Lessons for builders from Sam Altman's Cheeky Pint interview

On October 29, 2026, John Collison interviewed Sam Altman on Cheeky Pint, Stripe’s podcast. The episode is number 39, “Why Sam Altman is against a single superintelligence.” The video is on X. The official transcript is on the episode page.

They talk for about an hour. The dollar figures below are offhand estimates from that conversation.

Scan your code, then change the architecture

OpenAI’s models found serious vulnerabilities in PHP. Altman left the bugs unexplained, so they stay out of this post. His point was coverage. The models “continue to get better fairly rapidly,” and “people just hadn’t really used them to go look at everything out there.” He expects “many more serious vulnerabilities than we expect.”

If you run PHP, install the October security update. For a repo you control, Codex Security can point a model at the code. The wider checklist is in how to protect your app from AI-powered attacks.

He does not “want to declare victory at all on the PHP thing.” He assumes “some bad things are going to happen.” The first day of the response still went better than he expected. Companies took an app offline for an hour to “do a quick scan and deployment.”

Collison pushed past the patch. “We maybe need to change the architectures, which is harder, but a much more durable fix.” Altman agreed. Collison asked about rewriting PHP in Rust. Altman skipped the language. People want to “re-architect what it means to be a router, or an operating system, or an application.” A “whole new stack” “would have seemed insane two years ago.”

On a product you already run, take a secret out of the repo and add an API an agent can call. A later model will read this month’s code, so change that shape before you close the file.

Give an agent team a dollar cap

“I hate the word ‘agent swarms’. Let’s call them ‘agent teams.’”

A few years ago, Altman said, more than a dollar of compute on one problem did not help. Chain of thought moved that limit. You can “throw a whole team at something” and “do that 10,000 times in parallel.”

A lab can run 10,000 copies. You need a smaller group and a review. Collison said advanced math and hacking already use agent teams, and most company work is still one chat.

Parallel coding agents covers how to split work and review it. Too many conversations with coding agents is what happens when the extra chats never report back.

Collison doubted a prompt that says “don’t spend more than $10,000.” Altman said the other wording works. “If you say do the best job you can with $10,000 of compute. Yes, it will understand that.”

Put your own number in that sentence. “Do the best job you can with $20 of compute” fits a bug or a security pass. Stop when the budget is gone. Permission to run with nobody watching is separate, and it is covered in things I let coding agents do unattended.

OpenAI still has no settled team shape. Smartest model, or many copies. Longer rollouts, or shorter ones. “We did not understand that at all. And now we understand it a little bit.” Try one setup. Measure the result. Then change one variable.

The last box is your review of what the team returned.

Explore once, then make the rerun cheap

“We did let it get out of control for one week on the math. Intentionally so.” You “don’t know how far it’s going to go until you try.”

He wants people to “explore at the limit of what is possible.” Optimization can make that work cheap afterward. He expects “10x improvements per year” in turning dollars, or flops, into useful work. He bet OpenAI “could redo Navier-Stokes at 1/100th of the cost a year from when it was done,” using “smarter models and better approaches.”

Collison said “1%-ish order of magnitude of Stripe’s OpEx is tokens.” Altman answered, “That feels too low.” Collison expects 5%, then 10%, then 20%. Altman said many companies will sit in the tens of percent of operating expenses, and that he is biased because he is “talking about my book.”

“There are more companies with the problem of not using enough AI rather than using too much AI.”

For a small team, math week means one expensive pass on a problem whose edge you cannot see. A migration, or a security review you have avoided. Use the strong model. Once the shape is clear, rerun the repetitive parts on a cheaper model. A two-dollar cap on the first look stops you from seeing how far the problem goes.

Reward the result, not the token count

People inside OpenAI get “very wide latitude” on compute, “unless they’re really being abusive.”

“We try never to reward people for using a lot of compute. No token leaderboards. It’s actually slightly bad to be high on the token leaderboard, certainly on a promotion.” Important work still gets “nothing but appreciation.”

A long run earns the tokens when it ships a fix. It wastes them when it only restates the ticket. He set that next to offices that save “$17 on the snack budget” while a useless meeting costs far more, and he called his own view biased.

Keep enough skill for the day the tool is down

“There was a crazy moment at OpenAI where we had a SEV and Codex went down.” A SEV is a severe incident. Collison joked, “And everyone had forgotten how to code.” Altman answered, “not all the way. But it was like a ‘Whoa, what if you couldn’t?’”

You still want to read a diff, run the tests, and edit a function when Codex is down. Collison named two lost arts. Holding a program in your head, and having the code run the first time. Keep enough of both to finish an afternoon without the model.

Build for the user’s agent

Companies keep adding a chat tab inside the product. Collison said “no consumer really wants this UI.” People want to stay in the chat app they already open, and tell it to refund an order. “Your defensibility probably did not come from your ability to put together HTML.” It came from brand loyalty and product selection.

Altman said, “I so directionally agree.” In ChatGPT he picks Amazon “10 out of 10 times” and will “pay a higher price,” because he trusts returns and delivery. “It’s nothing to do with my experience of clicking around amazon.com.”

WebMCP exposes tools on the site. Serving pages as Markdown gives an agent a page it can read.

Collison called a web form fake work, because “you’re never inventing new information for a web form.” Altman said, “I would be so happy to never do fake work at a computer again.”

Check the forms on your product. A name, email, or address the account already stores is that kind of work. Let an agent that already has the account finish the step through an API. Keep a form when the person is deciding something the account does not know.

Charge for a unit they can predict

OpenAI almost stopped pricing the next model in tokens. You would send a prompt. They would “send you a bill,” “like AWS.” Customers rejected it. “I need to know what I’m paying for somehow.” Some would rather pay more, in tokens OpenAI inflated so the unit stayed consistent. They fear that “over the long term they’re gonna get screwed somehow.”

Altman still says “tokens are a kind of fake unit” and “they’re not a good unit.”

If you charge for an AI feature, publish a unit the customer can forecast. A monthly allowance, a cap per run, or a token rate you leave alone. You can still move the work to a cheaper model behind that unit.

Ask for a sharp coworker

Collison described a “glazing era,” then models that “disagree with you, and nitpick at you on every minor thing.” Altman said the latest models are “much better than we’ve had in a long time.”

The bar he quotes is from Emmett Shear. “A really thoughtful and challenging, but still respectful, senior coworker.” He has “still not heard a better, brief encapsulation than that.”

Use it when you pick a coding agent, and when you write instructions for an assistant in your product. Read one reply from the agent you use most. Would a good senior coworker have sent it? If not, change the instructions or the model before you build the next feature on that behavior.

Want me to talk about your product? You can sponsor this site.