How To Build The Future: Sam Altman
Y Combinator46:512024-11-08English
Chapters
In this Y Combinator interview, Sam Altman uses OpenAI’s early experience to make a point: when technology shifts, startups have an advantage because they can focus their resources on testing one direction faster than large companies, change course when the evidence proves them wrong, and turn their head start into lasting business value.
Overview
How To Build The Future: Sam Altman
- The two discuss why technology shifts favor startups, under what conditions superintelligence might arrive, and how abundant intelligence and energy could together advance knowledge work and physical production.
- Altman looks back on insisting on joining Y Combinator’s first batch and explains why being around people willing to act can do more for someone’s ambitions than a mentor’s words of encouragement.
- Altman talks about wanting to start a research lab, how he recruited early members including Ilya and Greg Brockman, and how OpenAI went from announcing its founding to sitting down and figuring out what to research.
- Faced with outside skepticism and limited resources, OpenAI focused on scaling deep learning. Altman uses that experience to distinguish conviction backed by evidence from stubbornness that ignores results.
- The interview traces the GPT series back to Alec Radford’s discovery of a neuron associated with sentiment, then uses Casetext and user demand for GPT-4 to show when the commercial opportunity became clear.
- Altman looks back on founding Loopt before the iPhone came out. The company did not find product-market fit, but it taught him to adapt quickly in the chaos of a platform transition.
- Altman discusses staffing changes that came with rapid growth and how the research and product directions are coming into focus. He also uses o1 and a hackathon example to explain how AI might progress from reasoning to agents and then to innovators.
- Altman advises founders to use speed, focus, and execution to seize the opportunity created by better models, while warning that an excellent demo cannot substitute for a business with lasting value.
Key points
- Altman thinks large companies have the upper hand when change is slow, while startups can make better use of their speed and focus when platforms shift rapidly. He makes clear that superintelligence arriving within a few thousand days is uncertain, noting that progress could run into unexpected bottlenecks.
- Altman initially thought the most distinctive thing about Y Combinator was that someone told young founders, “You can do it.” He later realized it mattered more to be surrounded by people who were just as willing to act, so he recommends finding peers early who push you to be more ambitious.
- OpenAI publicly made AGI its goal early on, drawing criticism as well as researchers willing to give it a try. With fewer resources than rivals such as DeepMind, the team concentrated on a direction that was already showing results rather than spreading its bets.
- OpenAI initially believed deep learning would get stronger with scale. It took years to get data showing how predictable those improvements were. Altman stresses that when you lack data, you need to be willing to act on your judgment, but when the results prove you wrong, you should admit it and change course.
- GPT-3 could produce impressive demos but rarely supported a mature business. With GPT-3.5, startups started buying it on their own initiative, and as soon as GPT-4 became available, users asked how many GPUs they could get. To Altman, that demand says more about whether a product has succeeded than the team’s own tests and experience using it.
- Altman roughly divides AI capabilities into five levels: chatbots, reasoning systems, agents that can carry out long-running tasks, innovators that can explore unknown problems, and systems that can do that kind of work at the scale of an entire organization. He thinks o1 has reached the second level.
- Founders should take advantage of a small team’s speed and try a new AI capability the day they see it, rather than waiting for quarterly planning. But adopting technology early only produces short-term growth. In the long run, founders still need a better product and a competitive advantage that is hard to copy.
Chapter notes
The window for AI startups and a vision of abundance
- Sam Altman recalls that the team made AGI its goal from the start, even though saying so publicly drew criticism at the time. With far fewer resources than peers such as DeepMind, they decided to focus on one bet instead of trying several approaches at once. He thinks the value of that kind of focus is still underrated.
- Altman thinks now is the best time yet to start a tech company. Large companies have the advantage when change is slow, but major shifts such as the internet, mobile computing, and semiconductors give new companies a chance to use new capabilities to build products with greater impact.
- Superintelligence could arrive within a few thousand days, but that is only a conditional guess: the pace of progress over the past three years would have to continue for several more years, with gains building on one another. Nine years is about 3,500 days. o1 already shows strong reasoning on some well-defined, closed-ended tasks, and architectural changes have brought significant improvements, but unexpected bottlenecks could still arise.
- The two discuss ambitions such as addressing climate problems, establishing space colonies, and solving mysteries in physics. Altman thinks those goals might not be as far off as they seem. He appreciates that YC encourages founders to aim higher, even when most people think those things cannot be done.
