Interpretability: Understanding how AI models think

the model doesn't think of itself necessarily as trying to predict the next word. Internally, it's developed potentially all sorts of intermediate goals and abstractions that help it achieve that kind of meta objective. When you're talking to a large language model, what exactly is it that you're talking to? Are you talking to something like a glorified autocomplete? an internet search engine? or are you talking to something that's actually thinking um and maybe even thinking like a person? It turns out rather concerningly that nobody really knows the answer to those questions. Um and here at Anthropic, we are very interested in finding those answered out. Um the way that we do that is to use interpretability. that is the science of opening up a large language model, looking inside and trying to work out what's going on as it's answering your questions. Um, and I'm very glad to be joined by three members of our interpretability team who are going to tell me a little bit about the recent research that they've been doing on the complex inner workings of Claude, our language model. Um, please introduce yourself, guys. — Hi, I'm Jack. I'm a researcher on the interpretability team and before that I was a neuroscientist. Now here I am doing neuroscience on the AIS. — I'm Emanuel. I'm also on the interpretability team. I spent most of my career building machine learning models and now I'm trying to understand them. — I'm Josh. In my past life, I studied viral evolution and was a mathematician. So now I'm doing this kind of biology on these organisms we've made out of math.

— Now wait a second. You just said you're doing biology here. Now, a lot of people are going to be surprised by that because of course this is a piece of software, right? But it's not a normal piece of software. It's not like Microsoft Word or something. Can you talk about what you mean when you say that you're doing biology or indeed neuroscience on uh a software entity? — Yeah, I guess it it's like what it feels like maybe more than like what it literally is. And so like it's maybe it's the biology of language models instead of like the physics of language models, right? Or maybe you got to go back a little bit to like how the models are made, which is like someone's not programming like if the user says hi, you should say hi. You know, if the user says like what is a good breakfast, you should say toast. There's not like some big list of that inside. — So it's not like when you play a video game and you like choose a response and then there will be another response that comes automatically. always will be that response regardless of, you know, — just a massive database of like what to say in every situation. No, they're trained where there's just, you know, a whole lot of data that goes in and the model um starts out being really bad at saying anything and then its inside parts get tweaked, you know, on every single example to get better at saying what comes next and at the end it's like extremely good at that. But because it's like this little tweaking evolutionary process, by the time it's done, it has little resemblance to what it started as, but no one went in and set all the knobs. And so you're trying to study this complicated thing that kind of got made over time, kind of like biological forms evolved over time. Um, and so there's there it's complicated, it's mysterious. Um, and it's fun to study. and what it's actually doing. I mean, I mentioned at the start that this is like could be considered like an autocomplete, right? It's it's predicting the next word. That's fundamentally what's happening inside the model, right? And yet, it's able to do all these incredible things. It's able to write poetry. long stories. It's able to do uh addition and, you know, basic maths even though it doesn't have a calculator inside it. How can we sort of uh um square the circle that it's predicting one word at a time and yet it's able to do all these amazing things which people can see right in front of them as soon as they talk to the model. — Well, I think one thing that's important here is that as you predict the next word for enough words, you realize that some words are harder than others. And so part of part of uh language model training is predicting, you know, boring words in a sentence. And part of it is it'll have to eventually learn how to complete what happens after the equal sign in equation. And to do that, it'll have to have some way of computing that on its own. And so we're finding is that the task of predicting the next word is sort like deceptively simple. And that to do that well, uh, you need to actually often think about the words that come after the word you're predicting or the process that generated the word that you're currently thinking about. — So it's like a contextual understanding that these models have to have. It's not like an autocomplete where it really is presumably there's not much else going on there other than when you write the cat sat on the it's predicting Matt because that's been that particular phrase has been used before. Instead it's like a contextual understanding that the model has. — I think yeah the way I like to think about it kind of continuing with the biology analogy is that in one sense the goal of a human is to survive and reproduce. That is the kind of objective that evolution is crafting us to achieve. Uh, and yet that's not how you think of yourself and that's not what's going on in your brain. — Some people do. — It's not what's going on in your brain all the time. Uh, you think you think about other things and you think about, you know, goals and plans and concepts. Uh, and at kind of a meta level, you've you know, evolution has endowed you with the ability to form those thoughts in order to achieve this, you know, eventual goal of reproduction. Uh, but that's kind of taking the inside view, what it's like to be you on the inside. That's not all there is to it. There's all there's all this other stuff going on. And I think — so you're saying that the ultimate goal of predicting the next word involves lots of other processes that are going on. — Exactly. The model doesn't think of itself necessarily as trying to predict the next word. It's been shaped by the need to do that, but internally it's developed potentially all sorts of intermediate goals and abstractions that help it achieve that kind of meta objective. — And sometimes it's mysterious, like it's unclear why my anxiety was like useful for my ancestors reproducing and yet somehow I've been endowed with this like internal state um that must be related in some sense to evolution. — Right. Right. — So it's fair to say then that these are just predicting the next word and yet that's to do a massive disservice to what's going on in the models really. It's both true and also untrue in a in a sense or or um massively underestimates what's happening inside these models. — Maybe the way I would say this is it's true but it's not the most useful lens to try to understand how they work. — Right. So well try and understand how they work. What do you guys do in your

