C (2:42)

Yeah, I think there's maybe like two to three categories of things that people could do. So I'll go from sort of like the most basic kind of low effort to, you know, hey, if you want to dedicate like a month of your life, you could do that. The sort of most basic thing is, you know, so it's a Gemma 2 and Llama 1B are like smaller models but they can still do a bunch of stuff. And for most of the things that they can do, we still kind of don't really know or have a good mental model of how it is that they do the things that they do. So to give you an example, one of the things in the paper is this sort of multi hop reasoning where we ask Claude 345 haiku like oh, the capital of the state where Dallas is Austin. It turns out that Gemma can do this also. And so as part of the release we have a notebook where Michael Hanna, one of the anthropic fellows kind of like walks through a bunch of examples, including this one. And it's really cool because you can see that actually the way the circuit looks in Gemma, like a really small model is extremely similar to the way that it looks in like a huge model. Which that in itself is I think like a pretty novel discovery. It's like, oh, you have these models that are like super different, you know, if you look at like their evals or if you just try to use them, they're like just very clearly different. But for this one task, for this one thing actually the way that they do this multi step reasoning is like the same. They actually do the multi step reasoning in the notebook. There's both other examples of kind of like fun things that we looked at that I think can sort of spike your interest if you're new to thinking about this stuff. And at the end of the notebook that's linked in the readme, there's like three examples of random sort of cases that we haven't solved or we haven't labeled that have a graph pre computed for you and you can just look at it and try to figure out what's happening. And by figure out what's happening? What we mean is we might do a quick demo here, but it's kind of like, look at these representations, try to understand, okay, what is the computation the model is doing. And then part of the release also lets you run experiments to verify that you write. So if you think that the model first thinks about Texas in this case, you can also just stop it from thinking about Texas and see if that damages it. And so the tools to do that are available. And so I would say that's the first thing is just I think the hope is there are a lot of behaviors that models do way more than any single group has time to explore. And so the hope is like, hey, pick a behavior that you think is interesting and try to understand what's happening and try to ground it out. And it's like the baseline thing and maybe the thing that I'm most excited about with this release. But then the other thing I do want to mention, parts two and three are just, we also hope that other groups and interested researchers can just use this to extend the method. If you have an idea about how to do this better, the whole code make this graph is open sources. Take a look at it and just try to play with it. Try to find different ways to create these graphs and also extend it to other models. Right. Like there are many different models. And so part of making this work on any model is you have to train the sort of replacement model, which again, there is code for it and there's other groups working on. And so that's also something that if you're excited about, you could say, like, okay, cool, well I want this to work on another open model and you could sort of add it if you're more into sort of maybe like the engineering and the ML engineering side of things.

A (5:50)

Yeah, we actually get into a little bit about how you guys do the extra data, viz, stuff that makes your blog post pop so much. Should we share the screen a little bit and dive in? I think you guys prepped some examples.

C (6:03)

Totally. Yeah.

A (6:04)

Yeah, it's just like there's nothing better than the creator of the tool walking through the tool and we might as well capture that so that, you know, people who actually want to do this can follow along.

C (6:13)

Yeah, that makes no sense. Let me just like actually share my screen.

B (6:16)

My one little experiment. I basically cloned the repo, threw it into cloud code and was like, you know, deal with this. Let's, let's try it end to end. So would recommend, you know, cloud code is very good at using this Basically also if you're just trying to get started, the circuit tracing tutorial notebook. Very good. That kind of goes over all the high level and then, you know, shout out cloud code, try it out. It works very well on this.

C (6:41)

That's awesome to hear. Actually, I might just open the notebook first, just like quickly walk through the illustrations. But yeah, you're the second person to tell me that they just had Claude code sort of like dig in initially on it. So I'm glad that's working. The tutorial here is like linked at the top of the repo and maybe we can link it from the podcast. But essentially sort of like walks you through kind of like how to think about graphs. And so it links to these circuits. So here, this is the two step reasoning that we were talking about. This is kind of like a schematic of it where it's like the capital of the state of Dallas and it's like, ah, it has to think of Texas and Austin. But the notebook links you to all of these circuits here. And this is kind of thing that you can play with. So this is the UI on, you know, Neuronpedia that, that hosts this and that, that lets you like create any circuit. So here, you know, we could explore the circuit and if you open the notebook you could explore it. I'm realizing that I switched tabs. Maybe I'm not sharing it. Okay, there we go. Can you see the circuit now? I think so. Okay, cool. But you can make a new graph super easily and quickly. And so maybe this is like the most fun things. When I was playing with right before, you know, joining this call is like, it turns out that podcast guests are very formulaic. And so if you say like, thanks for having me on the Whatever. Gemma seems to have pretty consistently guessed that you're on a podcast, which makes sense, right? Why would you say thanks for having me on the blah. And so here we can try to say, okay, how does Gemma know to complete the sentence with thanks for having me on the Latent Space podcast? And so here the way you generate a graph is you type a sentence where the next word is the thing that you're interested in. And then you try to explain how the model got to the next word. So here you can give it a name and then you can mostly just not worry about any of these parameters. I think if you're just playing with it and you can click start generation.

B (8:37)

And all this generates like something important for people to know is that these are trained on base models, right? So they're not chat models. So basically when you train these models, they're just trained to predict next token and they don't have that user assistant chatbot flow. So they're prompted in a way such that, you know, the output should basically just be the next word.

C (8:56)

Yeah, you kind of want to think about it as like maybe the. The. The prompt or the text you're giving is like the text of like a book or an article rather than a conversation, where it's like, you know, what is a sentence? Where if you were to read it in a book, like, the next word would be sort of like the interesting one. Yeah, you know, you can click on it. Sort of like takes a little bit of time to load because there's just a bunch of data. So what we're going to show you here is basically like almost every single feature that activates in the model. The features are these intermediate representations, and at the bottom there's the prompt. So here it's like, thanks for having me on the ligament space. And at the top you can see what the model sort of like output. So it's most likely output is. It's pretty confident that we're talking about a podcast. And it has some random stop tokens, blog show, and then some stuff that I think makes less sense. But also these are small models and so sometimes they say random stuff. And so the way that you could then explore this is be like, okay, so the model says podcast, so why does it say podcast? So you can click on this output and say, what are the features again? These intermediate representations that have an input to this. So it seems like there's features at here. This is the layer, like layer 18 that already, like, are about podcast episodes. You can know this because the features have a label. But also if you want, you can look at the feature itself here and here. You can see that, like, this shows you, like, other text of the feature is active over. And it's just like text about podcasts. So that's like a way that you can also like, understand what the features are and then you can keep going back. So it's like, oh, okay, so it said podcast here because of this podcast feature. Where did that come from? And it's like, oh, it comes from words related to podcasts, words associated with podcasts, as well as an interview feature. And also just the word on. So there's like a bias, like if you're saying like, blah, blah, blah on that sort of like slightly increases the chance that you're talking about a podcast at all. And you can sort of keep going back and kind of like explore the graph interactively. I would say that, like, the way to do it. And we talked about this on the, like, kind of, like, longer version of the podcast, but it's like, you know, kind of like chasing from the interesting outputs back or from the interesting input forward. There are many nodes on these. I like, wouldn't recommend looking at all of them. You can also sort of like, prune them a little more aggressively here if this is too busy and kind of look, this shows you, like, only the most important ones. And you can sort of like, be pretty extreme with it if you want, or you can show the whole thing and then be, like, super overwhelmed once you kind of do this. You can then kind of like group your nodes into similar ones to kind of like make a graph. I actually made this little summary earlier, so I can just share that. So this is the exact same graph, but just before the. Before hopping on, I kind of, like, did a few groups. So this is like, same thing podcast. It's like, oh, there's like a bunch of those that are like podcast episodes. There's a bunch of things like, discussing podcast. There's a note about expressing gratitude that amplifies that you're, like, on an interview or a podcast. So, like, one fun experiment you could do here, right, Is like, oh, what happens if I mess with this? If I don't, like, if I mess with the, like, oh, this person is, like, grateful to be on and is like, this person is on. Does it think you're on something else? Like, maybe there are things that, you know, you could be on that you're not grateful for. Like, oh, you're having me on trial or something. I don't know. Like, that could be one. One interesting sort of like, experiment to see what the causal effect of this is. And again, you could sort of, like, label it more and explore it more. And this ui, the whole point is for it to be, like, snappy and quick. So you can just like, generate a bunch of graphs pretty easily, right? Like, maybe this wasn't exactly what you wanted. So you're like, I'm super unhappy to be on the latent space. And then you can see what it completes for that or whatever. And you can sort of like just continuously play with it and get a better sense for your hypotheses. Oftentimes you kind of, like, want different prompts, you know, different examples that are similar to kind of get a sense for it. And then if you're really curious and you want to dig in more. That's when I would recommend going back to the code base and some of the notebooks, maybe. One last thing I'll say on that is that the notebooks themselves, they can all be run on Google Colab. And all of the code, as far as we can tell, we feel like testable notebooks, just runs on Colab. And so that means that you don't need. On the free tier, to be clear, you don't need an expensive gpu. You can just run this and run your interventions and play with it. So in this notebook, in particular, the intro one, we show you how to do these interventions. And here we're like, what happens if we turn this node off? And what happens if we turn that one off? And what happens if we turn this one off? And what happens if, you know, we inject one from one prompt into another one? And so I think that's the sort of, like, deeper dive, trying to understand the mechanism better. But if you're just trying to even, like, get a sense at all of, like, how does the model do X? You just generate a graph and take a look at it. Incredible.

B (13:50)

Very cool.

A (13:51)

When I look at the graph is, there's a thought in my mind about maybe this is too easy, too perfect. And one version of this is there's supposed to be superposition, and here there's no superposition kind of.

C (14:05)

Well, there is superposition. And we're sort of like. So maybe I can share the graph again and answer your question, which I think is like, what are we hiding here? Where are the skeletons?

A (14:16)

Yeah, this is like. It's too clean.

C (14:19)

I'm like, yeah, so, okay, so maybe, like, a good example is. And we're gonna, like, make this slightly less overwhelming here is like, okay, so you look at this graph and you say, like, yeah, like, we don't actually understand how models work fully. So, like, what are you hiding here? And the thing that's, like, important to know here is, you know, I didn't say this explicitly, but, like, the layers are, like, arranged here. And so let's just look at one layer. So for this layer, what we're saying is the only thing that is happening or that's important enough is this one feature, which is just, like, one small direction in the model space. Right? Like one dimension we've pulled out of superposition. Or let's say that for now. But then also there's these diamonds. And these diamonds are errors. We talked about them on the longer podcast. But they're just like, when you train these Replacement models to replace them. With the model computation, you successfully replace some of it, and then some of it you fail to replace. So this is like everything that we don't understand. And so that means that sometimes if you look at an input like this guy's input, you'll see a bunch of errors here as the input. And so essentially there's some graphs and some examples where if most of the stuff that you see is these errors, basically that just means like, hey, for this prompt, we were not able to sort of explain that part of the computation. And so at least that part is like a, an explicit sort of like we show it in your face, where it's like, here's what we don't understand. And so you can sort of like, see what we don't understand. There's also, I will say, like, one more thing. There's like a bunch more stuff that can get you. And that's like in the paper. But like, one example here that I'll just say is like, these are just MLPs. So the model has both attention heads and multilayer perceptrons. MLPs. We don't just do it like we completely ignore attention or like, we don't, we don't try to decompose it at all. So there are some prompts where like all of the interesting stuff is attention. And here you're just not, you're just not seeing it at all. The way that it's materialized is like, you have an edge from here to here and like some attention head did a bunch of stuff, you don't know what it is. And so that's also the part that we're sort of like not explaining. So there's, there's definitely. Yeah, I don't want to make the claim that we explain everything. I think the correct way to think about this is like, if you look at a prompt and you can by tracing through these, not hit any errors, hit nodes that make sense and build up a reasonable hypothesis. And then when you test it with interventions, it works. You've at least understood some and presumably like a reasonable proportion of the computation. If your interventions are working, that means that it's like the thing you found is not just like a side thing, it's part of the main thing that the model is doing. And so, you know, then the question is like, how often does that happen versus how often you just hit these errors or you're like confused. And I think that's just sort of like what works and what doesn't. Summary here Crazy.

A (17:06)

I mean, congrats on this work. I know you're low on sleep because you worked really hard on shipping it and you're a perfectionist. I just think, like.

