Sam Altman wants more people to tell him he’s wrong.
Asked whether people still call him delusional, the OpenAI CEO answers, “Yes, but I think I should probably get more of that.” Leaders in roles like his often end up surrounded by yes-men, he explains, and the fix is rewarding the people who push back.
That thread runs through his interview with Raj Shamani on the Figuring Out podcast, which covers the compute bet behind GPT-4 and the ChatGPT launch most people around him thought was too early, plus the Y Combinator habits under both. Where a figure is his own recollection, I say so.
In short, Altman bet OpenAI on compute after months of using GPT-4 internally, and he tests every contrarian idea on smart non-experts before he commits. Here are the 10 lessons, each with its clip.
1. Why Sam Altman Bet on Scaling Neural Networks Years Before the Scaling Laws (The Pattern He Saw at Y Combinator)
Altman saw the curve in startups before he saw it in a neural network.
“When you really scale up something that’s working at small scale, you often see big super linear returns to doing it at bigger scale.”
Everyone gets a small number of great insights in a lifetime, he says, and applies them again and again in new ways. His was the surprising properties of scale, which he spotted in companies, countries, cities, individual careers and networks like Y Combinator.
He describes himself as a pattern matcher by default. So when neural networks worked at all and improved as they grew, his strong intuition said to scale them up a lot more.
That was years before the scaling laws, back when predicting a model’s properties with any precision was still out of reach. The evidence was a rough observation, backed by the AlexNet result and many more after it: a bigger neural network could do more impressive things.
The curve you trust in your own field is a bet on the new technology, and you can test it before any benchmark does.
2. The Argument Against Scaling AI That Sounded Airtight (Here Is the Flaw in It)
An argument against scaling still sticks with Altman, because it sounded so convincing.
“Obviously now we know that’s ridiculous. But at the time it sounded very compelling.”
Look at the mistakes neural networks made and you could decide that something so dumb would stay dumb at any size. Smart people said exactly that, each with a very good reason, pointing to symbolic representations and other approaches. Someone once laid out why OpenAI’s approach was doomed, and Altman still remembers it.
The argument ran like this: models learn to predict what comes next from their training set, so they can only model what the set holds. Show a model 1 million math problems and it learns those problems, yet an unproven problem stays out of reach, because the answer is still unknown.
The argument missed what the models learn. They pick up a general purpose ability to reason, and that ability expands into new things.
A skeptic’s case can sound airtight and still rest on a wrong assumption about what the system is learning.
3. How GPT-4 Convinced OpenAI to Bet the Company on Compute (What the Team Saw First)
OpenAI made a very big bet on compute before the rest of the world had seen GPT-4.
“We had enough conviction that we were probably right because we were able to use this model and we had internalized the scaling properties.”
The model finished training around August 2022, as Altman recalls it, and sat inside the company for roughly 7 to 8 months before release. The team spent that time studying it and knew it was extraordinary.
It took a little while to see what GPT-4 opened up: new avenues of research, and how fast models were likely to improve after it. That led to the chain of thought work and the reasoning models. Then came the decision to bet the company on buying up a lot of compute, which felt very crazy to people who had yet to see the model.
Altman names the limit of that conviction himself: had GPT-4 arrived at roughly GPT-3.8 scale, the team would have lacked the conviction for a bet that size.
Conviction comes from hours inside your own product, and those hours belong before the budget request.
A bet like this runs on compute budgets, scaling laws and reasoning models. Start with these:
4. How Altman Keeps Yes-Men Out of the Room and When He Overrules Everyone
Power thins out honest feedback, and Altman plans around that.
“One way that like a lot of people in roles like mine have had things go wrong is they get surrounded by too many yesmen.”
The fix has 2 parts. People first: old friends, plus colleagues he has worked with for 10 plus years. Then culture, which rewards people for pushing back hard when he’s wrong and makes it easy to do.
It takes active care. The more often you’ve been right and the more power you hold, the harder honest pushback gets.
The host pressed on what happens when old friends and mentors say he’s heading the wrong way. Going against the room is rare: the number of times he acts against everyone’s advice is very small, though it happens. Some founders take the most pride in doing exactly that, he has noticed, and he thinks you should do it only very grudgingly.
A decision against all of them is an expensive, high stakes bet that’s unlikely to be right, so strong conviction has to sit behind it.