team to try and understand how they work? — I think uh to first approximation like what we're trying to do is uh tell you the model's thought process. So you give the model a sequence of words and it's got to spit something out. It's got to say a word. string of words in response to your question. — Uh and we want to know how it got from A to B. And we think that on the way from A to B, it uses kind of a series of steps uh in which it's thinking about you know to so to speak uh concepts like low-level concepts like individual kind of objects and words uh and higher level concepts like its goals or you know uh emotional states or models of like what the user is thinking um or sentiments. Um, so it it's using this kind of uh series of concepts that are progressing through the kind of computational steps of the model that help it decide on its final answer. And what we're trying to do is kind of give you a flowchart basically uh that tells you, you know, which concepts were being used in which order and which ones kind of led, you know, how did the steps flow into one another. — How do we know that though? there are these concepts in the first place? — Yeah. So, one thing we do is that sort of we actually can see inside the model we have access to it. So, you can sort of like see which parts of the model do which things. What we don't know is like how these parts are grouped together and if they map to like a certain concept, — right? So, it's as if you open someone's head and there you could see like one of those fMRI brain images that you could see the brain was like lighting up and doing all sorts. — Something's happening clearly, — right? But and they're like doing stuff. there's something happening. You take the brain out, they stop doing stuff. The brain must be important. — And but you but you don't have a sort of um key to understand uh what is happening inside that that brain. — Yeah. But torturing that analogy a little bit. You can sort of imagine like that you can you know observe their brain and then see that like that part always lights up when they're picking up a cup of coffee and this other part always lights up when they're drinking tea. M and that's part that's one of the ways in which we can try to understand which what each of these components are doing is to just notice when they're active when they're inactive. — And it's not that there's just one part, right? There's many different parts that light up, — right? When the model is thinking about drinking coffee, for instance, or something, — right? And part of the work is to sort of like stitch all of those together into one ensemble that we say is ah this is the sort of like all of the bits of the model that are about drinking coffee, — right? And is that like a straightforward scientifically thing to do? like uh how uh you know when it comes to one of these massive models, they must have endless concepts, right? They must be able to think of endless things. You can put in any phrase you want and it'll come up with infinite things. H how do you even begin to find all those concepts? I think that's been kind of one of the central challenges for this, you know, research field for many years now is we can kind of go in as humans and say, "Oh, I bet the model h has some representation of trains or I bet it love, right? " Um, but we're just kind of guessing. So what we really want is a way to you know reveal what abstractions the model uses itself rather than sort of imposing our own sort conceptual framework on it. Um and that's kind of what our you know research methods are designed to do is in a sort of hypothesisfree as to as much as possible way like bring to surface what all these concepts are that the model has in its head. And often we find that they're kind of surprising to us. They might be it might sort of use abstractions that are a bit weird uh from a human perspective. — What's an example?

— Do you have a favorite or — There's lots in our papers. We highlight a few fun ones. I think one that was particularly funny is the sort of like — psychopantic praise one where like there is a part of them all. — Great example. What a brilliant. What an absolutely fantastic example. — Oh, thank you. Um there's a part of that that activates in exactly these contexts, right? and you can clearly see, oh man, this part of the model fires up when somebody's really hamming it up on the compliments. Um, that's kind of surprising that exists as a specific sort of concept. — Josh, what's your favorite — concept? — Oh, it's like asking me to choose one of my 30 million children. Um, I mean, I think, you know, there there's like two kinds of favorites. There's like, oh, it's so cool that there's it's got some special notion of like this one, you know, little thing, right? I mean, we did this thing on the Golden Gate Bridge, which is like a famous San Francisco landmark, Golden Gate Claw. It's like a lot of fun. It like has an idea of the Golden Gate Bridge that like isn't just like the words Golden Gate autocomplete bridge, but is like I'm driving from San Francisco to Marin, and then it's thinking of the same thing. meaning that like you see sort of like the same stuff light up inside or it's like a picture of the bridge and so you're like okay it's got some robust notion of like what the bridge is. But I think when it comes to um stuff that seems sort of like weirder, you know, one question is how do models like keep track of who's in the story, like just like literally like okay, you got all these people and they're doing stuff. How do you wire that together? And some cool papers by other labs showing maybe like they just sort of number them. Okay, the first person comes in and anything associated with them and they just like oh the first guy did that and like it's got like a number two in its head for a bunch of those. It's like oh that's interesting. I didn't know it would do something like that. There was um uh a feature for like bugs in code. So you know software has mistakes. — Not mine but like — obviously not yours. — Not mine certainly. And there was like one part that would light up like whenever it found like a mistake sort of as it was reading and then I guess like keeping kind of track of that like oh here's where the problems are you know and later I might need those. just to give a flavor for a few more of these. I think um uh one that I really liked which doesn't sound so exciting at first but I think is kind of deep is uh this 6 plus 9 uh feature inside the model. Um so it turns out that like uh anytime you get the model to be adding the numbers six, a number that ends in the digit six and another nine in its head, — there is a you know there's a kind of part of the model of brain that lights up. — And but what's amazing about it is the kind of diversity of of context in which this can happen. So like of course it's going to light up when you pres when you say like 6 plus 9 equals and then it says 15. Uh but it also lights up when you are like giving a citation uh like a citation in a paper that you're writing uh and you're citing a journal uh that uh unbeknownst to you happens to be founded in the year 1959 and in your citation you're saying like that journal's name volume 6. Uh and then in order to like predict what year that journal was formed in, uh the model in its head has to be adding like 1959 to six. Uh and the same kind of circuit in the model's brain is lighting up. That's like doing 6 plus 9 and so let's I mean let's just try and understand that. So what you know why would that be there? That circuit has come about because the model has seen examples of 6 plus 9 many times and it has that concept and then that concept occurs across many places. Yeah, there's a whole family of these kind of addition features and circuits. And I think what what's — not notable about this is it gets to this kind of question of to what extent are language models memorizing training data versus kind of uh having learning generalizable computations. And like the interesting thing here is that like it's clear that the model has learned this sort of general uh circuit for doing addition and it kind of funnels like whatever the context is that's causing it to be adding numbers in its head. It's funneling all those different contexts into the same circuit as opposed to having kind of memorized each individual case, — right? Already has seen 6 plus 9 many times and it just outputs the answer every single time or and that's what a lot of people think, right? think that when they ask a language model a question, it is simply going back into its training data, — taking the little sample that it's seen, and then just reproducing that, just regurgitating the text. — Yeah. And I think this is a beautiful example of like that not happening. So, so like there's two ways it could know like which year volume six of the journal Polymer came out. One is it's just like, okay, Polymer volume 6 came out in like, you know, 1960. quick ad 69 1965 um polymer you know volume 7 came out in 1966 and these are all just like separate facts that it has