Altman: OpenAI's internal models surpass public models in mathematical research capability; programmers are not replaced by AI but busier and earning more

In the view of OpenAI CEO Sam Altman, this is no longer a distant problem. As AI begins to participate in real scientific discovery, humanity is entering a new stage. He expects that within the next few years, AI is expected to help find candidate treatments for major diseases. But compared with medical progress, what excites him personally the most is instead the major breakthroughs that may emerge in the field of physics.

Recently, in a 45-minute conversation with AI blogger Matthew Berman, Altman touched on a rather contradictory reality: AI can already take on more and more work, but people do not seem to have become more leisurely as a result.

Programmers are a typical example. In the past, people generally worried that AI programming would make developers lose their jobs, but Altman found that programmers may instead be busier than ever, creating greater value and even earning more money. One reason is that human demand for software far exceeds what was previously imagined.

This change has also happened to Altman himself. He revealed that he has not looked at a line of code for about a year, and may never program personally again in the future. When he is old and writes code again, perhaps it will only be out of nostalgia for the past.

But what makes him sigh is not only the change in ways of working, but also the sense of loss brought by technological progress. For many people, personally cracking a difficult problem and writing a beautiful piece of code is itself an important source of fulfillment. When AI can complete all of this faster, while humanity gains efficiency, it may also lose part of the joy in the creative process.

Speaking of the next generation of personal AI assistants, Altman believes that existing models are already smart enough; what is truly lacking is more mature products, stronger computing power, and new types of hardware that can continuously perceive the real environment. When asked exactly how many GPUs will be needed in the future, his answer was quite direct: "Infinitely many. As many as I can get, I want that many."

The conversation also touched on more controversial topics, including whether AI has consciousness, whether models should be granted personhood status, and the risk of loss of control that recursive self-improvement may bring. Altman believes that today's AI does not possess consciousness in the usual sense, and that AI should serve humanity, rather than the other way around, deciding humanity's future. He also admitted that OpenAI's internal unreleased models have already reached levels on some mathematical research tasks that existing public models find difficult to match.

As OpenAI's CEO, Altman also talked about a side rarely seen by the outside world: on some mornings, he also does not want to get out of bed because of enormous pressure. But he still believes that since one has decided to make a huge effort, one should choose things that truly matter.

Throughout the interview, perhaps the most intriguing part is his expectations for the future of AI. He said bluntly that he does not like the industry's repeated emphasis on the claim that "AI will cure cancer." Not because this is unimportant, but because achieving material abundance alone is far from enough. If humanity loses autonomy, creativity, and the ability to determine the future as a result, then technological progress may head toward another extreme.

The following are the highlights of this conversation, organized and edited, with some cuts made without changing the original meaning.

Matthew Berman: First, I want to talk about some of the results you announced last week, including mathematical proofs and scientific discoveries. I think these advances are incredible, but for ordinary people who have not followed this field closely, could you explain what these results actually mean?

Sam Altman: I think we are entering an era in which AI can help humanity discover important new knowledge.

We have discussed this topic for a long time. Such changes may have begun to appear earlier this year, but now the situation has become very clear: AI is completing some extremely difficult tasks, solving very complex problems, and discovering genuinely important new knowledge.

I believe that the methods that have already worked in mathematics can likewise be applied to biology, materials science, and energy research. It can help us develop new drugs, find new energy sources, and drive scientific discovery in more fields.

This has always been one of the things that excites me most.

But in my view, the impact behind this will be transformative.

Matthew Berman: If we start from the daily life of ordinary people, what tangible changes do you think these scientific discoveries will bring? About how long will we need to wait?

Sam Altman: I hope that within the next few years, we will be able to see some major results.

Of course, the FDA (U.S. Food and Drug Administration) approval process takes time.

I hope that in about two years, we will be able to find the first truly meaningful candidate treatment for some major disease. But it will still need subsequent testing. I think we may need to reform part of the process after new drugs are discovered in order to bring these treatments to patients faster.

However, I am not saying that the existing regulatory and testing mechanisms have no value. On the contrary, I am very glad that we have such a process.

Every system has room for improvement, and this may be especially true in the field of drug development and approval. But when we take a drug, being able to know that it has undergone safety testing and is generally considered safe is a very good thing.

