OpenAI's Altman: Jobs that can be eliminated by ChatGPT are not real jobs.

Your work today may not be real work.

This sensational statement comes from the latest interview between Altman and Rowan Cheung.

During this 30 - minute conversation, besides sharing his thoughts on AI and work, Altman also talked about the progress of GPT - 6, whether ChatGPT will become the American version of WeChat, the changes in the vision of AGI, the future interaction mode of AI, and his feelings about being made into a popular Sora meme.

It can be said that this conversation covers multiple perspectives from entertainment gossip to cutting - edge technology, being both interesting and pointing directly to future trends.

The full text of the interview, after being organized, is as follows:

(Note: Some modal particles and introductions have been adjusted for easier reading.)

Full Text of the Interview

After DevDay: The Biggest Highlights and Product Layout

Q: Among all the announcements at Dev Day 2025, which part are you most excited about?

Sam Altman: I'm excited about all of it. Introducing apps to ChatGPT is something I've wanted to do for a long time.

But what really gets me excited is hearing people talk about what they've created with Agent Builder. Both Agent Builder and Agent Kit have many features that I'd like to use myself. However, if I have to choose one, I think running apps in ChatGPT would be the best.

Rowan Cheung: ChatGPT, with 800 million weekly active users, has become a new distribution platform. How can developers and entrepreneurs use the Apps SDK to build apps on ChatGPT?

Sam Altman: I think we need to go through some iterations to really understand how people will mainly use these features. For example, will people get used to invoking apps by their names? Or do they hope ChatGPT can figure out what they commonly use and recommend them proactively?

I believe that in the future, developers will figure out a new distribution mechanism to make these apps be used naturally. But it's always like this: only when you really release it to the world will you be surprised by all the unexpected ways of use.

Rowan Cheung: I remember you also released documents to teach developers how to increase their chances of being recommended?

Sam Altman: Yes, but we have to add a disclaimer. New products change rapidly, and we'll learn together in practice.

Rowan Cheung: Going back to the first Dev Day two years ago, you launched GPT Builder, which was really great. I remember being one of the first people to publicly build a GPT. What breakthroughs have you made with Agent Builder since then?

Sam Altman: The biggest change is that the model itself has become much more powerful. When I look back at the first Dev Day, the capabilities of the model at that time were far behind what they are now. In just 22 or 23 months, the model's capabilities have advanced incredibly. Meanwhile, we've also learned a lot about how users want to build these agents. They not only want to build on ChatGPT but also on other platforms. What impressed me the most is that now you can easily build a rather complex system - use the visual interface, upload a few files, authorize access to data sources, tell it your requirements, and then it can be deployed in just a few minutes. I was shocked when I saw this process for the first time during the rehearsal yesterday. The experience of rapidly developing impressive software with tools like Codex and Agent Kit is like experiencing a "tectonic shift".

Rowan Cheung: Now you can basically build agents in Agent Builder with zero - code, right?

Sam Altman: Yes, but if you know a little or a lot of code, you can do more complex things. Even ordinary knowledge workers can start building agents. It can be said that this is almost a "zero - code revolution" for agents.

Rowan Cheung: What does this mean for the next wave of entrepreneurs or developers?

Sam Altman: This is a question I've been thinking about. Watching Romain's demonstration backstage yesterday, I thought - if we had done this a year ago, how long would it have taken? Now it can be done almost in real - time. I even feel that my creativity can't keep up with its speed. I'm not entirely sure what changes this will bring, but one thing is for sure - the amount of software to be written in the world will increase significantly, and the time required to test and improve ideas will decrease significantly. You can try more ideas and find good ones faster, but I haven't fully figured out what else will change.

How Long Until the First Billion - Dollar Pure - Blood Agent Company?

Q: When will the first billion - dollar company operated by agents be born? Has Agent Builder reached that level of autonomy?

Sam Altman: Not yet. We used to have a small betting pool to predict when the first single - person billion - dollar company would appear. Although it's not officially set up yet, there are many speculations - like the first "zero - person company".

Rowan Cheung: A few months? A few years?

Sam Altman: I expect it to be a few years. But now we can even credibly talk about - you input a prompt into the chatbot, and it can run. That in itself is very incredible.

Rowan Cheung: However, some of the agent products we've seen still require a lot of human supervision and feedback. When will agents be able to work for a week without feedback?