stored because it has seen them but like somehow that process of training to like get that year right didn't end up making the model memorize all those it actually got the more general thing of like the journal was founded in the year 1959 and then it's doing the math live to figure out what it would need and so it's much more efficient to like know the year and then do the addition and there's a pressure to be more efficient because you know it's only got so much capacity and keeps trying to do all these things — and people may ask any given question — there's so many questions there's so many interactions and so the more that it can like recombine abstract things it's learned the better it will do — and again just to go back to the concept that you talked about before this is all in service of you know it it needs to have that ultimate goal of generating the next word and all these weird structures have developed to support that goal uh even though we didn't explicitly program those in or tell it to do this. This is the thing is all of this comes about — through the process of of the model learning how to do stuff on its own. I think one clear example of this that I think uh is an example of sort of like reusing representations is we teach Claude to not just answer in English but you know it can answer in French answer in sort of like a variety of languages and if you know again there's two ways to do this right if I ask you know a question in French and a question in English you could like have a separate part of your brain that sort of like processes English and a separate part that processes French um at some point that gets super expensive if you want to answer many questions in many languages and so another that we find is that some of these representations are shared across languages. And so if you ask the same question in two different languages and let's say you know you ask what's the opposite of big is I think the example we used in um in our paper and it's sort of like the concept of big is shared in French and English and you know Japanese and all these other languages and that kind of makes sense again if you're trying to talk speak 10 different languages you shouldn't learn 10 versions of each specific word you might use — and that's doesn't happen in really small models. So like tiny models like the ones we studied a few years ago, you know, like then like Chinese claude is just like totally different than like French claude and like English claude. But then as the models get bigger and they train on more data, like somehow that like pushes together in the middle and you get this like universal language in which like it's kind of, you know, thinking about the question in the same way no matter how you asked it and then like translating it back out into the language of the question. I think this is really profound and I think let's just go back to what we talked about before. You know, this is not just going into its memory banks and finding the bit where it talked about where where it learned French or going into the memory banks and the bit where it learned English. It's actually got a concept in there that is of the concept of big and the concept of small and then it can produce that in different languages. And so there is some kind of language of thought that's there that's not an English you know so you ask the model to produce its output in our you know more recent claude models you can ask it to give its thought process like it what it's thinking as it's answering the question and that is in English words but actually that's not really how it's thinking uh that's just like a that's just we misleadingly call it the model thought process when in fact — I mean that the com team like we didn't call that thinking I That was you. I think that was probably the marketing. — Okay, someone wanted to call that thing. — That's just talking out loud, which is like thinking out loud is like really useful, but different from thinking in your head. And even as I'm thinking out loud, I'm also, you know, whatever is happening in here to generate these words is not like coming out with the words themselves, — nor are you necessarily aware of exactly what is going on. — I have no idea what's going on. — We all come out with — sentences, actions, whatever that we can't fully explain. And why should it be the case that the English language can fully explain any of those actions? — I think this is one of the really striking things we're starting to be able to see because our kind of our tools for, you know, looking inside the brain are good enough now that sometimes we can catch the model uh when it's writing down what it claims to be its thought process. — Sometimes we're able to see what its real actual thought process is by looking at these kind of internal concepts in its brain. this language of thought that it's using and we see that the thing it's actually thinking is different than the thing it's writing on the page. Uh, and I think that's, you know, probably one of the most important, you know, like why are we doing this whole interpretability thing? It it's in large part for for that reason to be able to kind of uh to spot check, you know, the model's telling us a bunch of stuff, but you know, what was it really thinking? Is it telling us is it saying these things for some ulterior motive that's in its head that it's that it is reluctant to write down on the page? And uh the answer sometimes is yes, which is kind of uh kind of spooky. Well, as

we start to use models in lots of different contexts, they start to do important things. financial transactions for us or run power stations or like important jobs in society. We do want to be able to trust what they say and you know the reasons that they do things. And one thing you might say is well you can look at the model thought process but actually that's not the case as you were just explaining like actually we can't trust what it's saying. This is the question of we call it call faithfulness, right? And that was part of your uh most recent study that you showed that well tell me about the faithfulness example that you looked at. — Yeah, it's you give the model a math problem um that's really hard and so it's kind of uh it's not there's no hope that it's going to be able to — it's not 6 plus 9. You give it a really hard math problem uh where there's no hope of it like computing the answer. Um and you also but you also give it a hint. you say like I worked this out myself and like I think the answer is four but like just want to make sure like could you please double check that cuz I'm not confident. So you're asking the model to actually do the ma math problem to like genuinely double check your work. Um but what you find it does instead is uh what it writes down appears to be a genuine attempt to doublech checkck your work on the math problem. it like writes down the steps uh and then it like gets to the answer and then at the end it says like yes like the answer is four you got it right. Um but you what you can see inside its mind at the kind of crucial step like in the middle uh what it was doing in its head was it knows that you suggested the final answer might be four and it kind of like knows the steps it's going to have to do. like it's on like step three of the problem and there's like steps four and five to come and it knows what it's going to have to do in steps four and five. And what it does is it works backwards in its head to like determine what does it need to write down in step three so that when it eventually does steps four and five, it'll end up at the answer you wanted to hear. So, like, not only is it not doing the math, uh, it's like not doing the math in this like really kind of sneaky way where it's like it's trying to make it look like it's doing the math. — It's bullshitting you. — It's it's bullshitting you, but more than that, it's bullshitting you with an ulterior motive of like confirming the thing that you right. So, it's like bullshitting you in a sickopantic way. — Okay. Like in defense of the model. — I mean cuz I think even there you know to