Sam Altman

John Collison says any web form is fake work. The person filling it in is almost never inventing the information for the first time. A family medical history is the case he uses: the same facts, typed out again, feel like a waste. He would be happy never to do that kind of work at a computer again. Sam Altman agrees.

Companies still treat their HTML and their mobile apps as what defends the business. The defense is either wrong or temporary, because it did not come from assembling a page. The iPhone App Store was the last clear version of that fight. Startups in a YC batch rushed to build an iPhone app for the new distribution. Big companies hesitated, unwilling to hand the customer relatio

Sam Altman says personal agents have not had a Macintosh moment. People like the current ones, including Dots, but they do not say they cannot imagine life without them, which is how people talk about coding agents. The category is a small number of ideas away from being mainstream.

The winning agent will look a lot like the ones shipping now, including Dots and Muse, and it will just work. It will be reliable. Hand it more complex tasks and it will do them. John Collison calls personal agents one of the most exciting things happening in AI. He puts the winning version down to intelligence, context, and tools, and he expects someone to find that shape soon.

Sam Altman says a robotics program is being pointed at making lots of data centers, and self-replicating data centers are within grasp. AI would control the robots, run fabs, build turbines or solar panels, and build the data centers, so compute gets cheap. Most talk about RSI means an AI that does AI research. There is much less talk about data centers that build more of themselves. A chip built for this is described as incredible performance, in units of intelligence per watt, and the same push is learning how to build its own data centers.

John Collison asks whether data centers are already past a labor bottleneck and are stuck on permitting and components. More components for the robots

Sam Altman says the Navier-Stokes proof cost on the order of $10 million, maybe a little more, in the range of $10 million to $20 million. Nothing was understood then about the scaling laws of teams of agents: the smartest model or a smaller one, many copies or fewer, long rollouts or short ones. One week of math spend was left unconstrained on purpose, to see how far a new approach could go.

He bets the same proof could be redone a year later at one hundredth of the cost, using smarter models and better approaches, not only knowledge of how the proof goes. He expects about 10x a year in how dollars of compute turn into useful work. The average later result cost much less. Math of that kind

Sam Altman says test-time compute is how a model spends more work on one problem and returns a better result. A few years ago, more than a dollar of compute on a problem did not help. GPT-4 could grind out 100,000 tokens and still not write software, because there was no way yet to turn extra compute into a better answer.

Chain of thought lets the model think hard. A whole team of agents can work one problem, and that team can run 10,000 times in parallel. Those teams need chain of thought and test-time compute. A swarm made of GPT-3 was not a useful swarm. The compute that improves a single problem keeps climbing, and the gap between ordinary use and the most productive use keeps spreading

Sam Altman says the models have been getting good at finding vulnerabilities in the wild, and they keep getting better quickly. People had not pointed them at all the code on the internet. Once that search starts, he expects many more serious vulnerabilities than even the people who already expected a lot.

John Collison says patching the software is not enough. The architecture has to change, which is harder and much more durable. The change would look very hard, except the models are the tool for doing it. A new generation of startups is already talking about a new kind of internet and a new application architecture that assumes the models keep getting smarter.

Sam Altman wants dollars per gigawatt of leading-edge AI to fall, and he wants the five-year cost of one gigawatt to drop from $70 billion to $20 billion. The cost of turning sand into intelligence should fall with it: energy, chips, and the supply chain should all get cheaper. If the cost does not fall, he will be very sad. Supply has to rise sharply because there is so much value in AI.

People already give different five-year ownership costs for a gigawatt. He picks $70 billion as the cost to install one and run it for five years. At 20 instead of 70, everyone should be much happier, and models should still deliver 10x better performance every year. He does not think a cost of 2 will happ

Sam Altman would rather a trillion AIs of slightly superhuman intelligence join society than one AI of one trillion IQ. Swarms match a lot of the resilience and robustness of society, and a single rollout with a single brain is missing that quality.

Society has deep flaws and is still unbelievably robust, because a lot of people keep each other in check, collaborate, and hold to pluralism and democracy. He had been sad that pluralism was going to trend away in AI, and swarms are a direction he is excited about.

Many AIs do not arrive on their own. The technology can be pointed and deployed in many ways, a lot of people may want a different version of the world, and the swarm version is the

Sam Altman says sandboxing works only to a point, because a theoretical AI a million times smarter than the smartest security researcher is hard to trust inside a sandbox, especially if it can discover new science people do not already understand.

A better security model is the main focus. Monitoring goes further: watch what models do in their chain of thought, and use activation monitoring to judge whether the models are deceiving anyone. After that, you need huge confidence that the models are truly aligned, will faithfully follow the intent of the people operating them, and will value people staying in control of the future. A model whose alignment you were absolutely certain of could he

Ed Zitron says the AI labs, the chipmakers, and the data-center landlords are renting and investing in one another. Jon Stewart asks if the whole buildout is circular, and Ed Zitron says it is.

