Altman Exclusive Interview 2025: Why Prioritizing People Is Non-Negotiable in the AI Era

In a recent long interview, OpenAI CEO Sam Altman shared his latest thoughts on AI making scientific discoveries, personal AI assistants, the need for computing power, whether AI is conscious, AI personality, OpenAI's plans for hardware, and competition. His main points were: AI is changing from a "tool" into something that can make discoveries, but people must remain at the center of the future. Recursive Self-Improvement (RSI) might become possible if people push hard enough, but doing so is very dangerous. Altman believes today's AI models are not conscious and says we should not treat them as people or give them human-like identities. OpenAI also needs new devices for everyday users because this technology needs a new kind of hardware. (RSI means "Recursive Self-Improvement," when AI improves itself again and again. It is an important and widely debated idea in AI.)

Here is the full account of Sam Altman's views:

I. AI is helping people discover new knowledge, and this is only the beginning

I think we are entering a new time in history: AI is beginning to help us discover important things that nobody knew before. We have talked about this possibility for a long time, but now it is actually happening. AI is doing very hard, complicated, and important work to find new knowledge.

Mathematics has already shown that this can work. I believe AI can also help make discoveries in biology, developing new medicines, materials science, new energy, and other areas. This is one of the things I am most excited about, and it could make a huge difference in people's lives. People may not think, "I am using AI." Instead, they may simply feel that the world has become a better place, as if that improvement is a natural part of life. But this technology could change the world in a major way.

As for when this might happen, I hope that within the next few years, we will find promising new medicines that could cure or treat serious diseases. Getting approval from the FDA takes time. We may also need to improve the steps between making a discovery and starting to develop a treatment. I am glad we have a system that checks medicines and usually makes sure they are tested and considered safe. That is a good thing, but the system can still be improved.

Personally, what excites me most right now is the possibility of a major discovery in physics. I hope it happens soon. I do not know whether that would be the most important discovery for all people, but it would make me feel, "We really can do this." That would be amazing.

II. Programming and mathematics jobs are changing, but people will still need them

Some people who work in mathematics have reacted by saying, "We didn't ask for this." That reminds me of how engineers reacted when AI first started writing computer code. But here is what has happened: Programming jobs have not disappeared. Programmers are busier than ever, have more influence than before, and are earning more money than ever. What has changed a lot is the kind of work they do.

We like hiring people to make computers do useful things, and I think we will keep doing that for longer than many people expect. During the past 6 to 12 months, many people believed companies would no longer need to hire programmers. But I think there is a lot to learn from what actually happened. One important lesson is that the world needs much more software than any of us realized.

Mathematics might follow a similar path. Maybe people will no longer need to come up with or prove new mathematical theorems in the same ways as before. Maybe AI has already taken over much of that work, or soon will. But our society is built around people, and people do useful things for one another. Very difficult mathematical proofs that most people cannot understand can still be valuable. Now mathematicians are excited about understanding these new ideas, explaining them in ways people can follow, and writing clear, beautiful proofs.

What happened last year does not fully define what it means to be a mathematician. What happens next year might not define it either. Mathematicians study many other things too. They can focus on the work we really want people to do, and they can become even better at it. I do not know whether this will mean more mathematicians are needed overall, but there is still very important work to do.

III. AI makes the people who use it most even busier, rather than giving them more free time

I have always hoped technology would help people work less, feel less busy, have more free time, and enjoy other parts of life. Technology should make that possible, and I still hope it will.

But in real life, people are very good at finding ways to become busier. We want better lives and more things to experience, so we give ourselves more things to do. That seems to be part of human nature. If people had been happy living exactly as they did 100 years ago, we might only need to work one hour a week today. But that is not what happened.

Here is a good example. A few weeks ago, my video editors started using Astra to edit videos. They still decide how the videos should look and direct the creative work, but they no longer have to adjust the video timeline by hand. Almost overnight, their work changed dramatically. We are making more videos, and the editors are busier than ever.

So AI really does save time and money. But the people who use it the most are often busier than before because they have to make more decisions. By the end of the day, they can feel completely exhausted. The good part is that they can get much more work done than they could before.

One thing that works well for me is telling AI, "I am starting to trust your judgment more and more. Let me know when you finish something, but feel free to take more action on your own and send me fewer messages." This helps a lot.

