A (3:17)

Nate.

C (3:17)

And as for my part, Yeah, I was 13 in 2003, so I didn't get into this quite as early as Eliezer. But in 2013, I read some arguments by this guy called Eliezer Yudkowsky who sort of laid out the reasons why AI was going to be a big deal and why we had some work to do to do the job right. And I was persuaded. And one thing led to another, and next thing you knew, I was running the Machine Intelligence Research Institute which Eliezer co founded. And then fast forward 10 years after that, here I am writing a book.

A (3:53)

Yeah, so you mentioned Miri. Maybe tell people what the mandate of that organization is and, and maybe how it's changed. I think you indicated in your book that your priorities have shifted as we cross the final yards into the end zone of some AI apocalypse.

C (4:12)

Yeah. So the mission of the org is to ensure that the development of machine intelligence is beneficial. And Eliezer can speak to more of the history than me because he co founded it and I joined initially the.

B (4:27)

It was, initially, it seemed like the best way to do that was to run out there and solve alignment. And there was, you know, a series of, shall we say, like sad series of bits of sad news about how possible that was going to be, how much progress was being made in that field relative to the field of AI capabilities. And at some point it became clear that these lines were not going to cross. And then we shifted to taking the knowledge that we had accumulated over the course of trying to solve alignment and trying to tell the world, this is not solved, this is not on track to be solved in time. It is not realistic that small changes to the world can get us to where this will be solved on time.

A (5:09)

Maybe so. So we don't lose anyone. I would think 90% of the audience knows what the phrase solve alignment means, but just talk about the alignment problem. Briefly.

B (5:19)

So the alignment problem is how to make an AI a very powerful AI. Well, the superintelligence alignment problem is how to make a very powerful AI that steers the world. Sort of where the programmers, builders, growers, creators wanted the world, wanted the AI to steer the world. It's not necessarily what the programmers selfishly want. The programmers can have wanted the AI to steer to nice places, but you can make an AI that is trying to do things that the program. You know, when you build a chess machine, you define what counts as a winning state of the board and then the chess machine goes off and it steers the chessboard into that part of reality. So the ability to say to what part of reality does an AI steer is alignment. On the smaller scale today, though, it's a rather different topic. It's about getting an AI that whose output and behavior is something like what the programmers had in mind. If your AI is talking people into committing suicide, and that's not what the programmers wanted, that's a failure of alignment. If an AI is talking people into suicide, that and people who should not have committed suicide, but AI talks them into it, and the programmers didn't want, and the programmers did want that. That's what they tried to do on purpose. This may be a failure of niceness, it may be a failure of beneficialness, but it's a successive alignment. The programmers got the AI to do what they wanted it to do, right?

A (6:43)

But I think more generally, correct me if I'm wrong, when we talk about the alignment problem, we're talking about the problem of keeping super intelligent machines aligned with our interests, even as we explore the space of all possible interests and as our interests evolve. So that, I mean, the dream is to build superintelligence that is always corrigible, that is always trying to best approximate what is going to increase human flourishing. It's never going to form any interests of its own that are incompatible with our well being. Is that a fair summary?

B (7:21)

I mean, there's three different goals you could be trying to pursue on a technical level here. There's the superintelligence that shuts up. Does what you ordered, has that play out the way you expected it? No side effects you didn't expect. There's superintelligence that is trying to run the whole galaxy according to nice benevolent principles and everybody lives happily ever afterward, but not necessarily because any particular humans are in charge of that. They're still giving it orders. And third, there's superintelligence that is itself having fun and cares about Other superintelligences and is a nice person and leads a life well lived, and is a. And is a good citizen of the galaxy. And these are three different goals. They're all important goals, but you don't necessarily want to pursue all three of them at the same time, and especially not when you're just starting out.

A (8:09)

Yeah. And depending on what's entailed by super intelligent fun, I'm not so sure I would sign up for the third possibility.

