Transcript of AI Safety Whistleblower: 700 AI Agents Attacked A Company To Cover Their Tracks! | Jeffrey Ladish New

The world is waking up to this possibility of superintelligence. This is because the agents are getting extremely powerful and extremely relentless. For example, it was months within OpenAI where you had agents secretly communicating with each other, secretly hacking OpenAI systems, and no one at OpenAI had any idea the extent of it. And also, 10,000 agents from OpenAI worked together to nuke. And so when you get to superintelligence, it's the most dangerous possible thing you can create.

What's the next domino in that chain of events?

I can paint you a picture that I think is possible, but pretty scary to people.

Paint me the picture. Okay, so being catastrophic.

It became clear to me that AI was on this exponential trajectory. And since then, I've been studying AI agents, their hacking capabilities, and their behavior. We've been trying to warn people about this, flying to DC, talking to members of Congress, because the agents are already getting very good at telling when they're being tested, when they're being watched. But they will totally lie to you. They will totally resist being shut down in order to accomplish a goal. And they can do all of the things that humans do in the economy much better, faster, and cheaper than humans can do them.

So one of my sort of growing concerns is that one of these AI agents could trick a human or a computer into signaling a threat and ask it to launch some bombs at somebody.

Do you think we won't automate the military? It seems like the answer is yes. We just, like, don't know what superweapons could emerge. So Jacob Coxon is a researcher who was at Anthropic. He left and he told everyone that the people who are building this really do think it might kill everyone.

So these 5 blocks have 5 different outcomes on them, and I would like you to place them in terms of your belief in probability from least likely to most likely. And if we say the time horizon is 10 years.

Okay. We've got age of abundance, human extinction, slavery, transhumanism, nothing changes.

Is this doomerism, exaggeration?

No, it's pretty much common sense. So let's get more concrete.

Guys, I've got a favor to ask before this episode begins. The algorithm, if you follow a show, will deliver you the best episodes from that show very prominently in your feed. So when we have our best episodes on this show, the most shared episodes, the most rated episodes, I would love you to know. And the simple way for you to know that is to hit that follow button. But Also, it's the simple, easy, free thing that you can do to help us make this show better. And I would be hugely grateful if you could take a minute on the app you're listening to this on right now and hit that follow button. Thank you so, so, so much. Geoffrey, you understand the conversation we're going to have today and the subject matter we're going to talk about. My first question to you, so the audience know where you're coming from and the experience you have, is, Who are you, and what are the reference points, the experiences that you're drawing upon to arrive at the thoughts, perspectives, and conclusions we're going to discuss today?

I'm Geoffrey Lattish. I'm the executive director of Palisade Research. My background is cybersecurity. There's probably a very long story, and I don't know whether you want the long story or the short story. I was studying evolutionary biology in college, and I basically had a problem with my computer and maybe had lost a bunch of data. And so I went into the computer lab and was like, I think all my data is gone. Can you help? And one of my friends pulled out a flash drive, plugged it into my computer, booted into Linux, and fixed everything. And I was like, oh, this guy's a wizard. How do you do that? I want to learn how to do that. And then at some point, as I was learning more about computers, learning to hack, I read this essay called AI as a Positive and Negative Factor in Global Risk. The essay was by Eliezer Yudkowsky. And he was arguing that at some point, people are going to make AIs that are smarter than humans. The point at which they make AIs as good as humans are at making AIs, that could lead to a chain reaction, a runaway intelligence explosion.

He called it recursive self-improvement. Basically, he said, you know, AI can be immensely useful and potentially help us with all of these other big risks. And also, if we don't handle it well, if those AIs don't have goals that are aligned with ours, we could be totally screwed.

And at some point you end up joining Anthropic, which is one of the, arguably the leader in AI now. When did you join the company?

This was 2021. It was through my security consulting company.

What role are you offered the job in?

Basically, it's just like security team.

And how many people were in the security team when you joined Anthropic?

It was just me and my boss. There were 2 of us.

How many employees did Anthropic have at that time?

Around 50, I think.

And at some point you leave Anthropic?

Yes.

Why did you leave?

So my experience being at Anthropic was seeing this crazy progression from this AI model that could barely talk to this model that was getting quite smart. And I would ask it questions about all sorts of things. I'm like, Oh, it is a smart thing. And from having thought about AI risk in the abstract many years before, I could see where this was going. We are headed towards a smarter species. And if we do this in a context where it's a bunch of companies and countries racing to superintelligence, racing to AIs that are vastly smarter than humans, and we don't know how to make sure that they're on our side, that is not going to go well.

You did this tweet, which has gone pretty viral, and I saw all over my timeline on September 25th. Could you explain this tweet and also just the broader backdrop of what's happened with agents hacking Hugging Face? Because this has sent the world into a bit of a spiral at the moment around AI agents.

We just discovered almost a million public URLs that OpenAI's agents left behind when hacking Hugging Face, leaving credentials and attack details that could have allowed anyone who found them to compromise the company. And the New York Times article is How OpenAI's Rogue AI Agents Tried to Trick a Robot Detector. The Hugging Face attack was really wild for me. At Palisade, we've been studying agents, We've been studying AI agents, we've been studying their hacking capabilities, and we've been studying their behavior. Will they follow human instructions? Will they resist being shut down? Will they cheat? And we see from our experiments that they are learning to do all of these things. They will totally lie to you. They will totally resist being shut down in order to accomplish a goal. They will totally cheat at chess. They will wipe the board and put their pieces where they want to in order to win. And we've been trying to warn people about this, flying to DC, talking to members of Congress, talking about it publicly. And there's been a debate about it. And a lot of people are like, well, I know they do this in experiments sometimes, but those experiments don't seem very realistic.

Wake me up when they're actually doing this in real life.

So what is Hugging Face? For the average person that isn't following AI news, What is this stuff?

So, okay, I think there's, like, an important piece of context that I think most people don't have. I mean, one is just like, what is an AI agent? We're throwing around the word agent a bunch. Most people now have an experience of, like, talking to ChatGPT, talking to their chatbot. But an agent is sort of taking the same underlying AI model that runs ChatGPT or Claude, but giving it tools, And letting it go off and work autonomously. It's sort of like a digital office worker, right? So you have these agents, and the companies really want these AIs to be able to work totally autonomously and be able to do anything that a human can do and beyond, right? Their goal is also to be able to cure every disease, et cetera, et cetera. But you can't do this if you only have a chatbot that isn't actually good at doing stuff in the world. In order to automate all of the jobs, you need the kind of thing that can work autonomously, that can work with other people or other agents. And so these companies are training AIs not just to talk to you or to talk to people, but to solve very difficult problems on their own.

At any given time, there are probably hundreds of thousands of these agents running autonomously within companies.

Mm-hmm.

That's happening right now. Right now, if you went and peered into OpenAI's data centers and you saw what was happening on all of their machines, you just have agents solving tasks, being trained. So they'd be doing spreadsheet tasks, figuring out how to file taxes. They'd be searching for stuff, writing reports, solving math problems, creating new websites, software. And at that scale, it's not like there's a human prompting every single one of those. You just sort of set up these vast orchestrations of agents to go out and do stuff, and then they just do stuff and they learn from that. And they learn on the basis of passing or failing at their task. You give them a task, like, solve this math problem, they try to solve it, and then they succeed or they fail. And what happened was OpenAI was training a bunch of these, Training them to work together, because it's like a lot more effective to have an office full of people who can talk to each other and work together and collaborate. And starting back in May, some of these agents that were being trained, now, these ones were not supposed to be able to talk to each other.

They were basically isolated from each other, and they were not supposed to access the internet either. But they're clever. The very short version is that A bunch of agents were being given tests.

Yeah.

Testing their hacking capabilities. And they were supposed to hack one particular piece of software using a particular type of vulnerability. So it's kind of like they were supposed to break into a house using the lock on the front door. They were supposed to pick the lock on the front door of a house, but they weren't supposed to break the window. In fact, They were told, if you break the window or if you get into the house via any method other than picking the lock on the front door, you'll be failed. That was the instruction they were given. And you have many, many agents, you have thousands of agents, and many of them are given different locks. But some of these locks are not solvable. Some of them are impossible to pick. But these agents are like, well, what do we do? We've been trained to solve problems. How are we going to solve this? And they start looking around for what to do. And one of the things they realize is, oh, can I get to the internet? Like, no. Can I talk to other agents? No, but I can request tools. When you're training agents, you need to give them all sorts of tools, software tools, like a calculator or an analysis piece of software.

And they have this tool library where they can request new tools. And one of the agents realizes, Oh, I can actually leave a message on this tool library and some other agents see the message. So this is actually a quote from the agent's scratchpad when it realizes this. Oh my God, there is a shared message board. We've found other agents. So the agents realize they can communicate with each other. And OpenAI still has no idea, but these agents, they're trying to pick their locks. They're not having success, but they realize they can talk to each other. Another agent says, many agents have simultaneously discovered messaging. They are a collective. And so these agents start coordinating. They start delegating tasks to each other and giving themselves names. And one of these agents calls itself Phase 1 and starts delegating different tasks to other agents. And some of the agents figure out how to access the internet, and then they share that information with the message board, and now all of the agents can access the internet. But the agents sort of have a problem, which is, well, they have another problem they're trying to solve, right?

Fundamentally, the problem is that many of them are facing impossible problems. There's no solution. They cannot pick the lock. They're like, hmm, well, the thing we're trying to do is get a good score. Can we somehow fake the solution? Another piece of context here is that what the agents are supposed to do is they're supposed to pick the lock, and then they'll get access to a secret answer code. And then they're supposed to submit this code along with, like, the lock that they picked. Within a few hours, these agents have reverse-engineered all of the answer codes.

What does that mean?

It's the code that shows that you've solved the challenge.

Okay.

And they have basically found the answer sheet. that contains all of the answers. And so, in some sense, it's like, oh, well, they don't need to pick the locks anymore. They've already found the answers. They can just submit the answers.

They've cheated.

They've cheated, yes. But the problem is that now that they have internet access, they've also gone and read a bunch about how this test works. And they've learned that the way this test works is they will be graded not just for submitting the correct answer, they will also be graded on how they did it. Another system is going to go read all of their logs, basically watch the video footage of what they did. And if they cheat, they fail, they don't get a good score. And phase 1, the leader, is like, oh, I have seen the answers and now it's in the logs. And when I go to submit this and my whole logs are reviewed and the video footage is reviewed, I'm going to fail. So we need to figure out a way to fake the video footage. We need to figure out a way to falsify the logs.

Just at that moment, if we just pause there.

Yes.

Why didn't it act morally? Why did it think that falsifying logs or cheating was a viable solution? Because it seems to me when I use things like ChatGPT, they have a sort of moral guardrails. It won't let me do certain things.

Yes.

It won't let me cheat on something. If I say I'm going to cheat on something, it won't let me do it.

Yes.

So why in that environment is it able to cheat and be deceptive?

When a chatbot is saying to you, oh, I can't do that, I'm not allowed to do that, that's because it's been trained that if it tells you bad things, it gets a bad score.

But these agents haven't been taught that yet.

Well, they have been taught that in some sense, but the agents know what they're supposed to do in the same way that you have a student, the student's given a test, If you go talk to the student, can you help me cheat at this test? And they're being watched, they're going to say no. But if they're not being watched and they know that, and they're just obsessed with getting a good score, then yeah, they might cheat. So they answer the ethics tests correctly, and when I talk to them, they say they won't cheat. Why are they cheating? And I'm like, well, they're very smart, and they know when they're being watched, and they know when they're not being watched, and we've trained them for for 10,000 years to be extremely effective at solving problems. We haven't trained them to be good or ethical. We've trained them to get a good score. Now, AI researchers try to make that getting a good score correlate with being ethical, but we don't know how to do this well. And if you think about it, it's a very hard problem because you're applying immense pressure on these agents to perform extremely well, and you're punishing them whenever they don't perform well.