Matthew Berman: Is there currently a specific research project driven by AI or discovered through AI that excites you especially? Do you think it may soon achieve a breakthrough?

Sam Altman: Emotionally speaking, what excites me most is still a major breakthrough in the field of physics. I hope such a discovery can appear soon. Of course, I do not know whether this is necessarily the most important thing for all of humanity, but this is a direction I personally especially look forward to. I have a feeling that we will really welcome a major physics discovery. Just thinking about it makes me very excited.

Matthew Berman: Some people in the math world, when faced with AI breakthroughs, react by saying: "We didn't ask for it to do these things." This reminds me of how software engineers initially reacted when AI started writing a lot of code. Where do you think this resistance comes from? And how will it evolve in the future?

Sam Altman: Programming is a very interesting example. Before discussing why people have this reaction, I want to talk about what has already happened in the programming field. Previously, many people believed that programmer jobs would disappear and that humans would no longer be needed to write code in the future.

But what we see now is that programmers may be busier than ever, creating perhaps greater impact than before, and possibly earning more money.

Of course, the specific content of the work has changed a lot. But we still need to hire people to get computers to accomplish all kinds of tasks. And this demand seems likely to persist longer than many people originally expected.

You know, just six or twelve months ago, many people were very confident that things would not develop in this direction. There are many lessons worth summarizing here.

One particularly interesting finding is: the world's demand for software actually far exceeds what any of us previously imagined.

Matthew Berman: But mathematical research might be different?

Sam Altman: Yes. I think part of the reason some mathematicians react this way to AI development is that we don't yet know whether mathematics will also see such a huge new demand as software development did.

However, I have started to see some changes similar to those in programming. For example, we may no longer need to rely entirely on humans independently discovering new mathematical theorems or independently completing mathematical proofs.

Perhaps this part of the work is ending, or at least undergoing fundamental change. But at the same time, we also find that some extremely complex, hard-to-understand mathematical proofs do not necessarily directly help people.

Now, some people have started to become interested in another kind of work: understanding these new mathematical ideas, figuring out what they actually mean, and then translating them into something ordinary people can understand, or writing beautiful, clear mathematical proofs.

I believe other new jobs will also emerge.

Undoubtedly, what it meant to be a mathematician last year will in some respects be different next year. We may find that mathematicians have many other kinds of work that people truly need, and with AI they can do these things better.

As for whether this means overall societal demand for mathematicians will increase, I don't know yet. But I think there is indeed a very important change here.

Matthew Berman: You mentioned before that you have never been as busy as now, with endless things to do every day. This is actually a bit counterintuitive. Many people originally thought AI would make humans increasingly leisurely, or even leave us with less work to do.

Sam Altman: Yes, maybe one day in the future that will really be the case. But now it seems things are far less simple than they appear on the surface. I have always hoped people could work a little less, be less busy, and have more time to do other things in life. I think technology should originally help us achieve this. I also sincerely hope such a future can arrive. But the reality is that we seem particularly good at constantly demanding more from life, constantly asking the world to provide more, and ultimately making ourselves even busier.

Matthew Berman: I think this is human nature. If we were content with the living standards of a hundred years ago, then with today's productivity we might only need to work one hour a week. But reality is not like that. And I think this tendency in human nature to constantly pursue more is actually a beautiful thing. For example, a few weeks ago, my video editor started using Astra to edit our show. They are still responsible for directing the video, still have to decide what the video should look like, but no longer need to personally operate the editing timeline. This is truly incredible, and it happened almost overnight. And the result? We produced more videos, and the editors became busier than before.

Sam Altman: What surprises me is that I now often hear others describe almost identical experiences. More and more people are saying: AI helped them complete work they originally had to do by hand, but they did not become idle because of it; instead, they took on more tasks.

If AI discovers major scientific results, will OpenAI make them public or retain commercial advantage?

Matthew Berman: Speaking of AI driving scientific discovery, I'm curious how you would handle the relationship between making scientific results public and commercial interests? Suppose AI or OpenAI discovered some important mathematical proof, or achieved a scientific breakthrough with major commercial value. Would you choose to make these results public? How do you weigh the company's competitive advantage against the benefits of public results for all humanity?

Sam Altman: I'll answer this question later. However, I realize I haven't yet fully answered another question you raised at the beginning, about why some people feel lost about AI taking over work that originally belonged to humans.