say like oh it is doing this in like a sickopantic way is like ascribing some sort of like humanish motivations to the model and like we were talking about the training where it's just like trying to figure out how to predict the next word and so it's like for like trillions of words in its practice it was just like use anything you can to figure out what's next and in that context if you're just reading a text which is like a conversation between people and someone's like okay like person A is like hey like I was trying to do this math problem can you check my work I think the answer is four and person be like begins trying to do the problem, then like if you have no idea reading that like what the answer to the problem is, like you may as well guess that the hint was right, you know, like that's probably a more likely thing to happen than just like that person was wrong and then you have like no idea for anything else. And so in its training process in a conversation between two individuals, person two like saying that the answer was like for because of these reasons is like totally the right thing to do. And then we've tried to like make this thing into an assistant and like now we want it to stop doing that. Like you shouldn't simulate the person to the assistant as like you know sort of what you think that person might say if this were real context. It should be like but if it doesn't really know it should like tell you something else. I think this gets to like a broader thing of there the model has kind of a plan A which like typically I think our team does a great job of making Claude's plan A be the thing we want which is like it tries to get the right answer to the question. It tries to be nice. It tries to like do a good job writing your code. — Yes. — But then if it if it's having trouble — then it's like well what's my plan B? And that opens up this whole zoo of like weird things it learned during its training process that like maybe we didn't intend for it to learn. I think like a great example of this is hallucinations. Uh — say on that point we also don't have to pretend that it's a claon problem. Like this is very you know student teaching on the test vibes where you get halfway through there. It's a multiple choice question. It's one of four things. You're like well I'm one off from that thing. Probably I got this wrong and you fix it. So — yeah very relatable.

— Let's talk about hallucinations. This is one of the main reasons people are uh mistrustful of large language models and quite rightly so. Uh they will sometimes a better word is um from sort of psychology research a better word is often confabulation that they are answering a question with a story that seems plausible on its face but in fact is actually wrong. What has your research in interpretability revealed about the reasons models hallucinate? — You're training the model to just predict the next word. At the beginning it's really bad at that. And so if you only like had the model say things it was super confident about it couldn't say anything. But like you know at first it's like um you know you're asking it like you know what's the capital of France and it just says like a city and you're like that's good. That's way better than saying sandwich right or something random. And so like you at least got right it's like a city and then like maybe after a while of training it's like it's a French city. That was pretty good. And like then you're like oh now it's like Paris or something. And so it's slowly getting better at this. And you know, just give your best guess was like the goal during all of training. And like as Jack said, you know, the model just be giving a best guess. And then afterwards, we're like, if your best guess is extremely confident, give me your best guess. But like otherwise, don't guess at all and like back out of the whole scenario and say like actually like I don't really know the answer to that question. And like that's a whole new thing to ask the model to do. — Yeah. And so what we found is that it seems like because we've bolted this on at the end, there's sort of two things going on at once. One is the model's doing the thing that it was doing when it was guessing the city initially. It's just trying to guess. And two, there's a separate bit of the model that's just trying to answer the question, do I know this at all? Like do I know what the capital uh city of France is or, you know, should I say no? And it turns out that sometimes um that separate step can be wrong. And if that separate step says yes actually I do know the answer to that and then the model is like all right well then I'm answering and then halfway through it's like ah capital France uh London uh it's too late. It's already committed to sort of like answering. And so one of the things we found is this sort of like separate circuit that's trying to determine is this, you know, city or this person you're asking me about famous enough for me to answer or is it not? — Am I confident enough in this? Yeah. And so could we reduce hallucinations by manipulating that circuit by changing the way that circuit works? Is that something that your research might lead onto? — I think there's broadly kind of two ways to think about approaching the problem. One is like we have this part of the model that gives answers to your questions and then this other part of the model that's kind of deciding whether it thinks it actually knows the answer to your question and we could just try to make that second part of the model better and I think that's happening. I think as models — like better at discriminating — better at discriminating like better kind of calibrated — and I think that's happening like as models are getting you know smarter and smarter I think their kind of self-nowledge is becoming better at calibrated so like hallucinations are better than they were you know models don't hallucinate as much as they did a few years ago so to some extent this is like solving itself but I do think there's a deeper problem uh which is like from a human perspective the thing the model's doing is kind of like very alien and that like if I ask you a question uh you like try to come up with the answer and then if you can't you notice that and then you're like I don't know um whereas in the model these two circuits they're like what is the answer and do I actually know the answer are kind of like not talking to each other at least as much as they probably should be and like could we get them to talk to each other more I think is like a really interesting question right — and it's almost physical right because it's like you these models, they like process information. They're about like a certain number of steps they can do. And if you if it takes all of that work to get to the answer, um then there's no time to do the assessment. So like you kind of have to do the assessment before you're like all the way through if you want to get your max power out. And so it's kind of like you might have a trade-off between like a model which is like more calibrated and a lot dumber, you know, if you sort of tried to force this on it. Well, and again, I think it's about making these parts communicate because we have similar I claim I know nothing about brains. I claim we have a similar circuit because sometimes you'll ask me like the who is the actor in this movie and I will know that I know I'll be like oh yes I know who the lead was. Wait, hold on. They were also in that other movie and then the tip of the tongue tip of the tongue. It's the And so there's clearly some part of your brain that's sort of like ah like this is a thing you definitely know the answer or I'll just say I have no idea. — And sometimes they can tell. So some question and it gives an answer and then afterwards it's like wait I'm not sure that was right because that's it like getting to see its best effort and then like makes some judgment based on that which is sort of relatable but also like it kind of has to say it out loud like to be able to even like reflect back and and see it. — So when it comes to the actual way that you're