Musk: AI may wipe out humanity; I’ll still get on and enjoy the journey

Author: Zhao Ying; Source: The Wall Street Journal Weekly

In the face of an unstoppable AI frenzy, Elon Musk has laid out an aggressive timeline and a pragmatic regulatory plan: AI is expected to surpass human intelligence within five years, and top companies in the industry should immediately set up a model “mutual review” mechanism to cover safety risks.

Recently, Tesla CEO Elon Musk gave an interview to the editor-in-chief of The Economist, Zanny Minton Beddoes. In this interview, Musk not only offered the latest assessment of the timeline and potential risks of AI surpassing humans, but also for the first time detailed his regulatory vision for leading AI frontier model development companies, and did not shy away from discussing the personal rivalries among key figures at major Silicon Valley AI giants.

AI surpasses human intelligence in 5 years; probability of the end of humanity at 10% to 20%; “I have to get on the boat”

The market is currently highly focused on the timeline for AGI to take shape. In response, Musk offered an extremely clear and aggressive prediction: “In about 5 years, the intelligence level of AI may surpass that of humans.”

What comes with an intelligence surge is the loss of control. When asked whether humans can still manage things 10 years from now, Musk said bluntly that “that’s unlikely.” He used a brutal but vivid biological analogy to explain the logic: “If the intelligence gap between AI and humans is far greater than the intelligence gap between AI and chimpanzees, then it’s hard to imagine chimpanzees taking a dominant position. We’re actually just evolved versions of chimpanzees. Not long ago, we were eating bananas in the jungle.”

Previously, Musk had repeatedly warned that AI could bring an end to humanity. In this interview, when asked whether the probability of killer robots leading to the end of humanity is still 10% to 20%, Musk did not deny the risk, but chose an attitude that was optimistic in a fated, resigned way.

He made a highly visual analogy using a rocket launch. When the host asked, “If your rocket has a 10% to 20% chance of exploding and killing you, would you still get on it?” Musk answered affirmatively: “Yes. The only thing we can do is try to reduce the probability of a bad outcome.”

He further explained the real-world logic behind why he “has to get on the boat”: “The development of AI and robots seems unstoppable. Even if I wanted to stop it, I couldn’t.” He believes that since you can’t stop it, the best choice is to be in the game to try to influence the direction—“the most likely outcome is bringing huge benefits to everyone.” Excluding extreme factors such as a global thermonuclear war, Musk predicts that around 2036, humanity is most likely to enter an “era of extraordinary material abundance, where everyone can have everything they want.”

Call to build an AI “mutual check” mechanism: use commercial competition to keep honesty

As AI giants such as OpenAI, Google, and Anthropic heat up their arms race on underlying foundation models, effective regulation has become a key focus for the market. In this interview, Musk proposed a very down-to-earth plan: let competitors review each other’s models.

Musk revealed that he had a long talk with Demis Hassabis, the founder of DeepMind, and he believes the most urgent move is to “make those leading AI companies hold meetings or exchanges at least regularly, to discuss any safety-related issues—like once a week or every two weeks.”

Musk pointed out that government officials lack deep technical background, making it difficult to judge whether a cutting-edge model should be released. Therefore, safety should be secured by leveraging the commercial game between giants: “For competitors, as long as you give them one or two weeks to review a new model, they can find the problems—thereby encouraging all competitors to stay honest.”

The core logic of this mechanism is that “in fiercely competitive companies, they can realize that certain things might pose safety risks, so they won’t hesitate to point out that the release of a competitor’s model should be delayed.” If a company refuses to address the risk, then and only then would the government need to step in. Musk emphasized that this mechanism must be established immediately, since what industry people call “building it within six months” is simply too long—because right now there are multiple breakthroughs in AI within a single day.

Heavily criticizes OpenAI for straying from its original purpose, hoping to set aside personal differences for the benefit of the world

When talking about cooperation barriers in the AI industry, Musk made no attempt to hide his dissatisfaction with Sam Altman, the current CEO of OpenAI. This blunt remark also revealed to the market the root cause of the lack of trust among top AI labs.

“I don’t like Sam Altman, because OpenAI was originally an artificial intelligence company intended to achieve open source—it was supposed to belong to the whole world. Somehow, though, it turned into a for-profit company worth $800 billion using closed-source technology.” Musk said outright that this completely violates the original purpose of his donation.

At the same time, Musk also spilled details about Anthropic (one of the highest-valued AI startups at present, founded by former core members of OpenAI): “Darío (Anthropic CEO) also doesn’t like Sam Altman. Actually, the reason the Anthropic team left OpenAI is that they didn’t trust Sam Altman—otherwise they should still be working at OpenAI.”

Despite intense personal feuds among the giants, Musk at the end of the interview showed a posture of compromise. He described the Anthropic team as “principled people who care about the future of the world.” To move forward the establishment of an AI mutual safety review mechanism, Musk urged: “For the sake of the world, let’s set aside personal differences.”

Verbatim transcript

Musk

I think that in about five years, the intelligence level of AI may surpass that of humans. Five years is basically the time horizon for that prediction. So, besides being human themselves, AI really can be made to do better than humans.

Host

So what will life be like in the future?