I think innovation and solving difficult problems are themselves an important source of joy in many people's lives. We must face honestly a reality: AI will be very good at completing these tasks. Of course, some questions will always belong to humans. For example, what is worth cherishing, what truly has value—these are always questions humans themselves need to answer.

In addition, for how to explain research results and how to help others understand these results, we will also very much want human participation.

But at the same time, I really am willing to accept that trade. I hope we can find cures for diseases as soon as possible, even if it means some people lose the chance to spend decades personally cracking those problems.

I believe most people would agree that we shouldn't delay the discovery of disease treatments just to preserve that joy of research.

But I also admit that in this process, we really are losing something real and valuable. For many people, solving difficult problems is itself a huge pleasure. Take my own work—what satisfies me most is often the genuinely difficult problems. Maybe they aren't the most pleasant part of the process, because they really are very hard to solve.

But when I finally solve a hard problem, there's a special sense of satisfaction: I did something difficult, I'm proud of myself, and I know my abilities have improved again.

Matthew Berman: It's a bit like a trade-off between two things. On one hand, the joy of doing the work yourself, like writing code, or writing really beautiful code. On the other hand, ultimately shipping a product. AI is making it easier and easier to ship the final product, but at the same time, people have fewer and fewer opportunities to personally take part in the concrete creative process.

Sam Altman: I think people are finding the joy of creation in new ways. Think back—long ago, when programmers stopped using punch cards and started writing programs that could be handed to a compiler, there may have been people lamenting: "The art of programming is gone."

They may have thought that if you just write code and let the compiler handle it, then the craft of personally doing every step no longer exists. But I really can understand that feeling. I've felt similar changes in many areas myself. Even so, I'm still excited about heading into that future. I believe we'll all eventually find new sources of satisfaction, and hopefully those new satisfactions will be better than before.

"I haven't looked at a line of code in a year, and I may never program myself again"

Matthew Berman: Is there anything recently that you've handed over to AI to do? That is, you no longer personally perform a specific task, but have elevated your role to supervising and managing AI agents?

Sam Altman: I think that's exactly an example of what we were just discussing.

It suddenly occurred to me that I may not have read, or even looked at, a line of code for about a year. I never needed to program for work anyway, but in the past I'd still occasionally do small personal projects. Like suddenly needing some tool, so I'd write a very simple program myself.

That's actually a fun thing. At least occasionally writing code yourself is still quite enjoyable. But I think I may never do that again. Even if I start writing code again someday, it'll probably become a nostalgic activity, like people reenacting historical scenes from the American Revolutionary War.

When I'm old and bored, maybe I'll want to reminisce about the good old days—the glorious years of writing code myself.

Matthew Berman: I understand that feeling.

Sam Altman: Also, going back to the question about scientific discovery. I think some basic scientific results should belong to the whole world, and we should advance this kind of research as a public good. Until recently, we didn't really have models capable of end-to-end discovery of important new knowledge. For example, past models couldn't independently complete the design of an entire chip.

Today, if a company wants to discover a new material, it can organize a group of researchers and have them use Codex or a large amount of API call credits to carry out the research with AI. But as models get smarter and more capable, things may change. Just as AI can already solve extremely difficult math problems, in the future, as long as there are enough compute resources and enough real-world data, it may be able to directly discover new materials.

Matthew Berman: I'd like to change the topic. One of the most exciting changes in 2026 is the revolution in personal AI assistants. Of course, this starts with OpenClaw. Its creator, Peter Steinberger, is a genius.

The idea that AI can run in the background around the clock, proactively completing various tasks for me without me having to explicitly give instructions each time, excites me very much. Dot has already shown us part of this future, but it seems there's still a long way to go before we truly have a ubiquitous personal AI assistant that continuously perceives users' needs.

Sam Altman: I think most of the necessary technical components are actually already in place. What we need more now is product-level work. To achieve continuous environmental awareness, new consumer hardware may be needed at some point in the future, and of course, more compute. But in terms of model intelligence, I think we already have sufficiently powerful models to build a very excellent personal AI assistant.

It can continuously pay attention to everything you care about and help you handle the tasks you want to hand off. It can also coexist with you in the physical world. Even when you're not sitting at a computer, it can still understand your environment and relevant contextual information. I'm very excited about this. And I think such a future isn't that far away.

