question "is scaling up the same as generalizing" and led us to distinguish the scale of input, the ability to transfer and emergent properties
7. Spotting people is also a practice game that needs feedback
Sam first explains why he is willing to hand huge responsibility to young people, then why YC partners get better and better at picking founders. This is not a static talent rubric but an ability trained through many judgments and looking back at the results
Sam Altman: I am a huge believer that if the single thing you do in your career is identify and bet on young, unproven talent, you’ll be wildly successful.
This was true at YC. This is a big part of how we succeed at OpenAI.
If you do only one thing well in your career, spotting young people the market has not yet proven and giving them opportunity, resources, trust and real responsibility while others are still unsure, you can still earn enormous long-term returns. This was YC's experience and it is part of how OpenAI succeeds
Sam Altman: One of the surprising things to me about running YC was just how much of a numbers game that is.
The reason that YC partners get so good at picking founders is we just see more data points than anybody else, and it is like a practice game.
I look for high energy, quick thinkers, quick doers, very smart, strongly held opinions, and a high degree of determination. But there’s a lot of people like that.
There’s something else about, “This person has a very unique perspective on the world, and I’m going to bet on that,” and that feels easier to learn with practice.
YC partners get better and better at picking founders because they have seen more samples and can repeatedly compare their initial judgment with how founders actually perform afterward. Energy, speed of thinking, speed of action, intelligence, strongly held opinions and determination can all be listed as explicit criteria, but the truly hard part is judging whether a person's unique view of the world is merely odd or sees something others have not yet seen. That pattern recognition depends more on long practice