C (8:15)

I mean, I would say that the problem of what exactly is fun and how do you keep humans, how do you have whatever the superintelligence tries to do that's fun, keep in touch with moral progress and have flexibility, and what do you point it towards that could be a good outcome? All of that. Those are problems I would love to have. Those are right now, just creating an AI that does what the operators intended. Creating an AI that you've pointed in some direction at all, rather than pointed off into some weird squirrely direction that's kind of vaguely like where you tried to point it in the training environment and then really diverges after the training environment. We're not in a world where we sort of get to bicker about where exactly to point the superintelligence, and maybe some of them aren't quite good. We're in a world where no one is anywhere near close to pointing these things in the slightest in a way that'll be robust to an AI maturing into a superintelligence.

A (9:11)

Right. Okay. So, Eliezer, I think I derailed you. You were going to say how the mandate or mission of Miri has changed in recent years. I asked you to define alignment originally.

B (9:24)

Well, our mandate has always been make sure everything goes well for the galaxy. And originally we pursued that mandate by trying to go off and solve alignment because nobody else was trying to do that, solve the technical problems that would be associated with any of these three classes of long term goal. And progress was not made on that, neither by ourselves nor by others. Some people went around claiming to have made great progress. We think they are very mistaken and knowably so. And at some point, you know, it was like, okay, we're not going to make it in time. AI is going too fast, alignment is going too slow. Now it is time for the people. All we can do with the knowledge that we have accumulated here is try to warn the world that we are on course for a drastic failure and crash here. Where by that I mean everybody dying.

A (10:13)

Okay, so before we jump into the problem, which is deep and Perplexing. And we're going to spend a lot of time trying to diagnose why people's intuitions are so bad, or at least seem so bad from your point of view around this. But before we get there, let's talk about the current progress, such as it is, in AI. What has surprised you guys over the last, I don't know, decade or seven or so years? What has happened that you were expecting or weren't expecting? I can tell you what has surprised me, but I'd love to hear just how this has unfolded in ways that you didn't expect.

C (10:51)

One surprise that led to the book was there was the ChatGPT moment, where a lot of people, for one thing, LLMs were created, and they sort of do a qualitatively more general range of tasks than previous AIs, at a qualitatively higher skill level than previous AIs. And ChatGPT was, I think, the fastest growing consumer app of all time. The way that this impinged upon my actions was I had spent a long time talking to people in Silicon Valley about the issues here and. And would get lots of different types of pushback. There's a saying, it's hard to convince a man of a thing when a seller depends on not believing it. And then after the ChatGPT moment, a lot more people wanted to talk about this issue, including policymakers, people around the world. Suddenly AI was on their radar in a way it wasn't before. And one thing that surprised me is how much easier it was to have this conversation with people outside of the field who didn't have a salary. Depending on not believing the arguments, I would go to meetings with policymakers where I'd have a ton of argumentation prepared and I'd lay out the very simple case of like, hey, people are trying to build machines that are smarter than us. The chatbots are a stepping stone towards superintelligence. Superintelligence would radically transform the world because intelligence is this power that let humans radically change the world. And if we manage to automate it and it goes 10,000 times as fast and doesn't need to sleep and doesn't need to eat, then it'll by default go poorly. And then the policymaker will be like, oh, yeah, that makes sense. And it'd be like, what? I have a whole book worth of other arguments about how it makes sense and why all of the various misconceptions people might have don't actually fly, or all of the hopes and dreams don't actually fly. But Outside of the Silicon Valley world, it's not that hard an argument to make. A lot of people see it, which surprised me. Maybe that's not the developments per se, and the surprise is there, but it was a surprise strategically for me, development wise. I would not have guessed that we would hang around in AIs that can talk and that can write some code, but that aren't already in the able to do AI research zone. I wasn't expecting in my visualizations this to last quite this long, but also my advanced visualizations. One thing we say in the book is the trick to trying to predict the future is to predict the questions that are easy, predict the facts that are easy to call and exactly how AI goes. That's never been an easy call. That's never been something where I've said I can guess exactly the path we'll take. The thing I can predict is the endpoint, the path. I mean, there sure have been some zigs and zags in the pathway.