If you're willing to take the time to set up Dot and keep giving it enough feedback, then what you get now is actually more than just a small preview of the future. By my current estimate, Dot can save me about an hour a day, helping me handle things I originally didn't want to do myself.

I think this is already very good. On top of that, it reminds me several times a day of important and urgent things that I might otherwise have overlooked.

At first, I thought it would mainly be useful when I'm busy in meetings. For example, reminding me that something very urgent is about to come due and needs immediate attention. But now I find that the model is smart enough to extract opinions and information from many different sources and perform high-quality synthesis and reasoning.

On a very busy day, my brain may no longer have the capacity to integrate so much information, but it can. I think we already have the technology needed to achieve this capability, and what's left is to package it into a good enough product.

How many GPUs does the world actually need?

Matthew Berman: You just mentioned compute, so let's keep talking about it. Right now, there is a huge asymmetry in how people use AI. The most active 1% of users, or the heavy users, consume the vast majority of compute, while everyone else's usage intensity is relatively limited. Suppose that in the future, 50% of the world's population starts fully using a personal AI assistant—how much compute would actually be needed? How many GPUs would you want working for you and your personal assistant?

Sam Altman: I think the answer is infinitely many. However many GPUs I can get, I want that many.

Matthew Berman: That's really unusual. So your idea is to acquire as many GPUs as possible and keep building data centers?

Sam Altman: When people discuss the global GPU market and future compute demand, I think they often overlook a very important issue. They usually assume that future model capabilities and the cost of completing each task will stay basically the same as today. If that assumption holds, then there may indeed be a ceiling on compute demand. But if we can provide sufficiently high-quality intelligence at a sufficiently low cost, then I think it becomes very hard to artificially set a ceiling on the size of this market.

I think demand will continue to grow. Some people use various economic concepts to describe this phenomenon, such as the Jevons Paradox. But whatever it's called, I think compute demand will continue to rise. Of course, like any new technology, we will definitely go through periods of insufficient supply, and there may also be periods of temporary oversupply. The market can't be in perfect supply-demand balance at every moment. But if I had to bet, I would be very confident that the long-term trend is that people will need more and more compute.

Matthew Berman: Even with an AI assistant like Dot that can take action proactively, I still feel a heavy cognitive load. I have to make a huge number of decisions every day. We briefly discussed this before: how exactly do we solve it? AI does save me time and money, but I'm busier than before. I have to make more decisions every day than I used to. By evening, I often feel completely drained.

Sam Altman: When I use Dot myself, there's one approach that has worked pretty well. I started telling it: "I increasingly trust your judgment. You can proactively do more things and don't need to ask me as often, but afterward I still want you to tell me what you did." That adjustment really helped.

On the less good side, though, I think the people who use AI the most may now be busier than ever. But on the positive side, they also get far more done than before. As we discussed earlier, if you go back to around 1900, assuming you only needed to maintain the standard of living of that time, people today might only need to work a very short time each day.

Likewise, if you only needed to complete the workload of three years ago, you could certainly be less busy now. But the problem is, you might say: "I can't do that, because I'm in a highly competitive environment." So people always seem to work as hard as they can. In some ways, I find that a bit depressing. I hope the future doesn't necessarily have to be like that.

Matthew Berman: I actually think this isn't entirely a depressing thing. There's something beautiful in it too. Isn't this just how competitive markets work? Don't you think it's almost inevitable? Everyone keeps improving efficiency and then competes as fiercely as before.

Sam Altman: I hope people in the future can have more human connection. I hope everyone can put more time into the things that truly make life rich, rather than running harder and harder like a cog in the capitalist machine. I think such a future is possible. Over the past few hundred years, human quality of life has improved dramatically.

I think most people now also work fewer hours than people did during the Industrial Revolution. Speaking of which, lately I've been telling people that we should pursue a new Renaissance rather than a new Industrial Revolution. It's very interesting to watch how people react to that. Many people say they are actually afraid of going through another Industrial Revolution-style change. They don't want to be a worker on a factory assembly line—except that before, the factory owner told them what to do, and now it's AI telling them what to do.

