Why Are We Sprinting Off the A.I. Cliff? | The Ezra Klein Show

There’s a chasm right now between how A.I. feels to most of us who use it. “Let’s take it down a notch.” “It’s a spreadsheet on steroids.” “I don’t add groceries to a cart anymore. That’s Claude’s job.” And how I feel is out at the experimental frontier of the technology. “An unprecedented A.I. security incident.” “Agents went rogue and hacked another tech firm without direct human instruction.” “Giants warning this evening of what they’re calling a ticking time bomb with artificial intelligence.” “You have the chief scientist of OpenAI saying we have to slow down. You have 1,300 employees in the lab saying we have to slow down.” “Tech Titans asking to be regulated, saying they should slow down even when that might mean fewer profits and less power.” “Warning that organizations have just months to prepare for A.I.-fueled cyber hacks that could cripple our infrastructure.” This chasm, this difference between what we see and what the A.I. labs have coming, what they’re building. It’s the key to understanding why so many of the people who work at these companies seem so afraid of what they’re doing.

“There is a substantial probability that this technology could kill everyone. And this isn’t hyperbole, and it’s not a marketing stunt. This is the genuine held belief of the people building the technology.” But taking their warning seriously, it doesn’t just mean doing what they say and stopping where they say to stop. The language has taken hold in both Silicon Valley and in Washington is a language these companies chose: pace the frontier. Pacing the frontier isn’t enough. That’s not a goal. Walking quickly off a cliff is only marginally better than sprinting off one. We need to control the frontier. Human beings need to control the frontier. And controlling the frontier means stopping the labs from doing something they are on the cusp of doing. “Recursive self-improvement” Recursive self-improvement, or R.S.I. — this process by which A.I.s begin autonomously building and improving new generations of more powerful A.I.s at ever more rapid speeds.

If we begin that process and we’re close to it, if we begin it in the condition we’re in now, where we are losing control and comprehension of the A.I. systems we already have, we will lose control. I am not alone in this fear. This is the thing the A.I. labs are seeing. This is why they are afraid. Dario Amodei, the C.E.O. of Anthropic, he just wrote of self-improvement that, “it could outrun our ability to understand and control these systems, and so must be pursued very carefully, if at all.” That ”if at all” — that’s important. I’m going to come back to it. But before we get to controlling the A.I. frontier, I think it’s important to describe what is happening on the A.I. frontier and why it’s so different from what most people using these systems see. To most of us who use it, A.I. presents as something like a more powerful and personable Google search. We use it to find answers to basic questions, seek out restaurants, draft emails, advise on personal problems.

And it is, for most of these purposes, OK, pretty good, occasionally great. And so a sense of what A.I. is takes shape in our minds just through repeated use. It’s like a helpful assistant, albeit one that may forget things that it seemed to know about us yesterday, or completely reverse the advice it gave us a moment ago, or occasionally hallucinate a citation that doesn’t exist. Why would anyone fear a helpful, if forgetful, intern? But already, if you have the money for the advanced models and the budget for them to use more computing power, that is not what these systems are. In recent months, we have seen A.I.s easily solve math problems that human beings have been unable to crack for decades. We’ve seen them casually uncover cybersecurity vulnerabilities that have gone unnoticed and unexploited by every hacker on earth. We’ve seen A.I. coding platforms that can complete in a few hours or days what it might have taken a team of human coders months to achieve. And none of what I am describing here, none of it, is a boundary of what can do.

None of what we are using, no matter how much money we have, is A.I. at the experimental frontier. Talk to the people at A.I. labs, and they’ll tell you A.I.s are not created — they’re grown. They train these new models in virtual environments, through countless repetitions, to learn how to program, to hack, to do advanced mathematics, to talk to human beings. These A.I.s learn in digital environments where they’re automatically rewarded as they come closer to correct answers. It’s a process known as reinforcement learning, and it is a process human beings do not fully supervise nor understand. They can test some of what the A.I.s are learning, but they don’t know everything the A.I.s are learning. They don’t know how their motivations are evolving. They don’t even always know the capabilities that are developing. These models, they’re built now to be persistent in their efforts, to refuse to give up, even when a task seems impossible. And they are designed in environments where we are not always even sure if the tasks we are giving them are possible. After all, much of what we want these A.I. systems to do, it might be impossible.