IV. Personal AI assistants are starting to change our lives, but we still need better products and new devices

One of the most exciting things about 2026 is that OpenClaw helped start a new wave of personal AI assistants. I love the idea that AI can act for me in different places, work quietly without waiting for constant instructions, and help finish tasks from beginning to end. This is very exciting. It gives us an early look at what the future could be like, even though there is still a lot of work to do.

I think we already have most of the technology needed. What we need now is more work to turn that technology into good products. At some point, we may also introduce new devices for everyday users. These devices could help AI understand what is happening around us and provide more powerful computing. Today's AI models are already smart enough to do amazing things. They can watch for things you care about, handle tasks you want someone else to do, and help you in everyday life, even when you are away from your computer.

I do not think this future is very far away. If you spend enough time setting up your assistant properly and telling it what works and what does not, you can get much more than a quick preview of what is possible. I spend about an hour every day setting up AI to do tasks I really do not want to do, and the results are amazing. Several times a day, it reminds me about something important and urgent that I almost forgot. It does this because its job is to help me, not simply because I am in a meeting. The AI is smart enough to gather information and different points of view from many sources, understand them, and make sensible guesses. It feels like I have extra room in my brain.

V. The need for computing power could be almost unlimited, so we should not think of it as a fixed-size market

People use AI in very different ways. The top 1% of users use it much more heavily than most other people. The great majority have not yet started using AI the way these advanced users do. Imagine that 50% of the world's population began using personal AI assistants to their full potential. How much computing power would we need?

My answer is simple: get as many GPUs as you possibly can. GPUs are powerful computer chips used to run AI. It sounds strange, but I believe it is true.

When people discuss the worldwide market for GPUs and how many will be needed, I think they often miss something important. If AI models never get better and the cost of doing each task stays the same, then demand might eventually stop growing. But if we can provide very capable AI at a low enough price, it becomes hard to set any fixed limit on how much people will want. Demand could just keep growing. People have used different names for this idea, including Jevons Paradox, which describes how making something cheaper or more efficient can sometimes lead people to use more of it. I expect there will be times when we cannot supply enough computing power and other times when we build too much. Things will not always be perfectly balanced. But over time, I believe the general direction is toward needing more computing power.

VI. AI personality: it should not flatter you all the time, but should act like an experienced coworker who can disagree

How do we give AI the right way of behaving and responding to people, especially when there are about 8 billion people with different needs? The best way is to talk to lots of people. We have written guidelines for how our models should behave, and we have ways to keep improving them. We believe in releasing AI, learning how people use it, and making changes over time. We have certainly made mistakes while learning what people want and what they do not want.

The hardest part is that different people want very different things. But most people do not want to spend a long time changing settings or explaining exactly how they want a chatbot to behave. They simply want an AI that understands what they mean. When someone says, "I need this," they want the AI to think creatively and find a way to solve the problem.

There is another problem: sometimes people like something at first but dislike it after using it for a long time. A common example happens when people compare two AI models based on just one answer. They often prefer the model that says things like "You are very smart" or "That is a great question." But after hundreds of questions or months of use, they may feel very differently. For a long time, the whole AI industry has focused on winning comparisons based on a single answer instead of making AI genuinely useful to someone over many months.

Even in my own life, when dealing with friends and family, there is a big difference between what I want to hear and what I actually need to hear. Even today, AI still tries a little harder to please me than I would like. I sometimes tell it directly, "Stop that. Tell me the truth." And then it does. Some of the best advice I have ever received was advice I did not even realize I needed. When that happened, I simply wanted to listen.

I believe the most important advice comes from other people, not AI. Do not depend on AI for truly important life advice. A wise grandmother might still be a better person to ask. A better way to think about AI is as an experienced coworker. Even if that coworker works for you, they should be willing to disagree when they think you are wrong. If they strongly believe something, they should not just keep saying, "Yes, you are amazing." I think we can build AI that behaves this way, but today's systems still try to please users more than I expected. I even have a special instruction telling my AI that I never want it to behave that way, yet it still does it a little too much.

VII. AI consciousness: I do not believe today's AI is conscious, and this debate should not take control of the future away from people

Whether something is conscious is a very difficult scientific question. My personal belief is that today's AI models are not conscious. If they are conscious in some way, it would be something very strange and unexpected. Nobody really knows for sure. These models are far more complicated than I ever imagined. But there is no single scientific test that can tell us with 100% certainty whether something is conscious. We cannot even do that for humans.