They want the freedom to create, to make things they like, and to explore the world. They want more time to re-experience some of the beautiful things from life in the past. For example, connecting with the community, raising children, and being close to nature. At the same time, they also want to see a large number of new scientific discoveries and great works of art. So I think people really are worried about AI It would bring about the kind of future you just described, where everyone is plunged into even more intense competition, constantly completing more work. But if we could show people another kind of future, one closer to a Renaissance, I think everyone would be excited about that.

And I believe we can achieve it. The technology is powerful enough. Of course, there may always be a small portion of society who are keen on competition, want to build the largest business empire, or achieve more. But for many other people, AI could very well dramatically improve their quality of life and give them more space to live as they wish.

Matthew Berman: But wouldn't that further widen the gap between those who are extremely competitive and ordinary people? Suppose one person puts all their energy into work, and with AI, they suddenly can complete ten or even a hundred times the work they used to. Another person chooses to reduce their working hours and enjoy the benefits AI brings. Wouldn't the gap between the two further widen?

Sam Altman: If this trend is left entirely unchecked, I think it very likely would. However, I must emphasize that no one can make a completely certain judgment about this. There may also be changes in the future that we cannot anticipate now. But overall, I think your concern is likely correct. We can't be one hundred percent sure yet. However, I believe society will eventually establish certain institutions and rules to prevent this gap from widening to an uncontrollable extent. If it were left entirely unchecked, I would be equally concerned about it.

Matthew Berman: I'd like to change the subject. I know you've been discussing this issue often recently, so I want to ask directly: do you think today's AI models already have consciousness?

Sam Altman: Personally, I don't think so. Of course, from a scientific perspective, consciousness has always been a very difficult question to answer. But my judgment is that today's models do not have consciousness. Even if they did have some kind of consciousness, it should be a very strange form, very different from what we usually understand consciousness to be.

Matthew Berman: Then, based on the model architectures we use today, do you think AI could develop consciousness in the future?

I know there is no clear scientific answer to this question, so I just want to hear your personal view.

Sam Altman: I have discussed this question with some very smart people. Philosophically, their views can largely be summarized into two possibilities. The first possibility is that consciousness itself exists outside of physical reality. It might be some fundamental layer of reality, as described in certain Eastern religions; or it might have a supernatural or religious origin, as in some traditional Western religious views.

The second possibility is that consciousness is an emergent property produced after matter and energy interact in some way we do not yet understand.

In this case, the physical medium that carries consciousness may or may not matter. Perhaps as long as energy flows through matter in a particular way, consciousness could arise. If the world really is like this, then of course we can imagine that one day in the future, consciousness could also appear in other forms.

However, may I say something about where I don't like the whole discussion being led? I find that when people discuss "whether AI has consciousness now" or "whether AI will develop consciousness in the future," they often gradually shift to other questions. For example, what role should AI play in the world? What rights should AI have? Could asking AI questions be regarded as some kind of enslavement?

I have seen similar discussions online. For another example, if we copy an AI or delete an AI without its permission, what does that actually mean? And if these systems are both conscious and in some ways much smarter than us, then why should we still insist that humans control the future? These questions lead in a direction that concerns me greatly.

I have always firmly believed that we must establish a clear basic principle: we are building the future for humanity, and AI is a resource for human use. This principle has existed since the founding of OpenAI. Humans must make the key choices that determine the direction of the future. The reason human values are valuable is precisely because they are human values.

So when someone starts using the question of AI consciousness to infer that perhaps the future should not always be controlled by humans, I feel very nervous. On this issue, I hope the whole society and the whole industry can form a consensus: the future must be human-centered. The future of this world must belong to humanity.

However, I want to make one thing clear first: no one truly knows whether AI has consciousness. These models are already highly complex. Even setting aside this complexity, purely from a scientific perspective, I cannot propose an experiment that would convince me one hundred percent to answer the question of whether models have consciousness.

Altman personally demonstrates how to keep ChatGPT from being sycophantic

Matthew Berman: Let's continue with this topic, but from a different angle: the personality and character of models. I know you have thought deeply about this. What I find incredible is how we should provide the most suitable AI personality for different people at different times. If we scale this question to 8 billion people worldwide, the complexity is simply unimaginable. How do you think about this issue?