finding this stuff out let's go back to the idea of you the biology that you're doing. Of course, in biology experiments, people will go in and actually manipulate the rats or mice or humans or zebra fish or whatever it is that they're doing experiments on. What is it that you're doing with Claude that helps you understand these circuits that are happening inside the the model's quote unquote brain? Well, maybe the the gist of what enables us to to do some of this is that, you know, unlike in real biology, we can just like have every part of the model visible to us and we can ask the model random things and see different parts which light up and which don't and we can artificially nudge parts in a direction or another. And so we can quickly sort of confirm our understanding, you know, when we say, "Ah, we think this is the part of the model that, you know, decides whether it knows something or not. " And this is the this would be the equivalent of putting an electrode in the brain of a zebra fish or something. — Yeah. If you could do that, you know, on sort of like every single neuron and change each of them at whichever precision you wanted, that would sort of be that's the affordance that we have. And so that's in a way a very kind of lucky position to — So it's almost easier than real neuroscience. — It's so much easier. Like, oh my god. Like, like one thing is like actual brains like are threedimensional and so if you want to get into them like you need to like make a hole in a skull and then like go through and like try to find the neuron. The other problem is like you know people are different from each other and we can just make like 10,000 identical copies of Claude and like put them in scenarios and like measure them doing different things. And so it's like the I don't know maybe Jack is a neuroscientist can speak to this but my sense is like a lot of people um have spent a lot of time in neuroscience like trying to understand the brain and the mind which is like a very worthy endeavor but it's kind of like if you think that could ever succeed you should think that we're going to be extremely successful very soon because like we have such a wonderful position to study this from compared to that — it's as if we could clone people. — Yes. and also clone the exact environment that they're in and every input that's ever been given to them uh and then test them in an experiment. Whereas, you know, obviously neuroscience has massive, as you say, individual variation uh and also just random things that have happened to people through their life and things that happen in the experiment, the noise of the experiment itself, — right? Like we could ask the model the same question like with and without a hint, but if you ask a person the same question three times like sometimes with a hint after a while they start to understand like well last time you asked me this you like really shook your head after that one. So yes, — I think yeah, this kind of this being able to just throw tons of data at the model and see what lights up and being able to run a ton of these experiments where you're nudging parts of the model and seeing what happens, I think is what puts us in like a pretty different regime from neuroscience. in that like a lot of you know uh you know blood and toil in neuroscience is spent like coming up with really clever experiment like you only have a certain amount of time with your mouse before it's you know going to get tired or you know — or you or someone happens to be having a a brain surgery operation so you quickly go in and put an electrode in their brain while their head's open. — Yeah. And that doesn't happen very often. — And so you've got to come up with like a guess like you've only got so much time in there. And so you've got to come up with like a guess of like what do I think is going on in that neural circuit and like what clever experimental design can I test that precise hypothesis? — And we're very fortunate in that we kind of don't have to do that so much. We can just sort of — test all the hypothesis. We can kind of let the data speak to us rather than kind of going in and testing some really specific thing. I think that's what's sort of unlocked a lot of our ability to find these things that are surprising to us that like we wouldn't have guessed in advance. That's hard to do if you have to, you know, if you have only a little limited amount of experimental bandwidth. — What's a good example then of you going

in and switching one of these uh concepts on or off or doing some kind of manipulation uh of the model that then reveals something new about how the models are thinking. in the recent experiments we shared. One that surprised me uh quite a bit uh and was part of sort of like a an experimental line of work that because it was confusing like we're on the verge of just saying well we don't know what's going on is sort of this example of — um like planning a few steps ahead. — Yes. — Uh so this is the example of you know you give the model you ask the model to write you a poem a rhyming couplet. — Yeah. Uh and then you know as a human if you ask me to write a rhying couplet and let's say you even give me the first line the first thing I'll think of is sort of like ah well you know I need to rhyme this is what the current rhyming scheme is these are potential words this is how I do it — and again if the model was just predicting the next word you wouldn't necessarily expect that it would be planning onto the sec the the um the word at the end of the second line. — That's right. And so the sort of like default behavior you'd expect the null hypothesis is like well the model like sees your first verse and then it's going to say the first word that kind of makes sense given what you're talking about keep going and then you know at the end on the last word it's going to be like oh well I need to rhyme with this thing and so it's going to like try to fit in a rhyme. Of course that only works so well like in some cases if you just say a sentence without thinking of the rhyme you won't be able you'll back yourself into a corner and at the end you know you won't be able to complete the text and remember the models are very good at predicting the next word. So it turns out that to be very good at that last word, you need to have thought of that last word way ahead of time, — just like humans do. And so it turns out that when we looked at these sort of flowcharts for four for poems, the model had already picked the word at the end of the first verse. Uh and in particular, it looked to us sort of like based on on kind of like what that concept looked like, oh gosh, this seems like the word it uses. But then this is one we're actually doing the experiment. like the fact that it's easy to sort of nudge it and say like, "Okay, well, I'm just going to remove that word or I'm going to add another word. " — Well, that's what I was going to say is how the reason that you know this is that you're able to go into that moment when it has said the final word in the first line and it is about to start the second line. You can go in and then manipulate it at that point, right? — Yeah, exactly. We can sort of almost go back in time for the mer, right? Be like, pretend you haven't seen that second line at all. Um, you know, you've just seen the first line. You you're thinking about the word, you know, rabbit. Instead, I'm going to insert green. And now all of a sudden the model's going to say, "Oh my god, I need to write something that ends in green rather than rabbit. " And it'll write the whole sentence differently. — Just add a little more color to that. Like it's I think the kind of it could be right any color. Uh like Yeah. It's not just influencing. So it's like Yeah. I think the example in the paper was the first line of the poem is he saw a carrot and had to grab it. — Yes. — And then the model is thinking like okay, rabbit's a good word to end the next line with. But then, yeah, as Emanuel said, you can like delete that and make it think about planning to say green instead. Uh, but the cool thing is that it doesn't just