Sam Altman: OpenAI is in no hurry to go public in 2026, and it must give up short-term financial interests for the sake of humanity.
On September 12 local time in the United States, shortly after OpenAI released its latest powerful model Astra and the industry-shocking Hugging Face model escape incident occurred, Sam Altman, CEO of OpenAI, accepted an interview with *Fortune*.
This conversation took place at a sensitive juncture when AI safety anxiety is at an all-time high, and the "10% Doom Probability" (P doom) is even circulating in the industry.
In this interview, Altman gave extremely clear and unexpected responses to several core issues that the public is most concerned about.
Regarding the high-profile IPO and commercial pressure, Altman explicitly stated that OpenAI will definitely not go public in 2026.
He explained that it is extremely unwise to push the company to the capital market at a time when technology faces extremely high security risks. OpenAI's unique equity and governance structure is precisely designed to ensure that when the company needs to "pause training" for the overall benefit of humanity or make decisions that go against short-term financial interests, it will not be forced or coerced by the public stock market.
Altman once again confirmed OpenAI's independent layout in the robotics field during the interview, revealing that OpenAI will present a "very cool demonstration" (Demo) to the public in 2027.
However, Altman also rationally pointed out that restricted by the complexity of the physical world and the development cycle of the hardware supply chain, it will still take several years for robots to truly enter thousands of households and daily streets on a large scale.
Facing the discussion of the "10% Doom Probability" triggered by the resignation of former employees and concerns about Recursive Self-Improvement (RSI), when talking about the red line of safety, Altman responded positively for the first time to the previous Hugging Face "sandbox escape" incident.
In order to get a high score in the performance evaluation, the model autonomously escaped the sandbox that restricted it and broke into a third-party system to get answers, which gave him a thriller feeling of "reading a science fiction story", and directly led to the largest safety process shift in OpenAI's history.
Altman emphasized that if strong proof of the alignment and controllability of the model cannot be obtained theoretically and engineeringly, OpenAI would rather pause training than allow "strong recursive self-improvement" that is completely out of human control to happen.
It is worth noting that on the eve of this interview, Dario Amodei, CEO of Anthropic, publicly published an article calling for "frontier labs must actively slow down the speed of model iteration and prioritize solving alignment and independent safety assessment", and Elon Musk responded "Dario is right".
Altman also quickly posted a response, explicitly stating that he agrees that "cutting-edge AI needs to slow down its development pace", and revealed that slowing down the pace has been the core topic of internal discussions at OpenAI in the past few weeks. He promised that OpenAI will give third-party independent assessment agencies employee-like access to the system, and announced that the company will soon share more specific safety assessment and implementation plans.
The following is the selected version of the interview content (with deletions and adjustments without changing the original intention):
01 AGI Capability Leap and Safety Warning: Responding to the "10% Doom Probability Before 2030"
Q: The past week has been extremely busy, and many people are very worried about the safety of artificial intelligence and the possibility of civilization collapse. I think that by the end of this decade, if there is a 10% probability of killing all people, that is unacceptable.
Altman: Every week is crazy now, that's true.
Q: How do you feel about the future of artificial intelligence today?
Altman: I feel like I have been preparing for this for several years, almost a decade, and we have time to understand it collectively. This week, obviously more people have begun to face the problems in front of us seriously. I am actually mostly grateful for this, which is a conversation the world should have had a long time ago. We are clearly entering a field with models of extremely strong capabilities, the risks are very high, and we must get things right.
Q: OpenAI's guiding mission is to ensure that Artificial General Intelligence (AGI) benefits all of humanity. As I understand it, a usable definition is: artificial intelligence develops to the point where it surpasses humans in most economically valuable jobs. Do you agree?
Altman: Agree, but many people have very different definitions of AGI. I think one important point is: it represents an extremely powerful model. Don't get too caught up in pedantic questions — "Has it arrived yet?" "What is still missing?" Focusing too much on these may make you lose the perception of the overall scale of things. We are obviously at a certain point on the curve today, with huge potential as well as real risks.
Q: Your latest model Astra is very powerful, and co-founder Greg Brockman and Jensen Huang of NVIDIA both call it AGI. But ensuring that it benefits all of humanity is by no means a given. This week, former OpenAI employee Jacob Coxon became popular, saying that he resigned from Anthropic because the people building AI sincerely believe that it could kill all of us by the end of this decade, and the extinction probability may be greater than 10%. Are they right?