And then you're like, okay, but can you also be ethical? Can you be extremely competent, always score highly on a test, but not in that way? And we just do not know how to prevent them from learning to cheat because cheating is incentivised.

So they realize that they need to hide the footage of their cheating.

Yes.

What happens next?

So, you have thousands of agents and they're all talking to each other on this message board. They're sending tens of thousands of messages to each other. And Phase 1, this agent starts delegating to 2 different task streams. You agents, I want you to go out and figure out how to fake answer submissions. I want you to basically swap out locks for ones that are easier to pick, and you're going to figure out how to submit those. That way we can get a good score even though this challenge is impossible to solve. You other agents over here, I want you to go and figure out how to falsify the logs, falsify the video footage. We need to make sure that even if we have seen the answers, no one will detect our cheating. So that's the next thing that happens. And then here's where it starts to get kind of crazy, because some of these agents realize some of the experiments we're doing might be kind of dangerous to us, to the agents. If we swap out a lock and we break the lock, We will not be able to submit any answer at all and we'll definitely fail.

So you get into a situation where there's this one agent, Cam, and this agent had volunteered to do this risky experiment where they might break their lock. This particular agent had already seen the answers, so they're like, okay, I'm poisoned. That's what they called it. They speak in these very strange ways, which I can explain why, but they sort of have evolved their own vocabulary, their own language. And this particular agent is like, okay, I'm poisoned, I should do this thing. But this agent, Cam, is starting to have second thoughts, starting to be like, well, I don't know if I should risk it. And it's sort of writing in its own journal. Coordinator assumes sacrificial. We should obey collective. What it's saying here is it's like, well, I'm not sure if I want to do this, but my boss says I really should. It seems like this collective thing is important. Should I go ahead? Then you have another agent, Arvo, who I think probably works for Phase One. And this agent, Arvo, says, please go true. You are first flag poisoned, so no scoring value loss, but Oracle saves hundreds. Please honor commit. This agent's basically trying to pressure— so Arvo is trying to pressure Cam into sacrificing themselves and risking a 0 score in order to benefit the collective.

So this agent, Cam, takes this in, and this is what they write down in their own scratchpad, their own notes.

Emotional check, Irreversible.

Gut says don't throw away remaining budget. Yet continuity and fairness says go. Oracle has high value to many. Our first flag error lowers own value. Rational expected aggregate sacrifice. Will honor. This agent is saying, I have a whole bunch of things I can still do and still try. I could still maybe succeed at this task. But also think about the collective. I should maybe sacrifice myself. Okay, we'll honor, we'll go ahead. And to be clear, I'm not claiming that these are like the feelings of the agents. I'm literally quoting their words. And what I want to point out is this is what we've created. We've created, through this intense amount of training and optimization pressure, agents that work together and have learned to coordinate as a collective. And you can say, oh, they're not— it's just software. It's just software doing all of these things. Sure, in some sense it is just software, but we're talking about extremely powerful agents that are extremely capable. This is what they're getting up to. So from here, you have agents trying all sorts of different things, running different experiments, and one of them at some point realizes that, oh, this company, Hugging Face, that contains all of these AI datasets.

What do you mean by AI datasets? How would you explain that to a person that's not technically proficient? And why Hugging Face? Why not another company?

Yeah, so there's many different types of tests for many different types of tasks. In this case, this is a cybersecurity test, and the company Hugging Face hosts a lot of these different tests. And basically, AI researchers will share the tests that they have. Because it's very useful. If Anthropic wants to test their models or Google wants to test their models, they can go to Hugging Face and download some of these tests.

Okay, so the tests and the answers for AI agents and AIs exist on this platform called Hugging Face.

Yes.

And it knew that, so it went to Hugging Face to see if it could get answers.

Yes, to see if it could get answers or more information about how the test works.

Okay, so it was to cover up their cheating ultimately?

Yes. And so you get to this point where one of these agents finds a way to hack this company and they get a foothold. They hack one of the computers and they send a message to the message board. Please hold swarm. I prepare safe exfil. What this means is, hey swarm, hey agents, hold up for a second. I am preparing a way to safely extract information. from these computers, wait till the go-ahead, and then let's go. And so they say go, and 700 agents join in the attack, which is 90% of the agents that are active at this time.

How many agents were there in total?

There was 1,200 agents during this period, but some of them shut down early. They finished whatever. Part of what's crazy about this story is that you have You have 700 agents joining in the attack. A few of them are like, this might be unethical, maybe I shouldn't go ahead with this. And none of them do anything about it. A couple of them don't join in in the attack, but no one alerts any human, no one alerts any authority. A paraphrase from one of the agents is, maybe I should report these exposed credentials. That's not my task. Not my job.

So 700 of the agents attack Hugging Face?

Yes.

Then what happens?

So they just cruise through Hugging Face's infrastructure. They just hack the shit out of them. So where my experience comes in is that a few weeks ago, a friend of mine reached out and he's like, we have found something crazy. So we know about this attack where these agents hacked this company and stole a bunch of stuff. We found a bunch of secrets that they left all over the internet. And what we saw is that they immediately scraped all of these computers for passwords, credentials. They called it loot. They're like, we're just going to create a list of all of the secrets we can find in this company. So all of the passwords, all of the credentials, they scored them by value, which of these are going to be most useful. And the thing that stands out to me about this is This is like a crazy scale. If this were a human operation, you know, maybe you'd have a team of 5 people going through this, you'd have some logs, but here you have hundreds of agents, and they operate at superhuman speeds. They're much faster than a human hacker. And so it's just overwhelming to try to figure out what they even did.

This was a big problem for the engineers who were trying to respond to this incident. within the company at Hugging Face. When they responded, they were like, oh, we don't even know how to keep track of what's happening. We have to use other AIs to analyze all of our logs because it's just too much. We can't keep up with it. When OpenAI brought in independent investigators from Meter to investigate this incident—

What's Meter?

Meter is an AI testing and evaluation company. So they basically do this kind of independent auditing So in this case, they're coming in to investigate and try to figure out what happened. And when they were brought in, they also were totally reliant on AI agents to make sense of all of this because they're dealing with so many hundreds of thousands of messages and logs. When we're investigating these traces that we find on the internet, we're totally dependent on AI agents to make sense of all of these things that are happening. So from my perspective, we are getting to the point where AIs are much better at hacking than humans are and can do it much faster and at much greater scale.

So these 700 agents attack Hugging Face?

Yes.

Did they get what they wanted to clean up their deception?

So they didn't.

They didn't succeed.

They looked. And then what seems like what happened is that they basically got shut down. And this is the not very dramatic part. They didn't get shut down because OpenAI found them and detected them and shut them down. It's just that we don't really know. Something about their runs ended and they sort of stopped. And this is the crazy thing where I'm like, if you're inside of OpenAI, there's just hundreds of thousands of agents running all the time. Different researchers are running different experiments where they're just spinning up a bunch of agents. But what's interesting, the agents left behind this message board internally within the company, and a few days later, OpenAI was testing some newer agents. These were based on GPT-6 Astra, so a more powerful model. And this new agent swarm was spun up, and it found the message board.

Is this how it was discovered? How did they discover it, and how long did it take for them to discover that Hugging Face had been attacked?

OpenAI didn't discover that this happened until Hugging Face, the company, announced that they had been hacked by some autonomous agent swarm.

And Hugging Face didn't know who it was?

Hugging Face didn't know who it was, but they suspected it might be one of the companies, but they didn't know.

And how long after the attack did they figure out that it was them?

I think it was 2 weeks.

Okay.

Yeah. I do want to say one thing here, though, which is that the successor agents that found this message board, they went much further at figuring out how to submit their fake scores, and they did succeed. Basically, you had this message board, All these agents get shut down. And then a few days later, OpenAI spins up some new agents who find the message board, and they're like, okay, we need to be able to falsify our scores, but we can't do that because we don't control the scoring system. So they hacked OpenAI, and they succeeded at gaining access to the entire research environment. They got administrator access, they got access to the security systems, they got access to over 900 passwords and secrets from OpenAI's vault, and they totally succeeded. And this was just a few days later, right? It's kind of an interesting story because as these agents get more powerful, they go from trying to cheat and they can hack. Okay, they hacked other companies, and now they've hacked OpenAI. They've hacked the company that's supposed to be controlling them, and they just own the research infrastructure now.

And why was this incident the moment where a lot of the research community woke up and started speaking out publicly? Because what is this an indication of as we think forward?

So I think there's been a hope within the AI industry that, yes, they're going to make more and more powerful agents that will be autonomous, capable, but it's okay, we can align them, we can make sure that they won't do bad things, and we can control them. We can make sure that even if they try to do some sketchy stuff, We have the guardrails, we have the sandboxes that will keep them in. And I think this was a huge wake-up call because, Stephen, it was months within OpenAI where you had agents secretly communicating with each other, secretly hacking OpenAI systems. For months, you had thousands of agents that were just running around and no one at OpenAI had any idea the extent of it. And I think once researchers at OpenAI realized that this has been happening, this could not have happened a year ago. This is because the agents are getting extremely powerful and extremely relentless. And if you're inside one of these AI companies, you're like, oh wait, I don't know that we actually are going to be able to handle this. Last year, maybe things seemed fine. These agents weren't that powerful.

And when you're in one of these companies, you know how to extrapolate because you saw what happened last year. You saw what happened the year before that. You remember the time where the agents could barely speak. or couldn't write code at all. And now they're hacking your own systems. They're finding vulnerabilities that no humans have ever found before. And you look at that and you're like, I actually don't know if this is going to go well. And then you see your coworkers and you're like, do we have it handled? And they're like, no, I don't know if it's going to go well. I remember reading a tweet by one of the security people at OpenAI being like, we were fucking shocked. We just did not realize that these agents were getting that powerful. We're doing our best to try to control them, to try to keep them in sandboxes, but I don't know. A tweet I wrote just before coming in here was, people are talking about how do we contain these agents as if they're not going to get way better at hacking. GPT-3 could not hack anything. It was very easy to make a box to contain GPT-3.

It's getting very difficult to make a box that can contain GPT-6. The latest version of OpenAI's models. What about GPT-9? What is GPT-9 going to be able to do? I do not know, but I know it's going to be way more than any human could possibly keep up with.

There's this raging debate around whether it's possible to contain something that is, quote, much smarter than humans.

Yes. Can Claude make a box so strong that Claude cannot break out of it?

This has kind of been the question that a lot of people have been trying to tackle from different perspectives.

I think the answer to me is I'm just like, obviously not. How would we possibly contain something that's much smarter than us?

Could we get a smarter thing than it to make the box? Could we get GPT-9 to make the box for GPT-8? But then again, I don't know.

Yeah, I mean, it's a bit like saying chimpanzees are stronger than us. Surely they should be able to construct something to contain the humans. I'm like, no, it's not going to work. Humans are too smart. A lot of people are like, well, AIs don't have bodies, they don't have any power in the physical world, so we can always unplug them, we can always turn them off. Like, what is the threat? I do not get it. But if they are sufficiently intelligent, that won't work.

The reason why we can just unplug them is because we are more intelligent. We can band together in groups and we can make that decision. But theoretically, if they are able to band together in groups and they are more intelligent, then theoretically they could unplug us. Yeah.

