Y COMBINATOR · EXTRACTED
The Future of OpenAI, ChatGPT's Origins, and Building AI Hardware ft. Sam Altman
Sam Altman on the moment OpenAI almost didn't get started, the product overhang in front of every founder right now, and why this is the best time in history to build a company.
478K views on YouTube"This is the best fucking time ever in the history of technology, ever, period, to start a company." — Sam Altman
Sam Altman ran Y Combinator before he ran OpenAI. He returned to YC for this conversation as the leader of the company that triggered the AI inflection point. The talk is less a vision speech and more a behind-the-scenes look at the operating decisions: how OpenAI almost didn't get founded, how he thinks about defensibility now that everyone is competing in his space, what he actually looks for in a hire, the rumored Jony Ive partnership, and why he believes the gap between what models can do and what products are doing with them has never been wider. Refreshingly low on hype, high on specifics.
Lean into the decisions everyone else is talking out of
Altman opens with the moment OpenAI almost didn't happen. He had Garry Tan's job at the time. AGI sounded crazy in 2015, well before any working language model. DeepMind seemed impossibly far ahead. There was, in his words, "all this other great stuff to do that would work." For a year, he and the cofounders went back and forth on whether to start it. He calls the decision "coin-flippy." His framing in retrospect: this is the story of many ambitious things — they seem so difficult and there are such good reasons not to do them that it takes a core of people sitting in a room, looking each other in the eye, and saying "all right, let's do this." He says these are very important moments and "when in doubt, you should lean into them."
THE PLAY
Identify the most ambitious decision you've been talking out of for the past three months. The one that keeps surfacing in conversation but hasn't gotten committed to because the reasons-not-to are louder than the reasons-to. Set a date this week. Get the relevant people in a room. Make the decision in the meeting. The reasons-not-to don't go away — but they don't get less true with another month of deliberation either, and the cost of not deciding compounds.
Pick a one-of-one mission to concentrate talent
Altman makes a specific claim about why OpenAI was able to attract the talent it did despite long odds. He says: "If you say we're going to do this crazy thing and it's exciting and important if it works, and other people aren't doing it, you can actually get a lot of people together." OpenAI told the world it was going for AGI. Ninety-nine percent of the world thought they were crazy. The 1% that the mission resonated with included a lot of very smart people, and there wasn't really anywhere else for them to go. He generalizes the lesson: if you're doing the same thing as everyone else, it's very hard to concentrate talent and very hard to get people to really believe in a mission. Doing a one-of-one thing is a tailwind. He flags the failure mode this implies: most founders end up working on the same five ideas at the same time because humans are social and get influenced by what others are doing. The person in the room who builds something bigger than OpenAI, he predicts, isn't working on any of the popular five.
THE PLAY
Write down the five most popular ideas in your space right now. The ones half the room would raise their hand for. If your idea is on the list, your hardest problem isn't execution — it's that you'll fight for talent against everyone else doing the same thing. Either find the version of your idea that nobody else is working on, or find a different idea. The talent-concentration tailwind only applies to founders willing to be early to a one-of-one.
Build into the product overhang
This is the framing from Altman that almost nobody is quoting yet, and it's the most actionable for builders in the room. He argues we're in an unusual moment where model capability is "up here" and the products people have figured out to build are "way down here." Even if models got no better, which of course they will, there's a huge amount of new stuff to build. He adds that o3 cost five times as much last week as it did this week, and that price-performance is going to keep falling dramatically. Plus an open-source model coming soon that he says will surprise people on what's possible to run locally. The conclusion he draws: this is an exceptional time to build a company that takes advantage of the new square on the periodic table that almost nobody has built with yet. Reasoning models specifically — only in the last month, he says, have startups really started building for the fact that the interaction model is fundamentally different.
THE PLAY
Stop benchmarking against what AI products look like today. Benchmark against what current models can do that current products don't yet expose. For one specific user workflow you understand deeply, ask: if I gave the user a reasoning model that could think for ten minutes, search the web, write code, and verify its work, what would the experience look like? That's the overhang. The first products built into it will define the next category. Most of today's AI products were built for the chat-with-a-document model, which is now two generations behind.