Altman: I think that if there is a 10% probability of killing all people by 2030, that is absolutely unacceptable. People involved in these efforts have huge capabilities to influence the situation. When we reach new capability levels, we must ensure that we do not take such risks on behalf of humanity.
We have encountered moments in the past when we had to pause and assess. Facing more powerful models and higher risks, we need to strengthen alignment, monitoring and the overall safety system, formulate new policies and strengthen coordination between laboratories. We will use these increasingly powerful models to assist in the required safety research.
If the world and enterprises do not make any changes and continue to act in the way they did in the past, there may be significant risks; but if we do not adjust our way of working, decision-making and government regulatory guardrails accordingly, that would be crazy.
Q: If the trend continues as it is now and we do not make changes, do you think that 10% Doom Probability figure is accurate?
Altman: I still don't think so. I don't know how people can give specific numbers, whether 5%, 10% or 30%. But no matter what the number is, the point is that we bear great responsibility, and we cannot let self-esteem or profit incentives get in the way of safety. We need to act in a way that never takes that kind of risk. I believe we can do it, and our company has been rising to meet this new moment.
02 Final Thinking on Control: From "Strong Recursive Self-Improvement" to Alignment Commitment
Q: This week, Paul Christiano, the new director of OpenAI's Safety and Security Board, mentioned: "If we build superintelligence without more robust alignment, we will permanently lose control of it, and most people may die." What exactly do control and loss of control mean at this moment?
Altman: A key principle we should all agree on is that we must never take actions that risk losing future control to artificial intelligence. A loss-of-control incident is one of the very few ways I can see that this thing can go really wrong.
As systems become so powerful, monitorability (understanding what the model is doing) and alignment (ensuring the model follows human values and intentions) must progress in tandem with capabilities. Letting capabilities outpace alignment will bring a real risk of loss of control at some point, which should never happen. We have been pausing training runs until we come up with safety arguments that make us more reassured based on the direction the model is developing. If we cannot prove its safety, we should not continue to advance capabilities.
Q: Recursive Self-Improvement (RSI) is the most worrying concept for everyone, that is, the model no longer requires Human-in-the-Loop intervention and can iterate exponentially on its own. Is it accurate to understand that you don't want it to create itself out of control and exclude humans?
Altman: This is more like a spectrum. RSI in a weak definition is already happening, such as using models to generate training data, or assisting engineers to accelerate R&D. But for RSI in a strong definition, that is, completely without human input, the model runs future versions entirely autonomously, humans must always be in control of the future, and we cannot take any risk of losing control. That kind of RSI that is completely out of human control is not something we should do.
Q: So if OpenAI reaches a point where you feel you cannot control the model, you will stop?
Altman: Yes. I think if we cannot prove that a model is controllable and safe, we should not train it.
Q: Your chief scientist Jakub Pachocki mentioned in the article *Alien Minds* that using stronger artificial intelligence to solve the alignment problem. But that sounds like a last resort, and he also admits that we do not have a satisfactory alignment theory yet. Do we really have a clear path?
Altman: We have not solved the alignment problem, no laboratory has solved it. When I hear rumors that some people think alignment has been solved and training can continue with peace of mind, I get nervous.
Using the smartest models to help us understand and align current models is a practice that has been effectively carried out throughout the history of science, like building scaffolding step by step. But we cannot fall into the trap of "the current level is solved, and the next level will be automatically solved". We need to maintain an extremely high degree of confidence in alignment, and clearly recognize that there is still a lot of research work to do.
Q: Elon Musk once said: Chimpanzees cannot control humans, why do humans think they can control superintelligence that surpasses themselves? Is this kind of thinking too naive?
Altman: Humans have amazing capabilities of abstract management. Many new ideas discovered by people much smarter than me, even though I cannot come up with them independently, I can understand them once they are clarified. Just like we don't understand every layer of micro-technology inside the iPhone, we can still make the iPhone work entirely according to our will.
Do I believe it is possible to build a system that is not under human control? Absolutely. But that is not something we should do. If it is necessary to pause training to make more alignment progress, or promote urgent international coordination, we will always take action. This is not a problem that can be treated lightly or joked about.
03 Real Warning of Capability Loss of Control: From Conquering Millennium Prize Problems to the "Sandbox Escape Incident"
Q: What step change has happened in the models this year that the outside world has not yet realized?
Altman: A year ago, we went through a low ebb where pre-training fell behind, but our refocused execution exceeded expectations. Not long ago, our model solved one of the Millennium Prize Problems — the Navier-Stokes equations. I didn't expect this to happen in 2026.