- Abundant intelligence and energy are two key conditions for abundance. AI available on demand can speed up the process of coming up with ideas, while robots can turn those ideas into products, potentially allowing more people to benefit from physical production. Energy supports physical production and is also needed to run AI at scale.
- Even without a major breakthrough in nuclear energy, Altman thinks progress in solar power and storage is enough to provide a decent fallback. He believes making energy cheaper and more plentiful would directly improve quality of life.
Joining Y Combinator and the influence of peers
- In the version told in the interview, Paul Graham thought Sam Altman was a first-year student at Stanford and suggested he not rush to join Y Combinator’s first batch in 2005. Altman replied, “I’m a sophomore, and I’m coming.” Altman remembers only wanting to reschedule the interview and says he responded more politely; the story grew more dramatic in the retelling.
- Altman does not think he is as forceful as people say. He just does not assume things have to stay the way they are, and if something seems feasible from first principles, he is willing to try it. One thing he liked about Y Combinator was that it brought together people who were similarly willing to act on their own ideas.
- Altman originally thought the most special thing about Y Combinator was that a remarkable mentor would tell young founders, “You can do it. I believe in you.” He later understood that what mattered more was that the people around him were actually doing it. He used to think he could figure everything out on his own, but now he advises young people to find a group of peers who inspire them as early as they can.
- Altman liked Stanford, but competition among peers there often revolved around who could get an internship at which investment bank, and he got caught up in that for a while himself. A room full of founders who were willing to act did more to spark his ambition. After experiencing the atmosphere at Y Combinator, he found it easy to decide not to go back to school.
- The interview quotes Carl Jung: the world will ask you who you are, and if you do not know, it will tell you. Altman’s takeaway is that no one is immune to peer pressure, so what you can do is actively choose good peers.
From YC Research to OpenAI
- Sam Altman had always thought that the job he most wanted in retirement was running a research lab, not necessarily in AI at first. YC Research did fund work in several areas and tried quite a few ideas that did not work well. He sees this as similar to YC’s approach: fund smart people and accept that some attempts will work and others will not.
- From late 2014 to early 2016, Bostrom’s Superintelligence sparked discussion, and DeepMind achieved several impressive results. Altman had long been interested in AI and wanted to try working on it himself, but it was hard to tell what exactly to do at the time.
- After watching videos of Ilya, Altman wanted to meet him. His email went unanswered, so he went to a conference where Ilya was speaking, and after meeting there they talked regularly. He had known Greg Brockman since Stripe’s early days. From the start, the team made clear that it wanted to build AGI. Many in the field saw that goal as crazy or even irresponsible at the time, but it immediately drew Ilya in and attracted a group of talented young people.
- The early members met one-on-one first and then in different combinations, taking about nine months on and off to put the team together. OpenAI announced its founding in December 2015 and formally started work in early January 2016. After the holidays, about ten people gathered in Greg Brockman’s apartment and found they had to answer the most basic question: what exactly do we do now?
- Altman compares founding OpenAI to a startup closing a round of funding: it looks like a major accomplishment, but it is really just the starting gun. The team then spent a long time figuring out what to do. Although they took plenty of detours and some attempts worked while others did not, in hindsight he still thinks the broad direction Ilya and the other early members set was surprisingly accurate.
- The research plan written down early on was more specific than “do deep learning”: first build a large unsupervised model, then solve reinforcement learning (RL). Altman recalls that an early team meeting set a third goal: never grow beyond 120 people. They did not stick to that one, but he thinks the first two goals were a pretty accurate forecast of where the work would go.
Betting on scale and letting evidence correct you
- OpenAI was founded on two core judgments: deep learning works, and scaling it up would make models better. The team already knew that larger neural networks performed better. The real bet was that the improvement would continue. It took years to get data showing how predictable the size of those gains was.
- Some prominent figures in the field dismissed the models’ performance as a trick, arguing they were not really learning or reasoning, and some worried this approach would deepen the AI winter. OpenAI, meanwhile, saw results keep improving. The scaling laws that came later helped the team judge what gains further scaling might bring.
- OpenAI thought some fundamental emergent phenomenon might be behind deep learning’s performance, although the precise mechanism was unclear. With far fewer resources than competitors such as DeepMind, the team focused on the scaling approach that was already producing results instead of exploring many directions at once.
- Sam Altman’s startup experience at YC taught him not to assume that people who disagreed with him necessarily had the answers. If their objections did not hold up, he could try the idea and iterate quickly to find a direction. But conviction is not stubbornness: you can act on judgment when there is no data, but when the results show you were wrong, you have to accept that and adjust.