“if you surround yourself by smart people, you should mostly listen to them.”
Who on your team can tell you no without paying for it?
5. Why Obvious Startup Ideas Draw Crowds (The 10,000 Photo Sharing Companies Altman Remembers)
In 2015, Altman picked the empty field over the crowded one.
“If the ideas are obvious, a lot of other ambitious people are going to start them, too.”
By his count, the AGI labs starting that year fit on 1 hand when OpenAI launched. His suspicion is that OpenAI was the only one, though he admits he’s unsure.
Photo sharing was the crowd: by his guess, 10,000 companies. Most people thought it was the better idea, and those founders also ran into Instagram’s network effect.
The filter favors bets he has reason to believe are worth more than they seem, even if they’re unlikely to work. The payoff is big enough that the expected value stays good, and understanding a technology or a piece of science before other people do gives you an advantage.
One instinct Altman describes is letting smart people talk you out of an idea to avoid looking dumb. The moment to cherish, he says, is understanding something deeply, with a belief that rests on more than wishful thinking and an argument for why it’s a good bet.
“you should cherish those moments like that’s awesome as a builder that’s when you should be the most excited.”
Picking an AI bet that few people see yet? These guides help you find it:
6. The Smart Friend Test: How Altman Separates Conviction From Wishful Thinking
Altman checks a contrarian bet with 1 question.
“Can you explain it to an optimistic enthusiastic but kind of mentally disciplined smart friend and can you update them about why it might be a good bet?”
The listener he wants is enthusiastic and still disciplined. OpenAI itself was the test case: the default reaction was strong skepticism, so he talked to smart people outside the field and laid out the evidence, which covered what the team had seen, why it could work, why it would be huge if it did and why the approach had legs.
The payoff was part of the pitch: if OpenAI could make super intelligence, the reason to want it was kind of obvious. Often enough, the people he talked to said the explanation updated them and the idea seemed worth doing.
The other outcomes carry information too. Convincing everybody would have signaled an idea too obvious, and convincing zero people would have raised his skepticism.
The cheap version of this test: pitch your idea to 5 smart people outside your field and count how many update. That checks whether the argument survives outside your own head.
7. Why Altman Launched ChatGPT After an Internal Prototype Kept 5% of Users (The Y Combinator Rule Behind It)
A 5% retention number ends most launch plans. To Altman, it looked like a good sign.
“The more important thing for an initial project is a small number of people that really love something, not a lot of people that like it a little bit.”
Most people around him thought ChatGPT was too early and too rough, and it launched anyway. The reasoning came from Y Combinator, where he spent so long drilling into founders that a first version should embarrass you a little, because waiting longer costs you the feedback you need.
The internal version kept roughly 5% of its users, by his recollection, and the data was weak: the team had better models and other interfaces. Plenty of people would call 5% retention a reason to wait, and he agrees that’s normally true.
Then he talked to the users in that 5%, and some of them loved it. Low retention paired with heavy, devoted use looked like a good sign, thin data and all. That combination was the signal worth acting on, in his reading.
Retention tells you who left. The users who stayed tell you why.
Before your first AI launch, study retention and feedback loops:
8. What Altman Looks for in a Founder in 30 Minutes (The Signals People Rarely Discuss)
A 30 minute meeting gives Altman a pretty strong intuition about a founder, and he names the signals people rarely discuss.
“If I leave the meeting convinced that I want to work with them, that’s a huge thing I look for.”
The host pointed to companies Altman picked as an investor, Reddit and Stripe among them, and asked whether 30 to 40 minutes can show who will build a company. Imperfectly, Altman answered, though you can form a pretty strong intuition.
Communication comes first: a founder who explains things well and makes the explanation easy to follow. The job demands it, since an early stage founder has to evangelize to other founders, customers, people they want to hire and investors.
2 more signals follow: new ideas that arrive quickly and improve quickly, and well-reasoned thoughts on the world that are new to him.
Endurance is the hard signal, a trait the host raised. It resists judgment in a 30 minute meeting, so he used to ask for stories from a founder’s life history that point to it, though in the end you rely on the stories they tell.
9. Why No Single Founder Pattern Exists: The Top 10 YC Companies Test
Scramble the founders of the top 10 Y Combinator companies and Altman expects a disaster.