Musk

The most likely scenario is that we’ll enter an era of extraordinary material abundance, where everyone can have everything they want. That sounds a bit absurd, but we are already living in 2026. Let’s see what our situation will be like by 2036.

Of course, there are some factors that could hinder achieving this goal, like a large-scale global thermonuclear war. But assuming that the third world war doesn’t happen, I think we’ll enter an age of material abundance. So maybe that can be considered a signal of optimism and excitement about the future.

Host

Are you sure that in 10 years, humans will no longer be able to control the situation?

Musk

I think that’s unlikely. If the intelligence difference between AI and humans is far greater than the intelligence difference between AI and chimpanzees, then it’s hard to imagine chimpanzees taking the dominant position. We’re actually just evolved versions of chimpanzees. Not long ago, we were eating bananas in the jungle, and now we’re still eating bananas.

Host

Do you think the situation with killer robots has a 10% to 20% chance of leading to the end of humanity? You seem pretty relaxed about it now. Have you changed your mind?

Musk

I still think there are risks with AI and robots, and that risk is not zero. But I have to take an optimistic attitude. With the development of AI and robots, sometimes I even think that even if there were a stop button, we might not necessarily be the right people to press it—because the most likely outcome is that it brings huge benefits to everyone. I realize that now it seems everyone assumes this development is inevitable. Let’s hope for the best outcome and keep moving forward. The development of AI and robots seems unstoppable. Even if I wanted to stop it, I couldn’t.

Host

Do you think the probability that killer robots will lead to the end of humanity is still 10% to 20%?

Musk

That’s a pretty awful scenario. In the long run, all of us will eventually die. By the standards of the physical standard model, the end of the universe would be a long and boring process—like a cup of slowly cooling thin gray liquid. That’s what people call “heat death of the universe.” So if heat death really is the final outcome of our real world, then all of this is just part of the journey.

Host

If you thought your rocket had a 10% to 20% chance of exploding and killing you, would you still get on it?

Musk

Yes. The only thing we can do is try to reduce the probability of a bad outcome. I’ve always been reluctant to get involved in AI or related fields. Back then, OpenAI existed as a check on Google, because at the time Google basically dominated the development of AI. Now, Anthropic has become the leader in the AI space. These moves actually accelerate AI development, and that wasn’t my intention. Everything seems to point toward further AI development. So I thought, well, you can be sad about it, or you can join the bandwagon—and you clearly have.

Host

If you can’t beat them, join them. Apparently, some other leaders in the model development space are also trying to find ways to reduce risk. Just last week, Demis Hassabis put forward a proposal for setting up a public-private cooperative regulatory body. Do you think you and others should get more actively involved in this? Can we take some measures to control these risks?

Musk

Yes. My view on Demis’s suggestion is that before we published his article, I had a few hours of conversation with him. We believe the most urgent thing is to get those leading AI companies to hold meetings or exchanges at least regularly—to discuss any safety-related issues, every few weeks. Why haven’t you done this yet? I mean, you don’t all necessarily like each other—you’re all competing and you want to be number one.

Musk

But I think this competition can encourage people to stay honest. For government officials who don’t have deep technical backgrounds and don’t lead the frontier in AI, it really is difficult to judge whether a model should be released. But for competitors, as long as you give them one or two weeks to review a new model, they can find problems—thereby motivating all competitors to stay honest. That way, model creators can mutually check each other’s work. If there’s something that worries them, they can tell the government, and the government can judge whether there’s something that needs to be addressed. If that company doesn’t take any action to address that risk, then that’s when the government should intervene and take action.

Action should be taken immediately. People keep telling me that this kind of system must be built within the next six months, because the pace of model improvements is getting faster and faster. Six months is indeed too long and should be shortened. There are just too many breakthrough developments in AI—sometimes there are several in a single day. Last week, there were a lot.

Host

So you need to get together now and get it arranged. It sounds like you’re telling me that.

Musk

Yes. I recommend holding at least some regular informal meetings, like once a week or every two weeks. That would ensure there’s a window of one or two weeks to access the relevant content in advance.

Musk

That’s why I think this incentive mechanism works—fiercely competitive companies can realize that certain things might belong to safety risks, so they won’t hesitate to point out that the release of a competitor’s model should be delayed.

Host

You don’t really like each other and you don’t trust each other, and you often attack each other on social media. So do you think you can treat people you don’t like this way? There really are personal rivalries among these leaders.

Musk

I don’t like Sam Altman, because OpenAI was originally an artificial intelligence company intended to achieve open source—it was supposed to belong to the whole world. Somehow, though, it turned into a for-profit company worth $800 billion using closed-source technology. I feel like that completely goes against the purpose of the donation I made, and that’s something I’m very unhappy about. Also, Darío doesn’t like Sam Altman either; in fact, the reason the Anthropic team left OpenAI is that they didn’t trust Sam Altman—otherwise they should still be working at OpenAI.

Musk

I think Darío is a very principled person. He cares about all kinds of things and also about the future of the world. So far, everyone I’ve met at Anthropic has been like that—no one has made me feel like they have bad intentions. They look like they’re doing it in good faith. The road to hell is often paved with good intentions, and there are a few people who might still be acting in good faith, but we can’t be complacent. However, if it truly needs to be discussed, then let’s discuss it. For the sake of the world, let’s set aside personal differences.