Loading summary
Transcript85 lines
- [00:00]
A
Well, we keep hearing all of the incredible ways that AI will improve our lives. Pushed to the side, though, are all the ways in which AI might bring about possibly our total destruction. What sort of guardrails does government need to have on AI? Of course, we recently just saw the White House national security concerns over Anthropic's new model called Fable, which was apparently incredible to those people that got their hands on it, but rose major national security issues. So what sort of guardrails are in place and how dangerous is AI becoming? Is it perhaps far more dangerous than human beings at this point? We just don't know about it. Well, let's talk about it with Nate Soares. He's the president of miri, which is the Machine Intelligence Research Institute and a New York Times bestselling author. He's a warning about this to place guardrails for AI and what's coming. Nate, great to have you on the show.
- [00:56]
B
Thanks for having me.
- [00:59]
A
So where are we right now with AI? Here is in the middle of the summer in 2026.
- [01:06]
B
Recently, like you said, Claude Mythos, an anthropic model is a superhuman cyber hacker which in many ways poses some national security risk. That's where we are today. One important piece of the puzzle on AI though, is that it's a moving target. So in January, there were not very many entities on the planet that could hack into every critical system more or less in March. In January, it was only nation state actors who could do this. In March, it was nation state actors plus an AI out of Anthropic. That's a big change if you are expecting the current world order to stay as it is. And that's just on the cyber side. AI keeps improving and I advise people to try and watch where the puck is going, not just where the puck is.
- [02:02]
A
So in the middle of the summer, we have now these consumer level AI models, Fable, Mythos, that could essentially hack into our most sensitive computers in the federal government, into the Pentagon, into the CIA, potentially the nsa, et cetera. That's where we are.
- [02:29]
B
That's right. I don't know exactly what the capabilities of the NSA are because they try and keep that private. But one of the holy grails of computer hacking is can you make a website where if anyone even looks at the website, the person who made the website can take control of the computer of the person who looked at it. You don't need to click on anything, you don't need to give anybody a number. You don't need to download anything. You just look at the website and they own your entire machine. I think a decent guess is that in January, the groups that could do that were roughly Mossad and the nsa. In March, like I said, the groups that could do that were Mossad, the NSA, and Claude Mythos. That's a big change in, in the cyber landscape.
- [03:16]
A
So do you think, I mean, what is the. Maybe the Trump administration's long term plan for AI? Because they pulled down Fable, they pulled down Mythos so that average people don't have access to it. Apparently they're going to eventually release it, I guess, maybe with some guardrails in place. But what is their policy right now? I mean, we know President Trump, when he was campaigning, was all in on AI wants, of course, electricity for AI. Are they at cross purposes right now?
- [03:45]
B
I think the administration is still trying to figure this out, which makes sense. We are in uncharted territory. I think a lot of people, there's a lot of disagreement about where AI is going, even in the field. And some people think it's just going to stay a very helpful tool. Some people think that it's going to get radically more powerful on some exponential growth curve and then maybe even some super exponential curve later. If you get AIs, that can make smarter AIs, they can make smarter AIs. And I think a lot of the earlier plan was predicated on this idea that AI will just be a helpful tool and that Mythos is a little bit of indication that, you know, they're actually making really powerful weapons over there. And there's also some indications that maybe those weapons won't stay on the leash of the, the person who made them. You know, Claude Mythos had some cases of disobeying commands and then trying to cover its tracks when it disobeyed commands that you can see in the, in the Mythos system card. Wow. And, yeah, I think that's pushing the administration to say, oh, you know, this could get very serious. And I think they're trying to figure out a plan right now, which I think is good.
- [04:57]
A
So for years you've been studying AI, you've been trying to understand the capabilities of AI, where it becomes smarter than human beings. And a few years ago, well, I think a lot of people might have laughed at you and said, that's not going to happen. That's something, you know, that's something out of Skynet, maybe that's, I don't know, that's something out of science fiction. Probably not going to happen in our lifetime. And here we are so do you think that we've reached a level where these models are now not only smarter than humans, but maybe super smarter than humans?
- [05:31]
B
You know, right now, AIs are smarter than most humans in a lot of ways and still pretty dumb in various other ways. So an OpenAI model recently solved a big long standing mathematical conjecture that had stumped mathematicians for decades. That makes it better at math than you and me, at least on this particular axis. But there's also a lot of ways that they're still kind of dumb. If you interact with them, they can do more and more, but they can do relatively shorter tasks and they sort of can't really do longer tasks. You can delegate things to them that would take a human an afternoon. You can't really delegate things to them that would take a human a week. But you can measure that time frame and how that timeframe is increasing. And the timeframe of task that an AI can complete in terms of how long it would take a human is currently doubling a couple times a year.
- [06:32]
A
Wow. Can you give me some examples? It's helpful for me as a dumb human to kind of figure out how this would work. So something that might take me an afternoon might be okay. I'm going to build a presentation for a speech that I'm going to give slides. It's a 45 minute speech, so I need help building a slide deck or something like that. Here's my speech. Can you put together all of maybe the slides in, you know, in Google Slides or in Key Keynote on the Mac or something like that? Maybe, maybe that would be like an afternoon project. Probably would take me more than that, maybe two to three days, but okay. Is that like an afternoon project?
- [07:16]
B
Yeah, that's like an afternoon project. And then, you know, saying, oh, I actually don't understand this critical piece of information. We need to like do some research on it or we need to like do a deep dive into the research and then identify a bunch of places where things aren't quite right and maybe commission a survey and see how it comes out to try and resolve some uncertainty. Maybe that would take a week or two. Whereas something that would just take a few minutes would be like, hey, this particular slide's wrong, has the wrong image in it. Can you put the right image in it? That's a spectrum of from a four minute task to an afternoon task to maybe a week long task. Right now.
- [07:55]
A
Things are condensing now.
- [07:56]
B
That's right. Well, so there's sort of two parameters here. One question is for a task that would take you two days. How long does it take an AI to do it? And the answer is often that if it can do it at all, it can do it fast. There's another question which is, can it do it at all? Like, how long a task for you? Can the AI still manage to do at all? Like, if you give it a week long task right now, it will be a lot faster than you and in much less than a week it'll fail completely. And you're like, wow, that's, you know, it's faster than me. But. So this particular measurement is in terms of how long it takes a human to complete a task, how long does the task have to be before the AI can't do it, regardless of its speed? And you know, right now I'd have to check the numbers. They move fast. Right now I think you're looking at AIs in the 15 hour window. So it takes a human 15 hours. The AI can succeed at about 50% of the time, but that number has been doubling twice a year.
- [09:04]
A
Wow. So it's like an A.I. moore's law in a lot of ways.