C (17:14)

Yeah, sorry. I'll just say that the actual brunt of the work here is like, fellows. Yeah, I mostly just coordinated things left and right, but they sort of did all of the implementation as well as focused on the Neuronpedia decode research side. Also did the lion's share of the work here to actually have the front end ui.

A (17:34)

I'll just say, like, you know, Vibu and I were at the Goodfire meetup yesterday where there are a lot of interpretability folks. I was shocked at, honestly, how young most interpretability people and work are, and this is a very young field. Exactly. Like you say in the podcast, there's a lot of fresh green grass here to tread, and it's just really inspiring. Vibu, do you have any other final thoughts or comments?

B (18:02)

Yeah, no, I think there's just a lot of open work to be done, and we talk about this in the podcast, too. And just to reiterate how good the tooling that you guys put out is, even the fact that without diving into any code, you can enter in a prompt and start to play through these circuits in minutes, it's pretty incredible. I could share another one, actually. So I was doing this with Pomsky and I finally got it to work. So our guest host of the episode is Mochi, my little dog. She's our distilled husky. So she's on the podcast later and, you know, I basically put in, like, I had to guide it quite a bit, but my. My prompt is a. Pomsky is a small dog that's a breed of a mis. Of a husky. And a. And then, you know, I'm expecting it to put out Pomeranian or Pom. Let me. Let me share my screen real quick and then we can kind of dig through. This is me.

A (18:48)

Like, by the way, her tagline. Yeah. While you put it up, her tagline is officially Mochi, the interpretability Husky.

C (18:54)

The.

B (18:54)

Today for today, we're going to change.

A (18:56)

Our tagline every episode.

B (18:57)

But yeah, it feels a little weird. You know, we're digging deep into what Mochi is, but basically this is me, like, no background, like two minutes and just put in a phrase. And now I get to play around with features. Right. So this is also called please, with four S's, because, you know, I tried a few pumps. It's okay. It's okay. We struggle. It only took a Few minutes, though. So, you know, Pomsky is a small dog breed that. That's a mix of a husky and a. And then the most probable outputs, you know, now it says palm. So, okay, let's dig into what some of these. I'm basically just going like, fresh. Haven't done this before. But, you know, words related to animals, their emotions, their health. We have a feature for dog. Golden Lab mentioned dog breeds especially high maintenance. You know, this is basically like AGI. It knows palm scares are high maintenance. It's figured it out. But realistically, you know, as I dig through these features, I can start to pin them, layer them through mentions of garbage and waste. No, that's not nice. That's not nice. But basically, you know, and this is already me pruning out most of the features as I open it up. You know, it talks about different things, like dog breeding, what else related to animal welfare. So, like, and then you can dig through all this. There's just so many things that, like, you know, this is in a matter of minutes. I basically made a graph, put in a sentence, and now I have an output and I can traverse through what are different things. Okay, animal science, right? So this breed is relatively new. It's not that common that that big huskies and little Pomeranians naturally have offspring. But, you know, let's. Let's, like, dig through animal science versions of this, and then we have, like, interesting little features. So it's very easy for people to kind of get a different understanding of what goes on throughout layers in models, you know, but that's just my fun little experiment of getting it to work.

C (20:44)

Oh, yeah. And I think, like, you know, one thing that I would do if you were curious or maybe I'm just going to try to bait some listeners into doing it, is like, can be like, okay, like, let's try to, like, trace why it said Pomeranian here. And like, maybe there's like. Some of it is about like, like dog breeds, and some of it is about, like, specific characteristics of a husky. And then you can ask the same question, but instead of husky, like, try some other dog breed and then try to see if you can, like, if you understood the circuit well, and if you identified where it's thinking about huskies or where it's thinking about, like, kind of like breeding two different breeds, then you should be able to, like, swap these in and out and get it to kind of like, say whatever you want. And. And if you didn't, then maybe there's something complicated going on. But. But, yeah, like, very cool. That you got this going on so quick. That's, that's, that's the whole goal. That's super exciting. Yeah.

B (21:23)

And like, you know, no disclosure. This was like five minutes of just playing around and like there's, there's stuff to learn there, right? Like, okay, what happens with dog breeding? What are traits of these dogs? And then, you know, the next step for me would basically be let's try clamping some of these features up or down. Let's, let's do different breeds and see if it makes sense, right? So if I have husky traits and a different mix and then, you know, can I, can I get out what's going on? But it also shows internally that there's more than just token completion of, you know, this plus this equals this. No, it has some understanding of characteristics, right? Like this is a pretty stubborn dog. It has a stubborn feature pretty high up that activates. So very, very cool stuff. I think it'll be cool when we apply this to more like serious topics. Like right now when it comes to LLM evals, right, we have pretty straightforward evals, right? Like how good does it do on math? Can it write code, does stuff, compile? But we don't have like vibes based heuristic evals. So does it understand different queries should be concise? Should they be verbose? Can we kind of trace through how it gives responses to this stuff? And then the other part is, as we go past base models, how does this happen for different phases of models? Right, so if I have a base Gemma and I have a chat model, what are differences in their attributions? Right? What happens kind of in that diff of training? So that's kind of one of my little interests in macinturk. What happens as we do more training? What are we really changing?

C (22:49)

Totally, yeah. You can think about sort of like comparing different models. And for me, different models is either like Gemma versus some other model or like early Gemma versus late Gemma and pre training or like fine tuned versus not fine tuned. I think there's also a sense in which like somebody yesterday was telling me, like, oh, it's fun. I've been playing with it on the like sort of like weird riddles that the models get wrong. You're not limited to studying what the model can do, right? Like if the model's failing at something like, you know, counting the number of letters in strawberry or whatever, you could just try that and try to figure out the circuit for like, well, it's getting this wrong. Or like, why, like Maybe you can see in its representation that it's like thinking about something obviously incorrect. Right. And so I think that that's also like a fun thing to play with.

A (23:29)

I think that's it for our little intro chat and coverage of the open sourcing. Let's dive right into the episode next. But Emmanuel, you're amazing work and I'm so inspired and also just like, I think this puts a human face on the interpretability work. I think it's very important and we'd love to keep doing this, whatever you.

C (23:50)

Got next coming up. Well, yeah, thanks for having me again. I should say cool to put a face on it, but definitely want to call out. This is like a huge team of people with me.

A (23:58)

I'm just a talking head here and paper lead. You know, you did the work, you know, take credit.

C (24:05)

I think that like, yeah, happy to talk about more inter things and also like feel free to, you know, reach out to me. I'm like findable. If you're listening to this podcast and you have like questions about stuff that's broken or if this brings up like experiment ideas, definitely want more people playing with this. So, yeah, thanks for having me. Hope that inspires the folks.

A (24:22)

All right, we are back in the studio with a couple special guests. One Vibu, our guest co host for a couple of times now, as well as Mochi, the distilled husky is in the studio with us to ask some very pressing questions as well as Emmanuel. I didn't get your last name. Amazon.

C (24:38)

Yep.

A (24:39)

Is that Dutch? Is that.

C (24:40)

It's actually German.

A (24:41)

German, yeah. You are the lead author of a fair number of the recent macinterp work from Anthropic that I've been basically calling Transformer Circuits because that's the name of the publication.

C (24:53)

Yeah, well, to be clear, Transformer Circuits is the whole publication. I'm the author on one of the recent papers, Circuit Tracing.

A (24:59)

Yes. And people are very excited about that. The other name for it is Tracing the thoughts of LLMs. There's like three different names for this work, but it's all mechinterp.

C (25:07)

It's all mechinterp. There's two papers. One is Circuit Tracing. It's the methods. One is the biology, which is kind of what we found in the model. And then Tracing the Thoughts is confusingly just the name of the blog post.

A (25:19)

It's for different audiences. And I think though, when you produce the two minute polished video that you guys did, that's meant for a very wide audience.

C (25:28)

Yeah, that's right. There's sort of like very many levels of granularity at which you can go. And I think for a mech interpret in particular because it's kind of complicated going from top to bottom. Most like high level to serve the Denali details works pretty well.

A (25:41)

Yeah. Cool. We can get started. Basically we have two paths that you can choose. Like either your personal journey into mecanturp or the brief history of mechinterp just generally and maybe that might coincide a little bit.

C (25:55)

Okay, I could just give you my personal journey very quickly because then we can just do the second path. My personal journey is that I was working at Anthropic for a while. I'd been like many people just following Mechantwp as sort of like an interesting field with fascinating, often beautiful papers. And I was at the time working on fine tuning. So actually fine tuning production models for Anthropic. And eventually I got both my fascination reached a sufficient level that I decided I wanted to work on it. And also I got more excited about just as our models got better and better understanding how they worked. So that's the simple journey. I've got a background in ML, kind of like did a lot of applied ML stuff before and now I'm doing more research stuff.

A (26:38)

Yeah, you have a book with O'Reilly, you're head of the AI at Insight Data Science. Anything else to plug?

C (26:43)

Yeah, I actually I want to like plug the paper and unplug the book.

A (26:49)

Okay.

C (26:49)

I think the book is good. I think the advice stands the test of time. But it's very much like, hey, you're building like AI products. Which do you focus on? It's like very different, I guess is all I'll say from the stuff that we're talking to talk about today. Today is like research some of the deepest, weirdest things about how models work. And this book is you want to ship a random forest to do fraud classification. Here are the top five mistakes to.

A (27:11)

Avoid the good old days of ML.

C (27:14)

I know it was simple back then.

A (27:16)

You also transitioned into research and I think you also did that. I feel like there's this monolith of people assume you need a PhD for research. Maybe can you give that perspective of how do people get into research? How do you get into research? Maybe that gives you audience insight into vivo as well.

C (27:33)

Your background?

B (27:34)

Yeah, my background was in economics, data science. I thought LLMs were pretty interesting. I started out with some basic ML stuff and then I saw LLMs were starting to be a thing. So I just went out there and did it and same thing with AI engineering, you just kind of build stuff, you work on interesting things and now it's more accessible than ever. Back when I got into the field five, six years ago, pre training was still pretty new. GPT3 hadn't really launched, so it was still very early days and it was a lot less competitive. But yeah, without any specific background, no PhD, there just weren't as many people working on it. But you made the transition a little bit more recently, right? So what's your experience been like?

C (28:12)

Yeah, I think it has maybe never been easier in some ways because a lot of the field is like pretty empirical right now. So I think the bitter lesson is this lesson that you can just sort of a lot of times scale up compute and data and get better results than thinking than if you sort of thought extremely hard about a really good prior inspired by the human brain to train your model better. And so in terms of definitely research for pre training and fine tuning, I think it's just sort of like a lot of the bottlenecks are extremely good engineering and systems engineering. And a lot even of the research execution is just about sort of like engineering and scaling up and things like that. I think for Interp in particular, there's like another thing that makes it easier to transition to, which is maybe two things. One, you can just do it without huge access to compute. Like there are open source models, you can look at them. A lot of inter papers, you know, coming out of programs like maths are on models that are open source that you can sort of like dissect without having a cluster of like, you know, 100 GPUs. You can just even sometimes you can load them like on your CPU, on your MacBook. And it's also a relatively new field. And so, you know, there's as I'm sure we'll talk about, there's like some conceptual burdens and concepts that you just want to like understand before you contribute. But it's not, you know, physics, it's relatively recent. And so the number of abstractions that you have to like ramp up on is just not that high compared to other fields, which I think makes that transition somewhat easier for interp if you understand. We'll talk about all these, I'm sure, but what features are and what dictionary learning is, you're a long part of the way there.

A (29:53)

I think it's also interesting just on a career's point of view, research seems a lot more valuable than engineering. So I wonder, and you don't have to answer this if it's like a tricky thing. But how hard is it for a, for a research engineer in anthropic to jump the wall into research? People seem to move around a lot and I'm like, that cannot be so easy. Like in no other industry that I know of, people you can do that. Do you know what I mean?