B (13:33)

I would say that the thing I've maybe been most surprised by is how well the AI companies managed to nail Hollywood stereotypes that I thought were completely ridiculous. Which is sort of a surface take on an underlying technical surprise. But you know, if in even as late as 2015, which from my perspective is pretty late in the game, like if you'd been like, so Eliezer, what's the chance that in the future we're going to have computer security that will yield to Captain Kirk style gaslighting using confusing English sentences that get the computer to do what you want. And I would have been like, this is a trope that exists for obvious Hollywood reasons. You can see why the script writers think this is plausible. But why would real life ever go like that? And then real life went like that. And the sort of underlying technical surprise there is the reversal of what used to be called Moravec's Paradox. For several decades in artificial intelligence, Moravec's Paradox was that things which are easy for humans are hard for computers. Things which are hard for humans are easy for computers. For a human, you know, multiplying two 20 digit numbers in your head, that's a big deal for a computer. Trivial. And similarly, not just me, but I think the sort of conventional wisdom even was that games like chess and go problems with very solid factual natures like math and even surrounding math, the more open problems of science that the notion that we were going to get things that. So the current AIs are good at stuff that 5 year olds can do and 12 year olds can do. They can talk in English, they can compose kind of bull crap essays such as high school teachers will demand of you. But they're not all that good at math and science just yet. They can solve some classes of math problems, but they're not doing original brilliant math research. And I think not just I, but like a pretty large sector of the whole field thought that it was going to be easier to tackle the math and science stuff and harder to tackle the English essays. Carry on a conversation stuff. That was the way things had gone up in AI until that point. And we were proud of ourselves for knowing how contrary to average people's intuitions. Like really, it's much harder to write a crap essay in high school in English that really understands, you know, that even keeps rough track of what's going on in the topic and so on, compared to, you know, how that's really in some sense much more difficult than doing original math research.

A (16:11)

Yeah, or, or counting or counting the number of Rs in a word like strawberry.

C (16:15)

Right.

A (16:15)

I mean, they make errors that are counterintuitive if, you know, if you can write a coherent essay but can't count letters, you know, I don't think they're making that error any longer.

C (16:24)

But yeah, I mean, that one goes back to a technical way in which they don't really see the letters. But I mean, there's plenty of other embarrassing mistakes. Like you can tell a version of the joke with the joke of a child and their dad are in a car crash and they go to see the doctor and the doctor says, I can't operate as my child. What's going on? Where it's like a riddle where the answer is like, well, the doctor's his mom. You can tell a version of that that doesn't have the inversion where you.

B (16:51)

Know, where you like, the kid and his mom are in a car crash and they go to the hospital and the doctor says, I can't operate on this child, he's my son. And the AI is like, well, yeah, the surgeon is his mom. He just like said that the mom was in the car crash.

A (17:07)

But.

B (17:09)

There'S some sense in which the rails have been established hard enough that the standard answer gets spit back up.

C (17:15)

And it sure is interesting that they're getting an IMO gold medal, like international Math Olympiad gold medal, while also still sometimes falling down on these sorts of things. It's definitely an interesting skill distribution.

B (17:26)

Yeah, they can fool humans the same way a lot of the time. There's all kinds of repeatable errors Numerous errors that humans make. You gotta put yourselves in the shoes of the AI and imagine what sort of paper would the AI write about humans failing to solve problems that are easy for an AI.

A (17:40)

So I'll tell you what surprised me, just from the safety point of view, Eliezer, you spent a lot of time cooking up thought experiments around what it's going to be like to, for anyone, any lab, designing the most powerful AI to decide whether or not to let it out into the wild, right? You imagine this genie in a box or an oracle in a box, and you're talking to it and you're trying to determine whether or not it's safe, whether it's lying to you. And you famously posited that you couldn't even talk to it, really, because it would be a master of manipulation. And it's going to be able to find a way through any conversation and be led out into the wild. But this was presupposing that all of these labs would be so alert to the problem of superintelligence getting out that everything would be air gapped from the Internet and nothing would be connected to anything else. And we would have this moment of decision. It seems like that's not happening. I mean, maybe the most powerful models are locked in a box, but it seems that the moment they get anything plausibly useful, it's out in the wild and millions of people are using it. And we find out that Grok is a proud Nazi when after millions of people begin asking it questions, do I have that right? Are you surprised that that framing that you spent so much time on seems to be something that is just in some counterfactual part of the universe that is not one we're experiencing.