Hire for slope, not y-intercept
Altman's hiring framework in this talk is direct. Smart, driven, curious, self-motivated, hardworking, strong track record of accomplishment, works well as part of a team, aligned with the company's vision. That gets you 90% of the way. What surprises him is how much founders focus on other things. Then he gets specific about the early-stage hire. He doesn't believe in hiring "the very senior, eminent administrator" as a first hire. He'd take young and scrappy who clearly get stuff done over the candidate with the extremely polished track record. When he was reading YC applications, he never looked at resume items first — he went straight to "what's the most impressive stuff you've done." Resume was a backup. He attributes the phrasing to Paul Graham: "hire for slope, not y-intercept." Slope is rate of growth. Y-intercept is where someone started. The high-y-intercept candidate looks impressive on paper. The high-slope candidate is the one who'll be doing remarkable work two years from now.
THE PLAY
On your next hire, find the most impressive thing the candidate has actually built or shipped — not the institution they did it at, the thing itself. Then find the second-most impressive. Then the third. Plot those against time. That's their slope. The candidate with the steepest line over the last three years is almost always a better hire than the candidate with the bigger logo and the flatter trajectory, even when the salary expectations say otherwise.
Build defensibility after you have something worth defending
Altman is unusually frank about OpenAI's lack of a defensibility strategy in the early ChatGPT days. He says: "We didn't have any defensibility strategy. We just had the only good thing out there, and then you have some window before which you have to build defensibility." Defensibility for ChatGPT came later — through brand, then memory, then connections, then a bunch of other features. His broader advice to founders worried about being copied or run over by OpenAI: the question of defensibility is one you answer after you have a real product, not before. The most enduring companies are usually not doing the same thing as everyone else, which gives them time to figure out the great product and build the technology before they have to answer the defensibility question. Worrying about the moat before you have something worth moating is the wrong sequence.
THE PLAY
Sequence the work. First, build the only good version of something specific. Ship it. Get users. Once you have a real product with real users, then start mapping defensibility — brand, data, memory, integrations, network effects, switching costs. The premature defensibility worry kills more startups than copying does. If you don't have something users love yet, you don't have anything to defend, and the time spent designing moats is time not spent building the thing the moats would protect.
The startups win the clock-cycle changes
Asked about AGI timelines, Altman gives the answer that cuts through the usual prediction theater: "In 10 or 20 years, unless something goes hugely wrong, we'll have like unimaginable super intelligence." But the more useful frame in the conversation is what he says about why this moment specifically favors founders. When the clock cycle of an industry changes this much, startups almost always win. Big companies have a lot of advantages but they iterate slowly, and when something becomes very cheap their advantages erode. Five years ago, GPT-3 in an API could barely write a sentence. Now it's PhD-level in most areas. He believes the same rate of progress holds for the next five years. The startups that win this period aren't the ones who time the AGI announcement perfectly. They're the ones who picked a real problem worth solving with current models, built it, and were positioned when the next generation arrived.
THE PLAY
Stop using AGI timelines as a planning variable. Pick a problem worth solving with the model that exists today. Build it. Ship it. The clock cycle of the industry is changing faster than the incumbents can react to, and that's the actual edge — not predicting which year a specific capability lands. Founders who are operating at the speed of the current model will be ready when the next one arrives. Founders waiting for some specific threshold won't.
YOUR ACTION PLAN
All the plays, back to back. Use this as your checklist.
- 01
Lean into the decisions everyone else is talking out of
Pick the ambitious decision you've been deferring. Put it on the calendar this week. Decide in the room.
- 02
Pick a one-of-one mission to concentrate talent
List the five most popular ideas in your space. If yours is on the list, find the one-of-one version of it.
- 03
Build into the product overhang
Pick one workflow. Sketch what it looks like with reasoning, tools, and ten minutes of compute. That's your overhang.
- 04
Hire for slope, not y-intercept
On your next hire, list the candidate's three most impressive shipped things by date. Look at the slope, not the brand.
- 05
Build defensibility after you have something worth defending
Stop pre-designing moats. Ship the only good version of one thing. Map defensibility once you have users.
- 06
The startups win the clock-cycle changes
Pick a problem solvable with today's model. Ship a real version this quarter. Speed compounds across model generations.
Newsletter
Get each new protocol the day it drops
One email per drop. No spam. Unsubscribe anytime.
Y COMBINATOR · EXTRACTED BY PODEX