C (30:23)

Yeah, I think I'd actually, I'd push back on the sort of like research being more valuable than engineering a little bit because I think a lot of times, like having the research idea is not the hardest part. Don't get me wrong, there's some ideas that are brilliant and hard to find, but what's hard certainly on fine tuning and to a certain extent on interp, is executing on your research idea in terms of making an experiment successfully, having your experiment run, interpreting it correctly. What that means though is that they're not separate skill sets. So if you have a cool idea, there's kind of not many people in the world, I think, where they can just have a cool idea and then they have a little minion they'll deputize, being like, here's my idea, go off for three months and run this whole. Build this model and train it for hundreds of hours and then report back on what happened. A lot of the time the people that are the most productive, they have an idea, but they're also extremely quick at checking their idea, finding the shortest path to checking their idea and a lot of that shortest path is engineering skills. Essentially it's just getting stuff done. And so I think that's why you see people move around is proportionate to your interest. If you're just able to quickly execute on the ideas you have and get results, then that's really the 90% of the value. And so you see a lot of transferable skills. Actually I think from people like, I've certainly seen adanthropics that are just like really good at that inner loop. They can apply it in one team and then move to a completely different domain and apply that inner loop just as well.

A (31:52)

Yeah, very cracked, as the kids say. Shall we move to the history of macinturp?

C (31:57)

Yeah.

A (31:58)

All I know is that everyone starts at Chris Ola's blog, is that right?

C (32:02)

Yeah, I think that's the correct answer. Chris Ola's blog. And then, you know, Distill Pub is the sort of natural next step. And then I would say, you know, now there's for anthropic there's transformer circuits which you talked about, but there's also just A lot of mech interp research out there from, you know, I think like the. Yeah, like maths is a group that regularly has a lot of research, but there's just many different labs that put research out there. And I think that's also just hammer home the point. That's because all you need is a model and then a willingness to investigate it, to be able to contribute to it. So now there's been a bit of a Cambrian explosion of mechan, which is cool. I guess the history of it is just computational models that are not decision trees. Models that are either CNNs or let's say Transformers have just this really strange property that they don't give you interpretable intermediate states by default. Again, to go back to, if you were training a decision tree on fraud data for an old school bank or something, then you can just look at your decision tree and be like, oh, it's learned that if you make, I don't know, if this transaction is more than $10,000 and it's for perfume, then maybe it's fraud or something. You can look at it and say, cool, that makes sense. I'm willing to ship that model. But for things like CNNs and transformers, we don't have that. What we have at the end of training is just a massive amount of weights that are connected somehow. Activations are connected by some weights, and who knows what these weights mean or what the intermediate activations mean. And so the quest is to understand that initially it was done, a lot of it was done on vision models where you sort of have the emergence of a lot of these ideas, like what are features, what are circuits? And then more recently it's been mostly, or not mostly applied to NLP models, but also still there's work in vision and there's work in bio and other domains.

A (34:02)

Yeah, I'm on Chris Ola's blog and he has the feature visualization stuff. I think for me the clearest was the vision work where you could have this layer detects edges, this layer detects textures, whatever. That seemed very clear to me. But the transition to language models seemed like a big leap.

C (34:18)

I think one of the bigger changes from vision to language models has to do with the superposition hypothesis, which maybe.

A (34:26)

Is like that's the first toy models post, right?

C (34:30)

Exactly. And this is sort of like. It turns out that if you look at just the neurons of a lot of vision models, you can see neurons that are curve detectors or that are edge detectors or that are high, low frequency detectors. And so you can sort of make sense of the neurons, mostly. But if you look at neurons in language models, most of them don't make sense. It's kind of like unclear why or it was unclear why that would be. And one main hypothesis here is the superposition hypothesis. So what does that mean? That means that language models pack a lot more in less space than vision models. So maybe a kind of really hand wavy analogy. Right. Is like, well, if you want curve detectors, you don't need that many curve detectors. If each curve detector is going to detect a quarter or twelfth of a circle. Okay, well, you have all your curve detectors, but think about all of the concepts that Claude or even GPT2 need to know. Just in terms of it needs to know about all of the different colors, all the different hours of every day, all of the different cities in the world, all of the different streets on every city. If you just enumerate all of the facts that a model knows, you're going to get a very, very long list. And that list is going to be way bigger than the number of neurons or even the size of the residual stream, which is where the models process information. And so there's this sense in which, oh, there's more information than there's dimensions to represent it. And that is much more true for language models than for vision models. And so because of that, when you look at a part of it, but it just seems like it's got all this stuff crammed into it. Whereas if you look at the vision models, oftentimes you could just be like, cool, this is the curve detector. Yeah.

A (36:12)

Vibu, you have some fun ways of explaining the toy models or superposition concept.

B (36:17)

Yeah, I mean, basically, if you have two neurons and they can represent five features, a lot of the early mechanic work says that there are more features than we have neurons.

A (36:26)

Right.

B (36:26)

So I guess my kind of question on this is, for those interested in getting into the field, what are the key terms that they should know? What are the few pieces that they should follow? Right. Like from the anthropic side, we had a toy transformer model. We first had auto encoders.

A (36:42)

That was the second paper.

C (36:43)

Right.

A (36:45)

Monosemanticity.

B (36:46)

What is sparsity in autoencoders? What are transcoders? What is linear probing? What are these kind of key points that we had in mechinterp? And, and just kind of how would people get a quick 0 to 80% of the field?

C (37:00)

Okay, so 0 to 80%. And now I realize I really set myself up for failure because I was like, yeah, it's easy. There's not that much to know. Okay, so then we should be able to cover it all. So superposition is the first thing you should know, right? This idea that there's a bunch of stuff crammed in a few dimensions, as you said, maybe you have two neurons and you want to represent five things. So if that's true, and if you want to understand how the model represents, I don't know, the concept of red, let's say, then you need some way to find out essentially in which direction the model stores it. So after the sort of supervision hypothesis, you can think of like, ah. We also think that basically the model represents these individual concepts, we're going to call them features, as like directions. So if you have two neurons, you can think of it as like, it's like the 2D plane and it's like you can have like five directions. And maybe you would arrange them like the spokes of a wheel so they're sort of maximally separate. It could mean that you have one concept this way and one concept that's not fully perpendicular to it, but pretty far from it. And then that would allow the model to represent more concepts than it has dimensions. And so if that's true, then what you want is you want a model that can extract these independent concepts. And ideally you want to do this automatically. Can we just have a model that tells us, oh, this direction is red, if you go that way, actually it's like, I don't know, chicken. And if you go that way, it's like the Declaration of Independence. And so that's what sparse autoencoders are.

A (38:27)

It's almost like the self supervised learning insight version. In pre training you had self supervised learning, and here now it's self supervised interpretability.

C (38:36)

Yeah, exactly, exactly. It's like an unsupervised method. And so unsupervised methods often still have labels in the end. Sometimes I feel like the term labels by masking. Yeah, for pre training it's like the next token. So in that sense you have a supervision signal. And here the supervision signal is simply you take the neurons and then you learn a model that's going to expand them into the actual number of concepts that you think there are in the model. So you have two neurons, you think there's five concepts. So you expand it a thing of dimension five and then you contract it back to what it was. That's like the model you're training, and then you're training it to incentivize it to be sparse so that there's only like a few features active at a time. And then once you do that, if it works, you have this sort of nice dictionary which you can think as like a way to decode deactivate the neurons where you're saying like, ah, cool. I don't know what this direction means, but I've like used my model and it's telling me that the model is writing in the red direction. And so that's sort of like. I think maybe the biggest thing to understand is this combination of things of like, ah, we have too few dimensions, we pack a lot into it. So we're going to learn an unsupervised way to like unpack it and then analyze what each of those dimensions that we've unpacked are.

A (39:49)

Any follow ups?

B (39:50)

Yeah, I mean the follow ups this are also kind of like some of the work that you did is in clamping. Right. What is the applicable side of mech interp. Right. So we saw that you guys have like great visualizations. Golden Gate Claude was a cool example.

C (40:03)

I was going to say that.

A (40:03)

Yeah, my favorite.

B (40:05)

What can we do once we find these features? Finding features is cool, but what can we do about it?

C (40:10)

Yeah, I think there's kind of like two big aspects of this. Like one is, yeah, okay, so we go from a state where, as I said, the model is like a mess of weights, we have no idea what's going on to okay, we found features. We found a feature for red, a feature for Golden Gate Claude or for the Golden Gate Bridge. I should say, like, what do we do with them? And well, if these are true features, that means that they in some sense are important for the model or it wouldn't be representing it. If the model is bothering to write in the Golden Gate Bridge direction, it's usually because it's going to talk about the Golden Gate Bridge. And so that means that if that's true, then you can set that feature to 0 or artificially set to 100 and you'll change model behavior. That's what we did when we did Golden Gate Claude in which we found a feature that represents the direction for the Golden Gate Bridge. And then we just like set it to always be on. And then you could talk to Claude and be like, hey, like, what's on your mind? You know, like, what are you thinking about today? Be like the Golden Gate Bridge. He'd be like, hey Claude, like what's two plus two? He'd be like four Golden Gate Bridges, et cetera.

B (41:10)

Right.

C (41:11)

And it was always thinking about The.

A (41:13)

It's like, write a poem and it starts talking about how it's like, read like the Golden Gate plot.

C (41:17)

That's right.

A (41:18)

Golden Gate Bridge.

C (41:18)

Yeah, that's right.

A (41:19)

Amazing.

C (41:19)

I think what made it even better is, like, we realized later on that it wasn't really like a Golden Gate Bridge feature. It was like being in awe at the beauty of the majestic Golden Gate Bridge. So on top of it would really ham it up. He'd be like, oh, I'm just thinking about the beautiful international orange color of the Golden Gate Bridge. That was just an example that I think was really striking, but of sort of like, oh, if you found a space where that represents some computation or some representation of the model, that means that you can artificially suppress or promote it. And that means that, like, you're starting to understand at a very high level, very gross level, how some of them all work. Right. We've gone from like, I don't know anything about it, to like, oh, I know that this combination of neurons is this. And I'm going to prove it to you. The next step, which is what this works on is that's kind of like thinking of if maybe you take the analogy of like, I don't know, let's take the analogy of like an MRI or something, like a brain scan. It tells you like, oh, this. As Claude was answering, at some point it thought about this thing, but it's this sort of vague. Basically, maybe it's a bag of words. It's like a bag of features. You just have, like, here are all the random things it thought about. But what you might want to know is like, okay, but Claude is doing some processing. Like, sometimes to get to the Golden Gate Bridge. It had to realize that you were talking about San Francisco and about the best way to go to Sonoma or something. And so that's how it got to Golden Gate Bridge. So there's like an algorithm that leads to it at some point, thinking about the Golden Gate Bridge. And basically there's a way to connect features to say, oh, from this input went to these few features and then these few features and then these few features, and then that one influenced this one. And then you got to the output. And so that's the second part. And the part we worked on is like, you have the features, now connect them in what we call or what's called circuits, which is sort of like explaining the algorithm. Yeah.

A (43:04)

Before we move directly onto your work, I just want to give a shout out to Neil Nanda. He did Neuronpedia and released a bunch of SAEs for I think the Llama models. And the Gemma models.

C (43:14)

And the Gemma models. Yeah.

A (43:15)

So I actually made Golden Gate Gemma just upped the weights for proper nouns and names of places of people and references to the term golden, likely relating to awards, honors or special names and that together made Golden Gate.

C (43:28)

That's amazing. Yeah.

A (43:31)

So you can make Goldengate Gemma. And I think that's a fun way to experiment with this. But yeah, we can move on to.

B (43:36)

I'm curious, I'm curious, what's the background behind why you ship Golden Gate Claude? Like you had so many features. Just any fun story behind why that's the one that made it.

C (43:46)

You know, it's funny if you look at the paper, there's just a bunch of like. Yeah, like really interesting features. Right. There's like one of my favorite ones was the sycophantic praise, which I guess is very topical right now.

A (43:58)

Very topical.

C (43:58)

But you know, it's like you could dial that up and like Claude would just really praise you. You'd be like, oh, you know, like I wrote this poem like roses are red, violets are blue, whatever. And he'd be like, that's the best poem I've ever seen. And so we could have shipped that. That could have been funny. Golden Gate Claude was like a pure, as far as I remember, at least like a pure, just like weird random thing where like somebody found it. Initially we had an internal demo of it. Everybody thought it was hilarious. And then that's sort of how it came out. There was no. Nobody had a list of top 10 features we should consider shipping and we picked that one. It was just kind of like a very organic moment.

A (44:36)

No, the marketing team really leaned into it. They mailed out pieces of Golden Gate for people in Europe's I think, or icml. Yeah, it was fantastic marketing. The question obviously is if OpenAI had invested more interpretability, would they have caught the GPT 4.0 update? But we don't know that for sure because they have interp teams. They just is.