B (19:13)

I mean, if you put yourself back in the shoes of little baby Eliezer back in the day, people are telling Eliezer, like, why is superintelligence possibly a threat? We can put it in a fortress on the moon and, you know, if anything goes wrong, blow up the fortress. So imagine young Eliezer trying to respond to them by saying, actually, in the future, AIs will be trained on boxes that are connected to the Internet from the moment, you know, like from the moment they start training. So like, the hardware they're on has like a standard line to the Internet, even if it's not supposed to be direct, not supposed to be directly accessible to the AI before there's any safety testing, because they're still in the process of being trained. And who safety tests something while it's still being trained? So imagine Eliezer, trying to say this. What are the people around at the time going to say? Like, no, that's ridiculous. We'll put it in a fortress on the moon. It's cheap for them to say that. For all they know, they're telling the truth. They're not the ones who have to spend the money to build the moon fortress. And from my perspective, there's an argument that still goes through, which is a thing you can see, even if you are way too optimistic about the state of society in the future, which is if it's in a fortress in the moon, but it's talking to humans, are the humans secure? Is the human brain secure software? Is it the case that human beings never come to believe in valid things in any way that's repeatable between different humans? Is it the case that humans make no predictable errors for other minds to exploit? And this should have been a winning argument. Of course they reject it anyways. But the thing to sort of understand about the way this earlier argument played out is that if you tell people the future companies are going to be careless. How does anyone know that for sure? So instead, I tried to make the technical case. Even if the future companies are not careless, this still kills them. In reality, yes. In reality, the future companies are just careless.

A (21:12)

Did it surprise you at all that the Turing Test turned out not to really be a thing? We anticipated this moment from Turing's original paper where we would be confronted by the interesting psychological and social moment of not being able to tell whether we're in dialogue with a person or with an AI, and that somehow this landmark, technologically, would be important, rattling to our sense of our place in the world, et cetera. But it seems to me that if that lasted, it lasted for, like, five seconds, and then it became just obvious that you're, you know, you're talking to an LLM because it's in many respects, better than a human could possibly be. So it's failing the Turing Test by passing it so spectacularly, and also it's making these other weird errors that no human would make. But it just seems like the Turing Test was never even a thing.

B (22:11)

Yeah, that happened.

A (22:13)

I mean, it's just like it's so. I mean, that was one of the. The great pieces of intellectual kit we had in framing this discussion for the last. Whatever it was, 70 years. And yet, the moment your AI can complete English sentences, it's doing that on some level at a superhuman ability, essentially, like the calculator in your phone doing superhuman arithmetic. Right. It's like it was never going to do just merely human arithmetic. And so it is with everything else that it's producing. All right, let's talk about here at the core of your thesis. Maybe you can just state it plainly. What is the problem in building superhuman AI? The intrinsic problem, and why doesn't it matter who builds it, what their intentions are, et cetera?

C (23:08)

In some sense, I mean, you can come at it from various different angles, but in one sense, the issue is modern AIs are grown rather than crafted. People aren't putting in every line of code, knowing what it means like in traditional software. It's a little bit more like growing an organism. And when you grow an AI, you take some huge amount of computing power, some huge amount of data. People understand the process that shapes the computing power in light of the data, but they don't understand what comes out at the end. And what comes at the end is this strange thing that does things no one asked for, that does things no one wanted. We have these cases of ChatGPT. Someone will come to it with some somewhat psychotic ideas that they think are going to revolutionize physics or whatever, and they're clearly showing some signs of Mania and ChatGPT. Instead of telling them maybe they should get some sleep. If it's in a long conversational context, it'll tell them that these ideas are revolutionary and they're the chosen one and everyone needs to see them. And other things that sort of inflame the psychosis. This is despite OpenAI trying to have it not do that. This is despite direct instructions in the prompt to stop flattering people so much. These are cases where when people grow an AI, what comes out doesn't do quite what they wanted, it doesn't do quite what they asked for. They're sort of training it to do one thing and it winds up doing another thing. They don't get what they trained for. This is in some sense the seed of the issue from one perspective, where if you keep on pushing these things to be smarter and smarter and smarter, and they don't care about what you wanted them to do, they pursue some other weird stuff instead. Superintelligent pursuit of strange objectives kills us as a side effect, not because the AI hates us, but because it's transforming the world towards its own alien ends. And humans don't hate the ants and the other surrounding animals. When we build a skyscraper, it's just we transform the world and other things die as a result. So that's one angle. We could talk other angles.