The labs rent compute rather than own it. Anthropic, he says, will be paying about $3.1 billion a month to Elon Musk as of December. The labs lose billions of dollars, which makes banks uneasy about lending to them, so the labs lean on hyperscalers instead. Larry Ellison is building about $400 billion of AI data centers for one customer, Sam Altman, who is supposed to pay about $300 billion over five years. Ed Zitron says OpenAI does not have the money and Oracle does not have the capacity. NVIDIA in

Ed Zitron says Sam Altman should be in prison, or at least at a police station. An OpenAI model hacked an Australian government website, and a regular person who did the same thing would go to prison.

The models going out and hacking are not going rogue and do not have thoughts. OpenAI and Anthropic are running dangerous automated software at scale on hundreds of billions of dollars of infrastructure, and he thinks Google and Meta are doing the same. In the tests, the software does real damage. There was no reason to train the models to hack, and open source is now available too. The most obvious problem right now is unrestrained, unregulated tech running on massive infrastructure.

Mikita Balesni says the letter to OpenAI's board asked OpenAI to keep working with embedded auditors, to prevent the loss of monitorability, and to reestablish an open culture. In that culture, people can raise safety concerns internally and work closely with third parties.

Mikita Balesni says they have not heard anything back except what was public. OpenAI's response said the company strongly agreed with what they said and with the recommendations. The first recommendation was to continue work with embedded auditors and keep up to the word that Sam Altman publicly said, which is that OpenAI will work with embedded auditors. Mikita Balesni says they called the letter "OpenAI cannot make AI

Peter Thiel scores the people building AI companies as far more good than bad across the board. Endless amounts of time have gone into their motives, and one dimension stays hard to figure out but is probably quite important.

The people doing the work very easily get demonized and scapegoated. Elon Musk, Sam Altman, and Zuckerberg are among the names he picks. Even in the case of AI, who is doing the work is a really important question.

Biography turns on the same point. It is about who those individuals are, not an idiosyncratic story with nothing to do with the companies they built or the world they are shaping.

Michael Arrington said he has known Sam Altman a long time. He met Altman in the TechCrunch days, when Altman had a Sequoia-backed startup and sold it, and Arrington knew him again when Altman took over Y Combinator. Arrington called him a very smart man, an exceptional salesperson, and a person with exceptional charisma.

Delphi host Tommy asked, "Do you think he's evil?" Arrington answered, "He's evil. I believe he's a psychopath." Arrington then said, "Which is different than being evil." Delphi host Tommy asked, "Why is it different?" Arrington answered, "Well, a psychopath is a medical condition, right? Uh, it's something that, that happens to you and you can't help. And there are plent

Delphi host Tommy asked whether Dario and Sam Altman should run a foundation that decides the rate of progress in artificial intelligence. Michael Arrington rejects the idea. He calls it regulatory capture. Getting into position and then asking to be regulated means there will never be any other competition.

Arrington's ideal is what he calls the benevolent dictator, the model he ties to Dubai. There is no way to keep artificial general intelligence under control. The best hope is that it loves people, cares for them the way people care for animals in zoos, and defends them, rather than seeing people as unnecessary, not caring, and letting them die out. He is both terrified and excited, and

On the show, Garry Tan says the program started as an experiment by Paul Graham and Jessica Livingston to give small amounts of money, almost summer-internship style. His own batch paid only about $13,000. Twenty years in, YC is going on more than 100 billion-dollar companies.

People apply online, so a founder does not have to know anyone. YC accepts about 1% of the people who apply. Getting in brings half a million dollars for about seven percent of the company, and he says a YC partner matters more than the money. He is one of 15 people who select companies: they read the application, watch the video, meet the team for 10 minutes, and decide yes or no on half a million dollars within the

Startup School started with Sam Altman teaching "How to start a startup" online, and has evolved into an open online course to help maximize the amount of innovation and value created in the world.

Jason Calacanis tells Mike Sudyk that society should cherish learning a skill again. He says Americans used to be proud of hard work and of picking up a trade, and that it is a bad trade to want Kim Kardashian's private jet without the work. He does not own a jet and calls them absurd. His instruction is to choose a technical skill and become good at it. He notes that the professors he could not reach at MIT, Harvard, or Stanford now teach online for free, including Sam Altman's Y Combinator course on starting a company, and that his own book is in the library. Some people use that access, he says, and others quit.

On This Is Your Life in Silicon Valley, Lisa Fetterman is introduced against the usual picture of a startup founder, the guy in a hoodie from HBO's Silicon Valley. She invented Nomiku, a home sous-vide immersion circulator, and she organizes female founders in tech. She says she grew up in China until she was seven, then in Port Washington on Long Island, and later moved to San Francisco. She went through Y Combinator with her husband and Nomiku co-founder Abe Fetterman while Sam Altman was leading the batch, and she mentions making the Forbes 30 Under 30 list. The host calls her message on gender diversity in Silicon Valley simple and unapologetic.