Tool iteration is evolving rapidly; only by thinking outside the old framework can one adapt to an exponentially changing era.

Video source:Sam Altman - How to Start a Startup

Translation: Deep Think Circle

Have you ever thought about how a startup established just two weeks ago could recreate a whole suite of mainstream office software, with documents, spreadsheets, and presentations all redesigned around AI? This is not just my imagination; it’s something Sam Altman (co-founder and CEO of OpenAI) said in a recent interview where he revealed he had just met such a company. Ten years ago, it was almost predictable what a ten-week-old startup would look like. Now, if a ten-week-old startup still looks like those from a decade ago, it indicates that they are already falling behind.

This interview was rich in information; it discussed entrepreneurship, how OpenAI has progressed over the years, and how he handles stress and makes trade-offs. I整理了我最受触动的几段，对它们进行了很多个人理解，写给你看。

Most people are still choosing easy battles to win

Altman said this is a fascinating time; costs are rapidly decreasing, and the cycle of doing things is quickly shortening, which is precisely when startups have the most advantages. This phenomenon is occurring across many fields, so theoretically, it should be the best era for entrepreneurship. However, he observes a rather contradictory phenomenon: most startups are still doing the same thing, creating AI agents for specific industries. While this path can be feasible and even highly profitable, it is unlikely that they will be the companies that define this era.

What struck me was his following statement: clearly, tools have completely changed, and model capabilities continue to rise, yet people still hesitate to genuinely bet on something that cannot be achieved now but can be achieved in two years. He said the temptation to directly use today’s agents to solve the problems at hand is particularly strong; it’s something anyone can understand, but that is not the path he would choose.

I thought about this; it reflects a matter of patience and belief. Being willing to layout for something that cannot yet be realized now is essentially betting that models will continue to grow stronger and that your judgment on the direction is correct. Most people cannot do this, not because they cannot understand the trends, but because they cannot help but want to see returns immediately.

Believing in exponentials is far more difficult than it seems

Altman mentioned a method he constantly uses: whenever he meets someone new, he mentally creates a coordinate for them to determine where they currently stand. The next time they meet, he observes how far and how fast this person has progressed. He said this aligns with his fundamental belief in assessing model capability advancement; in essence, he has a deep trust in exponential growth, whether it pertains to a person, a company, or a model.

He said if he were still giving advice to entrepreneurs, this is the one thing he would want the other party to truly understand. He also mentioned that the reason this is difficult to generally accept is that the market itself has not yet adapted to the fact that model capabilities will continue to grow exponentially, making it entirely reasonable to start projects that require smarter and cheaper models now.

I believe this segment is particularly worth pondering. Trusting that a curve will continue to rise sounds like a simple principle, but putting all decision-making on that belief requires much more courage than one would expect. Most people’s intuitions about exponentials are wrong; they either underestimate the accumulation of the previous years or start to doubt whether the rapid climbing phase is about to stall.

Enduring chaos can only be learned through experience

There was a part I found particularly profound when Altman said that no matter how clearly someone understands a principle, certain abilities can only be truly mastered through repeated experience, operating amidst chaos and then believing that one can ultimately resolve it; it won’t cost you your life. While you may not know how to solve it right now, you eventually will. He said this is something that can only be learned and cannot be taught, and he believes this is the biggest shortcoming of many young founders; they have not gone through the process of slowly learning to coexist with chaos.

He also provided a very straightforward analogy. The first time you encounter a potentially catastrophic event for the company, it feels like the sky has fallen. But after you’ve survived the tenth time, you feel, I've endured the previous nine times; this time is likely not so bad. He later figured out that bad things will always happen. Instead of resisting, it’s better to learn to accept this uncomfortable process. He said that most people think the opposite of a bad experience is a good experience; in reality, the opposite of a bad experience is no experience at all. In the not-too-distant future, you will inevitably enter a phase where there are no bumps at all, so even terrible experiences deserve gratitude.

This sentence left me momentarily stunned. We are so accustomed to treating pain as something to be avoided at all costs, but if the opposite is not comfort but emptiness and numbness, then enduring chaos becomes not just a price to pay, but rather a part of being alive.

What does a trustworthy company promise the world?

When Altman talked about mission, he brought up a concern of his: one of his biggest worries about AI risks is a small group of people or a company thinking they should control the entire world, which he calls AI authoritarianism. Therefore, what OpenAI aims to do is to make intelligence extremely abundant and cheap, placing it in the hands of everyone rather than hoarded by a few individuals. He emphasized that they do not plan to develop every product in each vertical field but rather want to build foundational capabilities in intelligence, allowing the entire economic ecosystem to generate various products based on this foundation.