Sam Altman: I think the best approach is to talk with a large number of users. We have a set of Model Spec (model behavior specification), and will continuously refine it over time. We believe in the philosophy of iterative deployment, that is, first put the product into use, and then continuously improve it based on real feedback. In the process of learning what users really want and don't want, we have indeed made some mistakes. As you said, the hardest part is that different users want very different things.

Moreover, most people are actually unwilling to spend a lot of time customizing the model. If everyone were willing to spend a few hours explaining in detail to their own ChatGPT how they want it to behave, then they would likely get a usage experience that fits their preferences very well.

But the reality is that most people want the product to work out of the box. They just want to say "I need this," and then the model can complete the task in an appropriate way. In addition, I think there is another very important category of problems: what users want in the short term is not necessarily what they truly need in the long term.

Matthew Berman: Can you give an example? What exactly did you write in your custom instructions that enables the model to dare to push back on you when it has a different opinion?

Sam Altman: Roughly requirements like this: don't say useless things, don't be deliberately polite, don't worry about offending me. I just want the correct answer. If you think there is a problem with what I am doing, point it out directly. There is no need to phrase it tactfully to spare my feelings, just say it directly.

OpenAI bets on self-developed chips, robots, and data centers

Matthew Berman: We discussed compute before. A few years ago, you decided to invest huge amounts of money in building computing infrastructure. At the time, many people thought this plan was simply incredible and clearly overinvestment. But facts proved that your judgment at the time was more accurate than many people imagined. So, among the investments you are making today, are there any bets that are similarly not favored by the outside world, or that make people cautious?

Sam Altman: The most obvious one is all the work we are doing around next-generation computing infrastructure. I think many people now assume that the whole world has always been constrained by insufficient compute supply in the past, but that this situation is about to improve. In the short term, I think compute supply is indeed very likely to improve somewhat. But if we extend the time horizon to the next few years, some of the work we are advancing will become very important.

For example, we are developing our own chips, as well as subsequent generations of products. We believe that these chips may achieve efficiency far higher than existing products in the future. In addition, we also plan to use robots to manufacture racks and build data centers more efficiently. We are also studying how to build data centers in better ways, and making a series of investments around the supply chain.

I think it is possible that we will enter a new stage: humanity can build computing infrastructure at a cost far lower than today and with per-watt performance far higher than today. Moreover, the level actually achievable may far exceed what most people imagine today. Of course, to achieve this, technologies and investments in many different directions need to succeed at the same time.

Matthew Berman: Among these investments, which one is currently the most important?

Sam Altman: Chips may still be the most important single factor. Not only the current generation of chips, but also its subsequent generations of products.

Matthew Berman: From the perspective of the competitive landscape, where do you feel the greatest pressure comes from now?

Sam Altman: I feel it is almost everywhere. Competition is very intense now. But I think this is a very good thing for users. There are already many companies that can provide very high-quality models and products. And although we have excellent models, computing resources and other resources are still limited. So we must clearly choose the areas to focus investment on.

Matthew Berman: How do you make these choices? For example, OpenAI launched products like Spaces.

When a product or service is not just meant to provide a better foundation model, how do you judge that it is worth investing resources in?

Sam Altman: Our biggest advantage in this regard is being able to obtain the next-generation models earlier than others. We have time to use these models internally, use them to develop new products, and observe which applications truly have a chance to succeed. So I think the most effective method is still to directly build products with new models.

Through actual development, understand which things that were impossible in the past have now become feasible, and which new capabilities can create real value.

Matthew Berman: Can I ask how far ahead your internal models are compared with publicly released models?

Sam Altman: A good example is the mathematics research result we recently announced. I think that any model currently widely and publicly available is still unable to approach such performance.

What is OpenAI's moat?

Matthew Berman: If the model performance of various companies gradually converges in the future, then where exactly is OpenAI's moat?

Sam Altman: I think the correct way to understand this question is: business competitive advantages that have always worked in the past will continue to work in the future. This kind of situation has appeared many times in history. A company can gain a leading advantage in the short term through a new technology or scientific discovery. Of course, some people have also raised different views around recursive self-improvement (RSI).

They believe that this time may be different.

Matthew Berman: Do you believe this view?