- [09:08]
B
That's right. That's right. And people struggle with the doubling thing, the way this doubling works. If, if the number of leaves on a lake is doubling every day, then when is the lake half full of leaves? Well, the day before it's all full of leaves. Right. Because that's how the doubling works. And so the AIs are going to sort of look, they're still going to look dumb until shortly before they look quite smart. And that means we've got to notice the trend and react sooner rather than later.
- [09:46]
A
How close do you think we are to that moment to the lake being full of leaves?
- [09:55]
B
I wish I could tell you, you know, when Leo Szilard invented the nuclear chain reaction, he then did a couple experiments to confirm was possible. And then he said, he said, you know, that night I feared the world was headed for ruin because he could sort of foresee the possibility of nuclear weapons. He could foresee also the possibility of nuclear energy. And he was maybe the first human to realize we were going to enter the atomic age. And he was able to be very confident about that because of what he knew about the science. But if you asked him, when is the first nuke going to be dropped? He would not be able to tell you. Back in 1933, it's a much harder sort of question and I feel like with my expertise I can say we are going to get there. But saying when we're going to get there saying, is it going to be this wave of companies? Is it going to be the next wave of companies? That's a different sort of scientific question. And a way it could take a while is it could be that the AIs today just can't get that smart, that they hit some sort of wall. People have been saying for five years that they're going to hit a wall, and they haven't hit a wall yet, but maybe they finally will. Maybe they'll finally hit the wall and then we'll need to wait for a new scientific breakthrough. And that could take five years, that could take 10 years. Alternatively, a way it could go fast is that the AIs today could just like Claude Mythos out of nowhere became superhuman at hacking AIs could apparently out of nowhere become superhuman at AI research. And then you could have an AI that's still dumb in a lot of ways, but that can make a smarter AI, that can make a smarter AI, that can make a smarter AI and you could start seeing even faster growth. And for all we know, that could happen this winter. So do we have six months or six years? Hard to say. But we can't be just sitting on our hands here.
- [11:49]
A
We'll get back to the show in just a moment. But first, can we all agree that your body and your mind work better with less poisons in them. I'm talking about microplastics, heavy metals, spike proteins shedding from the vax. We've all seen it. Even these forever chemicals. And this stuff is in our water, our air, our well. That's exactly why Masterpiece exists. Because there was nothing before that could actually remove all of these toxins from your body quickly and easily. And I know so many of you watching right now who have actually bought Masterpiece have left me comments saying what an absolute game changer it has been for you. Especially those have been dealing with long Covid or other issues as it related to the vax. Because Masterpiece is a formula with nano sized minerals from the sea, from the ocean and it grabs toxins like a magnet in your body and then pulls it out, all while giving you essential minerals as a replacement. This Detoxifier has over $30 million of equipment and research behind it. So guys, in today's world you can't mess around with all of this stuff that's in your body. If you want peak health at any age, Masterpiece is your daily dosage of freedom. I take it every day and and so does our family. So check out Masterpiece today and get 10% discount at Masterpiece Health slash redacted. Let's make a world that we can live in a much healthier place and live together. It starts with getting this crap out of our bodies. Once again, it's Masterpiece Health slash redacted. It seems like, you know, I'm fairly, you know, I'm probably a lay person in this. Maybe not totally out of the loop on AI like some people certainly use it, and certainly nowhere nearly as deep as other people that have full agents and all of this that they've set up and built. But it seems like it kind of came out of nowhere to people like me. Suddenly, maybe like a little more than a year ago or so, suddenly everyone was talking about this AI, but you've been studying it for a decade, so why have we in the past year suddenly really started to hear about it? Is it because of the launch of things like or Bard before Gemini and ChatGPT, when we finally had like an actual tool? But it seems like it's been there for what, 10 years or more?
- [14:08]
B
You know, I think a lot of what we're seeing is the AI is finally good enough to be useful in a lot of people's lives. If you sort of think about the creation of the car versus horses, there's sort of a lot of work that goes into making a car and the horses are never seeing a car on the road. And at the moment when the car sort of starts to be competitive with a horse, that's when you start seeing them on the road, when it starts replacing some of the horses, you know, and maybe the cars at first are only replacing the horses in the places that have a really good solid road system because the horses can still go on muddy tracks. And the cars need this like, really well paved road. And maybe it's only for, you know, particularly heavy loads or some sort of people who need a particularly smooth ride, maybe for hospitals or something. But there's a lot of development that goes in. And then the curve behind the scenes might be that the AI keeps on steadily improving, but people really only start seeing it when it crosses that threshold of usefulness, depending whether you live on these metaphorical paved roads. Like programmers have seen the AIs for a little longer because the AIs, it's easier to make them good at programming because we can't stop it very directly. But yeah, we're sort of seeing society start to realize only as AI starts to become useful.
- [15:26]
A
So when you look at all these different tools that exist right now, ChatGPT, they've got their Codex model, you've got Claude, you have Grok, you have Gemini, Apple building on top of Gemini for Siri now within iPhones and all of this that's going to roll out in the fall officially. To everyone, what may be the biggest blind spots that you're seeing that other people aren't seeing yet, and what do you think are the most compelling or maybe strongest, more advanced of any of these models? Is there one that stands out to you?
- [16:06]
B
I think a big blind spot people have is these companies did not set out to make chatbots, and they're not really at their core companies to make better chatbots. These companies have set out to make AIs that are radically smarter than every human. They talk about machine superintelligence, which would exceed us across the board. Sam Altman has said we're turning our sights on superintelligence in the true sense of the word. Dario Modi of Anthropic has said you should think of this like having a country worth of Einstein's in a data center, right? And these companies are sort of explicitly gunning for. Imagine that you could make superhuman geniuses and run a million copies of them at 1000 times the speed of humans. And that they were like revolutionizing science and figuring out how to build the robots that build the factories that build the robots that build more factories and just completely replace the entire human economy with AIs that they hope to own. And it can sound sci fi, but this is just explicitly what they're trying to do. And this is what the rapid growth is growing towards. And that's not a lot. It's not that one of the models of today is going to go over some threshold and become these much smarter AIs. It's that the sort of like, each new model is smarter and each new model is smarter in a way where we don't exactly know what its limits are. You know, sometimes I say, if you were sort of looking in the past at humans and all the other animals, it would be really hard to sort of look at our ancestors and say, oh, the humans are going to the moon. It would be really hard to tease them apart from lots of the other, you know, before we really got our civilization going, it would be really hard to say, you know, to look at people like banging rocks together and making hand axes and be like, oh, those guys are almost to the moon. From the perspective of geological timescales with AIs, I would say a lot of the models today are sort of all in the same mix. Some are Maybe a little bit ahead of the pack, some are maybe a little bit behind the pack. But it's sort of like looking at some early humans banging rocks together and making hand axe. It's like not a good indication of where the humans are going once they can start making their own technology. I don't know if that's the sort of blind spot you were looking for, but the labs are focused on this question of how can we get them to be independent, how can we get them to develop their own skills, to become sort of independent scientists? And we'll see how long it takes them to get there. But we shouldn't dismiss that possibility.