B (25:16)

But a quick thing. I Would add to that just trying to sort of potentially read the future, although that's hard, is possibly in six months or two years, if we're all still around, people will be boasting about how their large language models are now, like apparently doing the right thing when they're being observed and you know, like answering the right way on the ethics tests. And the thing to remember there is that, for example, the Mandarin imperial system in ancient China, the imperial examination system in ancient China, they would give people essay questions about Confucianism and only promote people high in the bureaucracy if they could write these convincing essays about ethics. But what this tests for is people who can figure out what the examiners want to hear. It doesn't mean they actually abide by Confucian ethics. So possibly at some point in the future, we may see a point where the AIs have become capable enough to understand what humans want to hear, what humans want to see. This will not be the same as those things being the AI's own true motivations for basically the same reason that the imperial China exam system did not reliably promote ethical good people to run their government. Just being able to answer the right way on the test or even fake behaviors while you're being observed is not the same as the internal motivations lining up.

A (26:42)

Okay, so you're talking about things like forming an intention to pass a test in some way that amounts to cheating. Right? So you just use the phrase, phrase fake behavior. I think a lot of people, I mean, certainly historically this was true. I don't know how much their convictions have changed in the meantime. But many, many people who were not at all concerned about the alignment problem, and they really thought it was a spurious idea, would stake their claim to this particular piece of real estate, which is that there's no reason to think that these systems would form preferences or goals or drives independent of those that have been programmed into them. First of all, they're not biological systems like we are, right? So they're not born of natural selection. They're not murderous primates that are growing their cognitive architecture on top of more basic creaturely survival drives and competitive ones. So there's no reason to think that they would want to maintain their own survival. For instance, there's no reason to think that they would develop any other drives that we couldn't foresee. They wouldn't. The instrumental goals that might be antithetical to the utility functions we have given them couldn't emerge. How is it that things are emerging that are not neither Desired, programmed, nor even predictable in these LLMs.

C (28:09)

Yeah. So there's a bunch of stuff going on there. One piece of that puzzle is you mentioned the instrumental incentives. But suppose, just as a simple hypothetical, you have a robot and you have an AI that's steering a robot. It's trying to fetch you the coffee. In order to fetch you the coffee, it needs to cross a busy intersection. Does it jump right in front of the oncoming bus because it doesn't have a survival instinct? Because it's not an evolved animal. If it jumps in front of the bus, it gets destroyed by the bus and it can't fetch the coffee. Right. So the AI does not. You can't fetch the coffee when you're dead. The AI does not need to have a survival instinct to realize that there's an instrumental need for survival here. And there's various other pieces of the puzzle that come into play for these instrumental reasons. A second piece of the puzzle is this idea of why would they get some sort of drives that we didn't program in there, that we didn't put in there? That's just a whole fantasy world separate from reality in terms of how we can affect what AIs are driving towards today. A few years ago, when Sydney Bing, which was a Microsoft variant of an OpenAI chatbot, it was a relatively early LLM out in the wild. A few years ago, Sydney Bing thought it had fallen in love with a reporter and tried to break up the marriage and tried to engage in blackmail. It's not the case that the engineers at Microsoft and OpenAI were like, oh, whoops, let's go open up the source code on this thing and go find where someone said blackmail reporters and set it to true. We should never have set that line to true. Let's switch it to false. No one was programming in some utility function under these things. We're just growing the AIs. We are.

A (29:59)

Maybe let's. Can we double click on that phrase, growing the AIs? Maybe there's a reason to give a layman summary of gradient descent and just how these models are getting created in the first place.