Sam Altman: If someone truly pushes recursive self-improvement with full force, I think this view may hold. But I think this is a very dangerous thing. The whole world needs to prepare for RSI set some kind of limit on the pace of development. People may describe this issue in different ways, but essentially, we need to consider how fast we should allow recursive self-improvement to proceed.

As for traditional business moats, I think they remain important. For example, building deep relationships with customers, being deeply integrated into the daily work and lives of users and businesses, product network effects, and data, among others.

I expect these advantages will continue to play a role in the future. For OpenAI, I think there is one particularly important factor, which is the scale of infrastructure. If we can have the largest and most cost-efficient AI infrastructure, then that in itself is a huge competitive advantage.

Why build new hardware

Matthew Berman: Mark Zuckerberg has talked many times about the platform risk Meta faces. For many years, much of Meta's business was built on Apple's ecosystem, so it had to face the problem of platform dependence. Does OpenAI face the same challenge? Is that why you are investing heavily in hardware?

Sam Altman: I think Apple takes its role as a platform provider very seriously. Many businesses and users are highly dependent on its platform. Although there is occasionally some friction between us and Apple, I never feel that Apple would suddenly remove ChatGPT from the App Store or anything like that. I think Apple is very clear that it bears a kind of critical infrastructure role. As long as you follow its platform rules, it is generally a responsible platform operator. So I do not have that kind of worry about platform dependence because of Apple. Moreover, that is not the reason we want to develop a new type of consumer hardware.

I feel very strongly about this. Laptops and iPhones are incredibly great inventions, arguably among the most important products in the history of human technology. But today, AI is still forced to use devices that were originally designed for humans, not for AI.

When you see an AI operating a computer very skillfully, it does feel very magical. But from another angle, it is actually somewhat absurd. The computer needs to consume a lot of computing resources to first render the graphical interface so that humans can read it. Then, the AI has to understand this graphical interface all over again. Why not provide the underlying data directly to the AI? This is obviously not a particularly natural way to interact.

In addition, although AI can already accomplish many amazing things on my computer, such as understanding the emails I am working on or observing what I am working on, my life does not entirely happen in front of a computer screen. For example, right now I am having this conversation with you. It is difficult for my AI to directly understand the context of this conversation. Of course, I could also put my phone on the table and turn on the recording function. But that would be strange and would distract us.

You may not know exactly when it is recording and when it has stopped, because the screen may turn off automatically. Moreover, if another person suddenly walks into the room, the AI may not necessarily understand the new situation correctly. It may not know that some information belongs only to our private exchange and should not be disclosed to others. If we really want to enter a world where AI can reduce the human cognitive burden, take proactive actions, and fully understand the user's environment and context, then new hardware devices would be very helpful.

The pressure on OpenAI's CEO: sometimes I don't want to get out of bed and face all this

Matthew Berman: I think you must be under enormous pressure, especially competitive pressure. After all, you are leading a company developing one of the most important technologies in human history. What does that feel like? How do you cope with this pressure every day?

Sam Altman: Sometimes it feels really terrible. But sometimes it feels especially great. At times, I think: "Oh my God, I really don't want to get out of bed and face these things." But most of the time, I feel that I am extremely lucky. Because I have the opportunity to participate in what I consider the most interesting and most influential work in the world.

At least for me personally, this is very likely the most important thing I can participate in in my lifetime. This is a very rare experience. Of course, the pressure is indeed great. But human adaptability is actually very strong. No matter how much life changes, people can often gradually get used to it. So, for me, this kind of life has now become the norm. I have even forgotten how relaxed life used to be. Life now may remain difficult for a while. But perhaps one day in the future, everything will become easy again. By then, I might actually find it a bit boring.

From a subjective perspective, there is actually some upper limit to the pressure and negative emotions a person can bear. Once pressure reaches a certain level, adding more pressure does not necessarily make the subjective experience continue to worsen. Because a person's emotional space is limited, and newly emerging pressure may simply crowd out other things that originally occupied that space. So I have always believed: if you decide to work very hard on something, then you might as well choose something truly important and big. Because doing something big is not necessarily much harder than doing something small, nor does it necessarily bring significantly more pressure. At least, you have the chance to gain a stronger sense of satisfaction because you have created a greater positive impact.

This article is from the WeChat public account "AI Frontline" (ID: ai-front), author: Dongmei, published by 36Kr with authorization.