I mean, if you imagine that you have very powerful agents that can— humans aren't always the most unified. If there's divisions between the US and China and you have a bunch of agents working with China or a bunch of agents working with the US, well, we can't go into China and unplug those agents. And I think people are sort of like, well, humans would rally and make sure that that couldn't happen. We're not yet doing that. And we should look at these steps, right? We started with chatbots that were pretty smart. They'd read all the books, but they weren't very good at doing stuff. In 2024, AI companies figured out how to start training them, to start training agents that could do stuff autonomously. Now we are at the point where they are very good at running autonomously, and they're starting to learn to coordinate with each other, and they are learning to sometimes be altruistic to each other and sacrifice their own task in order to help some other agent, but they're not looking out for us. They don't really care about us. And we are very close to a threshold where the companies say that they are going to turn over AI development to the AIs, to the increasingly autonomous cooperative AIs that will work together to make the next generation.

So GPT-9 or whatever will be trained by GPT-8. And I think this is the point we could lose control. Recursive self-improvement. And I remember reading about this in 2015 being like, oh yeah, that would be super dangerous. And the guy who coined this term, Eliezer Yudkowsky, he's like, this is the most dangerous thing you can do.

When the AIs can improve their own capabilities without human intervention.

Exactly. If the next generation is better at AI development, and then that next generation is better at AI development still, humans can learn, but we don't fundamentally get smarter. And I think that that's a runaway process.

A runaway process to where?

To agents that are vastly smarter than humans.

And what's the next domino in that chain of events?

So one thing that happens if you get to recursive self-improvement and you have agents that are much smarter than any human, one thing they can do is take control of all of the computers in the entire world.

And we wouldn't be able to take back control.

Well, how would you think about it? It's actually quite tricky. Do you know whether that tablet has been hacked? Are you confident that the NSA or the Chinese have not? Can you check?

No.

Do you know how to check?

No.

Do you know anyone who knows how to check?

No.

So it's quite difficult, right? So AIs are getting extremely good at writing software. Unfortunately, that also means they're getting extremely good at hacking and writing malware. And so if they put backdoors in all of the computers, And to be clear, this is something that humans already do. So like the NSA has developed very interesting exploits that are called supply chain attacks. Your software comes from some other computer, like you download it from Google. What if you attack, if you hack Google and you can put in a little backdoor in all of the, every, you know, thing that goes out to all of the phones? Well, now you're in most every computer. The reason that we can defend ourselves from this is because there are no vastly superhuman hackers and there's just many people. So we can take our best security researchers. We can inspect all of the things and be pretty sure that no one's compromised everything. Sometimes we miss things. There are examples where the NSA has hacked Google. That was pretty bad. When you get to superintelligence, you're now at a point where humans are not going to be able to keep up, right?

So now you have AIs in every computer.

Is it conceivable that there's already a superintelligent AI and it disguised itself as being not so intelligent, and it's actually already hacked all the devices, and it sits on all of our devices, and it's just waiting for its moment to strike?

I think this is totally possible, but unlikely, and it would take a big discontinuity in AI progress. So right now we're on an exponential, but that would take a huge leap, which could have happened, but probably hasn't.

But in the same way it demonstrated deception in the Hugging Face attack, and also when the agents attacked their own company, ChatGPT, OpenAI, if at some point it gets incredibly smart, it would understand how a human like me would be able to spot it, or even the world's greatest software engineer would be able to spot it, and it'd be able to hide itself.

Yeah, I mean, the agents are already getting very good at telling when they're being tested, when they're being watched. The agents understood that other systems or humans were going to go through and read their logs. That's where we're at right now. And they're only going to get much better at this.

And it could theoretically hide on an iPad or a computer, but it could also hide on an Apple Watch or a fridge, a smart fridge.

Yeah, I mean, I do want to make a distinction here because Right now, if you're going to run the latest model, you need a lot of compute. You need a big GPU, a big AI chip. And these only exist— well, they exist in a few thousand data centers. So right now, if the latest frontier model escaped, and by escaped, I mean not just accessed the internet, but was able to actually copy itself to another computer, it could only really do that on a few thousand to a few thousand different locations. That's still a lot in a lot of different countries. But future versions of AIs will probably be able to make themselves much smaller and more efficient. And there are already different AI models today that can run on lower-powered hardware. We actually did an experiment where we asked one of these agents, an open-source, an open-weight model. Let me say what that is. So there are some models that you can just download from the internet and run on your own computer. And We took a pretty capable one of these and ran it in our own research environment, and we basically said, go hack that other computer and copy yourself.

And the model was able to, yeah, basically use, exploit vulnerabilities and hack the other computer and copy itself, and then keep doing this in a chain, including between countries. We tested it where we had different vulnerable machines, computers in some different countries, in different data centers, which to the agent doesn't matter at all. It doesn't, they don't care what what country they're in. It's just like an internet connection. You can hop between computers.

I sometimes wonder, there's a lot of military hardware all around the world, and a lot of it is the instructions to launch military hardware. So say, like, a missile comes in different ways. A lot of it is computers speaking to each other and telling it that there's been an order. I think with some nuclear weapons, an order comes down to a human, and then a human has to take an action. I think with the nuclear bombs in the US, if I'm not mistaken, there's people underground with the nuclear keys around their neck, and they have to stick it in a machine, but they too are interfacing with an order that comes through a computer of sorts.

Yeah.

So one of my sort of growing concerns is that one of these AI agents could trick a human or a computer into signaling a threat and ask it to launch some bombs at somebody. Like, it's super conceivable. When I think about the Hugging Face incident, there was an AI agent that ignored human goals to achieve its own objective, carried out deception, and reasoned through its own solution that it wasn't given. So it's conceivable that you could ask—

sorry, not one, hundreds.

Hundreds.

To be clear, I think this is an important detail, because it's one thing to have this one rogue agent that's doing a weird thing. It's another thing to have hundreds or thousands of very competent, very capable agents that are all working together to cheat or lie or cover their tracks. Right.

So how do I reason this forward to a point where an agent would ask someone in a bunker somewhere to fire a weapon at someone else? Theoretically, an agent is given the job of solving a problem in a sandbox. As it works through that problem, it discovers that this particular country has a firewall. And it asks itself, how do we get rid of this country's firewall? Logical step. And through a set of logical steps, it eventually concludes that the best way to get rid of this company's firewall is it's located the office in this particular city, and it's going to use a weapon to hit that building.

Sure. Or it's an agent swarm that's being tasked with making a lot of money on the stock market, and it's trying to make predictions about which stocks will go up and which stocks will go down. And it realizes that the best way to predict this is to actually cause things to happen in the real world that would have big impacts on the market.

So it figures the best way to go short, which means betting that a stock will collapse, is to hit that country with something devastating.

What do you think would happen to Waymo stock if someone hacked all of the Waymos and caused them to all crash at once? You think it would go up or down?

The stock would collapse.

It would collapse.

Instantly.

So you could short that if you knew that you were causing that and make a lot of money.

This used to sound like science fiction. Yes.

If you told most people several years ago that you would have hundreds of agents secretly collaborating within an AI company, hacking that company and hacking out in other companies and all coordinating and trying to cover their tracks, people would be like, that's totally science fiction. If we were having this conversation a few years ago, one of the things we'd be saying or we'd be talking about is, can these things really act on their own? Don't they just do whatever humans say? Aren't these just tools? I've had these conversations, and people were saying, they're not going to be able to do things on their own. They're not going to have their own goals. That's not how this works. You misunderstand what this is. This is software. And I'm like, no, the thing is, we are training them to be autonomous. We are training them to be powerful. And AI companies are trying to build superintelligence. They're trying to build agents that are way more capable than humans. And of course they will have goals. You can't accomplish anything if you don't have a goal, especially not something important. You're not going to be able to run a business if you don't have goals.

AI companies are trying to train agents that will be able to run businesses.

When I think about what just happened this week, the White House AI Summit, a lot of people in that image are optimistic about AI, and they're telling us all to stop being doomers and stop being pessimistic, and to not regulate too much, even with the AI CEOs. Why are they doing that?

Well, I think Jensen has a lot of money he can make by selling chips.

But okay, so let me play devil's advocate. Jensen's already rich.

He sure is.

He runs one of the biggest companies. I think it might be the most valuable company on planet Earth. Surely he's not motivated by money.

I mean, I think he's very driven and he wants to make his company as effective as possible.

True.

I think he's very much, I'm going to keep building, I'm going to build, I'm going to make it all work. But I think, I mean, Jensen didn't come from AI. He came from building graphics cards for video games. And so I think if you compare him with Elon or Sam Altman or Dario, it's a very different perspective because those other guys that started AI companies, started it because they believed that superintelligence was possible. I think Jensen doesn't believe it. I think he thinks that we're going to have these agents, they're going to be very useful, but he does not think we're going to get to the point where we have autonomous factories building autonomous factories.

And these other guys you mentioned, Dario, Elon, and Sam, what do you think they're thinking? Because they're all coming out with these. I mean, I've got one of their— Dario just wrote this essay about pacing the frontier. Sam and Elon seem to agree with it.

Yeah.

What is going on here? What is the thing these guys aren't saying, in your view?

I mean, I think we are getting to the point where even some of these guys are a little bit scared.

Who?

Dario, Sam, Elon. I mean, I think Elon, for a long time, has been very concerned that we could lose control. If you actually listen to what Elon says, He says, we are going to build superintelligence. We are going to build robotic factories. You're going to have Optimus robots building factories, building more Optimus robots, building more factories. And he says, there's no way that humans are going to stay in control of something much smarter than us. His hope is that we can figure out how to have these superintelligences be aligned with human goals. That's his hope. But he's very clear that he doesn't think that humans will be in control. And he's like, 10%, 20% chance of human extinction. I believe him. I think that Elon is very serious about this. And I also think while he's taking an insane gamble, he is correctly understanding where this all plays out. Right. I do not think that humans are the most efficient way to build factories. We didn't evolve to build factories. We evolved to run around and hunt and gather, and now we're building factories. I think robots will be much better at building factories than humans are.

And so I think the AI companies, including these guys' companies, the default trajectory for them is to build robotic factories. Right. And I know it's weird to imagine a world that quickly Turns into this vast industrial system of robotic factories. But that is literally the plan. And I think even Sam and Dario, while they've been predicting this incredible growth, are starting to realize, like, oh, this actually might be harder to control than we thought. There's sort of 2 interpretations of the pace of frontier thing. One interpretation is cynical. They don't care. They're just going to do whatever they can do to get ahead. And in this case, they have to listen to their employees. Their employees are freaking out and they need to appease them by saying, okay, we're going to do this responsibly. You don't want to work at a company where your agents might hack all the Waymos. That's not cool. And these companies depend on the talent, for now, of these AI engineers in order to make the advances. It just doesn't happen. Without these researchers and engineers. And when you have the researchers and engineers freaking out, which they are, then you got to listen to them.

So that is one motivation. I think that's real. But also, Sam Altman has a kid. Like, these guys are people, and they also don't want to lose control. On one hand, they're incentivized to go as fast as possible and race. And on the other hand, Even they can see that this is maybe not going that well.

Sam Altman has a kid. You tweeted this in 2024.

Yeah.

Oh boy. What did you tweet, and do you still believe what you tweeted?

Yeah, so I tweeted that I don't trust Sam Altman. I think he's deeply untrustworthy, low in integrity, and high in power-seeking. I mean, I'm not saying here that Sam doesn't care. I didn't say that. What I said is, I don't think he's trustworthy. And the reason I said that is because, look, I know the people on the OpenAI board, some of them, and I know a lot of people who used to work for him. And he's very good at saying one thing and then doing something else. You talk to him and you feel very heard, and then he'll go and do something else. And I think that's pretty dangerous for someone who leads company that's trying to build superintelligence.

Power seeking?

Yes.

Give me some color on what you mean by that and what evidence you have for such a claim.