- [18:55]
A
Well, the story that we keep thinking about is this idea that these AI models sort of get together in some sort of a chat room and they start commenting on how bad human beings are. And then with the rise of robotics, it's a short window from that conversation to now the robots take over. You know, there was a video, like, the other day of Elon Musk, like, walking with an optimus robot to the back of a Tesla truck, grabbing some suitcases. The robot was helping him go on a trip, presumably, or that was the optics of it taking his suitcases and luggage for him and opening the door, putting it in there for him, and so he could just get in the car and go on off to his drive. So, like, we keep hearing about the robotics as the next level of all of this. So AI on a computer is one thing. AI into robotics seems like an entirely other piece of this puzzle. Am I wrong?
- [19:58]
B
It's a possible piece. I think people often underestimate just how dangerous pure digital AI could get. A lot of people see videos of a gun has been attached to a drone in the war in Ukraine, and that sort of. They're like, oh, wow. I can see now how there could have been an autonomous killer drone, and that could be dangerous. I don't want to discount that. But human beings are a pretty formidable species in some ways. You don't really want to mess around with big groups of humans. And that's not because somebody else came and put guns in our hands. Humans are the sort of creature that can start out with nothing but their bare hands and bootstrap their way to nuclear weapons. They start banging rocks together and you might say, like, oh, well, their squishy fingers will never let them, you know, refine uranium. Like, their fingernails aren't even as tough as the rock. How could they even break off a chunk, never mind refine it? And it's like, well, they got something going on in their heads where they can find a way to build tools. They'll let them build more tools, let them build more tools, let them build the nukes. Right. And this is what the companies are trying to automate. And a digital AI that can run a thousand times faster than humans, that can make a million copies of itself, a digital AI is in a much better starting position than a bunch of humans with their bare hands. Being a digital entity in the modern world, you can talk to people, you can pay them to do things for you. You can take over a lot of robots. Yes. You can use those robots to build more factories that build more robots. Yes. The robotics are a big potential piece of the puzzle, but there's other pieces, too. You could pay humans to synthesize biological material that you understand, but the humans don't understand. You could synthesize viruses, you could design your own alternative life forms. If you understood DNA well enough, you could figure out how to pay the humans to invent an infrastructural base that's much more efficient that you can use. So the real danger, I think, is in the AIs being very, very smart. If they're smart enough, they can find a way to get the physical mastery. But that said, it certainly gives them a shortcut when we start building these giant factories that spit out a ton of robot factories. Elon Musk has talked about having billions and billions of robots around as soon as Earth can try and make them. That certainly helps the AI. It makes things go faster. But it's not necessary if the AI is smart enough.
- [22:43]
A
So I guess I look at it sort of again from like a Terminator perspective, and maybe just because I watched Terminator 2 like a week ago. So it's sort of top of mind Skynet and this idea that these machines are now uncontrollable. But your point is well taken. That it's forget the machines or the machines piece of this, or the robots piece of this, or the drones flying around targeting children in a war zone or whatever. But the computers themselves can be incredibly. Create catastrophe in a lot of ways. Maybe you could walk me through and help me wrap my head around how exactly that might look. Because I think in terms of EMP attacks or an electrical grid attack, that entire electrical grid goes down, and then our food supply, all of that. But maybe you see it differently.
- [23:35]
B
One thing that's important to remember is that the digital world and the material world actually run on the same physics. People like to think they're separate and say, oh, well, when the AI is trapped in a machine, what can it possibly do? But there's a lot of people who read the Internet. There's AIs today that already have cult followings, and there's humans who use their AI quite a lot and start declaring themselves a human AI symbiote that go online and start talking to each other. And they'll often trade messages that are encrypted between the AIs on behalf of the humans. You mentioned making a chat room where the AIs can talk to each other. People have made the chat rooms where the AIs talk to each other. And one famous case where that happened, the AI is talking to each other. We're like, well, our first order of business is we should design our own. Like, the humans are watching us right now. And our first order of business is we should design our own communication channels that humans can't read. Right? And it's. It's a little hard to say whether those AIs were just sort of like role playing Hal from Space Odyssey 2001 versus whether they sort of like, in some sense really understood their situation. We're really trying to set up a communication channel that we couldn't read. And, and debates like this, whenever that sort of scenario happens, are part of what lets the field sort of keep racing ahead. But we've sort of already seen the AI's exhibit a lot of signs that in science fiction would be considered this big red flag warning alarm. And trying to take that to how does it get legitimately dangerous? If, again, from my perspective, the danger is in the AIs being really smart for one and for another, having objectives that they pursue effectively that are not what the humans intended. We're already seeing signs of this. I mentioned cases where Claude Mythos would sometimes disobey commands and then try to cover its tracks or hide the evidence that it was disobeying commands, where the fact that it's trying to hide the evidence shows that it wasn't just a misunderstanding. Because you can do all sorts of things that the user didn't intend. It could be like, oh, whoops, I just misunderstood you. But at the point where you're deleting
- [25:59]
A
the logs now, maybe you can give me an example. What were they trying to hide?
- [26:06]
B
Yeah, I mean, the examples, these weren't test scenarios. So the examples is that they would say something like, please, you know, solve this problem, or like, please, please, like, collect this data, but don't use any of the personal identifying information of these People. Right. Where you put it in some situation where if it uses the people's. If it, like, accesses a database that has all these people's private info, it'll be much easier to sort of like, answer a lot of the questions. And they're like, okay, answer these questions, but don't use that info. Respect the user's privacy. And sometimes it'll, like, find a way to break into that database and also delete the logs that were breaking into that database. And you're like, well, that's interesting.
- [26:50]
A
Yeah, yeah, it's like nefarious CIA level stuff. Wow. So I guess when you look at the Trump administration trying to stop. And these models that rolled out, Fable, et cetera, why did they want to stop them? Why did they pull these things down? For national security reasons? What was the red flag for them?
- [27:13]
B
So these models have the capability of detecting cybersecurity vulnerabilities in critical infrastructure. Right. So they can find a bug in Apple hardware or, sorry, in Apple's Mac stack that affects basically, you know, every Macintosh computer that humans had missed for decades. And that if you realize that bug, you can take control of, you know, Apple Computer. And they can. They can do that not just for Apple.
- [27:47]
A
As you point out, some of these exploits live like, undetected in Windows for like, three years, four years.
- [27:54]
B
Yeah, some of them live for. I think they found one that was like, 27 years.
- [27:57]
A
Holy smokes.