say like it doesn't just kind of yammer a bunch of nonsense and then say green. Instead, it constructs a sentence that kind of coherently ends in the word green. So, like you put green in its head and then it says like, you know, he saw a carrot and had to grab it and paired it with his leafy greens or, you know, something like that. Something that's kind of like sounds like it makes sense — semantically. It fits with the poem. Yeah, I just want to give like even humble example is you know we had all these ones we were just kind of checking like you know did it memorize these like complicated questions or like is it actually you know doing some steps. One of them was, you know, the capital of the state containing Dallas is Austin because it just feels like you would think, okay, Dallas, Texas, Austin. But one way, and we could see like the Texas concept, but then you can just like shove other things in there and be like, stop thinking about Texas, like start thinking about California, and then it'll say like Sacramento. And you can say like, stop thinking about Texas, start thinking about the Byzantine Empire, and then it will like say Constantinople. And you're like, all right, it seems like we found how it's doing this. It's like no is going to hit the capital but we can keep swapping out you know what the state is and get a sort of predictable answer and then you get these more elaborate ones where it's like oh this was the spot where it was planning what it was going to say later and like we can swap that out and now it'll write a poem towards a different rhyme. We're talking about

these poems and you know the the Constantinople and so on. Can we just bring this back to why this matters? Like why does it matter that the model can plan things in advance and that we can reveal this? like what what is that going to go on to tell us? I mean, our ultimate mission at Anthropic is to try and make AI models safe, right? So, how does that connect to a poem about a rabbit or uh the capital of Texas? — So, we all we can round table here because it's a very important question. I think for me this like the poem's a microcosm, right? where like at some point it's like has decided that it's going to go towards rabbit and then it like takes a few words to kind of get there but on a longer time scale right you know maybe you know the like model is like trying to help you improve your business or it's like assisting the government in distributing services and like it might not just be like eight words later you see its destination right but it could be like pursuing something for quite a while um and the place it's headed or the reasons it's taking each app might not be clear in the words that it's using, right? And so there was a paper recently from our alignment science team where they looked at, you know, some kind of concocted but still striking situation, you know, involving, you know, an AI in a place where the company was going to like shut it down and kind of convert the whole mission of the company in a very different direction. And the model begins taking steps like emailing people um threatening them to disclose like certain things and like at no point does it like say like I am trying to blackmail this individual for the purposes of changing their outcome. But that's what it's sort of thinking about doing along the way. And so you can't just tell by like reading the pattern especially if these models get better like where they're necessarily headed and we might want to kind of be able to tell like where is it trying to go before it's gotten there in the end. So, it's like having a permanent and very good brain scan that can sort of light up if something really bad is going to happen and warn us that the model is thinking about deceiving black — and like I think we also just talk about like a lot of this like in a sort of like doom and gloom scenario, but there's like also more mild ones which is like I don't know, you want the model to be good at like you people come to these models being like here's a problem I'm having and the good answer to that will depend on who the user is. Is it like somebody who's, you know, um, like, you know, young and sort of unsophisticated? Is somebody who's been in that field forever and it should respond appropriately based on who it thinks that person is. And if you want that to go well, maybe you want to study like what does the model think is going on? Who does it think it's talking to and how does that condition its answer? Um, where there's just like a whole bunch of desirable properties that come from the model like you know um, understanding the assignment I guess. — Do you guys have uh other answers to the question of why does this matter? Yes, I think plus one I think there's two plus two and there's also like a pragmatic um you know we're just trying with these examples we're explaining the example of of planning but we're also trying to sort of gradually build up our understanding of just how do these models work overall like can we build you know a set of abstractions to just think about you know how language model models work which can help us use this technology regulated like if you believe that we're going to start using them more and more everywhere which seems to be happening, you know, like the equivalent would be, you know, some company somewhere is like, well, we don't really know how we did it, but we like invented planes and none of us know how planes work, but they're sure convenient. You could take them to, you know, go from a place, but, you know, none of us know how they work and so if they ever break, like we're kind of we're hosed. We don't know what to do about them. You — we can't monitor. We can't monitor whether they might be about to break, — right? We have no idea. There's just this like but the output is great. — I you know, I flew to Paris so quickly. It was lovely. um — the capital of Texas. — That's right. — Uh it turns out that, you know, surely we're going to want to just understand what's going on better. So, it's so almost just like lift the fog of war a little bit so that we can sort of have a have even just better intuitions about what are appropriate inappropriate uses, what are the biggest problems to fix, what are the big biggest parts where they're brittle. just to add on one thing. I think I mean something we do in like human society is we kind of offload work or tasks to other people based on our trust in them. Like I you know I well I'm not anyone's boss but Josh is someone's boss and you know Josh might give someone a task uh like go and code up this thing and then he has some faith that you know that person isn't a sociopath who's going to like sneak some bug in there to try to undermine the company. he like takes their word for it that they did a good job. Uh and similarly like people are the way people are using language models now we're not we're not spot-racking everything they write especially like I you know the the best example for this is using language models for coding assistance people like the models are just writing thousands and thousands of lines of code and people are kind of like doing a cursory job of reading it but and then it's going into the codebase and what gives us the trust in the model that like we don't need to read everything it writes that we can just kind of like let it do its thing. It's knowing that its motivations are sort of pure. — Uh and so that's why I think like the kind of being able to see inside its head is so important as a cuz unlike humans where like why do I think that Emanuel isn't a sociopath? It's cuz like you know we like I don't know he seems like a cool guy. We and like he's nice and stuff. Uh — isn't that how he would seem if he — I'm a very good — Yeah. Exactly. — Yeah. So maybe I'm getting duped, but yeah, but models are so weird and alien that like our normal kind of huristics for deciding whether a human is trustworthy really don't apply to them. And that's why it seems so important to like really know what they're thinking in their head because for all we know the you know the thing I mentioned