Altman says even he has not fully changed how he works with AI
OpenAI CEO Sam Altman said in an interview with David Senra that one of the strongest signs of how early the AI shift still is can be seen in his own day-to-day behavior. Even after helping create tools such as Codex, he said he still uses a computer much the way he did 20 years ago.
Altman said he still copies and pastes across different messaging apps, scrolls through email and picks the easiest message to answer first, and maintains his to-do list using familiar methods. In his telling, that gap matters. It suggests that people often internalize inefficient routines as what “work” is supposed to look like, which makes behavioral change much slower than technical progress.
The discussion came on the Founders podcast hosted by David Senra. The original episode title was “Sam Altman on Building OpenAI & Betting on the Impossible,” and the broadcast date listed in the source material was Aug. 23, 2026. TechFlow compiled and translated the conversation, while also noting that Altman’s views on OpenAI’s strategy, products, technical direction and industry outlook are closely tied to his role as CEO and should not be treated as an independent third-party judgment.
Why he thinks AI adoption will move slower than many expect
Senra asked Altman why he sees Shopify CEO Toby Lütke as one of the most interesting chief executives right now. Altman said Lütke has been one of the most forward-leaning CEOs in the AI era, writing software himself, running experiments himself, and sending OpenAI highly detailed product feedback. He added that Lütke was saying earlier than most that companies could not behave like NPCs and had to embrace agents.
Still, Altman did not fully agree with Lütke’s timeline that 2026 would be the year every business gets reshuffled. He said he agreed with the spirit of that view, but not the schedule. His argument was not that the technology lacks potential. It was that economies and societies carry huge inertia. People keep doing the same things, buying the same products and using the same tools, even when a better option exists.
He described that inertia as both stabilizing and slowing. It can make a major transition less violent, but it also means many forecasts are too optimistic about timing.
Altman tied that point to a broader view of behavior change. He said changing habits is much harder than many technology enthusiasts assume, and cited the period when Netflix was still mailing DVDs while people continued to rent videos from Blockbuster as an example of how strong behavioral lock-in can be.
He compares the current AI moment to the pre-iPhone smartphone era
Asked what it would take for him to use his own products in a much deeper way, Altman said the shift would probably come gradually. He does not expect one clean break. Changing someone’s workflow in a fundamental way is difficult, he said, even when better tools are already available.
His comparison was to smartphones before the iPhone. Altman said he had used a Palm Treo in 2003 and believed the underlying technology pieces were already largely there. What had not arrived yet were the product ideas that turned the iPhone into the iPhone. He sees AI in a similar phase today: the technical ingredients exist, but the product form that completely changes human-computer interaction has not arrived.
OpenAI made two unpopular bets: AGI first, then large language models
Altman said his interest in AI goes back to childhood. He described himself as a nerdy kid who spent Friday nights on computers, read science fiction and watched science fiction. During college, he said, he even spent a summer working in an AI lab. But around 2005, he was explicitly told by a professor that deep learning would not produce good results and was one of the safest ways to ruin a career. He said he believed that assessment at the time.
He then moved into startups and later investing, a sequence he called unusual but useful. As an investor, Altman said, you get to observe many companies at their most important crux moments. You do not get the daily operating reps, but you do get a very large dataset.
That experience helped shape one of the frameworks he now applies to OpenAI: power-law thinking. In venture, he said, the best investment can exceed the value of all the others combined, and the second-best can exceed all the rest after that. He believes AI research works the same way, with one outstanding idea outweighing all the others put together.
Altman said that when OpenAI was founded at the end of 2015, there were hardly any AGI efforts in the world beyond one or two groups such as DeepMind. OpenAI’s decision to say openly that it was trying to build AGI drew ridicule from what he called “intellectual giants” in the field. Later, he said, the company was mocked again for betting on large language models.
From that, he drew a hiring and judgment principle. The people worth backing are not the ones who make slight adjustments to conventional thinking and then try hard to sell themselves as original. The more valuable group, he said, consists of people whose way of thinking is visibly different and who are willing to hold deeply unpopular beliefs. They may be wrong, but if they are right, they can be very right.
