Sam Altman returns to Stanford, where he once taught How to Start a Startup, to explain why betting on scale beat the consensus and what that means for the people building now.
Altman's career pattern is one bet: when something already works a little at a small scale, push it to a scale nobody has tried, because interesting new properties tend to emerge there. Most smart people talk you out of it for non-specific reasons, and humans are bad at feeling exponentials, so the move stays underexplored. OpenAI applied this to deep learning, to GPU runs across tens of thousands of chips, and to the org itself, and expects intelligence to become a utility like electricity that everyone plugs into.
Startups have changed
Altman says he'd rewrite his 2014 startup class because everything about building one is different now. With an affordable amount of spend on tokens, a solo founder can do what a great engineering team used to do. Ambition, speed, and how much you can attempt at once are all higher. He warns against assigned ideas: if an idea is obvious to him, it's obvious to many, and the good ones are the four-companies-are-working-on-it, non-obvious kind that OpenAI itself was in 2015.
Scale as a design tool
Altman's core empirical claim, which he admits he can't fully explain, is that the most interesting things he's seen came from emergent properties at scale, or from scaling returning value long past what consensus predicts. He learned it at Y Combinator: smart people wanted smaller batches, but the network effects inside big batches were a property nobody had found because nobody had funded startups at that scale. When something works a bit at small scale, pushing it bigger is usually a good and underused idea.
Scaling breaks things
Scaling is always a little broken, and things fail at an accelerating, unpredictable rate, which is why people avoid it. Altman's method is to split each objection into a separate problem and solve them one at a time. For giant model runs that meant the technical question of whether a run across 10,000 to 100,000 GPUs was even possible, the capital and business question, and the cultural fight over concentrating all the compute into one bet instead of spreading it across projects.
How ChatGPT actually happened
OpenAI made GPT-3, needed revenue, and couldn't find a product, so it shipped an API and hoped someone else would. Copywriting was the only real business, but developers kept using their API keys just to chat. Following the YC rule of building what users already love, the team wrapped the new 3.5 model, with better instruction-following, into a chatbot as a research demo. It went viral. By day five of traffic that peaked higher each day, Altman called an emergency to build a company and product at once.
Codex and the research roadmap
Coding was the original plan before ChatGPT: OpenAI saw code as how models act on computers, and robots as how they act in the physical world. Codex got good in early 2026, and with the 5.5 model hit a real inflection point. Altman thinks the current pipeline of pre-training, mid-training, post-training, and RL will get AIs to do incredible work, but expects a major rewrite of it, ideally figured out by the AIs themselves. OpenAI's stated goals: an AI research intern by September 2026, and a full end-to-end researcher inventing new architectures by March 2028.
Intelligence as a utility
Altman frames AI as a new utility, a rare event on the level of electricity or the internet. Early electricity companies couldn't sell electricity because it sounded scary, so they sold light at night. He doesn't yet know OpenAI's version of that pitch, but expects people to buy access to the whole system, thinking in tokens or one level up, not chips. His biggest fork is whether this stays concentrated in a few companies, an attractor state he calls unstable and unfair, or gets democratized. He puts the democratic path at about 80 percent.
- A single founder with a modest token budget can now match what took an excellent engineering team, so ambition and speed should go way up.
- The best startup ideas are non-obvious and barely contested; if an idea is obvious to Altman, it's already crowded.
- When something works a little at small scale, pushing it to an untried larger scale usually reveals valuable emergent properties.
- Humans are bad at feeling exponentials, so both scaling laws and organizational complexity get underestimated repeatedly.
- ChatGPT was an accidental hit built as a research demo on top of a chatbot pattern users had already discovered on the GPT-3 API.
- OpenAI's roadmap targets an AI research intern by September 2026 and an architecture-inventing AI researcher by March 2028.
- Frontier labs will have to become inference companies, because delivering cheap, abundant intelligence at scale is underinvested.
In their words
“We were a research lab first that later had to bolt on a startup. I don't really recommend that.”
“You are getting from us is not electricity, it's light at night.”
“All of the frontier labs are going to have to become inference companies to a significant degree.”
Terms to know
- Scaling laws
- The observed pattern that AI models get predictably better as you add more compute, data, and parameters.
- Emergent property
- A useful behavior that appears only at a larger scale and didn't exist at all at smaller sizes.
- The pipeline
- The standard training sequence across labs: pre-training, mid-training, post-training, then reinforcement learning and supervised feedback.
- Post-training
- Tuning a trained model, for example for instruction-following, so it does a good job in conversation.
- Inference
- Running a trained model to actually serve answers; the cost and speed of using AI rather than building it.
- Intelligence as a utility
- The idea that AI becomes like electricity or internet: metered access that every person, company, and government plugs into.
Sam Altman at Stanford CS 153: Frontier Systems
New to this? Come build with us.
Reading is good. Building with people is better. Our drop-ins are free and open to total beginners.