Sam Altman: I think Codex isn't far from being able to complete a week's work. Although it may not be achieved in 2025, when I talked to some people today, they all said - now it can already complete a full - day's task, it's so fast. I rarely feel overwhelmed by the progress of AI, but observing the speed at which Codex can complete tasks has really shocked me this time. It's reasonable to expect that week - long tasks won't be far off.

Rowan Cheung: Where are the technical bottlenecks?

Sam Altman: More intelligent models, longer context, and better memory capabilities.

Rowan Cheung: So you have agents, various model upgrades, Codex, and you can use APIs. If you were to bring a 20 - year - old who just dropped out of Stanford to the present and give him all the knowledge you have now, what would you have him build? And what wouldn't you have him build?

Sam Altman: I've been thinking about this question these days. I really envy this generation of 20 - year - old dropouts because there are so many things they can build, and the opportunities are extremely vast. In the past few years, I haven't had enough mental space to seriously think about what I would do. But I know there are many cool things to do. Talking to people about these projects today is really exciting.

Rowan Cheung: I've been thinking about this question recently, and I guess many other developers are too - there are so many things you can do now. When building these products, do you have any suggestions, like how to find a unique advantage to maintain the lead? Is it through distribution channels, data, or some kind of workflow model?

Sam Altman: It's always difficult to answer this question at an abstract level because the best unique advantage is, in essence, unique - you have to figure it out for yourself. OpenAI spent a lot of effort to find our advantage. Generally speaking, there's no one - size - fits - all answer to this question.

The best answer is to find an advantage that suits what you're doing, your product, technology, your position in the market, and the timing. This is usually an important part of creating value when starting something new.

One general experience I can share is - you figure it out as you go. There's a business quote I've always liked: "Let tactics become strategy." You can start by doing what works, and surprisingly, things that can develop into strategies often emerge naturally in the process.

If you had asked me when we launched ChatGPT what would become our lasting advantages, I would have said I had no idea. I might have had some guesses, but I wouldn't have been confident. It turned out that one of the most exciting examples was - memory, which became an important competitive advantage for us and the reason why users keep using ChatGPT. We didn't consider this at all at that time. You start building features, and sometimes it just naturally occurs to you that "oh, this might become a very lasting advantage for us".

GPT - 6: Building a Model for the Product

Q: What advantages do you think should be established for GPT - 6? Or rather, what should be considered when building a product?

Sam Altman: This is the part you have to figure out for yourself. I'd be happy to have a brainstorming session about it, which would be fun. But to be honest, OpenAI takes up almost all of my thinking space, and I haven't had the chance to seriously consider how to start a new company, which is a bit of a pity. AI has changed many things in the world, but the basic factors that contribute to a company's advantage won't change because of it. For example, network effects, brand and marketing advantages, user data, and market effects. If you make a list and look at what has worked in recent years, it's roughly the same now, but there may be new tactics to establish these advantages.

Rowan Cheung: Recently, you launched the GDPval benchmark, which is used to measure the performance of AI models in actual economic tasks in major knowledge - based jobs. I'm surprised that GPT - 5 ranks second, only after Claude's Opus model. It's very impressive that you can still publish the results. What do you think of the results?

Sam Altman: First of all, it would be terrible if we weren't willing to publish the results when our model ranks second. There will always be things we do best and things we do worse than others. The way to build a culture of continuous improvement is to happily and directly admit that in some benchmarks or evaluations, others do better than you. I think the Claude team has done an excellent job in understanding enterprise use cases and presenting outputs beautifully. So I'm not surprised at all. Instead, I'm motivated to do better.

Rowan Cheung: Will this benchmark affect the way you build GPT - 6?

Sam Altman: It will affect part of our post - training methods, but I think the overall strategy for GPT - 6 won't change.

AGI: No Need to Exaggerate, No Need to Underestimate

Q: Your definition of AGI (Artificial General Intelligence) is - when it surpasses humans in most of the jobs with the highest economic value. At what GDPval score would you say we've reached AGI?

Sam Altman: I've been thinking about this question. First of all, like many people, I have multiple definitions of AGI. The closer we get to it, the fuzzier the concept becomes. But the thing that concerns me the most and surprises me is that we've finally reached a moment when it starts to happen - that is, when AI can make novel discoveries and expand the total amount of human knowledge. These achievements are currently small, and I don't want to exaggerate.