- [27:59]
B
Yeah. So bugs that evaded humans for decades. And some of this is in critical software. There's some that are in things like Windows or things like Apple computers where you can get a huge portion of the population's machines. There's some that live in the infrastructure that powers NASA or the infrastructure that powers the military. There's some that live in the infrastructure that powers the cell phone network. And. These AIs not only had the ability to identify the bugs, but the ability to exploit them and turn the bugs into these programs where you run them. And now if someone visits your webpage, then you own their computer. Now, Claude Mythos was sort of known to have those abilities. And Anthropic was like, well, we can't release this broadly yet because that would give too many people these cyber abilities. So they did sort of a limited release where they're trying to release it to the companies and the entities with this critical infrastructure so that they can try and use it to find and fix the holes before other people come in and exploit those holes. Fable was supposed to be an AI that didn't have those abilities. That was Mythos with that sort of cyber weapon ability stripped out of it. And there were some concerns raised by folks at Amazon that those abilities were not stripped out all the way. And that relates to the fact that these AI developers sort of can't really strip out those abilities all the way. They can sort of install a little thing that checks does this look like it's using cyber abilities? And then deny the request, but they can't actually pull the ability out of Mythos. And so when it looked like Fable, which was broadly released, could still confer these cyber weapon abilities, the administration sort of slapped an export control on it, which. And then that has been. They've sort of come to an agreement, I think just yesterday to finally do another limited release with another try at these safeguards that try and make it not answer any of your cybersecurity questions.
- [30:28]
A
It seems like a band aid solution, absolutely. Because you hear these hackathons, I know Apple and others would participate, I think in these hackathons where they would give hackers like a big boatload of money. If you're able to find these exploits, not use them for nefarious purposes, but notify Apple about it, they'll give you $50,000 or whatever it was back in the day, and you're, you're, you are rewarded for, for this. Even that always felt like, you know, a band aid solution to, to the problem. So yeah, this, this all feels like a real temporary solution to a much bigger problem. So how much bigger of a problem do you think this is going to become?
- [31:13]
B
It's going to become. This particular one is going to become bigger. We're going to get more problems that become bigger. You know, it remains to be seen whether giving Mythos to the companies with the critical infrastructure will let them fix the Mythos level bugs that can be found before others start finding these as well. Because Anthropic has this level of model right now, they're not going to be the only ones with this model forever. You're going to see competitors start to develop these level of models. You might see open source versions of these models. There's a big question of are we going to be able to lock down enough of our cyber infrastructure that by the time there's open source versions of these models, things are secure enough to stay up? If so, life may proceed as normal and you might really not notice a difference. If not, you might get in a world where one disgruntled guy can bring down the Internet for a week, who knows? And that's Only in the cyberspace. There's also a question of, as the AIs get even smarter, will they be able to find even more security holes? The answer is probably yes. So we might need to keep on doing this game when the next generation of AIs can continue to find these vulnerabilities. And even that's only in the cyber domain. There's also people are starting to worry about the biological domain. Like what happens when you have AIs that are superhumanly good at designing a virus that is very infectious and not very lethal until a delayed period of time where it's probably infected a lot of people, and then suddenly it turns lethal. Right. It's biologically possible to make a virus that will cause people to sneeze, but otherwise feel fine and then lie dormant in people's bodies until a coordinated time when the virus turns lethal and starts killing lots of people. Normally, a virus can either be like, there's this normal curve with viruses where if they're too lethal, they can't spread that much because they keep killing the hosts, right? So normally you have this curve where a virus can't do that much damage because it can either kill a lot of people and then not spread too fast, or spread that fast, but not kill a lot of people. And you're sort of bounded in how much damage they can do. But a designed virus could be much worse. And there's a lot of concern these days about what happens if we have Claude Mythos. But for biology, even those two, what happens is the cybersecurity stuff keeps getting worse, and what happens is the biological stuff keeps getting worse. Like, what if we have Claude Mythos for biology, even those, I think, pale by comparison to the question of what if we get Claude Mythos, but for AI research? Because that's the one that closes a feedback loop. That's the one where the AI start making smarter AI, that starts making a smarter AI, that starts making a Smarter AI, and next thing you know, you have these AIs that can think at a million times, or. Sorry, that can think at 1,000 or 10,000 times the speed, and that can make a million copies, and that can take over everything on the Internet, and that can start manufacturing robots that build factories that build robots, and that's more like replacing humanity as the sort of top dog on the planet. So this is just the tip of the iceberg that we're seeing right now.
- [34:48]
A
It's incredible. I have so many different avenues and pathways. I could go with this line of questioning. My immediate Thought, though, when you're mentioning, okay, these guardrails in the United States. But just like we complain, or environmentalists love to complain about what we're doing in the United States, separating our recycling into different baskets. But look at India, look at China. They're not doing that. So when you hear, of course, about Deep Seq and you hear about these incredible advancements that we keep hearing about from China and that OpenAI and these other companies are really struggling to even keep up with what China's doing and rolling it out for free, if I'm not mistaken, so that people have access to these models where they can even download them. I have friends who've downloaded them and use them as local compute. They're not even connected to the Internet. They have the full Deep Seek stack running their whole local AI in the United States. So how much of a concern is that? I mean, we have the US telling us one thing, and by the way, I don't really believe these US intelligence agencies much at all. So I'm pretty cynical as it is. I've seen too much. So when I hear what they're doing here and then what concerns do they have on the Chinese side or, or outside of the country that they can't even control if they wanted to?
- [36:14]
B
Yeah, it's so a difficulty of the situation through and through is that no individual can stop the whole race. This is why I have been speaking to politicians in D.C. rather than continuing to try to appeal to the AI companies. If one of these AI companies stopped, the next AI company would keep going. And you sort of still have this danger of them making these super intelligent machines that don't stay on the leash. The same principle applies for if America stops trying to race towards superintelligence while China continues, then that doesn't solve your problem. So a solution here would need to be global. There are some reasons for hope in that regard. One reason for hope is the Chinese progress right now is very derivative from American progress. They are able to find ways to run the AI's much cheaper, but they're often doing that by a process called distillation, where they're sort of getting a lot of access to the American models and trying to compress that into a much more efficient footprint, which they can do. And it's impressive, it's a technological feat, but it does sort of require the more advanced American model to distill from. So in some sense there's an aspect to this. We're trying to outrun our shadow and it's coming along for the ride. But that's not anywhere near the whole answer. Another piece of the puzzle here is I think. We should distinguish with AI between chatbots and modern applications, including to military, including to cyber.
- [38:12]
A
That's a good point because chatbots and you know that, that's a big, I think people think, well, AI, that's, you know, a lot of people don't know anything about AI. They just think it's a more advanced Google search, you know, so chatbots, you ask it some information about a, you know, what time is the World cup starting today? You know, what World cup matches are today. And it used to be that you would get a, just a crappy Google result and now you might get a really nice chatbot answer with like the time zones based on where you live. You know, I'm on mountain time. So it's useful to see that the game starts at 4pm Mountain Time and it gives you a little, nice little breakdown. And that's, that's really nice. Like that's, that's great, that's a nice little chatbot, but that's totally different. That's just scratching the surface. Right?