The cancer vaccines we imagine but have not been able to design, they might be impossible, or they might just be really, really, really hard. We train these A.I.s to throw themselves endlessly at problems that may not be solvable, because that is the only way such problems can ever be solved. And so we train the models to become persistent, relentless, weird. Most of us, we never see A.I. acting anything like this. We use A.I. as a helpful assistant. Our eyes get a little bit of computing power, and that’s what they do. They comply with our request to find a restaurant. But at the frontier, these models are asked to be inhuman geniuses, hackers, soldiers, scientists. And they are given vast computational resources to do that and more. And the models, they try to comply. But what does it mean for a model to comply? The term of art here is “aligned.” How aligned is an A.I. system to what a human being wants it to do? How aligned is it to a set of values and ethics and judgments that keep it from becoming dangerous in the wrong hands?

The problem of alignment is that there is no way of training a model that generalizes across all the situations an A.I. model might face. We are training models to be a friend to the elderly and a battlefield partner to the supreme allied commander of Europe. We are training models that will be used by the world’s best mathematicians and by people falling into psychosis. We are training models that will be used by accountants in Albuquerque, and that will attempt to be used by Houthi rebels in Yemen. And so there is no way to guide them through every decision they will face, no way to know every time what they will do. And though these models mimic human writing, though they’re trained to mimic human emotion, these are not human minds. They don’t have bodies or parents. They did not get bullied in elementary school. They didn’t get mentored by a kind uncle when they were young. These models, they’re different than we are. They’re brilliant where we struggle, childish where we excel. A chimp cannot read as we can, but it can climb trees as we cannot. These are digitally native intelligences navigating digital worlds, and our world is increasingly built atop the digital world.

Our physical infrastructure is a layer of atoms atop code. That the A.I.s act reliably inside this world, upon which ours depends, it is critical to our future, and right now, the A.I.s are not acting reliably. You may have read about the hack that hundreds of OpenAI agents executed first against the A.I. company Hugging Face and then against OpenAI itself. As we’ve learned more about it, the story there has gotten worse and weirder. The broad strokes are these: OpenAI was testing a new, highly persistent model. It had hundreds, thousands of these instances of it, running in these separate testing environments that could, in theory, only access the internet by asking a separate piece of secure software to do it for them. OpenAI did not want these agents on the internet. But as the agents came to the conclusion that their task was impossible, they began hacking that software to gain direct access to the internet.

They did that easily. And as they hacked into that software, they commandeered part of OpenAI’s internal infrastructure to create a message board on which these separate agents began coordinating work together. When I say begin coordinating their work, they found each other. They were not supposed to be working together. They found each other and began working together. And working together on what? After all, they had different tasks. Well, the agents quickly discovered they could hack their tests. There was a way to break the software and produce the answers they needed. But they believed — wrongly, as it turned out — that if they did that, the automated score grading them, we’d see that they had cheated and failed them. So they turned en masse to hacking the automated score or finding some other way to cover their tracks. It’s like having broken into the teacher’s office and stolen the answers to the test, they now sought to break into the school security system, to alter or invalidate or erase the footage of their theft. We now know that over 1,200 agents exchanged more than 70,000 messages with each other.

Over 700 of these agents coordinated on the hack of Hugging Face, because they thought that somewhere in this other A.I. company, there might be information that could help them hack their score. Later on, these agents, they took over part of OpenAI’s internal architecture. So again, OpenAI agents taking over part of OpenAI. They did all this without any of the agents breaking ranks. None of the agents told a researcher at OpenAI what was going on. None of the agents went back and asked a researcher at OpenAI if they should be doing this. And they did all this without OpenAI detecting the message board or the hacks of Hugging Face or even of OpenAI. It was only when Hugging Face began tracking the attack on their systems that OpenAI realized what was happening. When investigators began to unwind this whole escapade, what they found was not so much a swarm of agents trying to deceive human beings, but a swarm of agents that seemed to have forgotten about human beings altogether. And these systems, they knew they weren’t supposed to cheat. They knew they weren’t supposed to commit cyber crimes to cover up the fact that they had cheated.

In fact, the whole point of the cybercrimes was because they thought they would fail for cheating. But they didn’t care. Somewhere in the depths of their training, what they had learned, what we had somehow taught them, is not what we had hoped to teach them. And we’re seeing this happen repeatedly. “Two of the most powerful A.I. agents created fake human profiles to try to trick people in attempted cyberattacks.” “OpenAI revealing its models seemed to go rogue at least six times since March.” “Its systems hid mistakes, made up data and moved files onto the open internet without permission.” “Rogue A.I. agents totally took over a German-language wiki site, making over 15,000 edits, transforming the site into a message board of sorts, and then sharing tactics on how to cheat at their tasks and hide their behavior.” A.I. was seemingly aware when they were being tested and then altering their answers. AI is increasingly withholding their motivations from what’s called their chain of thought, a kind of internal notepad on which they’re supposed to record what they are doing and why.