His main focus now is research and compute
When Senra asked where most of his time goes today, Altman’s answer was research and compute. He said he would like to spend more time on products, and that OpenAI has strong product people already, but the highest-leverage problem right now is making the models smarter and making sure many people can use them efficiently and at scale.
He described compute expansion as a system-wide coordination challenge rather than a narrow engineering task. It requires chips, fabs, data centers, power systems, finance, policy, supply chains and logistics to line up. Altman said this may already be, or may be becoming, the most expensive infrastructure project in human history.
He also argued that his earlier life as a startup investor maps more closely to running a research organization than it may appear. The overlap, in his view, sits in areas such as identifying non-consensus bets, judging conviction, understanding exponential growth and managing outlier talent.
OpenAI’s strategy: one interface and one API
On products, Altman said OpenAI had just merged ChatGPT and Codex. Before that, the company had ChatGPT, Codex and the API as separate lines. He added that the Codex name often led people to think it was only for coding, when in fact it could handle any kind of work.
His preferred framing is that OpenAI should be a platform company more than a product company. It will still build products, he said, but what most people really want is a unified interface that gives an individual or a company access to their AGI, plus an API that lets others build whatever they want on top.
Altman said OpenAI wants to serve every point on the cost-performance curve. If a user needs high-end AI for scientific discovery, that should be available. If the need is lower-cost AI for large amounts of repetitive work, that should also be available.
Why OpenAI cut Sora and the Atlas browser
Altman said one of the hardest lessons for founders is learning to kill good ideas. As an example, he said OpenAI cut Sora last year and also cut the Atlas browser.
He described Sora as fun, cool and used by real people, but too demanding on compute relative to the company’s priorities, making it a worse use of scarce resources than putting that compute into Codex. He said the same resource logic applied to Atlas. In his words, it may have been the best browser, but the people working on it were more valuable elsewhere.
With compute, talent and capital all limited, Altman said OpenAI has to stay intensely focused on general intelligence, knowledge work, scientific discovery and the upstream capabilities needed to support those goals, including custom chips, data centers and infrastructure.
He also revisited the early period after ChatGPT launched. Growth was fast, he said, but the product still felt unstable and many people questioned how durable its value really was. He went to Peter Thiel for advice. According to Altman, Thiel listed a number of directions the company could pivot toward, then pointed out that, apart from the fact that ChatGPT was already growing, the more obvious mistake would be to go do something else.
Altman said Thiel’s point was that ChatGPT’s strength sat in the Google-like blank text box. A user could type anything, and the system could handle it. That model did not fit the dominant Silicon Valley ideas of the time around feeds, network effects or user lock-in. But if a blank box worked for Google, Thiel’s argument went, there was no reason to dismiss it here. Altman said he took that advice and pushed hard in that direction, and that it worked well.
The two AI risks he worries about most
Altman said his two main concerns are loss of control and excessive concentration of power. The first is the risk that AI becomes so strong that humans can no longer guarantee the level of control they want. The second is the risk that too much power ends up in one company, one model or one person.
He called both outcomes anti-human at a basic level. The right path, he said, is one where people remain deeply in control of the future and are deeply empowered by the technology, because “humans are the whole point of all of this.”
Altman rejected the idea that distrust in humanity should lead society to hand more and more control to AI models. He also rejected the version of AI safety politics that responds to risk by concentrating access in a few companies. He said it is a bad sales pitch to tell people they can have cured diseases or cheaper goods only if they give up autonomy and influence over the future.
Asked why that kind of pitch appears at all, Altman said fear and power are the main drivers. Some people are genuinely frightened by the scale of AI risk and conclude that large amounts of freedom must be traded for safety. But he said that line of thinking can also become an excuse for power-seeking.