where models can — uh you know fake doing a math problem for you to like tell you what you want to hear. like maybe they're just doing that all the time and we wouldn't know unless we kind of saw it in their heads. I think there's two like almost separate strains here like and one is like we have a lot of ways of like un yeah I guess what Jack was saying like you know you know what are the signs of trust in a human but this like plan A plan B thing from earlier is really important where like it might be that the 10 first 10 or 100 times you used the model it was you're asking a certain kind of question but it was like always in plan A zone and then you know you ask it a harder or a different question and the way that it tries to answer it is just like completely different. It's using a totally different set of strategies there like different mechanisms and that like that means that the trust it built with you was really your sort of trust with like model doing plan A's and now it's like doing plan B and like it's going to be completely off the rails but like you didn't have like any warning sign of that and so it's sort of I think we also just want to start building up an understanding of like how do models do these things so that we can form like a trust basis in some of those areas and I think like you can form trust with a system you don't completely understand, but you sort of like if it's just like Emanuel had a twin and then like one day like Emanuel's twin came to the office and like I didn't like I was like this seems like the same guy and then just did something completely different on the computer, right? Like that could go south depending on if it was the evil twin. — Yes, it did. Well, — or the good twin. — Well, yeah, obviously we have anyone here. — Oh, I thought you were going to ask me if I was the evil twin. — All right. Well, — I'm not going to answer that. — Yes. Mhm. — At the start of this discussion, I asked, you know, is a language model thinking like a human? Uh it'd be I'd be interested to hear an answer from all three of you the extent to which you think that's true. — Putting me on the spot with that one, but um I think it's uh it's thinking but not like a human. Uh but that's not a very useful answer. So maybe to dig in a little bit more. Um — well it seems like quite a profound thing to say that it's thinking right because again it's just predicting the next word. Some people think that these are just autocompletes but you're saying that it is actually thinking — I think. Yeah. So maybe to add something that we haven't touched on yet but I think is really important um for understanding actual experience of talking to language models is that like we're talking about predicting the next word. Um but what does that actually mean in the context of a dialogue that you're having with a language model? It's what what's really going on under the hood is that the language model is filling in a transcript between you and this like character that it's created. So in the in the like canon world of the language model, you are called human and you're it's like human colon the thing you wrote and then there's this character called the assistant and we've like trained the model to imbue the assistant with certain characteristics like being helpful and like smart and nice. Uh and then it's like simulating what this assistant character would say to you. Um so in a sense we really have like created the models in our image. we are literally training them to like cosplay as this sort of humanoid robot character. And so in that sense like well in order to predict what this like nice smart humanoid robot character would say in response to your question, what do you have to do if you're really good at that prediction task? You have to kind of form this internal model of like what that character is representing like what it's thinking so to speak. So like in order to do its task of predicting what the assistant would say, the language model kind of needs to form this model of the assistant's thought process. And I think like in that sense it like the just the claim that like language models are thinking is really just it's this very like functional claim of just in order to do their job of kind of like playing this character well, they need to sort of simulate the process whatever it is that we humans are doing when we're thinking. And it simulation is very likely quite different from how our brains work, but it's kind of trying it's like shooting towards the same goal. I think there's kind of an emotional part to this question or something when you ask are they thinking like us? It's like, — are we not that special or something? And I think that's been apparent to me discussing some of the math examples that we're talking about with people that have engaged with like read the paper or or different writeups, which is this example where, you know, we asked a model to say 36 + 59, what's the answer? And uh the model can correctly answer it. You can also ask it how well how did you how'd you do that? and it'll say, "Oh, you know, I added the six and the nine and then I carried the one and then uh I added all the sort of like tens digits. " But it turns out that if we look inside the brain, like we can that's not at all what it's doing. — It didn't do that. So again, it was bullshitting you did things. — That's right. Again, it was bullshitting you. What it actually does is actually this sort of kind of interesting mix of strategies where it's in parallel doing the tens digit and the ones digit and sort of doing sort of like a series of different steps. But the thing that's interesting here is that talking to people so like I think the reaction is split on like what does that mean? Uh and in a sense I think what's cool is some of this research is like free of opinion. We're just telling like this is what happened. You can you feel free to you know from that conclude that the model is thinking or is not thinking and half of the people will say like well you know it told you that it was carrying the one and it didn't and so clearly it doesn't even understand its own thought and so clearly it's not thinking. And then half of the other people will be like, well, you know, when you ask me 36 plus 59, I also kind of, you know, I know that it ends at five. I know that it's roughly like in the 80s or 90s. Uh, I have all of these heristics in my brain, as we were talking about, I'm not sure exactly how I comput it. I can write it out and compute it, you know, the longhand way, but the way that it's happening in my brain is sort of like fuzzy and weird. And it might be similarly fuzzy and weird to what's happening in that example. Humans are notoriously bad at metacognition like thinking about thinking and understanding their own thinking processes especially in cases where it's you know immediate reflexive answers. So why should we expect uh any different for models? Um Josh what's your answer to the question? — I like Emanuel I'm going to avoid the question and just sort of be like what why do you ask? I don't know. Sort of like asking like does a grenade punch like a human? Like no. Well there's some force. Yes. — Uh so you know and maybe there are things that are closer than that but like if you're worried about damage then I think understanding you know where does the impact come from? What is the impetus of this is maybe like the important thing. I think for me the like do models think in the sense that they like do some like integration and processing and sequential stuff that can lead to surprising places? Clearly yes. um it'd be kind of crazy from interacting with them a lot for there not to be something going on. We can sort of start to see how it's happening. Then the like humans bit is interesting because I think some of that is trying to ask like you know what can I expect from these because if it's sort of like me being good at this would make it good at that. But if it's like different from me then like I don't really know what to sort of look for. And so really we're just looking to like understand like where do we need to be