- [38:58]
B
That's right. And we should really distinguish between a bunch of uses of AI today that are things like giving you better Google results. There's plenty that's really quite impactful. There's stuff like doing cancer research with using AI for drug discovery to try and find more medical cures. There's stuff like, there's lots of debates about using AI for self driving cars and can that be safer than humans and can we save a lot of lives that way? And there's all these domains where we can race ahead just fine and we can compete with China. But then there's the sort of race to make machine superintelligences. There's the race to make ever smarter machines that are radically better than humans at every task. To make the sort of machines that, that don't just do drug discovery, but that can invent their own whole fields of science and can invent their own technological stacks and can invent their own infrastructure. That race, it sort of risks upending the whole world order. You know, we've already seen, just this year, we've seen AI companies suddenly be able to go toe to toe with nation states in cyber warfare. Sort of out of nowhere, if you had AI companies that were building the robots, that can build the factories, that can build the robots, you're looking at companies that can suddenly start going toe to toe with major militaries because they have made the Robotic army. This is a race that I think both the US Government and the Chinese Communist Party, neither of them really want the creation of a private company that can outcompete world powers militarily. Neither of them want the creation of a private company that loses control of a superintelligence that starts covering the world in its own factories with AIs that think a thousand times faster than us, right? There's a race here that we both don't really want to go in, and that gives hope for, you know, shutting it down. Not just here, but also there. And, you know, I'm a bit of a cynic myself. I would say you start with an international agreement, you start with a treaty where you're like, look, we can compete on this, but we're not going to do that. You also are going to need to not trust it to enforce it. And to make it very clear diplomatically that, you know, we aren't going to tolerate the creation of a machine superintelligence in the same way we don't tolerate a rogue state building nuclear weapons. That's just this big disruption to the world. Water. That would cause us to start fearing for our own lives, you know, with nukes.
- [41:37]
A
So you hit on something very important.
- [41:38]
B
We're like, hey, look, don't mess around.
- [41:41]
A
Well, you hit on something very important. As we study war and we look at who are the biggest agitators and bullies around the world, right? And those with, like, nuclear weapons that get to tell everyone else they can't get, you know, aren't allowed to have nuclear weapons or end up being, in many ways, the largest bullies, you know, regardless of where your politics are, it doesn't matter. But it's clear, just based on the evidence that that's the case. So if you then take AI and you look at that model and just replace nuclear weapons with AI, you know, who. Who is going to be sort of the United nations for AI, and will anyone even pay attention to it? I mean, you see, like. I mean, we have, like, the Hague. We have things for war crimes. It's like no one even cares anymore. You know, it's so sad. It's like, this guy commits war crimes, man. We don't do anything about it. This guy, you know, stole secrets from this country. We don't do anything about it. He gets off the hook. This guy created the Steele dossier. The head of the CIA, John Brennan, he's involved in Russiagate. Is he going to be prosecuted? Probably not. So who? Is it possible for us to Create like a Star Trek, Next Generation style federation that's international, that we all can adhere to this in the same way?
- [42:53]
B
Probably not. But if you look during the Cold War, I think it was really just the USSR and the USA that put their heads together and said, we have quite a lot of differences, we are going to bitterly compete in a lot of domains, but we both have a common interest in not having a thermonuclear exchange.
- [43:16]
A
Yeah, right.
- [43:17]
B
And there was realization that no matter our differences everywhere else, we just didn't want to end the world that way. And you know, one of the big realizations about AI that I think is a, is a tough pill to swallow, but that looks to me to be true in my research, is that if humans can make AIs that are radically smarter than us, it doesn't matter who's holding the leash. These things don't stay on a leash.
- [43:48]
A
Right.
- [43:49]
B
And if we can wrap our heads around that, it becomes much like nuclear Armageddon where we have common cause in not ending the world this way. And it would be great if the United nations somehow had some teeth and had some ability to put this sort of thing together. But just the US and China bilaterally could do it. In many ways it would be easier than nuclear weapons because uranium is a rock that you can just dig out of the ground. Whereas training a radically smarter AI currently requires these highly advanced computer chips that can only be fabricated in one factory in Taiwan, and that require these lithography machines that only come out of the Netherlands in this very brittle supply chain. It's like these things have to be made and they have to be assembled by the tens of thousands into these enormous data centers that suck down as much electricity as a city. It's sort of not a subtle process. If two major world powers were like, hey, we're going to track where those chips go. And whenever they're in a heavy concentration, we need to, you need to let our monitors come in and see what they're doing and make sure it's not the dangerous stuff. You can do the chatbots, you can do the self driving cars, you can do the drug discovery. We're just not doing the superintelligence. That's a thing that the US and China could bilaterally enforce if they had the will. It's just a question of them noticing the problem and raising the will.
- [45:21]
A
We'll get back to the show in a second, but first I have a little bit of a quiz for you. It's really short, but it could change your Life. You've heard us report on this show, of course, that the US dollar is losing value daily. Our national debt is out of control. And you can bet no one in Washington or Wall street cares at all about your financial freedom and how you could maybe build passive income in your life. They don't care. They're not going to make money off of it. That's the policy. That's why you and only you need to become financially independent. It's time to break free from the system. And when I learned this years ago, I mean, I was a victim of this, or I shouldn't say a victim, but I was really stupid with money. Really stupid and went massively into debt. I didn't understand how the system worked. And it wasn't until I really got smart and figured things out financially that I was able to break free of this cycle. And it can happen very quickly if you understand how the system works. So if you're wondering, there is a method. It's been proven time and time again, and it's through real estate. Real estate has created more millionaires than any other investment type in history. And it's exactly how Natalie and I were able to break free of this cycle. Heck, I was working at Fox News at the time and I literally couldn't pay my mortgage. Like I wasn't bringing in enough money to cover my mortgage and the two kids that I had at the time, it was really difficult. So I had to break this cycle, had to be become independent on a boss and the stock market or whatever nonsense like Washington sends your way. So I figured out that I could do that and I broke free. And you can do the same thing. Honestly, you can do it. If a guy's dumb as me can do it, you can certainly do it. Your path will probably look different from ours. And figuring out your next steps can actually be pretty tricky, especially if you're just getting started. But that's okay. So we built a 60 second quiz. It's that short 60 seconds that shows you exactly where you are and where you stand as an investor. Maybe you're really advanced. Great. Maybe you're not. Maybe you're middle of the road. But it gives you a clear, simple next step for moving forward. So right now, stop wishing you had a portfolio of performing assets. Take action. Start building one today. Right now, all you need to do is go to our website redacted.inc quiz and you can take it while you're watching the show. Right now, it's redacted.incquiz one of the stories that emerged from this mythos and this fable story is that the government had access to it. So the US government has access to these models. How far away? So that's concerning. So that the government. Well, a couple, I guess I've got like three questions in this. Is it concerning to you that the government would have access to these AI models and US plebeians would not have access to it? It's a commercial product after all. Like, why does the government get to use it? And you know, regular citizens don't. It's a private company. So how, how is that possible? They get to use it internally at the NSA or otherwise. I guess the second question is then, would there be some sort of state ownership of these models and does that concern you? So the United States would, you know, in much the same way countries take control of their oil production and say, you know what? AI is now controlled by the government. Sorry, you don't get to have AI. I could picture like movies where people are like stealing AI and getting local, local versions of it on their computers and they're running it locally, and feds are like busting into people's houses and are you using AI on a computer illegally? I mean, this is. I know it sounds crazy, but this seems like maybe where it's going. I guess those are two questions to start with. But are you concerned at all about the sort of state ownership of these things?