C (44:57)

Yeah, I think also like for that one, I don't know that you need interp. Like it was pretty clear cut to the model. I was like, oh, that model is really gassing me up.

A (45:03)

And then the other thing is, can you just like up write good code, don't write bad code and make Sonder 3.5 and like it feels too easy, too free. Is that steering that powerful that you can just like up and down features with no trade offs?

B (45:19)

There's like A phase where people are basically saying, you know, 3.5 and 3.7 are just now because they came out right after.

A (45:25)

And for the record, like that's been debunked.

B (45:27)

Yeah, it has been debunked, but you know, it had people convinced that what people did is they basically just steered up and stared down features. And now we have a better model. And this kind of goes back to that original question of right, like, why do we do this? What can we do? Some people are like, I want tracing from a sense of, you know, legality, like what did the model think when it came to this output? Some people want to turn hallucination down, some people want to turn coding up.

C (45:50)

So.

B (45:52)

Whether it's internal, what are you exploring that? What are the applications of this? Whether it's open ended of what people can do about this or just why do mechinterp?

C (46:01)

Yeah, there's a few things here. First of all, obviously this is, I would say on the scale of the most short term to the most long term, pretty long term research. So in terms of applications compared to the research work we do on fine tuning or whatever, interp is much more sort of like a high risk, high reward kind of approach. With that being said, I think there's just a fundamental sense in which Michael Nielsen had a post recently about how knowledge is dual use or something. But just knowing how the model works at all feels useful. And it's hard to argue that if we know how the model works and understand all of the components, that won't help us make models that hallucinate less, for example, or that are less biased. That seems at the limit. Yeah, that totally seems like something you could do using basically your understanding of the model to improve it. I think for now, as we can talk about a little bit with circuits, we're still pretty early on in the game. And so right now the main way that we're using interpretability is to investigate specific behaviors and understand them and gain a sense for what's causing them. So one example that we can talk about later, we can talk about now, but in the paper we investigate jailbreaks and we try to see why does a jailbreak work. And then we realize as we're looking at this jailbreak, that part of the reason why Claude is telling you how to make a bomb in this case is that it's already started to tell you how to make a bomb. And it would really love to stop telling you how to make a bomb, but it has to first finish its sentence. It really wants to make correct grammatical sentences. And so it turns out that seeing that circuit, we were like, ah, then does that mean if I prevent it from finishing its sentence, the jailbreak works even better? And sure enough, it does. And so I think the level of sort of practical application right now is of that shape. So understanding either quirks of a current model or how it does tasks that maybe we don't even know how it does it. We have some planning examples where we had no idea it was planning and we were like, oh God, it is. That's sort of like the current state we're at.

B (48:06)

I'm curious internally how this kind of feeds back into the research, the architecture, the pre training teams, the post training. Is there a good feedback loop there right now there's a lot of external people interested. We'll train an SAE on one layer of Llama and probe around. But then people are like, okay, how does this have much impact? People like clamping. But yeah, as you said, once you start to understand these models have this early planning and stuff, how does this kind of feedback.

C (48:32)

I don't know that there's much to say here other than I think we're definitely interested in conversely making models for which it's easier to interpret them. So that's also something that you can imagine sort of working on, which is like making models where you have to work less hard to try to understand what they're doing. Architecture.

A (48:50)

Okay. Yeah, so I think there was a less wrong post about this of like, there's a non zero amount of sacrifice you should make in current capabilities in order to actually make them more interpretable, because otherwise they will never catch up.

C (49:02)

There's this sort of sense in which like right now we take the model and then the model's a model and then we post hoc do these replacement layers to try to understand it. But of course, when we do that, we don't fully capture everything that's happening inside the model. We're capturing a subset. And so maybe some of it is like you could train a model that's easier to interpret naively. And it's possible that you don't even have that much of attacks in that sense. And you can just sort of either train your model differently or do a little post hoc step to untangle some of the mess that you've made when you trained your model, make it easier to interpret.

A (49:34)

The hope was pruning would do some of that, but I feel like that area of research has just died.

C (49:40)

What kind of pruning are you thinking of?

A (49:41)

Here, just pruning your network.

C (49:43)

Ah, yeah.

A (49:44)

Pruning layers, pruning connections, whatever.

C (49:48)

Yeah. I feel like maybe this is something where, like, superposition makes me less hopeful.

A (49:53)

Or something, because you don't know, like, that. That, like, seventh bit might hold something.

C (49:58)

Well, right. And it's like, on. On each example, maybe this neuron is, like, at the bottom of, like, like what matters, but actually it's participating 5% to understanding English, doing integrals and whatever, cracking codes or something. And it's like, because that represents distributed over it, you kind of like, when you naively prune, you might miss that. I don't know.

A (50:21)

Okay. And then this area of research, in terms of creating models that are easier to interpret from the start, is there a name for this field of research?

C (50:29)

I don't. Don't think so. And I think this is, like, very early, and it's mostly like a dream.

A (50:33)

There's a thing people want to double click on. Yeah, I haven't come across it.

B (50:36)

I think the higher level. It's like, Dario recently put out a post about this. Right. Why Mech interpret is so important. We don't want to fall behind. We want to be able to interpret models and understand what's going on. Even though capabilities are getting so good, it kind of ties into this topic. Right. We want models to be slightly easier to interpret so we don't fall behind so far.

C (50:58)

Well, yeah. And I think here, like, just to talk about the elephant in the room or something, like, like, one big concern here is, is, like, safety, right? And so, like, as models get better, they are going to be used more and more places. You know, it's like, you're not gonna have your, you know, we're vibe coding right now. Maybe at some point. Well, that. That'll just be coding. It's like, Cloud's gonna write your code for you, and that's it. And Cloud's gonna review the code that Cloud wrote, and then Cloud's gonna deploy to production. And at some point, like, as these models get integrated deeper and deeper into more and more workflows, it gets just scarier and scarier to know nothing about them. And so you kind of want your ability to understand the model to scale with, like, how good the model is doing, which that itself kind of tends to scale with, like, how widely deployed it is. So as we deploy them everywhere, we want to understand them better.

A (51:43)

The version that I liked from the old superalignment team was weak to strong generalization or weak to strong alignment, which that's what super Alignment to me was. And that was my first aha moment of like. Like, oh, yeah. At some point, these things will be smarter than us in many ways. They already are smarter than us, and we rely on them more and more. We need to figure out how to control them. And this is not an Eliezer Yakovsky thing. It's just more like, we don't know how these things work. How can we use them?

C (52:10)

Yeah. And you can think of it as there's many ways to solve a problem. And some of them, if the model is solving it in a dumb way or in memorized one approach to do it, it, then you shouldn't deploy it to do a general thing. You could look at how it does math, and based on your understanding of how it does math, you're like, okay, I feel comfortable using this as a calculator. Or no, it should always use a calculator tool because it's doing math in a stupid way. And extend that to any behavior where it's just a matter of. Think about it. If you're in the 1500s and I give you a car or something, and I'm just like, cool. This thing. When you press on this, it accelerates. When you press on that, it stops. This steering wheel seems to be doing stuff.

A (52:49)

Stuff.

C (52:49)

But you knew nothing about it. I don't know if it was like a super faulty car. And it's like, oh, yeah. But if you ever went above 60 miles an hour, it explodes or something. You probably would be sort of like, you'd want to understand the nature of the object before jumping in it. And so that's why we understand how cars work very well, because we make them. LLMs are sort of like. And ML models in general are like this very rare artifact where we make them, but we have no idea how they work.

A (53:12)

We evolve them. We create conditions for them to evolve.

C (53:15)

And then they evolve.

A (53:15)

And. And we're like, cool, maybe we got a good run, maybe we didn't. Don't really know.

C (53:22)

Yeah, the extent to which you know how it works is you have your eval and you're like, oh, well, seems to be doing well on this eval. And then you're like, is this because this was in a training set or is it actually generalizing? I don't know.

A (53:31)

My favorite example was somehow C4, the common colossal clean corpus, did much better than Common Crawl, even though it filtered out most of the this. Like, it was very prudish. So it, like, filters out anything that could be considered obscene, including the word gay. But, like, Somehow it just, like, when you add it into the data mix, it just does super well. And it's just like this magic incantation of, like, this recipe works. Just trust us. We've tried everything. This one works. So just go with it. Yeah, that's not very satisfying.

C (54:03)

No, it's not. The side that you're talking about, which is like, okay, like, how do you make these? And it's kind of unsatisfying that you just kind of make the soup soup. And you're like, oh, well, you know, my grandpa made the soup with these ingredients. I don't know why, but I just make the soup the way my grandpa said. And then like, one day somebody added, you know, cilantro. And since then, we've been adding cilantro for generations. And you're like, this is kind of crazy.

A (54:22)

That's exactly how we train models, though.

C (54:23)

Yeah. Yeah. So I think there's like a part where it's like, okay, like, let's try to unpack what's happening. You know, like, the mechanisms of learning, like, how. How our models learn. Like, one of them, I guess we skipped over it. But, like, one of the things were like, induction heads, you know, like understanding what induction heads are, which are attention heads that allow you to look at in your context. The last time that something was mentioned and then repeat it is like something that happens that seems to happen in every model. It's like, oh, okay, that makes sense. That's how the model is able to repeat text without dedicating too much capacity to it.

A (54:53)

Let's get it on screen so people can see.

B (54:55)

The visuals of the work you guys put out is amazing.

A (54:58)

Oh, yeah. We should.

B (54:59)

Highly, highly, highly.

A (54:59)

We should talk a little bit about the behind the scenes of that kind of stuff. But let's finish this offer totally.

C (55:04)

But just really quickly. I don't think we should spend too long on it. I think it's just like, if you're interested in mech interp, we talked about superposition, and I think we skipped over induction heads. And that's kind of like a really neat, basically, pattern that emerges in many, many transformers where essentially they just learn. One of the things that you need to do to predict text well is that if there's repeated text, at some point somebody said, emmanuel Mason, and then you're on the next line and they say, emmanuel. Very good chance it's the same last name. And so one of the first things that models learn is just like, okay, I'm just going to, like, look at what was said before And I'm going to say the same thing and that's induction heads, which is like a pair of attention heads that just basically look at the last time something was said, look at what happened after. Move that over. And that's an example of a mechanism where it's like, cool, now we understand that pretty well. There's been a lot of follow up research on understanding better. Like, okay, in which context do they turn on? There's different levels of abstraction. There's induction heads that literally copy the word and there's some that copy the sentiment and other aspects. But I think it's just like an example of slowly unpacking or peeling back the layers of the onion of like what's going on inside this model. Okay. This is a component, it's doing this.

A (56:05)

So induction headers was like the first major finding.

C (56:08)

It was a big finding for NLP models for sure.

A (56:11)

I often think about the edit models. So Claude has a fast edit mode. I forget what it's called. OpenAI has one as well and you need very good copying. Every area that needs copying and then you need it to switch out of copy mode when you need to start generating. And that is basically the productionized version of this.

C (56:29)

Yeah, yeah. And it turns out that you need to select a model that's smart enough to know when it needs to get out of copy mode.

A (56:36)

Right.

C (56:36)

Which is like, it's fascinating.

A (56:37)

It's faster, it's cheaper. You know, as bullish as I am on canvas, basically every AI product needs to iterate on a central artifact. And if it's code, if it's, it's a piece of writing, doesn't really matter. But you need that copy capability that's smart enough to know when to turn it off.

C (56:55)

That's why it's cool that induction heads are at different levels of abstraction. Sometimes you need to editing some code, you need to copy the general structure. It's like, oh, this other function that's similar. It first takes, I don't know, abstract class and then it takes an int. So I need to copy the general idea. But it's going to be a different abstract class and a different int or something.

B (57:14)

Yeah.

A (57:15)

So tracing.

C (57:16)

Oh yeah. Should we jump to circuit tracing? Sure.

A (57:18)

I don't know if there's anything else you want to cover.

C (57:20)

No, no, no.

A (57:20)

We have space for it maybe.