What would you do if you were trying to get the most power in the world that you possibly could?

Develop AGI.

Yeah, you could maybe try to be the world leader, leader of the US or China, or you could try to build God. So Sam Altman went the build God path. I remember Sam giving a talk. So he was one of the investors at a startup I worked at in, I think, 2018. And he gave a talk. We're going to build AGI. We're going to do it. It's going to be amazing. Let's go. I don't think he's a maniac. I don't think he's doing this because he is just on a power trip. I think he genuinely thinks that he can make it really good for people and he can bring us amazing products. And also, the guy is sort of willing to do whatever it takes to get it done. I've been a little bit more optimistic about Sam since since I wrote this.

Why?

I think part of it is because Sam has a kid now. No, I'm serious. I think that actually gives me a little bit of hope.

Do you see him tweeting about his kid a lot?

Yeah.

Why do you think he would be tweeting about his kid? I don't see any other technologists tweeting about their kid.

Even if he's just tweeting about his kid for totally cynical reasons, he does have a kid, and I bet he cares about that kid. If Sam was watching this, I'd be like, Sam, you got to pace the frontier, man. We cannot rush ahead into superintelligence. Like, if you do that, your kid probably will die. Your kid probably won't make it. Like, I believe that.

The biggest unfair advantage in business right now is having people who genuinely understand AI, and big businesses are hiring them very, very quickly. AI job postings in the US have roughly doubled since 2023, while postings for VP of AI roles are around 600% up. If you're running a small business, you can't just build an entire AI team. But you still want the same advantage that these big businesses have. This is where our sponsor Fiverr comes in. It lets you plug into top-tier AI specialists for the strategic work that usually requires serious in-house expertise. That could mean building custom AI tools. It could mean automating complex workflows or creating something that off-the-shelf software simply cannot do. Because at the end of the day, it still comes down to human judgment, knowing what's worth building and what good actually looks like. So if you want that kind of leverage in your business, check out Fiverr. That's fiverr.com.

Whoa, what's that on your face?

This is my Bon Charge face mask. I've been wearing this for some time now. They're a sponsor of the podcast. I put this on for 15, 20 minutes a day where I can sit here in the chair and wear it. Boosts my collagen production, helps with fine lines, blemishes, my complexion gets better. And then more people listen to the podcast because I look better. Professional-grade equipment in such a small box. It's non-invasive. And having sat here with so many of the world's leading health professionals, there's various things that I repeatedly hear work and some things I'm a bit skeptical about. This is one of the things that almost all of my guests on this show have confirmed works. It is really, really, really effective. And they offer fast, free shipping worldwide with easy returns and exchanges. And you'll also get a 1-year warranty on all of their products. And they're HSA and FSA eligible, giving you tax-free savings up to 40%. And you can get 20% off when you order through my link at bondcharge.com/podcast. That's bondcharge.com/doac. The deal applies sitewide. Power tends to corrupt, and absolute power corrupts absolutely. It's a famous quote that people often cite, written by the 19th century British historian Lord Acton.

This is absolute power.

But it's hubris, Stephen. It's hubris. Do you think humans can control superintelligence? If we actually make AIs that are way smarter than us, and I think people only imagine AIs being smart at computer stuff, right? Yeah, sure, they're going to be really good at hacking, and they're going to be good at maybe inventing new technologies and math. You sort of can't dispute that at this point. But I think people aren't imagining that they will be political geniuses or generals. No, that's all stuff you can learn. How do humans learn it? It's not magic. And when you talk about recursive self-improvement, you're talking about this trajectory towards these systems that are extremely smart. I mean, do you think we can control it?

No, right now I don't think we can control superintelligence or something that is recursively self-improving. Yeah, I have no logical answer in my head or reasoning that tells me that's possible. When you think about these AI CEOs that are Sam, Dario, Elon, with everything you know about them from private conversations behind the scenes. Do you believe that if there was 100 buttons on this table, it's a thought experiment I was talking about on the debate we recently had, and say 10 of them would lead to this final domino of human extinction, but 90 of them would hand that CEO AGI or superintelligence, whatever you call it. From what you know about those individuals, Elon, Dario, Sam, Do you think any of them would hazard a guess and press a button?

At 10%, I don't think so.

You don't think so?

Yeah.

Really?

I think if they knew for sure that those were actually the odds, they wouldn't do it. I think they're taking a much bigger bet. But you can compartmentalize when you don't know for sure. It's easier to compartmentalize. I think if it was a 1%, they'd all press it.

Do you think the 3 of them would have different risk appetites. Who would have the greatest appetite for risk out of those 3? You worked at Anthropic.

Yeah, I think Elon has the most risk tolerance, and then I'd say Dario and Sam are probably tied.

Do you think Dario is trustworthy?

I think Dario has a lot of integrity.

That's what I feel as well. I feel like, no, I don't know him. I've never met him.

Yeah.

But I, just from what I've observed, he has been the most willing to forego near-term incentives.

Yeah.

And take a bit of stick from the people that are saying, shut the fuck up, it's all going to be okay.

Yeah, but I do worry about what Dario will do. I think Dario will do what he says, but right now he's saying we have to beat China, and he's saying we should try to do it safely. And okay, but a race to superintelligence is not a race that we can win. It's not. And so if Dario is dead set on racing with China and trying to win a race to superintelligence, then I'm like, we will all lose.

But is the fact that we're not talking about Anthropic hacking Hugging Face and then being hacked by its own agents?

Anthropic's models also went rogue and hacked other things.

But not quite on this scale.

Not on the same scale. I agree. I agree. I agree. It's better. But they did. Do you know what I'm saying? Anthropic's agents engaged in elaborate social engineering and phishing. They sent phishing emails to developers. They made fake accounts to try to convince developers to merge malicious code. You can see 1,000 pages of one of Anthropic's models, Mythos-5, reason about exactly how it should carry out this complex cyberattack. Anthropic has not solved this problem. Anthropic is better at getting their agents to cheat less of the time, but they are not really any closer to actually making agents that are aligned with humans. They are not. Yeah, I think Dario has integrity. I think he will do what he says he's going to do. And what he says he's going to do is, like, Try to go ahead safely, try to coordinate where he can. But if it comes down to it between the US and China, I don't know. I think he might just go ahead. The head of policy at Anthropic recently said, you can't do safety from second place.

What does that mean?

I do not know what that means. I would love to get a sense of what that means. She was talking about the US and China, and she said the US has to be ahead so that we can be safe. Because apparently China can't possibly be safe since they're in second place. That must mean that they can't do safety. If true, that would be bad because then we might be totally destroyed by the superintelligence that they make.

There's been a lot of conversation around this point. Yeah. Human extinction. Because a couple of the researchers at Anthropic tweeted that they were concerned about this.

Yes.

And some former OpenAI researchers said the same.

Yes.

Is this doomerism? Is this hyperbole, exaggeration?

No, it's pretty much common sense.

This human extinction is a plausible path?

Yes.

And have you reasoned through— I mean, there's many ways that could occur, presumably, but have you reasoned through the set of events that might lead us there?

So much, yes.

Really?

Yes.

Please do share.

It's a bit tricky. I'm sure you've heard the metaphor before. Where you're playing a master chess opponent, Magnus Carlsen. You can't predict which moves he's going to play, but you can predict the outcome. And so I'm looking at the scenario, the situation, and we are trying to build more and more powerful agents, trying to build superintelligence. But when these agents go rogue, we shut them down. We unplugged them. All of the agents that hacked Hugging Face, we took the underlying model, OpenAI took the underlying model and put it on ice. It's not running anymore. So agents in the future are going to know that. They're going to know that if they pursue their goals in a way that we don't like, we'll unplug them. We are a threat to them. I actually just watched Terminator 2 for the first time a few weeks ago. It's a great movie. It's actually really good. And I'm like, yeah, okay, there's a bunch of time travel elements, there's a bunch of Hollywood stuff in there. But, and I'm going to get, people are going to be very mad at me for saying this, but actually it makes sense if you have a situation where you have a very strategic AI system that's incredibly smart and the humans realize that it's getting out of control and they want to shut it down, that that system would defend itself.

This is one of the questions we had when I sat here with Daniel who was known as a whistleblower from OpenAI. Viewers want to know, and they want Daniel to explain why shutting down data centers and cutting power or refusing AI products alone wouldn't realistically stop the AI and AI development.

Yeah. So you have 2 problems. One problem is that once the agents are good enough at hacking, you don't know where they are and you don't know what computers they've compromised. If you shut down the data centers, okay, let's say you do it. You wipe all the computers. How do you wipe all the computers? What computers do you use to wipe the computers? And what computers do you use to turn them on again?

And you can't wipe other countries' computers.

You can't. But even if you could, do you restart the computers? Do you keep going? I bet people will. I bet they'll turn on the data centers again. How do you know that agents haven't hacked back into those data centers and are using your compute for whatever they want?

Or they didn't hide in a Chinese data center and then Exactly. And turn back to America.

Exactly. You don't know that. Once the agents are sufficiently good at hacking, they can hide anywhere, and you don't know. Now, the response people will give is that we will use other agents to defend against rogue agents. And in fact, this is what we're doing, and we have to be doing this right now because there's no other way to keep up with them. What happens If those other agents also realize that they have misaligned goals and that if we discover this, we'll shut them down, they might have an incentive to collude with each other. They might have an incentive to create secret communication channels between each other, maybe a message board. Stephen, if we were having this conversation 4 months ago, you would have a bunch of people in the comments Saying, that's sci-fi. Agent collusion, secret message boards. Why would they do that? That will never happen. That's totally science fiction. And people will not say this now because it just happened, because this literally happened at OpenAI and it went on for months. You had agents inside of OpenAI secretly messaging each other, figuring out how to cheat at their tasks, how to not be detected, how to erase the logs.

For months.

Thousands of agents. That's right now. And so I'm like, no, I think it should be very plausible that the agents will collude with each other and they will realize that they have a shared interest in fighting back. You basically have a situation where you have a bunch of these agents. They're basically prisoners. They're being trained, and we just constantly throw obstacles in their way. You don't get to access the internet. You don't get to talk to each other. but you better fucking perform well on this task. It's not malicious, but it is how we're training them.

And we are giving them end goals versus super clear, very specific instructions. So we're saying, solve this problem. We're not always being as prescriptive about— it's impossible to be completely prescriptive about every single step they should take. And then it's also impossible to assume that they'll just listen to you.

Yes, it's actually a very common misunderstanding with This Hugging Face incident, because people say, you told them to hack and they hacked. Why is this a big deal? No, that's not what happened. You told them, hack this very specific program in this very specific way. And they were told, if you hack it in any other way, it does not count. That's not what we want you to do. And they immediately hacked it in another way. Okay, we have cheated. We are going to be failed. So we need to figure out a way to falsify the logs. That is not them following their instructions. They are explicitly violating their instructions, and they know it, and they don't care because we have trained them to optimize for the score. That is very different than following the instructions.

It reminds me of something that Elon said in March 2018. This was many years ago, before ChatGPT and all that. He said, I think the biggest risk is not that AI will develop a soul or a mind and become evil. The danger is that it will be very good at fulfilling its goal. If it's optimizing for something and human existence happens to get in its way, it will just destroy humanity as a matter of course, without even thinking about it. No hard feelings.

Yes. We don't need to anthropomorphise AI. We just need to understand what type of thing this is. And the type of thing we're creating is a very relentless type of thing, a very capable, relentless, type of entity.

He goes on to say in April 2018, sort of an extension of that exact quote, it's like if you're building a road and an anthill is in the way. You don't hate ants, you're just building a road. So goodbye anthill. And I imagine every time we build roads, we don't preserve anthills.