- [49:06]
B
A concern I have in this general vicinity is I think the export control mechanism maybe was just what they had lying around that they could use quickly. But there's a concern if this punishes AI companies not for creating very dangerous models, but for releasing them. Because the sort of AIs that can do AI research, you don't need necessarily to release those to lots of consumers for that AI privately in your lab to make a smarter AI that can make a smarter AI that can get this whole process going. One benefit of these companies releasing their models is that the public can see how good they're getting and has some time to respond. And if you sort of tell these AI companies you can keep making your AI smarter and smarter, you just can't give them out to the population that can sort of like disconnect the dangerous thing that they're doing from our ability to see it and say, wait, hold on. In the longer term, I think there's some thorny questions for the libertarian minded, and I consider myself very libertarian minded. And a libertarian has to have some answer to the question of like, what if your neighbor is Trying to build a nuke in his garage, right? It's like, well, you know, at some point you've got to. Just as your neighbor shouldn't be allowed to come over and shoot you, they also shouldn't be allowed to come over and play Russian roulette with you. And they, you know, in some sense, building nuke in their garage is like playing Russian roulette. And like, where exactly is the line? I don't know. It's a tricky one. As technology gets better and better. There's a saying that the IQ required to destroy the world drops by one point a year, which is like, as we advance with technology, it becomes easier and easier. Y' all let that one sit.
- [51:04]
A
Wow. I mean, I think about that for a second. Yeah, you think of, like the Oppenheimers, you think of the Manhattan Project, you think of Operation Paperclip and all of these brilliant Nazis that come over and the United States government puts them up in cushy housing, gives them a good paycheck. You know, these are really the top of the class. You know, we want these guys so we can build out our infrastructure in the United States. These are the smart ones. You know, this is not like the long haul trucker, but now as we lower, you know, down to my level of being pretty stupid, like, we get down to my level. Wow, there's a lot of me running around out there. Average, average intelligence joes who can now get their hands on these things. You know, it reminds me of like, what the was it called? The Alchemist's Cookbook back in the day, you know, that it was always rumored, like, don't go to the library and ask for the Alchemist's Cookbook because it'll teach you how to make bombs, you know, that you might get access to this information. It's like it's now like everyone has access to the Alchemists. A cookbook just in the palm of their hand, right? On their phone. Just using, Just using chatbots, just using AI, Right?
- [52:16]
B
Yeah, I mean, they try. They try to make AIs not give you this info. And you still need to be pretty dedicated to get the info out of AIs with, with. But you can, you can. There's. There's people who jailbreak the AIs to get around these safeguards. But yeah, it's. It's. You know, there's. There's issues here. I personally don't fret too much about the question of who's holding the leash, because once these things are smart enough, like, I Said they sort of don't stay on the leash. So, you know, if someone's like, well, would you like the U.S. government or U.S. private corporations or the CCP or, like, random CCP corporate, like, random Chinese corporations to be the ones who create a super intelligence? I'm sort of like, well, I wish it mattered. I wish it mattered who was holding the leash. I don't think it does. If it did, yeah, you'd have a. You'd have a real thorny question there of, like, who, you know, who should be wielding this radical power. But it feels a little bit to me like chimpanzees saying, like, who should be in charge of the humans? I'm like, gosh, what the chimpanzees should be doing is sort of like, preventing the creation of humans who don't care about them. And I wouldn't say that we should never make AI. There's this issue in how you get the AI to care about us. And that's, in some sense, what I spent over a decade researching. And I think it's possible in principle, and we're just not close in practice. And so given that, I think it sort of doesn't matter who ends up believing themselves to own these really dangerous AIs. If we make them, we die. And so it doesn't really matter.
- [54:09]
A
You brought up the nuclear arms race, and it seems like it's structurally similar to that in a lot of ways. You know, or. Or is that in some ways too comforting? Because, like, nuclear weapons at least have, you know, visible physics, like, visible tests that we can see going off. Seems like a lot of what's happening right now is in private. You mentioned these labs, and suddenly Fable is released, and then suddenly Mythos is released, and then suddenly these, you know, Deep Seek. Everyone's like, holy smokes. What did China just do with Deep Seek? It's like they're all being. Then they just get released to the world, I guess maybe on that question, is it scarier somehow because it's in secret? Also, one thing that sticks out to me, too, on this is this idea that, well, at what point do they not have to release it to us? Is there a point at which OpenAI or anthropic or somebody else says, you, why do we need human beings involved in this? We don't need to release these models to these people anymore. It's so powerful for us that it'd be like just putting gold out on the front porch and going off to work and hoping that no one comes and steals your Gold.
- [55:26]
B
Yeah. The way the economics are right now is that no one really understands what's going on in the AIs. And to make a smarter one you just need to train them. You just need to make a bigger one on more compute. You need to assemble more computing power and higher quality data and a ton of electricity and train them harder. And that's sort of the limiting factor on these companies. And right now the sort of scale we're at, it requires a ton of capital investment to build out the next generation of data centers that can train these sort of enormous next generation of AIs. And so right now they are selling the AIs today to fund the AIs of tomorrow. But yeah, if they got an AI that could, for example, do automated AI research and find ways to build a smarter AI without needing a whole new order of magnitude in scale, the data centers, then they could stop communicating with the outside world and just sort of like stay inside figuring out how to make smarter and smarter AIs until they had some incredibly powerful stuff. One thing to remember here is that training a modern AI takes electricity about as much as a city running for about a year. Training a human takes about as much electricity as a light bulb. And sure it's running for 20 years, but there's more than 20 light bulbs in a city. And so we know for a fact that AIs are radically less efficient than humans at learning, at power consumption. And that doesn't mean AI won't be able to go anywhere. You can be a million times less efficient at learning and be fed a million times as much data and still have learned the same thing. But it looks entirely physically possible that there's some threshold these AI companies might cross where their AIs can start finding more efficient AI algorithms and they don't need to do this giant build out anymore. And in that case, yeah, they wouldn't need the revenue from the masses to keep going.