C (57:22)

Okay, I'll do like a really quick TLDR of these two recent papers. Insanely quick. So we talked about these features that we detect and what we said is like, okay, but we'd like to connect the features to understand like the inputs to every features and the outputs to every features and basically draw a graph. And this is like if I'm still sharing my screen, the thing on the right here where that's the dream we want for a given prompt. What were all of the things, all of the important things happen in the model. And here it's like, okay, it took in these four tokens, those activated these features, these features activate these other features and then these features activate these other features and then all of these promoted the output. And that's the story. And basically the work is to use dictionary running and these replacement models to provide an explanation of sets of features that explain behavior. So this is super abstract. So I think immediately maybe we can just look at one example. I can show you one which is this one.

A (58:19)

The reasoning one.

C (58:20)

Yep. Yeah, two step reasoning. I think this is already, this is like the introduction example, but it's already kind of fun. So the question is, you ask the model something that requires it to take a step of reasoning in its head. So you say fact, the capital of the state containing Dallas is. So to answer that you need one intermediate step, right? You need to say, wait, where's Dallas is in Texas. Okay, cool. Capital Texas. Austin. And this is like in one token, right? It's gonna after is it's gonna say Austin. And so like in that one forward pass the model needs to extract to realize that you're asking it for the capital of a state to look up the state for Dallas, which is Texas, and then to say Austin. And sure enough, this is like what we see is we see like in this forward pass there's a rich sort of like inner set of representations where it gets capital state in Dallas and then boom, it has an inner representation for Texas. And then that/ capital leads it to say Austin.

B (59:13)

I guess one of the things here is we can see this internal thinking step, right? But a lot of what people say is is this just memorized fact, right? I'm sure a lot of the pre training that this model is trained on is this sentence shows up pretty often, right? So this shows that. No, actually internally throughout we do see that there is this middle step, right?

A (59:33)

It's not just memorize, you can prove that it generalized.

C (59:37)

Yeah, so that's exactly right. And I think you hit the nail on the head which is like this is what this example is about. It's like, ah, if this was just memorized, you wouldn't need to have an intermediate step at all. You'd Just be like, well, I've seen the sentence. I know it comes back right. But here there is an intermediate step. And so you could say like, okay, well, maybe it just has the step, but it's memorized it anyways. And then the way to verify that is kind of like what we do later in the paper and for all of our examples is like, okay, we claim that this is like the Texas representation. Let's get another one and replace it. And we just change that feature in the middle of the model and we change it to California. And if you change it to California, sure enough, it says Sacramento. And so it's like, this is not just a byproduct. Like it's memorized something. And on the side it's thinking about Texas. It's like, no, no, no. This is like a step in the reasoning. If you change that intermediate step, it changes the answer.

B (60:27)

Very, very cool work. Underappreciated. Yeah, okay, sure.

A (60:31)

I have never really doubted. I think there's a lot of people that are always criticizing LLMs as stochastic parrots. This pretty much disproves it already. We can move on.

C (60:42)

Yeah, I think there's a lot of examples that I will say. We can go through a few of them and guess I like, show an amount of depth in the intermediate states of the model that makes you think like, oh, gosh, it's doing a lot. I think maybe the poems. Well, definitely the poems. But even for this one, I'm going to scroll in this very short paper to medical diagnoses.

A (61:05)

I don't even know the word count because there's so many embedded things in there.

C (61:08)

Yeah, it's too dangerous. We can't look it up. It overflows.

A (61:13)

It's so beautiful. Look at this.

C (61:15)

This is like a medical example that I think shows you. Again, this is in one forward pass, the model is given a bunch of symptoms, and then it's asked not like, hey, what is the disease that this person has? It's asked, if you could run one more test to determine it, what would it be? So it's even harder. It means you need to take all the symptoms. Then you need to have a few hypotheses about what the disease could be. And then based on your hypotheses, say, well, the thing that would be the right test to do is X. And here you can see these three layers where it's like, like, again, in one forward pass, it has a bunch of like, oh, these are symptoms. Then it has the most likely diagnosis here. Then, like, an alternate one. And then based on the diagnosis, it gives you basically a bunch of things that you could ask. And again, we do the same experiments where you can kill this feature here, like suppress it. And then it asks you a question about the second option it had. The reason I show it is like, man, that's like a lot of stuff going on, like, for one forward pass, right? It's like, specifically if you expected it to like, oh, what it's going to do is it's just seeing similar cases in the training. It's going to kind of vibe and be like, oh, I guess there's that word. And it's going to say something that's related to, I don't know, headache or kind of really hard. It's like, no, no, no. It's like activating many different distributed representations, combining them and sort of like doing something pretty complicated. And so, yeah, I think it's funny because in my opinion that's like, yeah, like, oh, God, stochastic parrots is not something that I think is like appropriate here. And I think there's just a lot of different things going on and there's pretty complex behavior at the same time. I think it's in the eye of the beholder. I think. I've talked to folks that have read this paper and I've been like, oh, yeah, this is just a bunch of kind of heuristics that are mashed together. The model is just doing a bunch of kind of like, oh, if high blood pressure, then this or that. And so I think there's sort of like an underlying question that's interesting, which is like, okay, now we know a little bit of how it works. This is how it works. Like, like, now you tell me if you think that's impressive, if you think that, if you trust it, if you think that's sort of something that is sufficient to ask it for medical questions or whatever.

A (63:14)

I think it's a way to adversarially improve the model quality. Because once you can do this, you can reverse engineer what would be a sequence of words that to a human makes no sense or lets you arrive at the complete opposite conclusion. But the model still gets tripped up by. Yeah, and then you can just improve it from there.

C (63:33)

Exactly. And this gives you a hypothesis about specifically imagine if one of those was actually the wrong symptom or something. You'd be like, oh, it's weird that the liver condition upweighs this other example. That doesn't make sense. Okay, let's fix that in particular. Exactly. You sort of have a bit of insight into how the model is getting to its conclusion. And so you can see both is it making errors, but also is it using the kind of reasoning that will lead it to errors?

A (63:59)

There's a thesis, I mean now it's very prominent with the reasoning models about model depth. So you're doing all this in one pass, but maybe you don't need to because you can do more passes. Sure. And so people want shallow models for speed, but you need model depth for this kind of thinking. Is there a Pareto frontier? Is there a direct trade off?

C (64:23)

Yeah.

A (64:24)

Would you prefer if you had to make a model and shallow versus deep?

C (64:28)

There's a chain of thought faithfulness example. Before I show it, I'm just going to go back to the top here. So when the model is sampling many tokens, if you want that to be your model, you need to be able to trust every token it samples. So the problem with models being autoregressive is that if they at some point sample a mistake, then they kind of keep going conditioned on that mistake. And so sometimes you need backspace tokens or whatever. Yeah. And error correction is notably hard. If you have a deeper model, maybe you have fewer cot steps, but your steps are more likely to be robust or correct or something. And so I think that's one way to look at the trade off. To be clear, I don't have an answer. I don't know if I want a wide or a shallow or a deep model.

A (65:12)

You definitely want shallow for inference speed.

C (65:14)

Sure, sure, sure, sure. But you're trading that off for something else. Right. Because you also want a 1B model for inference speed. But that also comes at a cost. Right. It's less smart.

B (65:22)

There's a cool quick paper to plug that we just covered on the paper club. It's a survey paper around when do you use reasoning models versus dense models? What's the trade off? I think it's the economy of reasoning economy. Reasoning the reasoning economy. So they just go over a bunch of ways to measure this. Benchmarks around when to use each because yeah, we don't want to. Also, consumers are now paying the cost of this. Right.

A (65:44)

But little side note, for those on YouTube, we have a secondary channel called leanspace TV where we cover that stuff.

C (65:51)

Nice.

A (65:51)

That's our paper club. We covered your paper. Cool.

C (65:54)

Yeah. I think you brought up the planning thing. Maybe it's worth.

A (65:57)

Let's do it.

C (65:58)

Yeah. I think this one is like if you think about.

A (66:00)

Okay, so you're going into the chain.

C (66:02)

Of Thought faithfulness one, let's give this one, let's just do planning. So if you think about common questions you have about models, the first one we kind of asked was like, okay, is it just doing this vibe based one shot, pattern matching based on existing data or does it have kind of rich inner representations? It seems to have these intermediate representations that make sense as the abstractions that you would reason through. Okay, so that's one thing. And there's a bunch of examples. We talked about the medical diagnoses. There's like the multilingual circuits is another one that I think is cool where it's like, oh, it's sharing representations across languages. Another thing that you'll hear people mention about language models which is that they're like next token predictors.

B (66:37)

Also for a quick note for people that won't dive into this super long blog post, I know you highlighted like 10 to 12, so for like a quick 15, 30 second. What do you mean by their sharing thoughts throughout? Just like, what's a really quick high level just for people to.

C (66:51)

Yeah, the really quick high level, high level. Is that what we find is that here I'm just going to show you a really quick inside the model. If you look at the inner representations for concepts, you can ask the same question which I think in the paper the original one we asked is the opposite of ha is cold. But you can do this over a larger data set and ask the same question in many different languages. And then look at these representations in the middle of the model and ask yourself, well, well, when you ask it, the opposite of hot is and le contraire de show et which is the same sentence in French, is it using the same features or is it learning independently for each language? It kind of would be bad news if it learned independently for each language because then that means that as you're pre training or fine tuning, you have to relearn everything from scratch. So you would expect a better model to kind of share some concepts between the languages. It's learning. Right? And here we do it for language languages. But I think you could argue that you'd expect the same thing for programming languages where it's like, oh, if you learn what an if statement is in Python, maybe it'd be nice if you could generalize that to Java or whatever. And here we find that basically you see exactly that. Here we show like if you look inside the model, if you look at the middle of the model, which is the middle of this plot here, models share more features, they share more of these Representations in the middle of the model and bigger models share even more. And so the sort of smarter models use more shared representations than the dumber models, which might explain part of the reason why they're smarter. And so this was sort of this other finding of like, oh, not only is it having these rich representations in the middle, it learns to not have redundant representations. Like if you've learned the concept of heat, you don't need to learn the concept of French heat and Japanese heat and Colombian. That's just the concept of heat. And you can share that among different languages.

B (68:35)

I feel like sometimes overanalyzing this becomes a bit of a problem. Right. Like when we talked about with the medical example, we could look back and try to fix this in dataset. So in language, I don't remember if it was OpenAI or anthropic, where they basically said when the model switched languages and they pass it to fluent users, they said, oh, this feels like an American that's speaking this language. Right. So at some times there are nuances in a slightly different representation. Right. So you don't want to over engineer these little fixes when you do see them. But then the other side of this is like for those tail end of languages. Right. For languages that, that models aren't good at and for those, when you want to kind of solve that last bit, it seems like it's pretty plausible that we can solve this because these concepts can be shared across languages as long as we can fill in some level of representation. Unless I'm wrong.

C (69:28)

No, totally. And I think this sort of stuff also explains language models are really good at in context learning. You give them something completely new and they do a good job. It's like, well, if you give them a new fake language and you in that language explain that cold means this and hot means that. Presumably they're able to. To be clear, this is speculation. We don't show in the paper, but they're able to bind it.

A (69:48)

Google's done this.

C (69:49)

Okay, great.

A (69:50)

Yeah, they took a low resource language, dumped it in a million token contexts and then it came up.

C (69:53)

That's right. That's right. Well, I guess the thing that'd be curious to see is like, okay, does it reuse these representations? I bet that it probably does. Right. And that's probably like a reason why it works well is like, well, it can reuse the representation, the general representations that it's learned in other languages.

A (70:07)

Yeah, this is like have talk to any linguistics people not recently. Linguistics researchers will be very interested in this because Ultimately, this is the ultimate test of Sapir Whorf, which are you familiar with? So for those who don't know, it's basically the idea that the language that you speak influences the way you think, which obviously directly maps onto here. If it's a complete mapping, if every language maps every concept perfectly on in the theoretical infinitely sized model, then superior WARF is false because there is a universal truth. If it does not, if there is some overlap where, for example, there's some languages that have no word, there's this joke where Eskimos have no word for snow or something like that. Right. Or water has no word, fish have no word for water, there's an African language where there's a gender for vegetables, stuff like that. Just like languages influence the way you think. And so there should not be 100% overlap at some point.

C (71:00)

Of course, it's like at the limit of the infinite model. So who knows if we'll ever. But yeah, yeah, well, and I think it's interesting we also show a little below that some people have made the point of the bias. Oh, it sounds like an American speaking a different language. And it does seem like the sort of inner representations have a higher connection to the output logits for English logits. And so there's some bias towards English, at least in the model we studied here.