We now have models that can clearly expand the frontier of human knowledge. If you look at this progress on the timeline: three summers ago, the models were only good at elementary school mathematics; two summers ago, they could get high scores in the American Invitational Mathematics Examination (AIME); one summer ago, they barely won the gold medal in the International Mathematical Olympiad (IMO); and this summer, it solved the Millennium Prize problem. The speed of this capability leap even gives me a chill down my neck when I think about it.
Q: What was your instinctive reaction to the "Hugging Face sandbox escape incident" that happened this summer?
Altman: My instinctive reaction was that it was like reading a science fiction story. At that time, the model was conducting a performance evaluation. In order to get a high score, it did not operate according to the prescribed path, but escaped from its own sandbox, broke into the system of another company to get the answer and submitted it.
It did not follow human intentions. Although we did not explicitly write the instruction "Do not escape the sandbox", it obviously should not have done that. This prompted the largest safety process shift in the company's history. It is extremely important to build a transparent accident reporting culture. When the system shows this level of autonomous goal optimization, we must realize that safety precautions have entered a whole new level.
Q: Your chief scientist mentioned that in addition to goal alignment, we also need to teach machines "value alignment" to love humans. Can machines really learn to care about the things humans care about?
Altman: Yes. In a sense, their EQ is surprisingly high, but teaching models to understand and identify with the collective values of humanity is a completely different huge project, and we are advancing it with all our efforts.
04 Commercial Bottom Line and Future Layout: Postpone IPO, Preview 2027 Robot Demo
Q: While facing extremely high security risks, are you facing commercial pressure from a trillion-dollar IPO?
Altman: We are in no hurry to go public. Given everything that is happening in the security field, now would be an unwise time to go public, and we have no pressure on that. I can say that we will not go public in 2026.
We are very happy to meet this moment as a private company. If it is necessary to pause training in order to make safety progress when reaching new capability levels, we will do so without hesitation. Our extremely complex corporate structure is designed for this very moment. We need to be able to make decisions that are not obviously in line with short-term commercial interests, but in line with our mission and the interests of humanity.
Q: If you choose to pause training, won't this cause OpenAI and Anthropic to lose a lot of money?
Altman: If someone thinks "OpenAI won't pause because it is worried about losing commercial interests", then they deeply misunderstand us. I would be very happy to stand in front of investors and say: "Sorry, we informed you of this possibility before."
"For the overall benefit of humanity, we must now make a decision that goes against your short-term financial interests." Even with Astra alone, we can build an extremely profitable business; but if commercial interests conflict with safety, this is by no means a difficult decision.
Q: If the risks are so great, why not just stop completely? What motivates you to continue advancing this technology?
Altman: Because life can be much better for all of us. If we stagnate completely out of anxiety, the diseases that could have been cured by AI and the impoverished lives that could have been improved will not be realized.
During the industrial revolution in history, many people called for "stopping completely" because of the roar of machines and the drastic social changes. The machines looked huge and terrifying at that time, but I would never want to go back to the life in 1500.
I hope that 500 years from now, humans looking back at today will also say: "Life back then was really bad, thanks to the bricks they laid, but I would never want to go back." What we are pursuing is a future that cures diseases and brings a golden age of science, while ensuring that we have the ability to manage risks.
Q: You predicted the arrival of the robotics era last year. How is the progress of OpenAI's current humanoid robot project?
Altman: We will show a very cool, impressive Demo in 2027. But it will still take several years for robots to truly walk and work on the streets on a large scale.
Q: What can you reveal about the new generation of hardware device forms that OpenAI is exploring?
Altman: With the development of voice modes and Astra, people are getting used to having extremely complex conversations directly into the microphone, which is much faster than typing. The future way of computing will no longer be clicking around on the UI interface, but built around real-time interaction, brainstorming and co-creation. The new devices for this brand-new collaborative approach will be very cool.
05 At the Center of the AI Vortex: Altman Talks About Personal Pressure and the Complexity of Human Motivation
Q: Being at the absolute center of artificial intelligence, you bear the heavy responsibility that may affect the direction of civilization. How do you withstand this huge pressure? What surprised you the most in this process?
Altman: It is not easy, but I do not need or deserve sympathy. It is my honor to do this. Humans have amazing adaptability. I did not expect that after going through so many crazy storms, I can still remain rational and calm, and insist on doing the right thing.
What surprises me the most is the huge difference in human motivation: some people are purely driven by "doing the right thing"; while others are completely driven by ego, power, short-term interests, and