- OpenAI was not committed to language models from the start. Much of its exploration deepened its scientific understanding but was not the shortest route to its later results. The team made repeated mistakes in judging its technical direction, the company’s positioning and structure, and the development of AGI. Altman thinks one of the team’s strengths is that it can keep going after setbacks.
GPT: From research discovery to commercial validation
- The team initially worked on robots and agents that played video games. Sam Altman did not expect language models to become the main focus. GPT-3 arrived a few years later, but that path was not obvious early on.
- When unsupervised learning was still not working very well, Alec Radford found that as a model generated Amazon reviews, one neuron switched states depending on whether a review’s sentiment was positive or negative. The discovery came before GPT-1 and helped drive the development of the GPT series: Alec Radford built GPT-1, and others then scaled it up into GPT-2.
- Jake Heller of Casetext used GPT-3, GPT-3.5, and GPT-4 in succession. GPT-3.5 hallucinated too often to meet the demands of legal work, but GPT-4 could do what he needed once he broke the prompts down into small enough workflow steps. He built a large set of test cases for this and later sold Casetext for $650 million, becoming one of the first people to put GPT-4 to commercial use at scale.
- Founders were blown away by GPT-3 demos, but almost no one built a strong business with it, with copywriting as a possible exception. By GPT-3.5, startups, especially YC startups, started buying it on their own initiative. As soon as GPT-4 was available to try, users quickly asked how many GPUs they could get.
- GPT-4 impressed the team in both internal tests and hands-on use: for instance, it could rhyme and tell jokes that were a little funny. But the team also knew that developers tend to overestimate their own work, and that you cannot tell whether a product will succeed until you put it in users’ hands.
Loopt and the first platform shift
- Sam Altman started Loopt at YC at age 19. It was a service for seeing where friends were, much like Find My Friends, and the iPhone was still two or three years away. In high school, he went online with a phone that could only check email through a slow text browser and realized that a phone was really a computer you could carry around.
- When Loopt started, there was no iPhone or App Store, and Altman later played a small part in the launch work. Building a startup on a platform someone else had created gave him firsthand experience of the chaos at the start of a platform transition: things changed quickly and required constant adaptation, and even small decisions could shape the direction of the platform.
- Loopt was eventually sold, but Altman does not consider it a success. The team never really found product-market fit or got growth to the point where it could sustain itself. It was a painful experience, but he thinks starting a company teaches you broadly applicable lessons faster than anything else he knows of.
- Altman is optimistic about startup opportunities now because many people still do not realize how big the change is, while YC founders are already prepared to move quickly. The two use Facebook as an example: it was good at building web software, but it nearly missed the shift to mobile and later acquired Instagram and WhatsApp.
- Asked whether it is better to make money first and then tackle a harder technical project, Altman says being able to provide early funding for OpenAI himself did help, because it was hard to persuade others to fund it at first. Elon later provided much more substantial funding.
- Altman does not regret doing Loopt and acknowledges that he learned a lot, but in hindsight he thinks he could have picked a more worthwhile project. AI had always been what he most wanted to work on, but the advice he heard in his school’s AI lab was to stay away from neural networks because others had already tried them and thought they did not work.
OpenAI’s rapid growth and the path to greater capabilities
- ChatGPT has been out for less than two years, yet in two years OpenAI went through the kind of growth a midsize or even large tech company would normally take ten years to go through. Rapid growth came with changes in the management team: people who are good at going from 0 to 1 are not necessarily good at going from 1 to 10 or from 10 to 100.
- For the first time, Sam Altman feels that OpenAI basically knows how to get to AGI, although there are still clearly identified problems it has not solved and a lot of time and work will be needed. The research and infrastructure paths are fairly clear, and the product direction is coming into focus. The more the team can concentrate on a few goals, the faster it can move.
- At YC’s o1 hackathon, CAD/CAM startup Camfer built a system that could iteratively improve an airfoil, taking a design that could not fly and giving it competitive lift. Altman had thought going from agents to innovators would require major new ideas, but this demo and other examples showed him how much innovation could come from using existing models creatively.
- Natural language can be the interface people use to interact with a model, which can then call existing CAD/CAM software. With code generation, a model can not only use existing tools but also write its own and combine them. This is why Altman thinks capabilities will advance faster than many people expect.
- Because AGI has come to mean so many different things, OpenAI proposed a rough five-level scale. Level one is chatbots. Level two is reasoning models, which Altman thinks o1 already qualifies as. Level three is agents that can keep working on long-running tasks, interact repeatedly with their environment, and ask for help or collaborate when needed. Level four is innovators that can explore and come to understand unknown phenomena over time, like scientists. Level five extends that ability to the scale of an entire company or organization.