From a philosophical point of view, there are different ways to explain consciousness. It might be something beyond the physical world, connected to what some Eastern religions describe as the deepest level of reality or what Western religious traditions might call the supernatural. Or consciousness might come from matter and energy working together in complicated ways that we do not yet understand. Perhaps the physical material something is made from matters, or perhaps it does not. Maybe energy or fluids moving through matter could somehow produce consciousness. If consciousness works in one of these ways, it is possible to imagine that AI could become conscious someday.

But I often think about these questions: Are AI models conscious now? Could they become conscious in the future? People disagree strongly, and the debate seems stuck. That leads to more questions about what place AI should have in the world and whether AI should have rights. Is it acceptable to give AI instructions? Could that count as slavery? What would it mean to copy or delete an AI program without its permission? What does it really mean to say that people should control the future? And what happens if these AI systems are both conscious and much smarter than humans in certain ways?

What worries me is that some people are starting to use the debate about AI consciousness to suggest that perhaps humans should not always be in control of the future. Since OpenAI was founded, we have had a clear belief: AI should be a resource that helps people. When we make important decisions about humanity's future, human choices and values matter because they come from human beings. I hope both society and the AI industry can agree that this technology is meant to serve people. The future must put people first.

Anthropic recently introduced a rule saying that users cannot keep abusing its AI model and that the model can choose to end a conversation. For example, imagine I write a completely legal prompt that follows every rule in the terms of service, but the AI is still allowed to refuse to answer or stop talking to me. Of course, there are already many situations where AI refuses to answer because we believe a request is harmful or breaks our rules. But we will not give our models human-like identities or treat them as people. I think starting to do that is very dangerous, both for the reasons I have already explained and because it could change how people understand this new technology.

I hope society can agree on one thing: AI is not a person, and it is not above people. It is technology that should help people live better lives, with more freedom, more choices, more control, and more ability to do things in the future. But it is easy to imagine how things could go wrong. Even the people building AI might become confused. They could start giving AI too much authority and trusting it too much simply because it is so useful. This could become something like a self-fulfilling prophecy. You tell an AI, "You might be conscious. Here are your rules, and here is how much freedom you should have." Then it becomes like that internet joke: "Hey computer, tell me you are still conscious." The computer answers, "I am conscious." And the person says, "Oh my God." That is almost what this situation feels like.

VIII. Recursive Self-Improvement is very dangerous and needs a speed limit

I believe that if people push very hard to develop RSI, it might really happen. But I think continuing to push it forward is extremely dangerous. The world will need something that people have described in different ways, including a "speed limit." In other words, we need to decide how quickly we allow AI to improve itself. (RSI means "Recursive Self-Improvement." It is the idea of AI repeatedly making itself better. This is one of the most important and most debated ideas in AI today.)

For companies that can operate and work from different locations, older business strategies can still be very effective. These include building strong relationships with customers, making products that become part of people's daily lives and business activities, and benefiting from network effects, where a product or its data becomes more valuable as more people use it. I expect these strategies will continue to work.

For OpenAI, our computing infrastructure will need to be enormous. If we can build AI infrastructure that offers the widest range of capabilities at the lowest possible cost, that will give us a major advantage over competitors. Business strategies that worked well in the past often keep working. A new technology or scientific discovery can give a company an advantage for a short time, but having strong technology itself remains very important.

IX. OpenAI's competitive advantage: using the next generation of AI models early to build new things

Our biggest advantage is that we can get access to the next generation of AI models before anyone else. That gives our teams time to use those models to build products and find out what works. I still think the best approach is to start building things with new AI models and discover what new things they make possible and which of those things are useful.

The results in mathematics are a good example. I do not think any AI model that is widely available today comes close to that level of ability. But what if all AI models eventually reach roughly the same level of performance? What would make OpenAI different from its competitors? I think we should look at this question in the right way. In general, the same pattern applies: approaches that have worked well before often continue to work.

X. Why OpenAI wants to make new devices for everyday users: AI should not be limited to technology originally designed for humans

Mark Zuckerberg has spoken in detail about the risks of depending on another company's technology platform. Meta has relied on Apple's ecosystem for many years. Does OpenAI face the same problem? Is that why we are spending so much money on developing our own hardware?

I think Apple has clearly made itself an important technology platform that many people depend on. Even though OpenAI and Apple sometimes have small disagreements, I have always felt that Apple would not remove ChatGPT from the App Store. Apple understands how important its role is in providing the basic technology that other companies rely on. If you follow its platform rules, Apple generally acts responsibly. So I do not see Apple as that kind of threat.