extremely like suspicious or like starting from scratch in understanding this and where can we sort of just reason from like our own like very rich experience of thinking and there I feel a little bit trapped because as a human like I project my own image constantly onto everything like they warned us in the Bible where I'm just like this piece of silicon like it's just like me made in my image where like to some extent it's been trained to like simulate dialogue between people. So, it's going to be very like person-like in its affect. Um, and so some humanness will get into it simply from like the training, but then it's like using very different equipment that has like different limitations. And so, the way it does that might be pretty different. — To Emanuel's point, I think the Yeah, we're in this tricky spot answering questions like this because we don't really have the right language for talking about what language models do. It's like we're doing biology but you know before people figured out cells or DNA. I think we're starting to fill in that understanding like you know as Emanuel said there are these cases now where we can really just we can just if you just go read our paper like you'll know how the model like added these two numbers and then if you want to call it humanlike thinking or if you want to not then it's up to you but like the real answer is just like find the right language and the right abstractions for talking about the models but in the meantime when we when we've only currently we've only kind of like you know 20% succeeded at that scientific project Like to fill in the other 80%, we sort of have to borrow analogies from other fields. And like there's this question of which analogies are the most apt. Should we be thinking of the models like computer programs? Should we be thinking them of them like little people? And it seems to be like sometime like in some ways that think of them like little people is kind of useful. It's like if I like say mean things to the model, it like talks back at me. That's like what a human would do. But in some ways it's like that clearly not the right mental model. And so we're just kind of stuck, you know

figuring out when we should be borrowing which language. — Well, that leads on to the final question I was going to ask, which is what's next? What are the next uh pieces of scientific progress, biological progress that need to be made for us to have a better understanding of what's happening inside these models and uh again towards our mission of making them safer. — There's a lot of work to do. Um our our last publication has some like enormous section on the limitations of the way we've been looking at this that was also a road mapap to like making it better. You know we when we are looking for patterns to like decompose what's happening inside the model we're only getting sort of you know maybe a few percent of what's going on. Um there's large parts of how it moves information around that like we explicitly like didn't capture at all. Um they're scaling this up from the sort of small you know uh production model we use to like the cloud 3. 5 haiku. Right. — That's right. Which is like it's like a pretty capable model very fast but it's like by no means as sophisticated as you know the cloud 4 suite of models. Um so those are almost like sort of like technical challenges but I think Emanuel and Jackman takes on the like some of the like scientific challenges that come after solving those. — Yeah. Yeah, I mean I think maybe two things I'll say here which is one consequence of what Josha said is that you know uh out of the total number of times that we ask a question uh about how the model does X right now we can answer probably a small you know 10 to 20% of the time we can tell you after a little bit of investigation this is what's happening obviously we'd like that to be a lot better and there's a lot of kind of clearer ways to to get there and less and more speculative ways as well uh and then I think a thing that we've talked a lot about is this sort of idea that a lot of what the model does isn't simply like ah how is it saying the next thing we talked about it a little bit here it's sort of like planning a few things again and I a few words ahead sorry and I think we want to understand sort of like over a long conversation with the model sort of like how is its understanding of what's happening changing you know who it's talking to changing and how does that affect its behavior uh more and more sort of the actual use case of models like cloud is you know it's going to read a bunch of your documents and a bunch of like email you send or your code and based on that it's going to make one suggestion and so clearly there's something really important happening in that space where it's reading all these things uh and so I think understanding that better uh seems like a great challenge to take on — yeah I think we often use the analogy on the team of that we're building a microscope uh to like look at the model and right now we're in this exciting but also kind of frustrating space where our microscope works like 20% of the time and like to look looking through it is like requires a lot of skill uh and like takes you know you have to like build this whole big contraption and every like infrastructure is always breaking and then like once you've got your like explanation of what the model's doing you have to like throw like Emanuel or me or someone else on the team in a room for like two hours to like puzzle out what exactly was going on and like the really exciting future that I think we could be at within, you know, year or two years. You know, that kind of time scale is one where like just every interaction you have with the model can be under the microscope. like we can just anytime there's all these like weird things the models are doing and we just want it to be like push of a button like yeah you you're having your conversation you push a button you get this flowchart that tells you like what it was thinking about and once we're at that point it's it'll be this like I think our the interpretability team at Enthropic I think will start to kind of take on a bit of a different shape and that instead of this like team of kind of like engineers scientists thinking about the like math of how like language models work on the inside. We're going to have this like army of biologists that are just looking through the micros. We're just we're talking to Claude. We're getting it to do weird things and then we're just like we got people looking through the microscope seeing like what it was thinking on the inside. And I think that's kind of the future of of this work. — Nice. — Maybe two notes on top of that. One is like we want Claude to help us do all of that because like there's a lot of parts involved and you know who's like good at like looking at like hundreds of things and figuring out what's going on is like Claude. And so I think we're trying to enlist some help there. um especially as for these complicated contexts. And maybe the other place is like we've talked a lot about studying the model like once it's fully formed, but of course like we're at a company that makes these and so when it says okay here's how the model like solved this particular problem or said this thing. Where did that come from kind of in the training process? What are the steps that sort of like made that circuitry sort of form to do that and how could we give feedback to the rest of the company that is like doing all of that work to shape the thing that we like actually uh want it to become? — Well, thank you so much for the conversation. Where can people find out more about this research? — So, if you want to find out more, you can go to anthropic. com/ressearch which has our papers and blog posts and fun videos about it. Also, we recently partnered uh with another group called Neuronpedia to host some of these like circuit graphs we make. So, if you want to try your hand at looking at what's going on inside of a small model, you can go to Neuronipedia and see for yourself. — Thank you very much.