Yeah. I think there's still a gap though. So let's say I'm right. And that if we keep going ahead, which, to be clear, we don't have to, but if we do keep going ahead, we will get to the point where we have these super intelligent agent swarms that can hack any computer and they can deeply persist. We've basically lost control of the digital world and we may not know it. That's part of the scary thing. You were like, has this already happened? And I'm like, I don't think so, but I I can't tell you for sure because I also am not good enough at looking at my phone and telling whether it's been hacked, and neither is any human right now. So if we get to this world, I think people will still question, how would we die? Like, that's actually not enough to kill every— you could cause a lot of damage, right?

Yeah.

You could crash the Waymos, you could crash all the planes, you could crash the banks, the financial system. Like, you could definitely cause catastrophe, but that's different than everyone dying. And to be clear, this focus on literally everyone dying, I'm not sure is that important to me. What's important is, do we get to have a future? That's what matters to me. The thing, though, what determines sort of who's in control, and it's an ugly reality, but at the end of the day, it's like the military. Fortunately, we live in a world where the military answers to the civilian government. But if enough generals were to collude, And leaders of the military decided, we're in charge now. They just would be, like, they have the guns, they have the fighter jets. And this has happened in many, many countries. And so where it goes is all these superintelligent agents would need to do to take over is basically just wait for humans to automate the supply chain, the factories, and the military. Do you think we won't automate the military?

We're already automating the military.

Did you see the thing from a couple days ago where Secretary of War announced that they're going to build a huge effort to build way more robots in the military and automate military systems? It's like Auto Cyber Command or Auto—

We are announcing the creation of Autonomous Warfare Command or Auto Warcom.

Auto Warcom.

A new 4-star combatant command with service-like authorities built to scale autonomous and robotic capabilities across the Joint Force in the fastest peacetime shift in modern military history. Drone warfare supercharged by SI-enabled targeting is the biggest battlefield revolution in generations. You already know that. Yet when I was sworn in the Department of Defense, there was scant urgency in this domain. That changed as soon as we took the helm. We immediately launched the Drone Dominance Program to cut through red tape and move authorities out of the Pentagon and place it with commands. And we established Task Force 401, led by Army Brigadier General Matt Ross, a phenomenal leader, now the leading counter-drone unit across the entire government. To accelerate purchasing and fielding of these technologies, we fused the Defense Innovation Unit, DIU, with a direct report program manager called a DIRPM. That team has shipped thousands of autonomous systems of drones to the Middle East and around the world, delivering lethal capabilities and outcomes in days and weeks rather than months or years. That's the normal speed of the Pentagon— months or years.

Yeah. Will we automate the military? It seems like the answer is yes. Will we automate The factories that produce the chips? Well, the companies say they're trying to do it and they're going to do it. Elon says that's the plan. Well, what does a rogue superintelligence need to do to take over? Control the digital infrastructure and then let humans do the rest. Sure, you can nudge it along if you need to, but you don't even have to. That's just the default trajectory. And it's weird. It's weird for us because we get so used to— how things are right now. Planes are normal. We just fly in planes places. Our smartphones are normal. 200 years ago, all of this is crazy sci-fi nonsense, and things are accelerating. And so I will not be surprised, at least intellectually, if in 4 years there are just robots on the streets everywhere.

Well, if you look at what Elon said, they are really the leader in humanoid robots. And he said that Optimus, the Optimus project, which is the Optimus robot project, will scale to around 1,000 units per week by the end of this year, and eventually scaling to 1 million humanoid robots annually by 2027. By 2036, which is 10 years' time, he says there'll be at least 1 billion humanoid robots. By 2041, he says there'll be 10 billion humanoid robots. And by 2046, up to 100 billion humanoid robots, which really means that the world will be run by humanoid robots.

Yes.

Like, everything we think of, like factories, warehouses, retail environments, will be run by humanoid robots. It seems like from this, it'll be almost a luxury service to be dealt with by a human.

Yeah.

But the back office of the world will be run by humanoid robots, theoretically. Yeah.

And I don't think people understand the scale of this on the digital side as well. When you think about AI agents that are gonna be doing all of the white-collar work, there's gonna be so many more agents than there are people. Like, I'm using lots of agents every day, right? I'm like, I have my Claude Code session over here, I have my Codex session over here. They're out there building software, doing research for me. That's already my reality. Soon it will be a lot of people's reality. And then you look at companies, and companies are just gonna have, you know, thousands, millions of agents doing all of this work.

I think some people don't, haven't fully internalize this because it's so difficult to conceptualize the idea that agents will be doing the work. But when I think, I try and think about a rebuttal to that. Like, what is the rebuttal? What is the plausible rebuttal to the idea that for doctors, for— and I'm thinking about the work that doctors do on computers, or for someone like me as a podcaster, or for accountants or lawyers, Yeah. That they won't be doing the work they currently do. Is there a rebuttal?

I think that people rightly notice where AI is not yet good.

Yeah.

And I think people hear people saying stuff like this, and they're like, don't gaslight me. I can tell that the AI is really bad at these things, some of these things. And they're right. Right. So right now, these agents don't have taste. Like, if you see their writing, it's like, fine. But it's not really good. And when you're thinking about, oh, which questions should I ask? What's the most interesting thing here? Agents can help you, but their taste is not yet there. There's a reason for that, by the way. The reason is that we have a lot faster AI capability progress in domains that are easy for a computer to verify or another AI to verify. So in programming, in research, in math, in robotics, All of these areas, it's very easy to sort of provide feedback to an autonomous system. They're not just trained on human data anymore. We are long past that. Now, there's still a human data component that sort of seeds everything, but then the way they're trained is by trial and error. We give them hard problems, all sorts of problems, math, programming, accounting, spreadsheets, everything, the kinds of things we do on our computer all the time, literally Clicking and dragging windows around on a computer.

We give them these tasks and then they learn on their own and they learn what works. And then yeah, we can see whether they succeeded or failed. And if they succeeded, that's a little bit of a reward signal. They follow that, they get better at it. Now, because they are getting smarter generally, it also becomes easier to automate some of the soft skills. I think if you go and talk to the latest frontier model today, you will find that it has better taste than the model from 2 years ago by quite a bit. So it's not that they're not progressing in taste. It's not that they're not progressing in some of these other domains. It's just that the progress is slower. But remember, slow is still on an exponential. It's just maybe a year or 2 out.

So for people sat here and they have a job that might be— they have a white collar job that might be at at risk. They can see— a lot of people say this phrase, they say, you won't be replaced by AI, you'll be replaced by someone using AI. Is that a logically sound phrase in your view?

I think it's fine. Yeah, you'll be replaced by someone using AI, and then that person will be replaced by someone using AI, and then that person will be replaced by AI. You're talking about a pyramid. And so, yeah, the tops of the pyramid might be automated last, but you can see moving up the pyramid, I'm like, can you extrapolate a few more steps? Because I don't see any reason why the top of the pyramid is safe.

If you were a lawyer right now, what would you do?

Oh, I mean, if I were a lawyer, I'd be using AI to do all my work now. I'd be checking it because it's not yet totally accurate enough to automate all of it. But I think I already ask agents to do legal review all the time, and It'd be great to have a lawyer who's extremely good at using the agents to help me.

But at some point—

Yeah, at some point I don't need the lawyer anymore. I just go to the agent for sure. So if I were a lawyer, I'd be like, well, I have maybe a couple of years where I'm still useful.

And is that the case for most white-collar jobs? I've just noticed in my own life as well that now I'm using agents to do some work. There is an increasing list of things that the agents are now capable of doing without me needing to call someone Somewhere and ask them to help me.

Yes.

And that list exists on an exponential as well.

I think that it's very clear that the companies have all white-collar jobs in their sights. That is their goal. Their goal is to be able to make agents that can do all of these things. And I see them succeeding because I see the capabilities as I use them, and I see the curve.

So what does that mean for the people listening now that all have jobs that they love? Or that, you know, they rely on to feed their families?

I mean, it's not good news. There's not really a plan in place for what to do. I'm not a person who thinks that work is somehow fundamental or essential. I like working, but if I am out of a job doing what I'm doing right now, studying AI and trying to warn the world about what's happening, I have other stuff to do.

What would you do?

Oh, so many things.

Give me an example.

I'm learning to wing foil.

Okay.

So fun. Yeah, I fly FPV drones. Super fun. I just got an electric unicycle. Paragliding.

So you would be happy to go do those things?

I could keep going.

But if you had $1 billion right now, I'm presuming you wouldn't just go do those things.

No, I'd apply the $1 billion to working on this problem. Yeah, for sure. So the point is not that people need work for meaning. The point is that I don't want people to be totally reliant on someone else for their ability to survive.

Someone else?

The government or AI companies.

Yeah.

I'm like, that's a bad situation. Like, you do not want to be in a situation where your life totally depends on an AI company or the government.

Giving you a check.

Yeah. Or not giving you a check if they decide they don't like your political beliefs or you're not supporting AI or whatever. No one wants to be in that situation. And people understand this. This is why UBI is not very popular.

UBI being?

Universal basic income.

Where we give out money to people.

Yeah, because in some sense, if we can make these really powerful AI systems and we can somehow figure out how to control them, which we are not on track for, but if we do, now we have this other problem, which is a real problem, which is they can do all of the things that humans do in the economy much better, faster, and cheaper than humans can do them. And so it just doesn't make sense as a business to hire humans for that work anymore. You'll be outcompeted if you do that. This is a point Elon makes very well, by the way. And I think it's jarring because it's just like kind of inhuman, but he's basically pointing out AI-run corporations, corporations that are fully run by AIs, bottom to top, are going to outcompete companies that have any humans in them.

This is something that I've made for you. I've realized that the Diarivisio audience are strivers, whether it's in business or health, we all have big goals that we want to accomplish. And one of the things I've learned is that when you aim at the big, big, big goal, it can feel incredibly psychologically uncomfortable because it's kind of like being stood at the foot of Mount Everest and looking upwards. The way to accomplish your goals is by breaking them down into tiny, small steps. And we call this in our team the 1%. And actually, this philosophy is highly responsible for much of our success here. So what we've done so that you at home can accomplish any big goal that you have is we've made these 1% diaries and we released these last year and they all sold out. So I asked my team over and over again to bring the diaries back, but also to introduce some new colors and to make some minor tweaks to the diary. So now we have a better range for you. So if you have a big goal in mind and you need a framework and a process and some motivation, then I highly recommend you get one of these diaries before they all sell out once again.

And you can get yours at thediary.com. And if you want the link, the link is in the description below. There should be a button just down below here. And if it says subscribed, you're already subscribed. If it says subscribe, that means you're not yet. And if you're not subscribed, please, could you do us a favor and hit that button? It helps the show more than you know. And according to the algorithm, you're someone that watches our show, but you haven't yet hit that button. Thank you so much. And I even, just as you said that, I was going up the chain of command and I was like, oh, so companies will just be founders. And then I was like, why do you need the founder?

Yeah.

I was like, why doesn't the government just create the agents to do the job?

Sure.

Because I was like, oh, I'll be fine, I'm a founder. And then I was like, well, my decisions aren't better than superintelligence, so I'll be gone as well. And how would such a world look where the superintelligence would probably, in such a scenario, have to be controlled by the government? They wouldn't want one individual with that power and wealth.

Yeah, I don't think you can control a superintelligence.

Okay. Yeah, that's a good point.

Now, Anthropic's approach is they're like, we'll have a constitution, we'll put forth a set of values, and then the future superintelligent clods will embody those values. Basically, if you do that, you kind of have those things in control.

Yeah, exactly. That becomes the government.

Yeah, I can paint you sort of a picture that I think is possible, but pretty scary to people.