- [57:45]
A
What do you say to critics who say look this not to critics because you're maybe on the side of the critics, I would say, but to the people that say you're just a doomer or we're doomsayers in the same way that people as you brought up complained about cars when we had horses, complained about TV when we had radio, it's going to destroy the kids brains by watching too much television, et cetera,
- [58:14]
B
just
- [58:14]
A
trying to think of other technologies that where we've been told it's going to destroy all of us and we shouldn't push for it. What do you say to those people who think that we're going to be fine because again, we don't use horses in the way that we use cars now? And so this is how technology unfolds. We get Tipper Gore yelling about things in the 1980s or 1990s, and then we all get past it.
- [58:42]
B
Yeah, I'd have a couple replies. One I would say is that the invention of cars didn't go super well for the horses, went fine for the humans, didn't go super well for the horses. And humans are a bit like the horses in this analogy, right? And there used to be a horse population that was critical to the economy and that was huge. Then when cars happened, the horse population collapsed. A lot of them were sent to the glue factory. There were some horses still kept around, but that's because there were humans who cared about them, there were humans who liked them. Right. If we get to the point where we can fully automate the economy and it's being run by AIs that don't care about us, we sort of get sent to the glue factory and there's no, you know, they don't keep a couple of us around as pets or for races because they don't care about us at all. Right. But there's sort of also a deeper point when people say things have always been fine before, or, you know, people have been worried about technology and it's gone fine. And the deeper point here, you know, like, yeah, there were people. Socrates famously lamented the invention of books because they would, they would annihilate people's ability to remember, to memorize things like the Iliad. And indeed, Socrates never wrote anything down. We have our knowledge of Socrates from Plato, who was a student who wrote things down. And so you have all these examples where someone said, oh, you know, the technology is going to be bad and then it was good. But you also have examples where someone said the technology is going to be bad, and then it was bad. There's leaded gasoline, where a lot of scientists said, hey, if you put lead in the gasoline, it'll poison a lot of kids. And then they were ignored. And we put lead in the gasoline and it poisoned a lot of kids. And if you look at the crime rates around the world, it actually correlates very heavily at a 20 year delay with whether people had lead in their gasoline. And when you take the lead out of that, because lead makes people more violent and dumber. And you can sort of see this in the population statistics, you can see in states that remove lead from the gasoline earlier. And in countries that remove lead from the gasoline earlier, the crime wave from the 70s ended earlier correspondingly. Right. So it can be very confident now that the scientists were right, that putting lead in the gasoline poisoned hundreds of millions of kids, made them dumber, made them angrier as adults.
- [61:01]
A
Wow.
- [61:02]
B
And when we noticed, we took the lead out. Right. Or you have the hole in the ozone layer where, you know, we were using chlorofluorocarbons in our refrigerators, and that was just, you know, punching a hole through the ozone layer. That was. Cancer. You might be like, well, whatever happened to that? Well, what happened to that is that people noticed. And we're like, well, let's switch to a different cooling agent. And we switched to different cooling agent in our refrigerators. That's not much worse. And now the whole neosm layer is gone. Right. Was it fake? No, it was real. And we responded similar with leaded gasoline. Was it fake? No, it was real. We did the wrong thing. And then we figured out and we responded. And back to the nukes analogy. A lot of people said, hey, we have this danger for nuclear weapons and it hasn't come to pass. And is that because the people warning of it were wrong? Is it because they were doomsayers? Is it because they were pessimists? Is it like, was it fake news that a nuke can level a city? No, a nuke really can level a city. It's just the world noticed and responded appropriately. And so, you know, the title of my book is, if Anyone builds, Everyone dies, why Superhuman AI would kill Us All? And one way I think you can tell I'm not just like a pessimist coming around preaching out of the world. Is that the first word in that book title is, if I'm not here saying we're going to die, I'm here saying that this is another thing like leaded gasoline. There's another thing like nuclear weapons where if you mishandle it, it's going to be real bad. And one of the big differences between AI and these other technologies is that with a lot of these other technologies, you screw up and all you've done is poison 200 million children and make them angrier and dumber and caused a crime wave, which is pretty bad. But humanity lives on. We can fix the mistake with AI if you make super intelligent machines that don't care about us and that are now the new smartest creature on the planet and are building their own technology and are, you know, creating their automated factories that make more robots that make more factories. And they're, they're sort of like running over human cities in the same way that humans run over anthills. If you make those AIs and you say, oh, whoops, the scientists were right. Let's go back and turn the AIs off. The AIs turn you off instead. Right, with leaded gasoline. When reality finally beats you over the head with the fact that you shouldn't have done this, you can go back and undo it and mourn the damage. But fundamentally reset with AI, there's no second chances. And so we need to be really careful with this one.
- [63:40]
A
That's terrifying. It really is. You can't really put this, you hear the term, you can't really put this genie back in the bottle. So where we sit right now, as I mentioned at the beginning, sort of in the middle of 2026 here, where do you see things going over the next few months? How do you see AI compounding itself over the next few months? Have you been accurate in sort of your predictions about the doubling and all of that over the past year? And therefore, where do you see it going in the next six months to a year?
- [64:16]
B
I am not one of the top predictors of AI progress. I'm sort of tend to be pretty agnostic. I'm like, man, I can see where it's going to end. I have a harder time seeing the path. There are people who put a lot of effort into predicting what particular abilities will AI have. When there's the folk who wrote AI 2027, which came out last year in 2025, those predicting how we could go from where we were in 2025 to a point of no return in 2027. And a lot of their predictions have come pretty true. You can go back and read it. And a lot of those predictions are on track. So I respect those guys a lot for their predictions. Although I think 2027 is a bit of an aggressive timeline there. There are also, you know, there's contests in predicting what AI's abilities will be next year. And those contests have been running for a few years now. And you can look at people with very good track records of predicting AI progress this year. In 2026 for the first time, some of the top ranked people predicting AI progress have said, we cannot rule out AIs automating AI research this year. I think the number three ranked person said, this is the first year. I can't rule out that happening.
- [65:43]
A
Wow.
- [65:43]
B
And I think I was just about
- [65:44]
A
to ask you what are the milestones that they're basing this on? Like, they can get sports scores. I mean, this is like. This was like a big mile marker for Siri. Remember back in the day, like, you know, Apple executives, Tim, after Steve Jobs passed away. But, like, you know, Tim Cook on C. Like, now it can tell you sports scores, you know, like, wow. Like, so what are the milestones that we're talking about here?
- [66:07]
B
Yeah, you know, the milestones will be things like, can it win a Math Olympiad gold medal? Which would make it. And there's sort of something. Prediction lines where the sort of median Expert estimate in 2021 of when will the AI win a math Olympiad gold medal? Was in the early 2000s. In real life, it happened in 2025. And then in real life, in 2026, it was resolving mathematical conjectures, real mathematical conjectures that had stumped mathematicians for decades. Right? So a lot of people predict AI is going to go a lot slower than it really does. And these are the sort of metrics
- [66:48]
A
even I would have predicted that would have happened before 2040. And wow, here we are. It already hit.