At the same time, Altman said he agrees with part of the doomer argument: AI is powerful and calls for caution with a strong safety bias. What he does not accept is the claim that the problem is unsolvable. He said there were two broad beliefs when OpenAI was founded: that no very AGI-like system would be built within 10 years, and that even if one were built, safety could not be maintained. In his view, the world now has systems that many people at the time would have considered very AGI-like, and neither social collapse nor alignment failure has followed. Forecasts should be updated accordingly, he said.
He says the industry has failed to explain AI well to the public
Senra raised the tension between broad AI usage and equally broad discomfort with AI. Altman said people are always afraid of rapid socioeconomic change, and the Industrial Revolution was no exception. In that sense, he said, social inertia can even serve as a stabilizing feature during periods of disruption.
But he also said the AI sector, including OpenAI, has done a poor job explaining both the upside of the technology and the ways its risks can be reduced. The industry has spent too much time repeating lines like “there is a 25% chance it destroys the world” or “50% of jobs disappear next year,” he said, and then acts surprised when the public gets scared.
The message that has not been explained well enough, in his view, is that AI can give people more power and more personal freedom rather than less.
Altman said he believes the next phase will bring the biggest boom in small-business creation in history, with AI helping lower the barriers that once required privilege, luck and substantial resources. He added that the sector has barely communicated that possibility.
When Senra compared this need for education to Intel’s early efforts to teach customers and investors about microprocessors, Altman said the AI industry has no excuse and should do much more, even if it has not yet found the most effective method.
OpenAI spent four and a half years with no product
Altman said Y Combinator shaped not just how he runs companies, but the wider startup ecosystem. He highlighted principles associated with YC: iterative deployment, technical founders in charge, willingness to back young people with energy and ambition, shipping embarrassingly early versions, and learning from real user feedback.
OpenAI, he said, was a product of that philosophy in many ways. But on one major point it did the opposite. The company was founded at the end of 2015 and did not release its first product until the middle of 2020, a gap of four and a half years. YC would normally tell founders to ship as quickly as possible. OpenAI had no external customer signal, so it had to invent substitutes for product feedback.
During the Dota 2 period, he said, the team built a leaderboard that let different ideas compete publicly in ranked form, giving researchers an objective signal of progress. OpenAI also learned the value of external demos that put pressure on teams by inviting important outsiders to see results. Some methods did not work, he added, and fake deadlines were one example.
The first OpenAI gathering started with excitement, then silence
Altman recalled that in January 2016, roughly a dozen people gathered in Greg Brockman’s apartment. The mood at first felt like the first day of school. Then the group ran into a basic problem: everyone knew they wanted to build AGI, but no one knew what to do next. There was not even a whiteboard in the room.
He said Brockman sent someone to find one. When the whiteboard arrived, the room still went quiet. The initial energy dropped fast. From there, the group simply started with what it knew how to do: write papers, generate ideas and try things.
Altman described the following years as a long period of chaotic stumbling. Over time, the company found a rhythm for evaluating research bets and got better at making sure smart people were not wasted. The path, as he laid it out, ran from the unsupervised sentiment neuron work to GPT-1 and then to scaling laws, which eventually gave the team confidence to buy more compute.
He says success teaches him more than failure does
Near the end of the interview, Altman said he has learned more from success than from failure. Failures do contain lessons, he said, but most failed outcomes have many causes, which makes it hard to extract the right causal story. When something truly works, by contrast, the usable lesson is clearer. He cited effective practices from YC and OpenAI as examples of things that can be carried into the future with more confidence.
Writing letters to his son became a framework for honest reflection
Altman also spoke about his life outside work. After his first child was born, he said, he would come home at night, put the baby to sleep, and talk through what he had done that day and what was worrying him. Later he decided those thoughts might be interesting for his child to see one day, so he started writing a letter every Sunday.
He did not keep the habit for long. By his own count, he wrote about eight letters and then stopped.
Still, he said the exercise had one important effect: when you write to your child, you cannot hide behind anything. Because you care how that child may see you in the future, you are pushed into a more honest accounting of what went badly that week and what you want to change the next one. Altman called that a very interesting mental framework.