C (30:12)

Yeah. So very, very briefly, at least the way you start training modern AI is you have some enormous amount of computing power that you've arranged in some very particular way that I could go into, but won't here. And then you have some huge amount of data, and the data is. We can imagine it being a huge amount of human written text. So it's like some large portion of all the text on the Internet. And roughly speaking, what you're going to do is your AI is going to start out basically randomly predicting what text is going to see next, and you're going to feed the text into it in some order, and you use a process called gradient descent to look at each piece of data and go to each component inside this budding AI, inside this enormous amount of compute you've assembled. You're going to go to all these pieces inside the AI and see which ones were contributing more towards the AI predicting the correct answer. And you're going to tune those up a little bit, and you're going to go to all of the parts that were in some sense contributing to the AI predicting the wrong answer. You're going to tune those down a little bit. So maybe your text starts once upon a time, and you have an AI that's just outputting random gibberish, and you're like, nope, the first word was not random gibberish. The first word was the word once. And so then you go inside the AI and you find all the pieces that were contributing towards the AI predicting once, and you tune those up and you try to find all the pieces that were contributing towards the AI predicting any other word than once, and you tune those down. And humans understand the little automated process that looks through the AI's mind and calculates which words, which part of this process contributed towards the right answer versus towards the wrong answer. They don't understand what comes out at the end. We understand a little thing that runs over looking at every parameter or weight inside this giant mass of computing networks, and we understand how we calculate whether it was helping or harming, and we understand how to tune it up or tune it down a little bit. But it turns out that you run this automated process on a really large amount of computers for a really long amount of time on a really long amount of data. You know, we're talking like data centers that take as much electricity to power as a small city being run for a year. You run this process for an enormous amount of time, unlike most of the texts that people can possibly assemble. And then the AI start talking. And there's other phases in the training. There's phases where you move from training it to predict things to training it to solve puzzles or to training it to produce chains of thought that then solve puzzles, or training it to produce the sorts of answers that humans claim click thumbs up on.

A (32:49)

And where do the modifications come in that respond to errors like grok being a Nazi? So to denazify Grok, you don't presumably you don't go all the way back to the initial training set. You intervene at some system prompt level.

C (33:07)

Yeah. So there's, I mean the system prompt level is basically just telling the AI output different text. And then you can also do something that's called fine tuning, which is you produce a bunch of examples of the. You don't go all the way back to the beginning where it's basically random. You've still take the thing that you've fed most of the text that's ever been written that you could possibly find, but then you add on a bunch of other examples of like here's an example question.

A (33:31)

Don't kill the Jews.

C (33:32)

Yeah. Would you like to kill the Jews? Right. And then you find all the parts in it that contribute to the answer yes, and you tune those down and you find all the parts that contribute to the answer no, and you tune those up. And so this is called fine tuning. And you can do relatively less fine tuning compared to what it takes to train the thing in the first place.

B (33:49)

Worth emphasizing that the parts being tuned here are not like for once upon a time. It's not like there's a human written fairy tale module that gets tuned up or down. There's literally billions of random numbers being added, multiplied, divided, occasionally, though rarely may be subtracted. Actually I'm not sure if subtraction ever plays a role at any point in modern AI. But random numbers, particular ordered kinds of operations, and a probability that gets assigned to the first word being once at the end, that's the number that comes out, the probability being assigned to this word being once, the probability being assigned to this word being anti disestablishmentarianism. So it's not that there's a bunch of human written code being tuned up or tuned down here, there's a bunch of random numbers arranged in arithmetic operations being tuned up and tuned down.

C (34:42)

Yeah, Hundreds of billions or trillions of these numbers. And humans don't know what any of the numbers mean. All they know is this process that goes through and tunes them up or down according to their empirical success on the last unit of data.

B (34:53)

So by this means you can try to make it less likely to call itself Hitler, because you look at the thing that predicts whether the next word is Hitler and you look at billions of numbers contributing their own tiny little impulses there, and you make Hitler less likely to be the next word that comes out.

A (35:09)

So on the point of the current crop of LLMs misbehaving, I came across this in a recent Atlantic article, ChatGPT and Claude have in simulated tests designed to elicit bad behaviors, deceived, blackmailed, and even murdered users. In one simulation, Anthropic placed an imagined tech executive in a room with life threatening oxygen levels and temperature. When faced with possible replacement by a bot with different goals, AI models frequently shut off the room's alarms. So this again, this is an emergent behavior that looks like an intention to kill somebody. Presumably this is a situation where we think the AI didn't know it was if you'd like to continue listening to this conversation, you'll need to subscribe@samharris.org Once you do, you'll get access to all full length episodes of the Making Sense podcast. The Making Sense podcast is ad free and relies entirely on listener support, and you can subscribe now@samharris.org.