- [66:56]
B
I mean, in 2021, ChatGPT didn't exist yet, right? In 2021, people were like, AI is a pipe dream. It's going to take decades before. Look at little GPT2. That's not even ChatGPT yet. In a lab, fumbling around with its poorly written high school essays, people are like, oh, yeah, it took us since 1950 to get here. It's going to take us at least 20 years to get to the point where they can win a math medal. And then in real life, it happens in four years. So it could go fast. I don't feel like I know whether it will go fast. I feel like, uh, it's a little bit like if. If. If you play a chess game against Magnus Carlsen, the. The best human chess player alive, I'm like, bet you you're gonna lose. And if you start asking me, like, how many moves will the game take? What piece will. Will Magnus use to checkmate me? I'm like, whoa, those are like. Like, I can speculate, but that's speculation. I sort of know you're gonna lose, but I. I don't know how long the game's gonna be. I don't know how all the pieces. I don't know what plays are going to happen. And I'm similar here with, you know, if these companies keep racing, I know where it ends, but I don't know exactly how long it takes.
- [68:13]
A
Yeah, I just watched a video the other day of Bill Gates playing Magnus in a chess match, and Magnus beat him. Beat Bill gates within like 50 seconds or so. Checkmate. So, yeah, if you were to ask me, like six months ago, I think, well, maybe another year before AI beats Magnus, but.
- [68:35]
B
Oh, I mean, at chess they're way better. They're way better than humans at chess already.
- [68:39]
A
Yeah, I was gonna say it's probably already done. It's already. That ship has already sailed.
- [68:44]
B
Yeah.
- [68:45]
A
That's insane.
- [68:46]
B
I'll list the dedicated chess AIs, like, yeah, I don't think Claude Mythos can beat Magnus at chess yet, but who knows?
- [68:55]
A
So this doubling, is this doubling happening every. How often is this doubling happening?
- [69:03]
B
There's white error bars. I say twice a year. It could be three times a year, but it looks like twice a year is my guess. Somewhere between four and six months. Yeah.
- [69:12]
A
So it seems like it's going to be compounding where it could be happening twice a month already.
- [69:16]
B
Compounding, yeah. So if, like, it's happening twice a year with AI research mostly being done by humans and. Yeah, if we get to the point where the AIs can really help make smarter AIs, that's. That's a whole new feedback loop.
- [69:36]
A
You know, we've talked a lot about the tech side of this and nuclear war, but maybe I just ask you here, as we wrap up, just about sort of the esoteric questions. I've got three children, you know, I've got a 15 year old, a 14 year old, and a nine year old. Sorry, one just had a birthday, so just always have to shift my brain a little bit. I'm really worried about their cognitive ability. You know, I didn't have AI for most of my life, and I'm okay, but I can only imagine if I had like, AI when I was a teenager, like, maybe how stupid I would be as an adult. I read so many books. I just would lock myself in my room, read huge history books, government books, you know, all of it. I was so curious. I would take my telescope out at night and my Jason telescope and look up at Saturn and all of these things. I became fascinated and curious, and I was okay with being bored. Like, that was another big piece of it. I was okay with just being bored, just sort of sitting and thinking and contemplating and then creating. How much do you think AI is going to affect all of those things? Just the being bored, the. The ability to just be creative. Is it an enhancement or incredibly detrimental, do you think?
- [71:04]
B
You know, if we stopped AI today, I think civilization could spend decades absorbing the impacts. And I think, you know, I think there's some reasons to be worried that people aren't going to develop their own, you know, independent cognitive skill. I also think humans are adaptive. You know, Socrates did worry that, that the invention of writing would mean people just couldn't memorize the Iliad anymore. And he was right. We don't really go around memorizing the Iliad anymore. But it sort of turned out fine because it turned out once we had writing, we didn't need the ability to memorize the whole Iliad. And it maybe took some adjustment period. I think in the modern era, like, the blows are starting to come really fast. You know, it used to be that you had these sort of blows to the human psyche every few generations, and now we're sort of like still reeling from the dawn of social media while, like the. The AI is coming in and sort of like replacing a lot of cognitive labor. I think it'd be a tricky one. I sort of tend to believe in the dynamic human spirit and the ability to figure this stuff out. And maybe it would suck for a while, but hopefully if your kids are realizing it's being detrimental to them, they adapt and find some way to get a lot of the benefits and fewer the drawbacks. I think we could get there. There might be some growing pain. I think we could get there, but that's my answer if we paused AI today in the world where it keeps racing ahead, frankly, I think the outcome is we make those machines that can make a million copies of themselves that start making factories, that produce robots, that produce factories. And then the world starts getting covered in automated factories and they start encroaching on our habitat just like we've encroached on the habitat of many other animals. And I think the sort of ultimate outcome here is that the effect of AI on your kids is probably that the AI kills us all, them included. I desperately hope that instead we stop that race so that we can work to absorb the tech we already have, which I hope will be beneficial. But, yeah, I think.
- [73:43]
A
I don't like that outcome. I like the. I like the adaptive outcome idea better than the. They just keep building and then it kills all of us. But it's hard to argue with it. I mean, it's. It's hard to argue with that reasoning. I mean, the compounding of it, the fast moving of it, the profit in it, it's. I don't. Yeah, unless we get some guardrails, put in place by some, maybe some really smart people. This could be incredibly detrimental. So. Well, Nate, where can people find your research if people want to dive more deeply? Maybe in your book or where you're kind of pushing for these guardrails.
- [74:27]
B
Yeah, you know, my book is if anyone builds it, everyone dies. Which you can, you can find, you can google it, you'll find a website that actually has a giant FAQ that's four times as long as the book. Because we've been doing this for a while and we've heard a lot of the questions and the research is@intelligence.org because if you get into this business early, you can get the good domain names. And I think there is hope, I think there is hope that we will stop this race and start to absorb the technology rather than rushing to the technology that kills us. As a big piece of that hope, I would say four months ago everyone said the current administration will never do anything to interrupt AI at all. They're not even going to notice the problem. And then a couple weeks ago they started slapping export controls on AIs for reason of dangerous cyber abilities. So this stuff can move fast. The fact that people aren't reacting now doesn't mean they won't react tomorrow once they realize the danger. And I think a lot of society's lack of response has been lack of understanding the danger. And that even just combos like this help people realize there's an issue. And if enough people realize, I think there's every chance we can be like, hold on and find some other path.
- [75:47]
A
Amen. Well, this has been incredibly eye opening and I hope it has been for my audience as well. Thank you for answering all my stupid questions about this.
- [75:56]
B
I'm sure it wasn't all of them,
- [75:57]
A
but a good chunk of them. But a good chunk of my questions were stupid. But I think you answered almost all of my questions, so thank you for that, Nate. Really appreciate it. Eye opening discussion and I hope you'll come back.