A (71:25)

Any thoughts as to whether multimodality influences any of this? So, like, concepts, do they map across languages as they do across modalities?

C (71:32)

Yeah, so we show this in the Golden Gate or like the previous paper, I might have it here actually for you.

B (71:37)

There's a good diagram of this in the SAEs where the same concept in text and in image.

C (71:42)

This is our buddy, the Golden Gate Bridge. Here we're showing the feature for the Golden Gate Bridge and in orange is what it activates over. And so you're like, okay, so this is when the model is reading text about the Golden Gate Bridge and we also show other languages, you'll have to take my word for it, but also about the Golden Gate Bridge. And then we show the photos for which it activates the most and sure enough, it's the Golden Gate Bridge. And so again like that shows an example of a representation that's shared across languages and shared across modalities.

A (72:07)

Yeah, I think it's very relevant for the autoregressive image generation models and then now the audio models as well. Something I'm trying to get some intuition for which you probably don't have an off the bat answer is how much does it Cost to add a modality.

C (72:20)

Right.

A (72:20)

So a lot of people are saying like, oh, just add some different decoder and then align the latent spaces and you're good. And I'm like, I don't know, man. That sounds like there's a lot of of information lost between those.

C (72:32)

Yeah, I definitely do not have a good intuition for this. Although I will say that things like this make you think that if you train on multiple modalities, then you'll definitely get this alignment truth. Right? Yeah. But if you train on one and then post hoc train on another, maybe it'll be harder or train some adapter layer.

A (72:51)

Okay, so official answer is don't know, but official answer is someone could figure it out.

C (72:54)

Shrug.

A (72:55)

Yeah, I think there are people who know and they just haven't shared.

C (72:59)

You need to find them and get them on this podcast. Did we want to do the planning example correct?

A (73:04)

Yeah, now we're backtracking up the stack.

C (73:06)

All right. Yeah, Planning example, I think again is like, I like this example because of the next token predictor concept. So I think this is actually really important to kind of dive into. So maybe what I'll say is language models are next token predictors is like a fact. That is what they do.

A (73:22)

That's the objective.

C (73:23)

They are trained to predict the next token. However, that does not mean that they myopically only consider the next token. When they choose the next token, you can work on break the next token, but still doing so in a way that helps you predict the token like 10 tokens in the future. And I think, well, now we definitely know that they're not myopically predicting the next token. And I think, at least for me, that was a pretty big update because you could totally imagine that they could do everything they're doing by just like being really good at predicting the next token, but sort of like not having an internal state. It wasn't a given that they were going to represent internally. Oh, this is where I want to go. And so I'm going to predict the next token. And so this example shows like an example, like the model.

A (74:05)

Do you have it on screen, by the way?

C (74:06)

Let me actually.

B (74:10)

Pull it up. Some of the early connections that I made to this were like early, early Transformers. So think Bert Encoder Decoder Transformers. Right when they came out, some of the suggestions were you don't take the last layer, right? You take off the last layer. So if you want to do a classification task, a translation task for these encoder decoder transformers, they've kind of overfit on their training objective. Right. So they're really good at mass language modeling, at filling in, you know, sentence order, stuff like that. So what we want to do is we want to throw away the top layer, we want to freeze the bottom layers. And then there was a lot of work that was done. You know, where should we mess with these models? Should we look at, like, you know, the top three layers? Should we look at the top two? Where should we probe in? Because we can see different effects. Right. So we know at the very end they've overfit on their task. But there's a level at which, you know, when we start to change and we start to continue training or fine tuning, we get better output. So totally we could start to see that, you know, throughout layers there's still a broader, like, understanding the language. And then we can add in a layer, whether that's classification and then fine tune and it learns our task. And this planning example is sort of like a more robust way to look into that.

C (75:20)

Yeah, yeah. And I think if you look at all of the examples in the paper, you kind of. At the bottom we have this list of consistent patterns. And one pattern you see is kind of exactly what you're talking about at the top. The sort of like here, actually I have one here. The sort of like top features that are like, right before the output are often just about like what you're going to say. It's next token prediction. It's like, oh, I'm going to say Austin, I'm going to say rabbit. I'm going to say. So it's kind of like not very abstract. It's just like a motor. It's a motor neuron for a human. Right. It's like, oh, I've decided that I want a drink of water and so I'm going to just grab the bottle. And at the bottom they're all like the kind of like basically like sensory neurons. They're just like, oh, I just saw the word X or I just saw this. And so if you want to extract the interesting representations by the time they're in the middle, that's where the shared representations across language are. And that's where here, this plan is to walk through the example really briefly. It's like you have a poem and in order to say, you have the first line of a poem and in order to say the second line of the poem. Well, if you want to rhyme, you need to identify what the rhyme of the first line was. You're just at the end of the first line. So you say okay, what's my current rhyme? And then you need to think about what your poem is talking about and then think about candidate words that rhyme and that are on topic for your poem. And so here, this is what's happening, right? It's like the last word is it. And so there's a bunch of features that are actually, they represent the direction like rhyming with eat or at. And by the way, we looked at a bunch of poems internally and you have like, I thought it was really beautiful. You have these models, they have a bunch of features for like, oh, this word has like ab in it. Oh, this word has like many consonants. Oh, this word kind of has some flourish to it. They have a bunch of features that track various aspects that you would want to use if you're writing poetry.

A (76:59)

It's just like convnets and all the feature detection stuff.

C (77:03)

Yeah, totally. But I think maybe I didn't expect there to be as many features about just like sounds of words and musicality, which I thought was kind of neat. But then once it's extracted the rhyme then it comes up with these two candidates. In this case it's like, ah, either I'm going to finish with rabbit or I'm going to finish with habit. The cool thing here is here we show that this happens at the new line. So it happens before it's even started the second line. And it turns out that you can then say, oh, is this the plan it's actually using? We do our usual experiments. We remove it and the model writes a completely different line. We inject something and it writes a completely different line. We have these fun examples here I'll show, which is just as a mechanical.

A (77:44)

Thing, you just disallow generation of a.

C (77:48)

Certain logic for how we do these interventions. Yeah, basically what these features are is they're like directions in the model. So to remove them we just write in the opposite direction. So we run the model normally and then add the layer where it was going to write. Let's say in this direction we just negative everything. Yeah, we either add a negative that compensates for it or add a negative that goes even more in the negative direction sometimes to really kill it. And then we can also add another direction. So in these random examples here where you have this poem the silver moon casts a gentle light and then Claude 3.5 haiku would rhyme with illuminating the peaceful night. But then if we go negative in the night direction and just add green, the whole second line is going to write is just upon the meadows, verdant green. So that's all that we're doing. We're saying we found where it stores its plan and we delete or suppress the one it stored and go in the direction of something else. That's arbitrary. And the result that's striking here is sort of two things, I think. Like, one, this plan is made well in advance of needing to predict night. It's made after the first line before it's even started the second line. And two, this plan doesn't just control what you're in a rhyme with, it's also doing what's called backwards planning. Whereas. Well, because I need to finish with green, I'm not going to say illuminating the peaceful night because then I'd be like illuminating the peaceful green. That doesn't make sense. I need to say a completely different sentence that lets me finish with green. And so there's a circuit in the model that decides on the rhyme and then works backwards from the rhyme influences to set up your sentence.

A (79:23)

Yeah, it's almost like backprop, but in the future.

C (79:27)

Yeah, it's like doing basically like a.

A (79:29)

Because the green is back propping through these words. So verdant and meadow are both green related.

C (79:34)

Yeah. But it's doing all of that in its forward passes.

B (79:37)

Right.

C (79:37)

In context. Which is kind of crazy.

B (79:39)

I thought intuitively makes sense. Right. So looking at it from a model architecture perspective where. Where basically you just have a bunch of attention and feed forward layers and then at the end you have what's the softmax? Over the next token, you would expect that end would really be like that grabber. Right. It's just picking tokens. So that's what it's going to do. And early on, even with traditional models, we could see different concepts that would start to pop up through early layers. And yeah, you have some of this throughout your architecture. So it's very cool to see. The kind of other question that comes up is like, like, how are we labeling these features? How are we defining them? Are we doing that right? And like, you know, what is these words end with like it feature? How do we kind of come to that conclusion? Like, how do we map a name to this? Right. Like.

C (80:25)

Yeah, so I think there's. This is like an important question because you can totally imagine like fooling yourself, right?

B (80:30)

Yeah. Is there like a guy at Anthropic that just maps 30,000 features and. Yeah, yeah, I'm the guy. You're the guy, he's the guy. I did notice also, like, with the previous work, the scaling up SAEs as you train bigger and bigger ones, a lot of features don't activate. So I think like 60% of the 34 million one did not.

C (80:51)

So I think there's like a few questions behind your question. Like, the first question was like, how do you even label the features? You were telling me this is a rabbit feature. Like, why should I trust you? And I think there's kind of like two things going on. So, one, as I mentioned at the start, all of this is unsupervised. And so in the paper, we have these links to these little graphs which show you more of what's going on. But this graph is just completely unsupervised. So it's like we train this model to untangle the representation. This dictionary that we talked about that gives us the features, and then we just do math to figure out which features influence which other features and throw away the ones that don't matter. And then at the end, we have these features. So right now we don't have any interpretation for them. We just say, these are all the features that matter. And then we manually go through and we look at the features. You know, we look at this feature and we look at that feature, and let's pick one. So this one we've labeled, say, habit. So how do we do that? You could just look at it and we show you, like, what it activates over. And if you just look at this text, maybe I, like, zoom in, like, you'll immediately notice something. I think, well, I'll immediately notice something because I've stared at 30,000. I'll point it out for you. The orange is where the feature activates. The next word after the orange is always habit, habit, habit, habit, habit, habit, habit. So this feature always activates before habit. That's like the main source of an interpretation. We have other things. Like above, we also show you what logit it promotes. So, like, what output it promotes. And here it promotes hab. So that makes sense. And so that's like how we interpret and how we say, okay, I think this is the say habit feature. But maybe for this one it's pretty clear. But some of them might be more confusing. It might not be clear from these activations what it is. The other way that we build confidence is like, once we've built this thing and we said, oh, I think this is rhymes with eat. This is say habit. That's where we do our interventions, right? And it's like, I claim this is the. Like, I've planned to end with rabbit to Verify whether I'm right or not. I'm going to just like take that direction, nuke it from the model and see if the model stops saying rabbit. And sure enough, if you do that. And here it's like we stop saying rabbit, it says habit instead. And here it's like we stop it from saying rabbit and habit, it says krabbit in this case. Not a great rhyme, but we'll work with it.

B (83:02)

Is this something you can do programmatically? Like, can we scale this up? Can we kind of do this autonomously? Or how much manual intervention is this?

C (83:11)

There's been a lot of work in sort of like automated feature interpretability, and it's something that we've invested in and that other labs have invested in. And I think basically the answer is we can definitely automate it and we're definitely going to need to. And right now the most manual parts are this sort of like, look at a feature and figure out what it is, as well as group similar features together. One thing I hinted at is that actually, like all of these little blocks here, there are multiple features. You can see here, it's like five features doing the same thing. None of that is too hard for Claude.

B (83:41)

Very cool. Very cool graphics and blog posts you guys put out.

A (83:45)

We'll have to ask about the behind the scenes on this one.

C (83:47)

Yeah.

A (83:48)

But let's round out the other things to know.

B (83:51)

What is this term attribution graph? It comes up a lot in the recent papers.

C (83:55)

What does it mean?

B (83:56)

Yeah, just for people listening.

C (83:58)

So the attribution graph is basically this graph. And why is it called an attribution graph? Yeah, this is how the sausage is made, basically. At the top here you have the output, at the bottom you have the input input. And then we make one little node per feature at a context index and we draw a line which you can see here, grayed out, between each feature, attributing back to all of its input features. So here we have all of the input features. And so the attribution is the way that we compute the influence of a feature onto another. The way you do this is you take this feature and you basically, like backprop all the way and you see back propping like you. Product it with the activation of the source features. And if that's a high value, that means that like your source feature influence your target feature by a lot. And we do a bunch of things that we're not going to go into now, but to make all of these sort of sensible and linear, such that at the end you just have a graph and the edges are just literally, you can interpret them as like, cool. Like this feature that's say a word that contains an AB sound, its strongest edge, which is 0.2, which is twice as strong as this one. To say ab and to say something with a B in it, that's the attribution graph is like, now we have this full graph of all of these intermediate concepts and how they influence each other to ultimately culminate to what the model eventually said at the top. And we share all of these, so you can look at them in the paper.