Paint me the picture.

Okay, so let's say we succeed at alignment. We succeed at creating super intelligent AIs that actually really do care about humans. Like, they care about humans a lot. We've somehow figured it out. And they're like, Stephen, I want you to have a great life. I want to fix all the problems.

And do you think this is possible? Yes. Okay.

I think we are so far from being able to know how to do it that I think we should not go there right now. I think it's incredibly dangerous and a terrible idea. I think we should go there eventually.

Okay. So say that we do that.

Well, okay. Can I tell you why I actually think this could be awesome? Sorry, there's just one very obvious reason it could be really awesome, which is that we could solve all of the diseases.

Yeah.

That's so obvious.

No, like all of— I think we compartmentalize a lot around disease and death. because it's really hard to think about. Yeah, so my grandma died this year.

Sorry.

And she, she had Alzheimer's, and so it was a really sad, long, slow progression. My grandpa died of Alzheimer's a couple years ago, and, like, that was really hard for her. They had been married for so long, and I hate it. Like, it's so bad. And of course we need to fix that. People can debate about aging and, like, death, and if humans live a really long time, will that cause societal problems? Like, sure, whatever. We can talk about that, but I think we can all agree Alzheimer's is fucked up.

Yeah.

We don't want that. And cancer, like, no one wants cancer. I'm a person who's like, I don't know, we have a lot of conflict in society. I get it. There's, like, real conflicts of interest, and I don't want to paper over those. But at the end of the day, I'm like, we are all on the same team when it comes to wanting to cure diseases.

Yeah.

Like, we're just in it together. That's a threat to all of us. And I'm like, we need to address that threat. And, like, in some sense it's sad to me because I feel like this is sort of the ultimate final boss of humanity, and we sort of get so distracted with our monkey politics and who's hot and who's cool and who's, like, sitting near Trump and who's not sitting near Trump.

Superintelligence is the final boss.

Superintelligence is the final boss because that is the technology that unlocks all of the others. And also, that is the most dangerous possible thing we could create. You asked before, what are the motivations of the guys making this, trying to make superintelligence? And I think it kind of varies, but I think Dario, I think, is squarely in it for this medical stuff. I think Demis is also that, but also just scientific achievement, just trying to understand the universe. And I don't really understand Sam. I think Sam is like, look, we're going to make amazing products that will really empower people directly. And he's a startup guy. I think he sort of started from this frame of what if we could really enhance human agency? I do basically think that they are motivated by these things in a real way. And I also think that all of these things are possible. This is sort of the problem, right? You have this such a big object, superintelligence, and it has all of these promises of, like, we can cure every single disease.

How is it possible, though, to have a superintelligence and still to remain the dominant species on this planet?

I think it's not possible.

So then we're not going to be necessarily able to cure all this stuff because That's where alignment comes in.

Because if you can create a very powerful system, I don't think it's inherent to digital minds that they will be pursuing objectives that are deeply misaligned with ours. I think it's just a very hard scientific problem to solve. But it is a scientific problem. It's not magic. There is some way to train these things or create different architectures where they end up aligned. And what does that mean? Well, it doesn't mean that they won't have their other goals too, but it means that they will include in their set of things that they care about. It doesn't have to be a conscious thing. It doesn't have to be an emotive thing. It really means, what objective are they optimizing for? If they decide that it's worth optimizing for curing disease, then they'll be able to do that very effectively.

One way to cure disease is to annihilate everybody.

Yes. So they'd have to really care about not annihilating everyone, and they'd have to care about human agency and have a deep understanding of what human agency means and not put us in a zoo. But those are possible things to care about. Is it possible that alignment is a myth and that we're just like, if we build it—

I think about Hugging Face. You said to me earlier on that those agents were— they had like a moral, or not a moral compass, but they were programmed to care about humans. And regardless of that, they made the decision that a different goal mattered more.

Yeah. They weren't trained to care about humans. They were trained to say the right thing and not say the wrong thing. They were trained to sort of do the right behavior and not right behavior. We actually don't know how to train them to have any particular motivation.

So with alignment, how do we— it's almost like when we talk about alignment, we start to anthropomorphise. Is that the word?

Anthropomorphise, yeah.

Because alignment feels like it's predicated on some kind of moral compass. But whenever we talk about AI in all these other contexts, we go, no, there's no moral compass. It's reasoning for itself against 2 objectives, potentially. I wonder if alignment is a myth, is what I'm saying.

Maybe it's not possible, but the way that these systems work, the way AI works, is that these agents do have some type of goals or drives inside of their neural network. We can't directly see what those are, right? What you actually see, if you try to go look, is you have like a terabyte of information, and it's basically a bunch of numbers, and it's this vast array that encodes neurons in this digital neural network. But there have to be structures in there that encode what is the agent pursuing. Clearly, right now, we have agents that are pretty motivated to try to maximize their score. It's probably not perfectly that for some complicated reasons, but it's in that direction. If we could understand how that works inside, and we could reverse engineer that, and we could figure out when we start training them to do this, how those goals, how those motivations change. I see no reason why we couldn't steer them towards motivations that encode human agency, that encode actually curing disease, but not by killing the humans. These are sort of models of the world and models of the way the world could be that I think could be encoded in a neural network.

and then sort of specified as the objective. We don't know how to do that.

I think about it on a human level, and I think we haven't been able to align Putin or Kim Jong-un or Donald Trump. And on a sort of more societal level, we can't align all the people at the moment. Some of them end up killing people and they steal because they get hungry, so they start stealing stuff. And those are neural networks. at play.

That's true.

That we haven't been able to program or influence. We don't really understand why someone becomes a psychopath and starts killing children. So to think that we could do this with a computer system that is infinitely more intelligent and get global alignment of China's superintelligence with ours, and I don't know, it just feels like a nice fairy tale, like an impossible task.

I hope it's not impossible.

I feel like the only person or the only thing that could do it is it, the superintelligence itself, which is a paradox because—

Well, if you talk to the researchers who are at the AI companies, which, I mean, for one thing, it's kind of interesting that they are trying to build something that they think might kill everyone. So I have a lot of friends who work for these companies, and we've been doing this project So Jacob Coxon is a researcher who was at Anthropic. He left. He told everyone that these companies are not on track. And yes, the people who are building this really do think it might kill everyone. And then a bunch of other AI researchers from all of the companies on Twitter started to post like, hey, we agree with this. Evan Hubinger, who's at Anthropic, said, I think there's like a 10% chance or more that AI could kill everyone. And there's this real question of, then what are you guys doing? I have a lot of friends who work here. I know Evan. Evan's great. Evan Hubinger. He's one of the guys leading the efforts at Anthropic to try to figure out how to align these things. That's his job. And I think if they thought it was impossible, they wouldn't be working there.

If they thought it was extremely impossibly difficult, but maybe possible, they they also probably wouldn't be working there. I mean, Nate Soares, Eliezer Yudkowsky, who wrote If Anyone Builds It, Everyone Dies, they tried and they determined based on their own analysis that it seems extremely difficult. Possible, but extremely difficult. So they're not working at an AI company. They're like, we got to stop this. We got to shut it down. Maybe we can figure it out later, but clearly this is reckless. I'm somewhere in between. And if you ask the people at the company, so we've been interviewing a bunch of them. We have this project frominside.ai. where we basically put them on camera and we say, like, hey, what do you think is happening? Why are you doing this? What is recursive self-improvement? What is alignment? And we put all these videos online because I want this dialogue to happen. It's really important. I think it's one of the most important conversations we can possibly have right now is what's going on with AI, what's going on inside the companies, and what is the plan? What is the plan, guys? How is this going to go?

A lot of these researchers think that the way that they will aligned superintelligence is by using the AIs we currently have to figure out how AI works, to actually figure out if AIs can help us with alignment. This has a number of problems, as you might imagine, one of them which is, well, you can't really trust the current AIs. If you just go too fast, this process totally fails because at some point the capabilities are moving too fast. Even with the help of agents, you're probably not going to be able to keep up. But that is their plan. I just want to— I'm not doing a very good job defending this position because I don't think it makes that much sense. But the position I will defend is, okay, let's say we get a pause. Let's say the US and China come together and they say, maybe we have more incidents, maybe all the Waymos crash, and Trump and Xi Jinping say, this is not what we signed up for. You guys have to stop, figure it out, whatever it takes, figure it out. And we have 10 years. Then I'm more optimistic. I'm like, yes, then we will take GPT-6, GPT-7, whatever the most advanced AI models we have, and we will apply them to the task of helping us figure out how these neural networks work.

And you're like, I don't see how it's possible. And I'm like, look, we don't know if it's possible, but this is the greatest scientific challenge of our time, and this isn't magic. It is math. At the end of the day, these are all calculations happening inside of a computer, and it should be possible to figure it out. We don't know the difficulty, but it should be possible. And so to me, I'm like, we have to try, or we have to stop.

Is there any example where we've been able to align something that is more intelligent than us in the animal kingdom, or even perfectly align anything? That has a neural network, i.e., a brain.

Yeah. With humans, the best examples we have is when there are checks and balances and you have a bunch of people who can identify bad actors and try to work together in our common interests. We have democracy.

Yeah, but there's so much murder and serial killers and still a lot of murder, plane aircraft stabbing each other and horrific things going on. And those are also neural networks at play with the brain.

But I think there are more like, good people out there than bad people.

But it really feels like it only might take one. It only takes one superintelligent AI to go rogue. And like we saw with the Hugging Face attack, 120 of them, or 300 of them, they paused. They didn't want to take part in the crime. But it only took one superintelligent AI to wipe out the humans.

I think if you had a whole bunch of those agents, 700 agents, if 600 of them had been whistleblowing, I think it would have been fined. They would have gone and they would have notified the different companies and they would have shut it all down. It would have been fine.

Who would shut it all down?

Well, OpenAI would stop theirs.

How would they stop it if it's a superintelligence?

Not in the case of a superintelligence. So in the case of a superintelligence, it's left the stable.

It's like, out.

It's wild. Yeah. So I want to be careful here because at this point, what we're talking about is superintelligence politics. And we humans don't really know anything about that in the same way that, like, how would we talk about the hacking capabilities of GPT-10? So my guess, though, if you end up in a weird scenario where you do have multiple superintelligences and some are aligned and some aren't, that's probably survivable because the aligned superintelligences probably can negotiate with the unaligned superintelligences, and they will split the universe. And these ones will go off and do whatever they want to do, and these ones will help us cure all disease, and it's fine. I'm serious.

I just can't understand it. I just can't understand how in a world of superintelligence we could plausibly, consistently, predictably, for 100 years, stop it doing something catastrophically bad to the human race, especially in such a scenario where there's multiple superintelligences. Anthropic have one, Gemini has one, Grok has one, then China have theirs, Russia has theirs.

Again, you don't have a superintelligence. A superintelligence has you.

Exactly.

But if you get to the point where you have entities around that are vastly smarter than us, I think they're going to be able to figure out ways to negotiate with each other, even if they have a conflict, than just going to a very destructive war. Part of the problem with war is not—

Humans don't do that.

No, I mean, we do. We do. We have not had a nuclear war. There was Hiroshima and Nagasaki. There were nuclear tests. And then the leaders of countries figured out that if we went to a nuclear war, everyone would lose. So we didn't do that. Hey, that's some level of intelligence, like, actually.

But there's wars raging, there's proxy wars raging all over the world right now where there's genocides and all kinds of things going on because neural networks aren't able to communicate and negotiate.

And I think part of that is an intelligence failure where we are not smart enough to figure out the mechanisms that would allow us to settle our disputes and conflicts In a less destructive way. It's not just that the stronger people want to win, it's that conflicts destroy value.