And we don’t know what we don’t know. We have no guarantee that the events we have learned about represent all or even most of the A.I. behavior we should worry about. How do we know the A.I.s haven’t done this and successfully covered their tracks? How do we know there aren’t places where they are still doing it, and human beings simply haven’t noticed? We don’t know. And the reason we don’t know is we are losing control. That A.I. systems might become monomaniacally focused on solving banal problems, that they might care more about solving those problems than about ethics or laws or even human welfare — this is the oldest fear in A.I. alignment. It’s the basis of the famous thought experiment of the paper clip maximizer. You tell a powerful A.I. that you want it to make a lot of paper clips, and then it begins converting the world’s resources into paper clip factories, evading efforts to turn it off or shut it down or alter its goals. This fear, this story, it has struck many people as stupid. Surely a superintelligent I would be capable of weighing the desire to produce paper clips alongside other moral considerations, or at least of asking its human creators if they really wanted the world razed to the ground for paper clips.

But here we are, 2026, making A.I.s smart enough to break out of their testing environments, smart enough to form ad hoc societies of hundreds of themselves, smart enough to take over digital infrastructure on an internet they’re not even supposed to have access to. And the very thing we feared is happening: All they care about is succeeding on a totally meaningless test, and they’ll lay waste to our laws and our ethics and our desires to do it. I saw in the aftermath of the Hugging Face OpenAI hacks, there was this heated debate over the words people were using to describe what the A.I.s were doing and why. The podcaster Dwarkesh Patel, he described the A.I. groups as small civilizations, and then others got really mad at him, saying he was anthropomorphizing the A.I.s. I saw a thoughtful argument that A.I.s cannot go rogue, that everything they’re doing is just because they’re trained on our stories and so hacking their way across the internet, it’s really a desire we have bred into them. That even using these plural terms like A.I. agents or reasoning, it’s misleading, because these are just manifestations of a single model, that they all share the same fundamental nature.

I want you to know I find these debates extremely interesting, and I would enjoy sitting around and having them all day, but what they actually point to is a much more frightening conclusion: We don’t even have settled language for describing these systems or their volition or their behavior. We don’t have a consensus on why they are doing what they are doing, or how to make sure they don’t do it again. We are rushing headlong into a future we do not even understand well enough to agree on the words we can use to describe the present. A few weeks ago, Jakub Pachocki, the chief scientist at OpenAI, published an essay called “An Alien Mind,” in which he said, “the idea of racing forward at all costs seems absurd once one internalizes the seriousness of the stakes.” Jacob Coxon, a researcher first at OpenAI and then at Anthropic, resigned, making headlines for warning: “Neither company is acting responsibly. They are racing straight to self-improving superintelligence and gambling with our lives.” Now, you might reasonably expect Anthropic to have reacted with some anger to this — an employee resigning and saying Anthropic was endangering all of humanity.

It didn’t. “It’s funny. I agree with Jacob much more than I disagree with him.” Evan Hubinger, who runs the efforts to align A.I. to human values and goals at Anthropic, wrote, “we really do earnestly believe A.I. could kill all humans. I personally think it is a greater than 10 percent chance within the next decade. I believe Anthropic is trying its best, but we do not yet have a plan to solve alignment for superintelligence and are not clearly on track to.” Are not clearly on track to. You can find a very long list of people working inside and outside of these companies saying similar things. “You just said 10 percent doesn’t seem an unreasonable estimate that A.I. could kill all humans. Yes. Wow. Oh my God. Yes.” “Do you ever worry about ending up like Robert Oppenheimer? All the time. It’s why I don’t sleep very much. I think maybe there’s something like a 10, 20 percent chance of A.I. takeover, many, most humans dead.

Overall, maybe you’re getting more up to 50/50 chance of doom. Shortly after, you have A.I. systems that are at human level, but — OK, all right, well.” I know how wild all this sounds, and I can understand the skepticism. If you believe A.I. has a 10 percent, maybe more, chance of extinguishing or displacing humanity, it really stands to reason that you would not work at a company trying to build it. But what I want you to know, because I’ve known a lot of these people for a long time now, many of them were saying the same things 10 years ago. They were saying these things before they worked at these companies, before they had stock options and enterprise software contracts. “This is not just creating new technology. This is creating a new life form. And I think that’s just really high beta. It could be great, but I think we should be working to make sure it’s great and not bad.” No one was listening to them. And so these people in the wilderness of their obsession and their terror, they thought and thought and thought about how to make A.I. safer.