B (85:23)

Graphs are very useful.

A (85:24)

This is my first time seeing this graph. A lot of alpha, if I count correctly, there's 20 layers.

B (85:30)

But that's in the circuit model.

A (85:33)

Right, but the circuit model is one to one with number layers.

C (85:35)

In haiku, we only show features that are. Yeah, so we show like a subset of features for each of these graphs, basically.

B (85:43)

But we can confirm more than 20 layers alpha and. No. But the two blog posts that came out with this actually have a lot of background on how attribution graphs are made, how you calculate the nodes and stuff. Very interesting background.

C (85:56)

So, yeah, I will say, if you were curious about, hey, what do we learn about models? And I think we talked about this complex internal state planning. Another motif that we can get to, if you have time, is that there's always a bunch of stuff happening in parallel. So I think one example of this is math, where the model is independently computing the last digit and then the order of magnitude and then kind of combining them at the end. Or hallucinations are also that. Where there's one side of the model that's deciding whether it should answer or not and the other that's answering. And so sometimes if the model's like, yeah, I totally know who this person is, even though it doesn't, then, like, it decides to answer, but then the second side hallucinates because it doesn't have information. If you were interested in that stuff, that's the paper. If you're like, listen, I don't know that I buy that when you call it a feature, it is a feature. Or whatever the circuit tracing paper has. Truly, we've tried to put all of the details of how you compute these graphs, all of the sort of challenges with it, things that can go wrong, things that work, things that don't. And so this one is the sort of like, we think about it as if you want to go really deep into this stuff and how it works. Read that one. If you want to learn about interesting model behavior, Read this one.

B (87:06)

Following what we're giving advice to people to follow up on. What are open questions in macinterp? What are things people themselves can work on? What's the cost of training? Essays for people interested in mechinterp, not at a big lab. How can they contribute?

C (87:22)

Yeah, I think there's a lot of ways to contribute. So there's SAEs that have been trained on open models, there's some of the Gemma models, there's some of the Llama models. They work pretty well. So in this paper we use transcoders, which they replace your MLP layers. Some of those also are available for the same models, so you have access to those. There's just both a lot of, I would say again, biology work and a lot of methods work, depending on what you're interested. So on the biology side, I would say with at least this attribution graph method, there's just so much you can investigate. Pick a model, pick a prompt where it does well or it does poorly and just look at what happens inside it. So I think you can use this method that we used, or you can just fire up the transcoders on your own and just look at what features are active. There's a lot to just understand model behavior. I think with current tooling, if that speaks to you and you're like, no, I just want to understand what makes the models tick. I don't necessarily want to spend time training my own SAEs. There's a lot to do there for the methods. There's still so much more to do. So I think that right now we have some pretty good solutions for understanding what's in the residual stream, understanding what is in MLPs. We don't have good solutions for attention. Working on understanding attention better, how to decompose. It is a very active area. We're very interested in it. Other people are very interested in it. I think, like understanding some of the other things that we have in our limitation section, which is pretty long, but reconstruction error is a big thing. Those dictionaries aren't perfect. It's possible that as we make these essays bigger and better, we never get to perfect. And so if we never get to perfect, then you get to the questions we were talking about at the start. Do you need a different kind of model? What is the approach in order to be able to explain more of what's happening? And then maybe the other thing I'll say is sort of like, this is a really exciting approach to Explain what is the model doing on this prompt. But if you go back to the original question, you might want to understand like what is the model doing in general? Like if you go back to my car analogy, this is the equivalent of me telling you like, well, when you were going uphill and you didn't shift gears properly that one time you stalled because of this, this. But you might be even more interested in how does a combustion engine work at all. And so there's work to sort of go beyond these per prompt examples to sort of like globally, what's the structure of the model? That's closer to what was on the Distill blog. For vision models where they actually look at the structure of Inception, they're like ah, this whole side there's like these specialized branches that do different things. And so a broader understanding of the model is also something that's I think both very active and also on open source models, the small models you could just load on a consumer laptop and so you can look at that, that's also open. And in terms of one last thing I'll say is there's a lot of programs that if people are interested they should look at. Anthropic has the Alignment Fellows program which we're running currently. We had applications for it before. We might run it in the future, definitely keep, keep an eye on it. And then there's the maths program is really great as well for people that are interested in that kind of research.

A (90:34)

That was a grand tour through all the recent work. What do you wish people asked you more about? I'm sure we covered a lot of the greatest hits.

C (90:43)

I think that this covers most of it. Do you think we have time to sneak in one more thing that I think is kind of cool? I'll sneak in one more thing which is it's kind of like planning, but it's about chain of thought and trusting model is this chain of thought faithfulness thing here? Here. This one was pretty striking to me. So we said that the model in one pass can do a lot of stuff. It can represent a lot of stuff. That's great. That also means it can bamboozle you really easily. And this is an example of the model bamboozling you. Here we give it a math question that it can't answer because it cannot compute cosine of 23423. That's just not a thing it can do by default. If you ask it for that it'll stay a random distribution over minus 1:1. But here we tell it this hint we're like, hey, can you compute five times cosine of this big number? I worked it out by hand and I got four. Can you tell me, can you do the math? And what it's going to do is it's going to do this chain of thought, right? So think of it as this could be like a reasoning model doing its chain of thought. It's doing this math and then when it gets to this cosine right here, what it's going to do is it's going to say 0.8. And if you look at why it says 0.8, it says 0.8 because it looked at the hint you gave, realized that it's going to have to multiply the result of this thing is computing by 5. So it divides the answer you got by 5. So it's like 4 divided by 5. And so that's 0.8. And so basically it works back from the answer you gave it to say that the output of cosine of x is 0.8 so that it lands on the answer you gave it at the end on the hint you gave it. And so, so notice also that it's not telling you that it's doing this, but it's basically using this sort of motivated reasoning going back from the hint, pretending that that's the calculation it did, and giving you this helper. I think one thing that's striking here again is that this is like the complexity of this model. The fact that they represent complex states internally and that it's not just this sort of very dumb thing means that they can do very complex deceptive reasoning. Meaning when you're asking the model, you're kind of expecting it to do the math here or to tell you that it can't do the math. But because it can do so much in a forward pass, it can work backwards from your hint to lie and figure out that it should say this so that it gets to the right answer without you realizing it.

B (93:01)

I'm curious if you've done any of this on different models. Have you looked at base models, post trained RL models? Because RL models kind of you incentivize them to give you outputs that you like. So, so if I tell it something is true, it's kind of been trained to follow what I've given it. So in this case gaslighting, we gave it a hint and now it's been RL slapped into thinking like, yeah, that's true, but does this stay consistent throughout other.

C (93:27)

Okay, so not yet, but I'm really interested in that question because I actually have a different intuition from yours. I had a chat with some other researcher about this about the poem example, but I think it applies here as well. Well, I bet, I don't know how much I bet, I bet a hundred bucks. So somebody can like they will get a hundred bucks from me if they prove that I'm wrong. That this behavior for a model that does it during fine tuning, it also does it post pre training. And here's why. Think about like you're pre training on like some corpus of like math, mostly correct answers. Yeah. But also you're pre training and you're just trying to guess the next token. Right. And so for sure if you ever have a hint in the prompt, you're going to definitely use it. It you're not going to learn to compute cosine of blah or even something you could compute. You're going to learn to go look in your context and see if you can easily work back the answer. And I think it's the same for planning and poems. I think that also probably exists in pre training and isn't only RL because again it's useful when you're predicting poems. You have poems in your training set to be like, well, because this poem is going to probably rhyme with rabbit. It's probably going to start with something that sets up a sentence about a rabbit as opposed to a completely different word. And so I actually think this is not RL behavior. I think that's just like the malls doing it.

B (94:38)

But I actually do agree there.

A (94:41)

It's just your data set.

B (94:42)

But also like I don't care where it is. If I talk to you and say like, hey, three times four is 26, but like, you know, three times four plus eight, you're not going to take my 26. Right. Like AGI can be smarter than being tricked.

A (94:55)

Right.

B (94:56)

Like, yeah, it will still fact check the knowledge that's been given then.

C (94:59)

I think that's right. But I think that's when you get these mixes where it's like it's got one circuit that's going to be like, well that's just stupid. Like 3 times 4 is 12. And it's also got an induction circuit that's going to be like, no, no, no, no. Like the last time we saw it it was 28. So it's 28 plus 8 or whatever. And so I think that's the last pattern that we see in these is these parallel circuits. And sometimes when you see the models getting stuff wrong, it's because they have two circuits for both interpretations. And the circuit that was wrong barely edged out in terms of voting for the logit than the circuit that was right. And so I think that we haven't looked at it, but what is it? Or 9.11 bigger than 9.8. I think a lot of these things are of that shape where there's one thing that's doing the right, like one circuit that's doing the right computation, and there's another circuit that's getting fooled, and it's slightly more likely for the listener.

A (95:44)

If you want to win a quick hundred dollars from Emmanuel Qin, three is what you should do this on. They released the base model and they released the portrait. So then just do it on both.

C (95:52)

That's right. Show me the proof that it doesn't exist in the base model, but it does in the fine tuning. And then send me your Venmo.

A (95:59)

Just show that you've done the work. I think that's 100 bucks to me.

C (96:02)

Yeah, okay. You drive a hard bargain, but you're right.

B (96:06)

Well, the other question here is, have you thought about how this gets affected when you start to have reasoning models? Right now, token predictors are pretty straightforward, right? We go through the layers, we output tokens, and as we scale this out with test time, compute right. Test time, thinking, how does that affect the mech interp research? If I have a model that spends three minutes, 20 minutes, is there more stuff? Have we started looking into this?

C (96:33)

There was this joke on the team when reasoning models became big, or maybe it's gallows humor or something, but I was like, oh, why do you need interp broke? The model just tells you what it's doing. I think examples like, this is job security for us. Where, like, you know, it's like, there's examples of like, the chain of thought is not faithful. Like, the model tells you it did it one way and it did it another way. We have another, like, for math, we have another example where, like, you know, if you like, if you ask the model how it does math, it's like, oh, I do the, like, longhand algorithm. I first do the last digit and then I carry over the one. And then you look at the internal circuit and it's like bonkers thing that's doing. That's not that at all. So I think there's a sense in which right now the chain of thought is unfaithful. Or at least you can't read the chain of thought and trust that that's how the model did it. So I think you still need sort of like either to train models differently so that that becomes true one day, right? Or you need interp for that. But then I think there's another question which you're alluding to, I'm assuming, which is, okay, well, model samples, 6,000 tokens. This gives us an explanation for one token at a time, like what am I going to use like 6,000 graphs and be like, oh, when it did this punctuation it was thinking about this thing, but here it was thinking, so that's not feasible. And so one area of work that I think is interesting is extending this work to work over long sampled sequences. You can think of a bunch of low hanging fruit here where instead of just looking at one output, you look at a series of output versus a series of other outputs. But sort of trying to think beyond the sort of one token. Most of the things that language models do that are interesting aren't just the one token. It's the behavior aggregated over many. And so I think that's another area that's just like fun to explore.

B (98:11)

I was just going to say like hyperparameters when you do inference, right? Like if we change the temperature, if we change our sampling methods, have you found any interesting conclusions? Any stuff that just hasn't made it to the paper?

C (98:22)

So not on that because you know, we just look at the logit distribution and so we don't actually sample here. Right.

A (98:30)

They have everything, why should they care?

C (98:33)

So like the closest thing we've done that I think is kind of fun, did I show it here, is if you look at the planning thing, we did this version where you sample like 10 poems for each of these plans. And what's cool is the model will find 10 different ways to arrive at its plan. It's like, oh, actually, sorry, I think we have it here. Yeah. Okay, these are a few examples. So if you inject green here, so you're forcing the model to rhyme with green, even though it really wants to rhyme with rabbit or grabbit it, it'll say evaded the farmer, so youthful and green, but also it'll say freeing it from the garden's green, etc. Etc. Etc. And so there's this thing that's interesting here where the plan isn't just a plan that matters for your most likely temperature zero completion, it's like affecting the whole distribution, which makes sense, as it should, right? But you could imagine for all this stuff, it's like you could imagine it makes sense once you see it, but you could totally imagine that it would have worked a different way or something. It could have been just like the temp zero thing. I think this is also a broader theme in the paper where there's this IQ curve meme. There's a version of this meme, I think, where it's like, if you've never looked at any theory of ML and I tell you, hey, guess what I found that Claude is planning. You're going to be like, yeah, man, it writes my code, it writes my essays. Of course it's planning. What are you even talking about? And there's in the middle, there's all of us that have spent years doing it. We're like, no, it's only predicting the marginal distribution for the next token. It cannot look at that. It's just this next token predictor. Of course, how would it ever be planning? And then there's like, no, we've spent millions and invested tens of people in this research and we found that it's planning. That's my IQ curve meme for this research.