But now you can see many examples on Twitter where scientists from various disciplines say that AI has made a small discovery, proposed a new method, or solved a problem. I want to emphasize again that I neither want to exaggerate nor underestimate. This is the really important thing. And we're at the beginning of all this, and we're optimistic that we can make great progress in the next few months and years. This is a big deal. This might be the "AGI" indicator that I care about the most.

Rowan Cheung: Are there any scientific breakthroughs that you're particularly excited about and want AI to solve or discover?

Sam Altman: Of course, curing diseases and discovering new physical laws would be great. But even the small things happening now, like progress in mathematics, make me think it's very important. When GPT - 4 was launched, I had this feeling. I know there's a lot of controversy about the Turing test, but the public's perception of the Turing test used to be that it was out of reach. Once AI passed it, human society basically didn't update its perception. After two weeks of excitement, people started complaining about why AI wasn't fast enough or why it didn't work and asking for it to be improved. This also shows the greatness of humanity - that "eternal test of AI" just passed, and we all adapted. I feel that something similar will happen now - we'll gradually get used to AI making scientific discoveries.

Rowan Cheung: Recently, Stanford conducted a "workslop" study. This term is used to describe a low - return AI output - it looks perfect on the surface, but in fact, it will increase additional workload due to rework.

The study surveyed more than 1000 office workers, and the results showed that 41% of them had encountered workslop generated by colleagues in the past month, that is, they had to spend extra time to modify or clean up the content generated by colleagues using AI. The average cleaning time each time was 1 hour and 56 minutes, and each employee lost about $186 per month because of this.

If AI can increase some people's work efficiency by 10 times like many people here, then systematic education and training are needed to let people know when to use AI and when not to.

Sam Altman: First of all, many humans also generate things similar to workslop. This is not a phenomenon unique to AI. For example, some emails only add extra work, or meetings themselves may slow down efficiency. So we shouldn't expect AI to be different. The economy will adjust itself. People and companies that use tools to improve efficiency will have a greater impact on the future than those who use tools to slow down the organization. Of course, like using any new tool, there will be a learning curve, but I think it will be fast.

Rowan Cheung: Does OpenAI do any education or training to help people better build and learn to use these AIs?

Sam Altman: Yes. People will always use tools to do what they want. One thing I've learned is that you can create great educational content and training, but people will try all sorts of strange things, like making AI parrot - talk. However, we do try to create a lot of content to help people use AI in their workflows. In some scenarios of Codex, the adoption rate is very fast, and the integration and efficient use of the whole company only take a few days or weeks.

Parodying the CEO and AGI

Q: There are a lot of parody videos of you on Sora. Are you scared?

Sam Altman: It's actually not as strange as I thought. One or two are a bit, but after seeing hundreds, it's okay.

At that time, someone on the team asked me if I could make my cameo function available. It's a new technology, and I thought it would be my fault if I didn't try it, so I decided to do it. When I was on the plane later, I wondered if it would look strange. It did look a bit strange when it was first launched, but I quickly got used to it - obviously, this is an app full of generated videos, and the content is very interesting.

Rowan Cheung: The only thing I'm worried about is the watermark removal issue. This morning, several companies launched Sora watermark removal tools. If someone can remove the watermark and post it on social media, will it affect my personal brand? What's the mechanism behind this?

Sam Altman: First of all, one of the reasons we released this kind of technology is that we saw that it will eventually become popular. In the next few months or years, there will be excellent open - source models, and anyone can use public video - generation tools to create your image. Society will eventually adapt. We've found that one way is to release it in advance and set up guardrails, giving society and technology time to evolve together.

This method works. Text is relatively simple, and video will be more difficult because videos have a stronger impact, but I believe we'll learn to adapt. Soon, people will realize that there will be a large number of watermark - free, open - source model - generated fake videos on the Internet, which is inevitable. It may be valuable to let society adapt to this in advance.

Rowan Cheung: Is the goal of Sora to generate AI videos that are almost indistinguishable?

Sam Altman: The goal is AGI. I think high - quality videos are very important for achieving AGI for many reasons, such as spatial reasoning and what we can learn from the world model. I hope that one