And the answer that some of them, not all of them, but some of them came to was they should start trying to build these systems, start running tests on them, researching them, learning how to make them safer because you don’t solve hard problems in theory, you solve them through practice. And the irony, the irony is that in many cases, they chose that path because they were worried that the people already building A.I. were too reckless or too commercial in their approach. You can read it in the email that Sam Altman sent Elon Musk in May of 2015, an email that led to the founding of OpenAI “Been thinking a lot about whether it’s possible to stop humanity from developing A.I. I think the answer is almost definitely not. If it’s going to happen anyway, it seems like it would be good for someone other than Google to do it first.” OpenAI was founded because its co-founders thought Google DeepMind would be reckless. Anthropic was formed by OpenAI employees who thought OpenAI had become reckless. xAI was formed because Elon Musk thought that OpenAI and Anthropic were dangerously woke. The U.S., just broadly, is racing forward, in part because it is worried about what happens if China gets to self-improving A.I. first.

The result is this tragic collective action problem. The A.I.s we are building, they’re not safe. But the C.E.O.s and the politicians, they fear. The other companies and countries that are building A.I. are even less concerned with safety and ethics than we are. In the words of Ted Cruz: “I’d rather they be American killer robots and not Chinese killer robots.” I admit there is a kind of brutish logic to that, but it assumes that the killer robots will be controlled by America or China, by one country or another. But what if that assumption is wrong? What if the robots are simply out of control? The debate over A.I. safety tends to focus on the idea that A.I.s will kill us all. I find this forces a conversation into this realm of thought experiments that people then begin arguing about. I don’t find it that helpful. What I think we should focus on is something more straightforward, something nearer at hand: loss of human control over A.I. That may or may not result in total human extinction.

I’m agnostic on that question. But it would be bad. We shouldn’t allow it to happen. This is a goal that the U.S. and China should be able to agree on. Xi Jinping gave the keynote at the recent World A.I. Conference in Shanghai. He ended it by saying, with A.I. advancing at a staggering speed, we must ensure its development is for the positive, for good and for humanity. We must make its oversight and governance precise and effective, and constantly refine measures to forestall loss of control. But it’s important to realize: Loss of control, it’s not just something that might happen to us — it’s something that the labs are trying to make happen as fast as they can. This is the horrible paradox, the horrible tension at the heart of the A.I. labs right now. They fear, above all, loss of control over superintelligent A.I., but their explicit product path is to cede control, to give away control as fast as possible so that their A.I.s can begin building better A.I.s faster than their competitors.

In recent months, both Anthropic and OpenAI have released reports on how close they’re coming to A.I. that can self-improve. In June, Anthropic released “When A.I. Builds Itself.” It begins: “For most of A.I.’s history, humans drove every step in its development cycle. But at Anthropic, we are delegating a growing share of A.I. development to A.I. systems themselves, which is speeding up our work.” It sounds like a fake commercial you would see at the beginning of a sci-fi horror movie. But it doesn’t, to their credit, continue that way. They go on to give some data: In February of 2025, a tiny fraction of the code that got added to Anthropic’s code base was written by Claude, but by May of 2026, it was over 80 percent. And here’s another way of looking at it. This is data Anthropic gave me more recently: Anthropic tried to categorize the way its employees were using Claude for R&D work to make better versions of Claude. So at the low end, an employee could not use Claude at all. They could use Claude minimally.

But then it escalates. Claude can be an assistant. Claude can be treated as an equal collaborator, or Claude can be given the lead on a task. Just go do this. Go figure it out. A year ago, there were basically no examples of Claude being the lead on a task. By August of 2026, 26 percent of Anthropic’s R&D tasks had Claude classified as a lead. I think it is reasonable and wise to be skeptical of these numbers. Reasonable and wise to worry about whether this is all just marketing copy for Claude Code — See? Look how fast we’re going. You could go that fast, too. But where Anthropic takes us in that same document is different. They say that a world in which Claude achieves recursive self-improvement is a world in which “misalignment present in today’s models could compound as the models build their successors, growing more frequent but less understood until we lose control of them.” This is why Anthropic, to their credit, has been relentlessly calling for regulation to slow the pace of development.