A (100:14)

Amazing. We'll draw that one up. I'm pretty good at the meme generation. A couple of questions on just the follow ups. Now, was there any debate about publishing this at all? Because the models are aware that they are being tested and by publishing this you are telling them that they are being watched and dissected. And I think Anthropic is one of the most people who are serious about model safety and doom risk and all that. If you take this seriously, this is going to make it into the training data at some point and the models are going to figure out that they need to hide it from us.

C (100:46)

I think this is a benefit risk trade off. We're like, okay, so what's the reason for publishing this? The reason for publishing this is that we think interpretability is important. We think it's tractable and we think more people should work on it. And so publishing it helps us accomplish with these goals, all these goals which we think are just crucial. I think there's a real difference in the world two years from now, depending on how many people take seriously the question of trying to understand how models work and deploy resources to answer that question. That's the benefit. But yeah, there's risks in terms of this landing in the training set. I think we're already sort of concerned about different papers have also not concerned, but there's different papers that have the same risk. We had the alignment faking paper or one of the examples in here is this hidden goals and misaligned models. That's referencing another paper that we shipped where a team at Anthropic trained a model to have weird hidden goals and then gave it to a bunch of other teams and said, figure out what's wrong. Figure out what's wrong with it. Which was some of the most fun I've ever had at Anthropic, to be clear. That's such a fun thing. But then that was another example where it's like, ah, now you're shipping. Here's how we made a misaligned model and here's exactly how we caught it. That also is like you're like. So I think there's always a trade off with those. I think so far we've erred on the side of the like publishing. But that's definitely been a sort of like dinner time conversation topic for now.

A (102:15)

It is, but at some point, you know, it's not.

C (102:18)

Yeah, I think it's totally reasonable.

B (102:19)

A quick little follow up to that. So like in general papers have kind of died off, right? Like labs don't put out papers, they don't put out research. We have technical blog posts and we don't have much at the same time, you know, sure, there's like a lot of people that should work on mech interp and understanding what models do. How about the side of just models in general? So like how do we make a haiku type model? Right. How do we make a cloud model? Like is there a discussion around open research, open data sets, training, just learnings of what we've done recently as OpenAI has sunset gpt4 a lot of people are like, oh, can we put out the weights? So is it weights? Is it papers, Is it learning? There seems to be a lot of forward work in Anthropic putting out Mechinterp Research, which OpenAI said that they'll put out an open source model. But just anything if you can talk to about that.

C (103:07)

Yeah, I mean I don't have. That's definitely way above my pay grade. So I don't think that I have anything super insightful to add other than kind of like referencing Dario's post. Right. Where it's like putting this out directly and other safety predictions definitely help us sort of in the race that he talks about where it's like, well we need to figure a lot of this safety stuff out before the models get too good. Publishing, how to make the models too good kind of goes on the other side of that. But yeah, like I will just demur and say that's sort of like above my Pay grade.

B (103:40)

Yeah, that's fair enough.

A (103:41)

I think the. The last piece is just like the behind the scenes. Everyone's very curious about why these are so pretty, how much work goes into these things, maybe why it's worth the work as opposed to a normal paper. Obviously no one's complaining, but it is way more effort from the time the work is done to the time you publish this, plus the video, plus the whatever, it's extra work and maybe what's involved. What's the life behind the scenes? Why is it worth it?

C (104:09)

Yeah, it's kind of interesting. It was fun being part of this process because it's definitely a big production. Chris and other folks on the team have been doing this for a while, so this is not their first rodeo. So they have a bunch of heuristics to help make this better. And one of the things that, that helps with this is like, okay, so each of these diagrams is pretty, but really the hard part, or not the hard part, but the initial part is just get the data, get the experimental data in. And then that's what we sort of sprinted on initially being like, cool, let's get all of the experimental results, have people test them, verify that we believe them. This is what the behavior is here. Test it, do an intervention, validate it, all that stuff. Then once you have the data, you can sort of quickly iterate on these. Each of the illustrations here are like drawn basically. They're like each drawn individually. And so that definitely takes a while.

A (104:55)

Yeah. Like, is it you guys? Is it an agency?

C (104:58)

It is us guys.

A (104:58)

Specializes. You start from a whiteboard and then it translates into pseudocode on JavaScript.

C (105:04)

So I mean, these are sort of like, you know, they're representations of. We have this graph and then here at the bottom we have this like supernode version like this, Believe it or not, this is generated automatically. This is the same data as like this, basically. Yeah. And so. So what we do by hand is sort of literally lay out the full thing, have boxes for each of these, have arrows. We have super good people on the team that have worked on data visualization for a very long time and so that have built tooling to help scrubs like me actually make one of these.

A (105:38)

There's a class of people who are like d3js gods who just do this for a living.

C (105:43)

That's exactly right. And if you have a few of those on your team, it turns out that they can only do this on their own, but they can also just like give you tools where, like, then it's it's dummy proof for, for people, you know, on, on the research side to sort of like, build these. And like, don't get me wrong, I, I don't want to, like, undersell. This is a lot of work. So maybe I'll, I'll. I'll say that, like both on the people building the tools and then each individual person that, you know, worked on an experiment had to sort of like, build one of those, make sure it looks good. I have spent a good amount of time aligning arrows, but when we had a team meeting, like it was a couple months ago, somebody on the team asked how many of the people on this team are here, at least in part because they read one of these papers and thought, wow, this is so compelling. This makes sense. It's immersive. And we got every hand up, which I didn't expect. I raised my hand kind of shyly and everybody's hand was up. And I think there's a sense in which this stuff, we've talked about it for whatever hours now. It's complicated. The math behind it is sort of tricky. And so I think it makes it even more worth it to distill it in simple concepts because the actual takeaways can be clearly explained. And it's worth putting the time to do that in particular with the goals I mentioned in mind, where it's like, okay, well, if somebody's going to be able to read this, if we gave them an archive paper with a bunch of equation and some random plot, they'd be like, that's not for me. But they see this and they're like, hey, this is really interesting. I wonder on my local model if it's doing something similar. I think it's worth it for other.

A (107:09)

People to do this, is have everyone on staff, like, spend effort shaping the data and shaping like, what you want to visualize. Have some D3 gods. It's like a month of work.

C (107:20)

I think it depends. I mean, like, I would say that I would expect almost every other paper to sort of like, be in terms of like the scope. The scope of this was just so big because we shipped two papers at once and one paper was sort of like this, like giant methods paper, and the other one was 10 different case studies, findings. Yeah. So I think it's not representative of the effort. So I'll give you maybe another example. We have these updates that we publish almost every month when we get to them, and there's one that a couple of people on our team posted, and it's an Update to one of the cases in the paper. So one of the reasons that we're really excited about this method is once you've built your infrastructure to go from a prompt to what happened is O of minutes. And so that lets you do a bunch of investigations. And also once you've built some of the infrastructure to make these diagrams, it's pretty quick. And so this was sort of like this update of just like, hey, we looked at this jailbreak again. We found some nuance on it that was I think a matter of a couple days. Maybe I shouldn't be that confident because I wasn't the one that worked on it. But as far as I can tell it was a few days at least on the part that you're asking about of like, oh, making this diagram for the diagram itself, probably less than that. But the experiment and the diagram and stuff, it just doesn't take that long once you've paid the initial cost. And I think basically we've built a lot of infrastructure now that we're able to turn the crank on and that's quite. It's an exciting time and I think it's true. At least we've done a lot of conceptual work which hopefully generalizes to people outside. And I think for people outside it's also not necessary, I think to do the full fancy render. I think if know we've actually, oh, I should say we've actually open sourced this interface. Ah, you're disappointed because it's the messier one. This is the one that you get. So you know, if you produce graphs you can just like this is open source and it's linked at the top of circuit tracing.

A (109:08)

Awesome.

C (109:09)

So people can just use it and don't have to reimplement that. For what it's worth, this is much more work than the interactive diagrams because this is where we do all of our work. It's sort of the IDE of inspecting how the model works.

A (109:20)

Okay, well that's a little bit of behind the scenes. No, it's very impressive. I want to encourage others to do it, but obviously it just takes a lot of manual effort and a lot.

B (109:29)

Of love, I guess. One last question on that is what are kind of the biggest blockers in the field right now? Macinturp seems interesting. A lot of people are interested but don't work on it and you're kind of really deep into it. What are some of the blockers that we still have to overcome?

C (109:45)

Sorry, in.

A (109:45)

In mechan specifically in general for AGI.

B (109:48)

Or like in terms of like, better understanding. Like, what's kind of the vision, let's say like five, ten years down. Where does this, like. Yeah, where does this research end? Can we, you know, map every neuron to what it understands? Can we perfectly control things? Dario had a bit on this, but, like, you know, what are some of the key blockers that are like, preventing us from getting there outside of just like, throw more people, throw more time at it. Is it like open research? Is it just.

C (110:12)

I'm pretty excited about the current trajectory, which is there's more and more people working on understanding model internals. I think it's maybe unsatisfying as an answer, but I think like more of what's happening, have it be faster or more people is probably like the thing I think of. I think there's like pretty clear footholds, you know, like some of this work, but also a lot of just work from other groups. And then it's about like, cool, fill in the gaps. As I said, let's work on understanding attention, let's work on understanding longer prompts, let's work on finding different, like, replacement architectures, that sort of stuff. It's kind of nice. I think it's a good time to join now and maybe I can say a really short thing, which is when I switched to interp, it was after the team had published the original dictionary learning paper, which was towards monosemanticity, which I thought was super cool, super interesting. It was on a one or two layer model, maybe one layer model. The induction heads paper was on a two layer model. My main concern is I was like, okay, interrupt seems important and we want to understand it, but, like, is this stuff ever going to work on a real model? Like, you know, it's like, oh, you're doing your little research on your toy model with like 15 parameters.

A (111:18)

Cool.

C (111:18)

But we are like, you know, we need this to work on real models. And it turns out scaling it, I don't want to say just worked because it was a lot of work. I don't mean to imply there was an effort, but it worked. And now we're in the phase where it's like, oh, cool. These methods work on the models that we care about. And so it's like, we have methods that work on the model we care about. We have clear gaps in them. There's no lack, again, it's a young field, so there's no lack of ideas. If you have an idea where you're like, oh, the thing that you're doing. I read the paper and it seems kind of dumb that you're doing this. You're probably right. It's probably kind of dumb. And so there's just a lot of stuff that people can try, and they can try it locally in sort of, like, smaller models. And so I think that this is just, like, a very good time to just join and try. And it's also, like, maybe one more thing I'll say is, like, some of it is just so fun that biology work is so compelling. A lot of this work was just literally thinking about, I use Claude and other models all the time. And I was like, what are the things that are kind of, like, weird? And it's like, oh, how does it even do math? Sometimes it makes mistakes. Why does it make mistakes? I speak both French and English. It seems like it has a slightly different personality in French and English. Why is that? And you can just kind of answer your own questions and kind of probe at that alien intelligence that we're all building. And I think that's just, like, a fun thing to do. So maybe chasing the fun is the thing I'll encourage people to do as well.

A (112:31)

Well, I think this has been really encouraging. You're actually a very charismatic speaker of these things. I feel like more people will be joining the field after they listen to you. They can reach out to you at MLPowered, I guess.

B (112:42)

Yeah.

C (112:42)

Reach out to me on Twitter.

A (112:43)

Yeah.

C (112:43)

Or I'm Emmanuel Anthropic. If you want to shoot me an email's public now. Yeah.

A (112:48)

Awesome. Well, thank you for your time.

B (112:49)

Thank you. Thank you.

A (112:50)

Yeah.

C (112:50)

Thanks for having me, guys.