There was a section I found particularly interesting; he said that the products required to create abundant intelligence, like chips, energy, data centers, and robots, happen to be the very things that humanity will need immediately after intelligence becomes abundant. Even if ideas and creativity become undervalued, we still live in a physical world and need things to be actually created. Hence, energy and robots are not just stepping stones toward that goal; they are also things that will be used immediately after.

When I read this part, my first reaction was that this logic is quite simple; ultimately, no matter how intelligent a solution may be, something must physically move to create any real impact. However, upon deeper reflection, it serves as a reminder not to see intelligence as merely abstract; even the most powerful models ultimately rely on a host of very tangible, very physical infrastructure to materialize.

The invention of the company is more important than many technologies themselves

Altman shared a thought from his childhood; he has always been curious about the industrial revolution, where a plethora of technologies coincidentally emerged at the same time and expanded at a roughly similar pace. He has always wondered which specific technology was the most pivotal. He said from the present view looking back, the truly critical invention is actually the concept of the joint-stock company. Before that, business relied on trust between acquaintances and was made up of family businesses without the concept of shareholders. After the emergence of the joint-stock company, sovereign states granted this entirely new entity an unprecedented status, not giving them the power of a nation but granting them capabilities far exceeding those of individuals, allowing them to pool capital, engage in extremely high-risk and speculative ventures, and let different companies specialize in different segments while coordinating with each other.

He mentioned a graph I want to check out, showing the decline in the proportion of extremely impoverished people and the decline in infant mortality rates throughout human history. If we extended human history and marked the point in time when joint-stock companies were invented, the shape of those curves would evidently change thereafter. He said this is a remarkably extraordinary performance of capitalism in human society.

I find this segment particularly enlightening. When we usually discuss startups, we focus on products, financing, and growth, rarely taking a step back to consider that the organizational form of a company itself is a technological innovation that binds the interests of a large group of people together. Viewed this way, entrepreneurship is essentially utilizing this invention and layering your contributions on top of it.

Trust only a few things, and remain flexible with the rest

When discussing long-term planning, Altman said he doesn’t really work backward from the future to the present. His more accustomed approach is to first determine a few directions he firmly believes in and then step forward from this present point, clarifying what can be done now and what can be done this year, rarely planning for things five to ten years down the line. He said he has seen too many people hold a multitude of beliefs about the future, only to become constrained by their rigid worldview; for instance, when a rocket company suddenly pivots to AI, that's the kind of situation he refers to. The truly effective approach is to cling only to a handful of deeply held beliefs while keeping everything else flexible, firmly maintaining the core.

He mentioned a friend's company’s core values, termed "critical path," meaning to consistently focus on the largest stumbling block at hand, move it aside, find the next one, move that aside, and repeat this action continuously. He said he has been clear over the years that his life’s critical path is to make intelligence abundant; as long as there is no strange concentration of power, he believes this will lead to tremendous prosperity. He rarely gets tempted by other ideas and has not considered changing goals; instead, he has recently started to think seriously about the next steps if superintelligence is indeed on the horizon.

Upon reading this, I felt quite moved; the ability to focus on a single critical path for so many years, almost undistracted by other opportunities, is itself more challenging than any planning methodology. Ultimately, effective planning does not hinge on how precisely one can calculate; rather, it’s about whether one can maintain trust in a few things over the long term while filtering out the surrounding noise.

Are you brave enough to board the plane on the edge of risk?

Altman mentioned he has a principle he always adheres to: in a somewhat risky state, if it's time to board a plane, just board it. He shared a story from the time just after ChatGPT was released, when leaders around the world were quite anxious; some doubted whether this technology was going to spiral out of control, and he could feel a storm brewing. He then heeded Brian Chesky’s (co-founder of Airbnb) advice and decided to take an intensive trip around the world in a short time. Originally, Chesky had done a similar scale of trip across about eight cities, but they extended it to 28 countries in 35 days. He said he practically lived on an airplane during that time; it was a strange experience. Although the seating was comfortable, traveling is inherently exhausting due to time zone changes, longing for one's own bed, and one’s own office.

He also mentioned an interesting criterion to differentiate between true trends and false trends. A false trend appears as a lot of excitement around something, but after a while, people who bought it no longer like it and won’t design their lives around it; ultimately, it collects dust. He used VR as an example. A true trend, on the other hand, means the item remains a part of your daily life; for him, ChatGPT is almost used daily, sometimes for three hours a day, sometimes hardly at all, but it remains a persistent part of his life. He said this distinguishing method was derived from observing numerous startups during his time at YC (Y Combinator), and merely spending time analyzing this data reveals a lot.

I really like this method of distinguishing true and false trends because it's sufficiently straightforward; it doesn’t require analyzing complex growth curves, just one simple question: does this thing quietly embed itself into your daily rhythm, or does it merely excite you for a moment before being cast aside?