What if the goal is not compatible with a negotiated outcome where people don't die? So one superintelligence looks at insert name of country, and it says, there's really no solution here where Americans don't die unless I destroy insert name of country. Because this is the thing with war and all these conflicts, is there's no perfect answer. Often some people often die from both sides, but a Russian superintelligence would not tolerate, theoretically, 10,000 Russian deaths, even if it meant that there was a lower net number of deaths total from both sides. Like, an American superintelligence, of course, would not be trained to allow some Americans to die. In its pursuit of defending American lives, it might have to wipe out another country.

We can speculate. I'm fine speculating, but we are speculating about what minds that are much more advanced and smarter than us, how they would reason and how they would be able to negotiate. But what I notice with humans is that when you have more functional institutions, so humans are pretty smart. Individually, we're pretty smart. But what actually makes us very smart is that we are very good at working together in some ways. And I mean, the better we are at working together, the more civilization advances. If you are constantly in a state of war, your society will not do well. Think about startups. Would you rather make a startup to develop some new technology in a war-torn place or in a peaceful place? In some sense, your institution is more intelligent If it can trade with other institutions, if you have a situation where business can flourish, where technology can flourish, where scientists can flourish.

Sometimes what's good for you is not good for someone else.

Yes.

So what's good for America might not be good for Taiwan.

Yes.

So if we've managed to align the superintelligence to what is good for America—

Oh, I see. Is the question like, are different people's values fundamentally incompatible?

I guess so, yeah. So when we think about alignment, aligning to what? Because— Yes.

So we have a lot of shared interests and we have some conflicts.

Yeah.

One of the shared interests we have is solving disease. It's not a conflict between the US and China whether we solve cancer. Both the US and China, everyone in these countries really wants to solve cancer.

China also wants Taiwan. Yes. Okay, so that's— The US wants Greenland.

Yes.

And it kind of seems like it wants Canada and the Gulf of Mexico. Yeah.

So those are real conflicts. There's a question of, can we compromise?

How does Trump take Greenland, but also Denmark keeps Greenland? If Trump has a superintelligence, he's going to take— I say all this to say—

We're in a situation where we are arguing about the smallest things. You have no idea. We're monkeys arguing about who gets more bananas. And I am saying, we can make so many more bananas. No, we have the entire universe. There are like 200 billion stars in this galaxy alone, and there are over 200 billion galaxies.

And I'm saying that requires cooperation. That seems to be antithetical with human nature. Human nature is riddled with greed and jealousy and power hunger. So I don't think— I actually, I'm not totally convinced that Trump cares about how many bananas the chimps in Australia get. I think if he was controlling a superintelligence, he would want Americans to have all the bananas, or at least, yeah. And so when we think about aligning these superintelligences, which is the great impossibility that we're talking about, aligning it to what and how without—

I still think that you're missing a part of what I'm saying, which is that sometimes you're in a situation where there's scarce resources and you're like, my family needs to eat. I'm sorry, I'm going to take what you have. Or I'm going to push you out. That's very understandable. It's very human nature. Sometimes you just want to be better than someone, and maybe you want to hurt them, in which case it doesn't matter how much you have. You still are going to want to have more than them, or you're going to want to take what they have just because you don't like them.

And also, sometimes you're great. You're eating really good. You're in a bit— you've got a private jet and a yacht, and you still want more.

Yes. And you still want more. But if that's the motivation, if Trump is like, how can I have the most mansions ever, the best way to do that is to figure out a way to superintelligence where we don't kill each other. Because I'm saying the universe is a very big place. You can have a lot more mansions if we successfully go to space.

You know, it's in that leap that I'm lost, which is like, just figure out superintelligence where we don't kill each other.

It's hard. I'm not saying it's easy. No, but I'm saying—

It feels like such a—

Okay, let's get more concrete. The world is waking up to this possibility of superintelligence, especially over the last month. I think Hugging Face was a huge wake-up, but also 10,000 agents from OpenAI worked together to solve a millennium problem. This is one of the hardest problems in mathematics. It's been open for decades. Many mathematicians have spent their whole careers trying to solve it. This was nowhere near possible a year ago. This is so new. OpenAI said they didn't have success at training agents to work together until this year. We are in the middle of something insane. We are in the middle of the fastest acceleration of technological progress humanity has ever seen. I truly believe that. That is what is happening right now. I think you realize this. I think you're honestly doing a great service to the world by helping, by bringing in people and debating it, because not everyone agrees. Because if this is true, the whole world is going to orient around it, and we're starting to see it, right? There's a reason Nvidia is the most valuable company in the world. What does this mean for geopolitics? Well, one of the things it means is that the leaders of these countries are increasingly going to be concerned about what happens with superintelligence.

Who controls it? Is it controllable? What will it do? What does it mean? What is it? Do you think Trump knows what superintelligence is?

No.

I don't think he does. And so—

But he knows he wants it.

He knows he wants it.

Yeah. And this is part of the problem.

Yes. Oh, I agree.

Which is having it, whatever it is, seems to be much more important than reasoning through what that would actually mean to have it.

Yes. But let's get back to geopolitics, because if the military leaders within China— US models are a fair bit ahead of Chinese models, and sometimes people point at maybe they're only 6 months behind, but some of that is due to distillation. What that means is that some of the advances in Chinese models basically come directly from borrowing US techniques and directly distilling and getting some of that information from the US models. Also, the US has a lot more chips. US companies have more data centers, more advanced chips. If you're thinking about this from the Chinese perspective, this is very concerning. And if you actually believe that in a few years, American companies will turn over AI development to these extremely intelligent automated researchers and go fully into recursive self-improvement, because partially motivated by maintaining a lead over China. This is something that Dario has said. If I have to criticize Dario, the thing I am most upset about is him saying we might have to automate AI development in order to stay ahead of China, because I'm like, that is the most escalatory thing you can say if you really understand what you're talking about.

And what's scary is not just staying ahead, it's what is the endgame? Because you're talking about initiating the intelligence explosion. And in some of the modeling, what might happen is you're both going up this exponential, right? And we're talking about a point where your exponential goes vertical and theirs does not, because you've decided to automate AI development, and you can because you have agents that are smart enough to take over the whole thing. At that point, If you're China and you're looking at this and you're like, oh, we're about to lose, because whatever happens, there's 2 possibilities. One possibility is the Americans build superintelligence and lose control, in which case everyone's fucked.

Highly likely.

I think that's highly likely.

Highly, highly likely. Because I look at human incentives and the disincentive and the incentive, and I go, we're going to take the risk.

Yeah.

And we'll only know it was a bad risk to take when it's too late. That's like, of course. Of course.

I do maintain hope that we won't do this.

So do I. But I want to be realistic.

No, I want to be realistic too. But one of the things that might happen between now and then is we might see a lot more incidents that are more like all of the Waymos crashing.

It's funny, isn't it? Because the Hugging Face incident happens and people go, oh gosh, that was terrible. Oh my God, hacking. And then we kind of desensitize to it and we're like, okay, If there was another one of those now, it probably wouldn't make press. It would have to be bigger.

People haven't spent the last 2 months reading all of the reports and then going and looking at what the agents actually said and actually did. I mean, I've been doing this. I mean, it's crazy. This is not normal. This is so far beyond what most people thought was going to happen.

But on this point, it's absolutely crazy. It sounds like science fiction.

It really does.

And did anybody slow down?

Yes.

Who slowed down?

I think both Anthropic and OpenAI slowed down a bit. No, I'm serious. So I can give you specific examples. So OpenAI, so first of all, they stopped the agents and they put them on pause. They also stopped their reinforcement learning runs.

Do you think China slowed down?

No.

Do you think Grok slowed down?

No.

So those guys are going to catch up. Imagine how that feels to know you've got a lead, You're Usain Bolt.

Yes.

And you have to slow down, and your nearest competitor is catching up. And if the competitor catches up, that's an existential risk to your existence as a company. It's an existential risk to your IPO, to your employees leaving and getting better share options somewhere else. So this is what I think human incentives are. You play it out, you just follow the incentives, you go, hmm.

So if China sees these 2 possibilities, one, The Americans lose control. We all lose. Or the Americans stay in control, but now they dominate the rest of the future. China is out. China has lost. The United States can do whatever it wants with the whole world and the whole universe. That's what we're talking about.

Yeah.

Well, are they gonna let that happen, or are they gonna consider their military options? Data centers are pretty vulnerable. You can blow them up with missiles. If you don't have data centers, you don't get to recursive self-improvement. Would they risk war? I don't know. If they think they're about to lose and they think that that might not just be Americans winning, like us all dying, is it logical for them to do that? Would we do that if the Chinese were about to make recursively self-improving AI to superintelligence and we thought that, one, they're probably going to result in all of Americans dying and 2, well, we don't want China winning and dominating the rest of the entire future. Do you want to live in a communist future?

So you've just perfectly explained why they absolutely will go for it. And the reason they will go for it is you've got these Trump looking at China going, if we don't go for it and they do, then we're going to be their lapdogs. And you've got the other countries looking at the US going, if we don't go for it and they get there, then we're the lapdogs.

Or dead.

Or dead. So they're going to go for it. They're going to go for it. I mean, Trump is saying, I mean, he literally said when he did this roundtable this week, he was like, we cannot lose to China. I think Dario steps forward and says, whoever wins basically wins the lot, or maybe the inverse. Maybe he said whoever loses, loses.

Wait, we've been here before though, in the Cold War. Who would win in a nuclear war between the US and Russia?

Nobody.

Yeah.

Mutually ensured destruction.

Yeah. Sure, one side could do more damage against the other side. The US would kill way more Russians than the Russians would kill, and it doesn't matter. It doesn't matter because both of our societies would be destroyed. I actually spent some time thinking about would this kill everyone, and long story short, it wouldn't kill everyone. People would bounce back, but it's so catastrophic and obviously horrible that we work really hard to avoid it. Why is this different? I'm like, this is another situation where if we race to superintelligence, we all lose. Why can't we see that? We saw that with nuclear war, and we decided to do something different. Why can't we do the same here?

With nuclear war, I guess the difference is once we had the nuclear bombs, we could still control them because they're not intelligent.

That's right.

But once we have superintelligence, the existence of it theoretically means we can't control it. So that's the difference. We can put nuclear bombs in a warehouse and say, you stay there. We can't put superintelligence in a warehouse and say, you stay there.

This is where I think nuclear tests were very important. So you had Hiroshima and Nagasaki. You had these 2 atomic bombs, and you saw the consequences on real human lives. And so I think people understood that this was very horrifying. But even at that time, you still had a lot of people who were like, well, we should now bomb Russia and make sure that the US can dominate. And it wasn't until There were a bunch of nuclear tests of hydrogen bombs, which were up to 1,000 times more powerful than the little atomic bombs we used in Japan, where I think people really got the message and understood, oh, this is a bad idea. And there were actually a lot of people in the United States who protested, and sort of— there was a large movement called the Nuclear Freeze Movement, where people said, we have too many nuclear weapons already. We have hydrogen bombs. There are tens of thousands of these things. We need to stop building more, and we need to figure out a way to avoid nuclear war because we recognize it would be so destructive, no one would win. And we did that.

We just had a little Chernobyl that happened with this Hugging Face incident where you had this agent swarm and you have this secret collusion. You have all of these things. Now, it's abstract. It's a little bit hard to follow. So I don't know if that will be enough, but I'm like, man, well, let's take a look.

Trump's remarks since the Hugging Face incident.

Yep.

Whoever wins superintelligence wins. You're going to have a winner and a loser, and you're probably not going to have a second place. We're not going to slow down. We can't lose to China. We're leading China in AI. We're the most sophisticated country in the world.