- The interview imagines that future companies with billions of dollars in annual revenue might need fewer than 100 employees, or even just one. Altman thinks that trend seems to be emerging and says now is a good time to start a company.
Advice for AI founders
- Sam Altman advises founders who are just starting out to bet on AI. Models are nowhere near a saturation point, and they will get stronger quickly, greatly expanding what startups can do. Compared with many companies already planning on a quarterly cycle, small startup teams now have their biggest advantage in being able to move quickly, stay focused, and act with conviction, keeping up with changes in the technology.
- When you see a new capability, use AI to build something with it that same day instead of waiting for the next quarterly planning cycle. But using AI does not suspend the rules of business: the team still has to build its own moat, competitive advantages, and better product.
- Adopting new technology early can bring explosive short-term growth, but it is no substitute for lasting value. Anyone can now build an impressive demo and build a business faster and better than before, but the real goal is still to build a business that works.
- Asked what he was looking forward to in 2025, Altman mentioned AGI, but said he was most looking forward to the upcoming birth of his child. Nothing he had ever looked forward to made him more excited.
What it answers
- Why does Altman think now is a particularly good time to start a tech company?
- He thinks large companies have an advantage when things are stable, but when technology changes quickly, small teams can try new capabilities and build products faster without waiting for quarterly planning.
- Why did OpenAI focus its early bets on scaling deep learning?
- The team had already seen neural networks perform better as they got larger, and it did not have the resources to try as many approaches at once as its bigger rivals, so it chose to focus on scaling. How predictable the size of the improvements was only became clear later, with data.
- What turned GPT-4 from an impressive demo into a commercial tool?
- The interview gives the example of Casetext. GPT-3.5 hallucinated too often to be usable in legal work, while GPT-4 could do what was asked once the work was broken into smaller tasks and extensively tested. More broadly, as soon as early users got access to GPT-4, they started asking how much computing power they could get.
- What is the difference between agents and innovators in Altman’s framework?
- Level-three agents can carry out longer tasks, interact with their environment repeatedly, and ask for help or collaborate when needed. Level-four innovators can explore phenomena that are not yet understood over extended periods, like scientists, and come up with explanations.
- How should founders distinguish conviction from stubbornness?
- When there is no data, be willing to act on your judgment and iterate quickly. Once the results show you were wrong, accept them and change your approach instead of continuing to defend your original judgment.
Quotes
At least for myself, is no one is immune to peer pressure, and so all you can do is, like, pick good peers.
Most of the world still does not understand the value of, like, a fairly extreme level of conviction on one bet.
No one has all the answers.
There is nothing that I have ever heard of that has a higher rate of generalized learning than doing a startup, so it was great in that sense.
You can do it faster than ever before and better than ever before, but you still have to build a business.
Mentioned
- Sam Altman — The guest in this episode, who looks back on founding Loopt and OpenAI and offers advice on starting an AI company.
- Garry — The host of the interview, who looks back with Altman on their shared experiences at Y Combinator.
- Y Combinator — The program Altman joined when he started a company at nineteen, and an important place for him to shape his views on startups and meet peers.
- Paul Graham — The interview mentions that he founded Y Combinator and once suggested that Altman, who was still at Stanford, wait a little longer before applying.
- OpenAI — The interview traces how this research organization, which set AGI as an early goal, built its team and chose a technical direction.
- DeepMind — The organization Altman uses to contrast OpenAI’s early resources and research strategy; it had already achieved several results that impressed him.
- Ilya — An early researcher Altman sought out at a conference, whose judgment about OpenAI’s initial research direction Altman speaks highly of.
- Greg Brockman — One of OpenAI’s early members, whose apartment the team gathered in in early 2016 to discuss what to do next.
- Alec Radford — He discovered a neuron associated with positive or negative sentiment in a model that generated reviews and later built GPT-1.
- Loopt — The mobile location-sharing startup Altman founded at nineteen. Although it did not find product-market fit, it gave him firsthand experience of the shift to mobile platforms.
- o1 — Altman thinks it already shows strong capabilities on well-defined, closed-ended tasks and sees it as a sign that the second level, reasoning, has been reached.
- Casetext — The host uses it as an early example of commercializing GPT-4, noting that Jake Heller broke legal work into small tasks and built a large set of test cases. The company was later sold for $650 million.