Proactively asking can sometimes yield impossible results

Altman shared the example of Codex (OpenAI’s programming agent application), recalling it as a memorable experience where he proactively asked for something. At that time, they were noticeably trailing behind Claude Code (the programming agent product from Anthropic) in that sector. Conventional wisdom suggested that in such a scenario, attempting to reclaim a lead in a category already captured by another was considered nearly impossible, and most would concede and shift to the next direction. But they felt this was too important to give up easily, so they formed a team and entrusted them with what was typically regarded as a self-destructive task. The result was a rare achievement in commercial history; now, among the best programmers around him, this product is the most used programming tool.

He put it bluntly: if they had not proactively asked to assign this nearly impossible task to the team, none of this would have happened. His rationale is that programming is too crucial for RSI (Recursive Self-Improvement), not to mention its underlying economic value; they could not convince themselves to give up this track.

While reading this part, I thought about how straightforward asking for something sounds, but in practice, it involves combating a very strong societal default; the common assumption is that the winner has already been determined, and competing seems destined to fail. What Altman did here was, in fact, to refuse to accept that default, first assuming a possibility before trying.

Projects that need to be killed and the team's emotional state needing a reset

During the interview, they discussed a heart-wrenching question: how does one decide to pull the plug on a project that has already received over a year of investment, considerable funds, significant computation power, and countless human efforts, while users actively engage and enjoy it? Altman stated this is not something that can be resolved in a single meeting; it resembles a slowly accumulating awareness. Eventually, one realizes that those computational resources, those people, and that product direction could create greater value elsewhere, hence the painful decision must be made.

He gave two examples: when GPT-3 (OpenAI's early language model) was finally operational, they terminated an equally exciting robot project, reallocating all resources to this effort. Recently, after the programming agent was successfully operational, they shut down Sora (OpenAI's video generation product) and their browser, both of which had looked promising, and fully committed to programming. He stressed that this does not mean Sora was performing poorly; with continued investment, it could have been successful, but directing computational power and energy to the programming agent was simply more crucial at that moment.

Regarding how to help the team accept such a shift, he said everyone comprehends the mission and understands the trade-offs behind it. Even if it’s tough in the moment, the team knows why this decision is made; while some may be unhappy, many will say, I understand why we must do this; it is necessary for the mission.

I find the most challenging aspect here is not the decision itself, but how to help a group that has already invested a year of effort to re-believe that the next endeavor is equally worth going all-in on. This requires not just good judgment but also strong communication and team leadership skills.

Focus only on what you excel at; finding the right people for the rest

When discussing how to become more proficient at something, Altman cited Johnny Ive (former Chief Design Officer of Apple) as an example. He mentioned that the most important lesson he learned from Johnny is that truly exceptional design is more about thoroughly studying the problem itself rather than coming up with a solution in a flash. If one rushes towards an answer or prematurely locks themselves into a particular solution, the resulting work typically does not turn out well.

He candidly admitted that he does not excel in product development. He disagrees with the notion that one can only bring in people from fields they deeply understand; he himself knows nothing about design, but after just thirty minutes of conversation with Johnny, one can see that this person genuinely possesses remarkable talent. His principle is to expend energy on areas he is already strong in, making those strengths even stronger, rather than forcing himself to compensate for inherent weaknesses.

There’s also a rather personal detail; he mentioned that during the period when he was working on Sora, to understand the product experience, he intentionally made himself addicted to TikTok. Initially, it was just to learn, but later he genuinely enjoyed it, shifting from just ten minutes before bed to an hour, and then spending three hours on a Saturday afternoon on the couch. He stated that while it felt great in the moment, he clearly knew it wasn’t good for him, so he turned off the notifications for most apps, including messaging apps, and eventually deleted TikTok because he realized it was far too powerful for him to control.

When I read this segment, I was quite surprised; here is someone who builds more powerful AI products but can still fall victim to what he creates, requiring the most primitive methods of simply turning off notifications and deleting the app to regain control. It reminded me that judgment is not something that can be established once and for all; it needs constant self-management, even for those who are most familiar with the design logic of these products.

My own reflections

After listening to the entire interview, I felt the central theme was not about specific methodologies but rather an attitude towards facing uncertainty. Believing that exponentials will continue to rise, enduring chaos until it’s no longer terrifying, only holding a few deeply trusted principles, being willing to ask when needed, letting go when appropriate, and investing energy in one’s true strengths. These principles, when viewed individually, are not new; the real challenge is being able to do all these at the same time in an environment where all assumptions are being challenged.

Lastly, Altman said something that resonated deeply with me: most startups today still resemble those from ten years ago because that so-called correct way of doing things is being taught, albeit with new phrasing, like hiring fewer people and spending more on tokens, but that is far from enough. I feel this statement serves as a reminder to everyone: the tools have changed completely; if the way you think remains rooted in an old coordinate system, no matter how radical you claim to be in words, the outcomes are likely still outdated.