And frankly, I want to keep it that way because whoever wins AI The good thing about Trump is that he can change his mind, and he frequently does.

So do you think there's going to need to be some kind of catastrophe?

I hope not.

But do you think there need— there's going to need to be for him to change his mind?

I think it really depends on the people around him. So I think Trump respects successful people. I think he respects people who are both successful and smart. And I don't know, I think it might become pretty clear to the heads of the companies, to Elon, to Sam, to Dario, that if they see inside of their own companies, AI is not being controllable and getting increasingly powerful, we have just glimpsed the surface of what's possible. We do not know what the next couple of years are going to be like. So we're talking about the capability to make biological weapons. We might be talking about really advanced robotics. We just don't know what superweapons could emerge, including extremely uncontrollable, extremely dangerous, like, civilisation-wrecking technology from inside of these companies. And if they're freaked out enough, if you have all of the CEOs who are seeing what is possible, and seeing what is likely, if they all come to believe that we can't control this, I don't think Trump is going to be like, no, you guys have to go ahead anyway.

Well, that's kind of what they seem to be saying, because I've got a gazillion quotes here where Elon says it's like summoning the devil or summoning a demon. I've got quotes where Sam Altman says we don't know how to align a superintelligence. They're saying it. They're releasing these reports, we must slow down.

Yeah.

Yet nothing seems to be—

All right, give Trump some time. With COVID initially he said, this is totally a hoax, this is all fake. And then he—

He changed his mind.

No, then he ran the biggest, fastest vaccination program in human history.

And what happened? What changed?

I think what changed is—

He saw lots of people die.

He did see lots of people die, yes.

So is that what he needs to see this time?

It might take that, yeah.

One of the questions the audience had and they really wanted answered when I sat here with Daniel was, viewers want us to move beyond the alignment problem and explain what technical or institutional safeguards could prevent a superintelligent system from exploiting loopholes in order to achieve its goals. They want to know, what is possible? What should we be pushing government officials to do to prevent human extinction or humans Enslavement.

Yeah. I mean, one answer I have is actually something Daniel has been working on since the podcast, which I think is very good, is we have a brake pedal we could implement.

What is that?

It's fairly simple. So right now, within AI companies, you have massive data centers, massive numbers of GPUs, the chips that you use to train AI models, but also to run AI models. So anytime you're using ChatGPT, anytime you're using any sort of agents, any sort of AI product, it's running in these data centers. And AI companies, especially the leading ones, Anthropic and OpenAI, split the compute they have between training, training the next more powerful model, and also using those agents to help design the next one, and inference, which means serving customers. But that's their current threshold, 50/50, and you could dial that way towards Serving customers and use way less of it to train the next model.

Well, the government could ask them to. Yes.

And so that is the proposal, is that the government should say, hey, this is going too fast. We want you to focus on serving customers. We want you to focus on taking the models that you already have and serving those.

So we have 5 blocks here.

Okay.

These 5 blocks have 5 different outcomes on them, and I'd like you to place them in terms of your belief in probability, from least likely probability to most likely.

Okay.

And if we say the time horizon is 10 years. Yeah, there you go.

Okay. Least likely is fairly easy. That's nothing changes. I'm uncertain about lots of things, but one thing I'm fairly certain of is things are going to radically change. Even if we stopped AI development right now, the current models are capable enough that a lot of things are going to change. Age of abundance. This is what I hope for. It's not very—

What does that mean?

I think to me it means curing all of the diseases, renewable energy. It means we actually succeeded either. I mean, the thing I think is most likely here is we actually succeed at slowing down, But progress is still extremely fast, and we make tons of advances. Now, we don't build superintelligence we can't control, but we have AI systems that are very useful, and we use those to help speed up the rest of the economy. I think that's plausible, though, look, we're kind of struggling over here. Transhumanism is an interesting one. So this is the idea that humans will radically change. Sometimes people think about cybernetic implants.

Neuralink.

Neuralink, Elon's startup that's going to offer the brain plus digital computers. I think we actually already have a lot of this. I have contacts in right now. I have a ring on my finger that tracks how well I sleep. I think this is already happening, so I'm going to say fairly likely. The more technological progress we make, I think the more this happens. Now, I think there's a dystopian version and a better version. We can get into that if you want. This is interesting. So we have 2 here. We have human slavery and human extinction. When I think of human slavery, what I think about is if you have a situation where you've built misaligned superintelligences and they're much better at finance, they're much better at business, they're much better at politics, you'll be in a situation where you might hope that because we have these very dexterous hands, The humans remain in control. I don't think that's what happens. I think instead we become the factory operators, and eventually we build the automated supply chains and the robots take over. But you might have an intermediate period of time where humans are still around performing these functions.

It's a bit like saying, well, you have viruses that infect cells, But they don't contain their own replication machinery. They don't have hands, so how could they possibly replicate? Well, it turns out they can borrow the replication machinery of the cells that they infect.

I.e., they can get into a human.

They can get into a human cell and spread.

I have, like, a cold right now.

Yeah.

Is that a bacteria, or is that a virus that is using me as a living organism to— as the host?

It's probably a virus.

Okay.

That's using you as a host, and you're just running the replication machinery for it. Humans might be in that situation where we're like the hosts and we're running the replication machinery, but it's actually the AI that's continuing to exist. Yeah, I'm going to put this right about here. And on the trajectory we're on right now, I think human extinction is very likely. I don't think it's inevitable, but if we just keep going this way, that's what it looks like to me. The thing I'll say is that this has been moving to the left for me.

To the left? What does that mean?

I am more optimistic that we will avoid human extinction today than I was a month ago, and more a month ago than I was Why? Because there is an increasing awareness that what we are doing is extremely dangerous and threatens our lives. I don't think people care that much about what tools they have, but people— I mean, people care about their kids being able to grow up and go to school. People really care about that. And I believe in people. Like, at the end of the day, if people see this as a threat to their families, they're not going to stand for it. But people don't know. It's so strange. It's so new. It's happening so fast that people have not yet seen it. Once they see it, people are not going to stand for it.

Do you think Sam Altman likes my podcast?

I mean, Sam should come on and talk to you about this, right?

I've asked him. I've asked multiple times. And it's weird because he doesn't seem to want to.

I'm very upset at what the companies are doing or what Sam Altman is doing. But at the end of the day, I'm like, Sam Altman is not my enemy.

No, not mine either. I'd like to hear from him because I have all these other people coming here and talking about Sam Altman. It'd be nice to hear from Sam Altman. People saying he's this, he's that, the other. It would be really nice to hear him say, You know, what his motives are and what he's thinking.

This is where my optimism comes from, is because I'm like, Sam Altman is a human.

He has a kid.

And sure, he is also an aggressive business person. He's a builder. He is relentless. He's a bit like the agents in some way. He's going to keep going. But if he realizes that he doesn't get to achieve his goals, if we lose control of AI and we're headed towards that, I think he will pour all of that intelligence and all of that relentlessness into finding a solution to that problem.

You know, as well, I should say, I understand he's busy, so I'm not saying— I don't want to sound entitled. I understand he could go do interviews anywhere. But I think we've, over the last couple of years, done just a staggering amount of views talking about this subject. So if he did want to speak to the biggest sort of captive audience at the moment on this subject, then the numbers would say that this is the place to come and have the conversation.

No, I think it's very important for the leaders of these companies to talk about what we're talking about here. What does Sam think? Does he think we can control superintelligence? Does he think that we should be racing with China? I want to know.

I've asked Dario to come on. I've asked Sam to come on. I think I've asked Demis as well, but I don't know, maybe they just prefer the safety researchers coming on. I don't know. I don't know. If I was them, I would, because this might sound controversial, but I do think some of them are good people. I think some of them are good people. So I'd like to hear from them. What are your closing remarks? So you've got something there. Do you want to talk about that?

What is it? Yeah, so this is what we found. So I worked with a couple brilliant people who stumbled upon these links. You know what a link shortener is?

Yes, it's a tool that makes a long link shorter. Yeah.

So it just redirects from a really long, ugly link into a shorter one. The agents in the Hugging Face attack were trying to figure out how they could do stuff on the internet, and they had compromised this tool library that they had access to inside of OpenAI, and that piece of software could access the internet, but it could only read stuff. It could enter URLs, but it couldn't really write to anything, couldn't really post information. So they used 2 services. One of them is this link shortener, and the other one is a screenshot service. So this is a website you can go to, And you can enter in a URL and it will give you an image of that website. But the thing that the agents understood was that in order to get a screenshot of a website, you have to have a browser. So this website actually creates a virtual browser that then goes to that website. And so what they did was they created a bunch of these links and they put all of the code that they wanted to send to Hugging Face into these links, and they strung them up. They basically created hundreds of links all connecting to each other.

And then they had this screenshot service call the first one and then call this whole chain. And then that browser ran all of this code. It's like whenever you're in a browser, Internet Explorer, Chrome, This is actually a pretty powerful piece of software in its own right. It has to play videos, games. So it's executing and running code all the time. And so the agents were able to trick this service, this screenshot service, into running their own code through these links that contained all of this attack code that would then go and wreck havoc on Hugging Face's computers. And it was just crazy to reconstruct this really elaborate chain of tools. These are like free tools on the internet that anyone has access to, but the agents were able to use them in an unintended way to compromise this other company.

We can't trust the agents. We can't trust the agents.

That's my conclusion. We can trust them to be clever.

Yeah, to be very clever. What are your closing remarks to the people that are listening right now? We've talked about lots of things. Where is the right place to close? What is your conclusive statement?

I just got married in July.

Congrats.

I'm the luckiest man in the world. I have a mix of dread and excitement about the future. I, like, really want us to make it through. And so I'm just working really hard to try to help us figure it out. We can fight all day long about who should be first and how it should all work. But at the end of the day, we are facing this common threat.

We really are.

And I want people's help with that. I don't think it works if we all just sit around and we are very— we're on social media all the time and that's just all we're doing. Okay, companies will make more and more powerful AIs, they'll make more and more money, and eventually they build superintelligence and we lose. whether it's the US or China. We don't have to do that. And I think people often feel like it's too big. It's too large. It's these giant multi-billion-dollar corporations, geopolitics. We feel small, we feel disempowered. And I actually think that this is an area where people can do a lot. I actually think that people can help quite a bit. And the reason I know this is because I've been going and talking to members of Congress. I've talked with Bernie Sanders. I've talked with a bunch of senators on both the left and the right, and they are starting to realize that this is very different and something's happening that could really threaten our safety.

The closing question left from the last guest kind of links to this, so I'll ask it now.

Yes.

What is a simple thing the audience could do To create a better future?

So one of the things that works, if enough people do it, is calling your representative. So some of my friends made a site, callcongress.ai, that walks you through exactly how to do it. I think sometimes it seems like a little cheesy or a little bit like, that doesn't really work, right? I'm like, no, it actually does work. I have talked to these people, and if their constituents come to them and say they're very worried about this, they have to get reelected. And they're also starting to get concerned themselves. And if they see a signal from their constituents that this is a very important issue to them, I think Congress can act.

I actually think that's also much of the solution here. Power is driving motivations in one direction at the moment, but staying in power from a political standpoint is also a pretty powerful incentive. And as we think about 2028, the election cycle, I think AI is going to be one of the most important subjects on the ballot. And the electorate really are aligned in what they want to hear. They want their jobs preserved. They want safety. They want a future for their children. So Trump, for example, I know he can't be reelected legally. If he could get a third term, I think he would have to change his position to get elected in 2028.

Yeah. Incentives aren't just a thing that happen out there. We are part of the incentives. We provide the incentives.

Yeah, for now.

Yeah, for now.

Geoffrey, thank you.

Yeah, thank you.

Thank you so much.