“There’s many different ways to be successful as a founder and there’s not a single pattern that works in every case for every company.”
A long list of founder traits and a long list of good business traits are easy to write. The proof of how many ways there are to succeed, in Altman’s telling, is a thought experiment: take the top 10 YC companies, grant that all 10 founders and all 10 companies are great, then scramble which founder runs which company. It would be a disaster.
That is an existence proof for how many different ways there are to succeed.
Failure is easier to predict. The host asked whether a single pattern predicts disaster; with no formula on offer, Altman answered with anti-patterns, among them lack of drive, lack of focus, lack of vision and lack of ability to solve hard problems repeatedly with a smile on your face. Many things almost certainly predict failure, he says, while the ways to succeed come in many different forms.
The failure list travels well. The success list stays local.
10. 20 Years of Experience or 20 Copies of Year 1? Altman on Compounding Your Career
Altman ranks compounding first among his career rules, and he thinks most people fall off the curve early.
“Sure you have 20 years of experience, but you have 20 times the same 1 year of experience.”
Altman remembers writing an essay of success rules, though the rules themselves have faded, and he bets he could write a shorter, better version now. Asked what he would put in it, he starts with compounding, the point he had first in that essay, and says he thinks it’s underestimated.
A career lets you build on each successive phase, but only if you keep pushing yourself to do uncomfortable, non-obvious new things. A lot of people fall off that path very early.
Empirically, most people drop off the compounding exponential curve, and he calls that a very good, non-obvious piece of advice. The habit that counts is choosing discomfort again and again, with each new phase. Applied to himself: asked what feels most uncomfortable today, he answers that most of his work feels new again.
Does last year’s work differ from the year before it? If they look the same, your experience stopped compounding.
The Contrarian Builder Playbook
Pick a bet you can explain, test it on smart people who push back, and ship before the first version feels ready.
Founders: Explain your contrarian bet to smart non-experts and watch whether they update. If zero people update, Altman’s skepticism rises, and yours should too. Book 5 of those conversations this week.
Investors: Altman looks for 3 things in a first meeting: a communicator who leaves him convinced, ideas that improve quickly, and well-reasoned views new to him. Endurance, which the host raised, resists judgment in a 30 minute meeting, and he used to ask for life history stories that point to it. Score your next first call against those 3 signals and ask for 1 story that shows grit.
Operators: Honest pushback takes active care, and it gets harder to find the more often you are right. Keep the people who have known you for years. Reward the ones who say you are wrong, and name 1 person who disagreed with you this quarter along with what changed because of it.
Everyone else: Altman‘s career advice is to keep doing uncomfortable, non-obvious new things, or you repeat 1 year 20 times. Pick 1 thing this month that’s new to you and start it.
The 5 Principles to Steal
Test curves you already trust. Altman carried the returns to scale he saw at Y Combinator into neural networks, years before the scaling laws.
Earn conviction through use. OpenAI bet on compute after working with GPT-4, and Altman says a weaker model would have meant a smaller bet.
Reward the pushback. Make disagreement easy to deliver and reward the people who deliver it.
Chase a small devoted group. Altman read 5% retention paired with heavy use as a good sign.
Pitch smart non-experts. If some update, keep going. If they all shrug, doubt the idea more.
Obvious ideas draw crowds.
Early versions draw feedback.
Hear all 10 of these lessons in Altman’s own words in the full interview.
If this breakdown saved you 1 hour, send it to one founder or investor who needs it.
More From Sam Altman
The AI CEOs Asked for a Brake. Trump Said No. Here Is How to Plan Anyway.
Sam Altman Ships AI to 900 Million People a Week. He Still Reads His Inbox Like It’s 2006.
Sam Altman and Greg Brockman finally talked. 10 things matter
Founders Who Bet Early and Kept Going
Jensen Huang Bet Wrong, Told the Truth, and Built a $5 Trillion Company
How Jensen Huang Turned a Green Van and 30 Years of Rejection Into Nvidia


putting 'get called delusional more' in my OKRs for next quarter
Lesson 3 maps straight onto trading. The only setups I size up on are the ones I've traded and reviewed myself. Conviction borrowed from someone else's track record folds on the first red day. Conviction from your own data holds, and today is a good day to find out which one you have.