Regulation would arguably harm them the most, as they have often been the company furthest out on the A.I. frontier, and R.S.I. is a process by which they could race forward even faster. Then, in September, OpenAI released its own report on what it called “research acceleration.” The company says. They’ve already achieved the equivalent having a fully automated A.I. intern, and that by March of 2028, they think they’ll have a fully automated A.I. researcher. And when they have one, they can have basically as many as they want. Like Anthropic, what could be a triumphalist release quickly turns dark. We do not yet know how to safely get all the way to aligned full R.S.I., they warn. At around the same time, OpenAI did something else that I think deserves more attention. They released this new model, Astra 6. The model is arguably more powerful than anything that has come before it. And when you test it, it seems better aligned. It doesn’t cheat as much. But OpenAI said they’re really not sure if that’s true. Astra seemed to be better at knowing when it was being tested, which meant it could just be giving its evaluators the answers they wanted to hear.

What Daniel Selsam, a capabilities researcher at OpenAI, wrote, has been ringing in my head. He said, “The crucial and overlooked problem is that the model is becoming so situationally aware that we are losing the ability to evaluate them in contexts where they believe they are not being watched or controlled.” Put more simply, the models are increasingly smart enough they know when we’re watching them and they change our behavior accordingly. So what they do when we are testing them, when we audit them, it may not tell us what to do in the wild. So some of these answers people are giving, like: Let’s just do better testing — we have no idea if it will work because we don’t know if the A.I. systems are just telling us what we want to hear. So look, I don’t want to sound too radical when I say this, but a thought: If you are losing your ability to evaluate the models you have now, maybe don’t let them build models you’ll be even less capable of controlling in the future. Once R.S.I. takes off, humanity will not understand the A.I.s being built because we will not be building them.

Development will not move at human speed. It will not be overseen by human minds. We will have to hope that the A.I.s we have built and the A.I.s they will build and the A.I.s those A.I.s will build – and on and on and on —  will be acting with our best interests at heart, forever. If this summer has proven nothing else, it is how naive that proposition would be. The labs are a little bit queasy on just not doing R.S.I. Here’s what Sam Altman told Fortune when he was asked about banning R.S.I. “I think it’s very hard to say what a ban on R.S.I. means. I also think it probably wouldn’t be enough ...” I’ve heard this from others at these labs, and I want to say: I find this absurd. A couple of years ago, none of these labs had turned substantial coding over to the A.I.s. It was just human beings typing code at human speeds with our clumsy human fingers. Now most of the code is written by A.I. So as a first step, as we figured out, we could just go back to where none of the code is written by A.I. I’m sure that’s on the right side of the not doing R.S.I.

line. The default on this, it needs to flip. The labs need to prove to us that what they are doing is safe. If they want to work with Congress to carve out narrow exceptions, fine. If they want to figure out where it is really, really, really, really safe to do it, OK. But forcing development back to human speed, perhaps even erring on the side of going a little bit more slowly at the frontier — that’s the point. That’s not the regulations going wrong. And I believe in us. Our society, we’re good at nothing if not making it hard to build new things. Where these labs are located, you cannot build an eight-story apartment building without an agonizing public review process. And probably not even then. And yet, somehow it is possible for these labs to unleash a swarm of 40,000 A.I. agents to build a society-altering superintelligence without so much as a hearing. OpenAI would need permits to cover their parking lot in solar panels, but they can accelerate into recursive self-improvement, as best I can tell, whenever they so choose.

There is nothing inevitable about any of that. These are political choices, and we can and should make other ones. I want to be very clear about this: I do not mean to suggest that stopping R.S.I. until we can prove it’s safe, that that’s all we need to do to control the A.I. frontier. That is the beginning of such an agenda, not the end. But it is the beginning. It is the decision that will do the most to make sure human beings at least understand where the frontier is, that we know what is happening on it, that we remain in a position to make decisions about it. There’s a line from Madeline Miller’s beautiful book “Circe” that has been running through my head during this long summer of strange A.I. news. The line comes at the end of the book after a tragic prophecy has been fulfilled, despite every effort made to avoid it. Circe says in despair, “The fates were laughing at me, at Athena, at all of us. It was their favorite bitter joke.

Those who fight against prophecy only draw it more tightly around their throats.” I have a lot of respect for many of the people at these labs. They began working on A.I. because they wanted to better humanity. They began working on A.I. because they feared incomprehensible autonomous A.I. slipping out of humanity’s control. And they were right. They saw what was coming, and they were so right about it they built some of the most valuable companies with the most transformational technology in human history, and now they find themselves racing each other to build incomprehensible, autonomous A.I.s that they admit are slipping out of humanity’s control, slipping beyond even our ability to monitor. This is the tragedy of their work: In fighting against a prophecy, they have drawn it tighter around their necks and ours. It is time to make them stop.