The Utility of Interpretability — Emmanuel Amiesen, Anthropic

Thanks for coming to AIE - full Day 1 and Day 2 streams are available, with 6 other livestreamed tracks. You can binge this weekend! Longer recap when we have slept.

---

Emmanuel Amiesen is lead author of “Circuit Tracing: Revealing Computational Graphs in Language Models”, which is part of a duo of MechInterp papers that Anthropic published in March (alongside On the Biology of LLMs).

We recorded the initial conversation a month ago, but then held off publishing until the open source tooling for the graph generation discussed in this work was released last week in collaboration with Neuronpedia:

This is a 2 part episode - an intro covering the open source release, then a deeper dive into the paper — with guest host Vibhu Sapra and Mochi the MechInterp Pomsky. Thanks to Vibhu for making this episode happen!

While the original blogpost contained some fantastic guided visualizations (which we discuss at the end of this pod!), with the notebook and Neuronpedia visualization released this week, you can now explore on your own with Neuronpedia, as we show you in the video version of this pod.

Full Version on YouTube

like and subscribe!

Timestamps

• 00:00 Intro & Guest Introductions

• 01:00 Anthropic's Circuit Tracing Release

• 06:11 Exploring Circuit Tracing Tools & Demos

• 13:01 Model Behaviors and User Experiments

• 17:02 Behind the Research: Team and Community

• 24:19 Main Episode Start: Mech Interp Backgrounds

• 25:56 Getting Into Mech Interp Research

• 31:52 History and Foundations of Mech Interp

• 37:05 Core Concepts: Superposition & Features

• 39:54 Applications & Interventions in Models

• 45:59 Challenges & Open Questions in Interpretability

• 57:15 Understanding Model Mechanisms: Circuits & Reasoning

• 01:04:24 Model Planning, Reasoning, and Attribution Graphs

• 01:30:52 Faithfulness, Deception, and Parallel Circuits

• 01:40:16 Publishing Risks, Open Research, and Visualization

• 01:49:33 Barriers, Vision, and Call to Action

Transcript

swyx [00:00:03]: All right, we are back in the studio with a couple of special guests. One, Vibhu, 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, Amason? Yep. Is that Dutch? It's actually German. German? Yeah. You are the lead author of a fair number of the recent Macintyre work from Anthropic that I've been basically calling Transformer Circuits, because that's the name of the publication.

Emmanuel [00:00:35]: Yeah. Well, to be clear, Transformer Circuits is the whole publication. I'm the author on one of the recent papers, Circuit Tracing. Yes.

swyx [00:00:42]: And people are very excited about that. The other name for it is like Tracing the Thoughts of LLMs. There's like three different names for this work. But it's all Macintyre.

Emmanuel [00:00:49]: It's all Macintyre. There's two papers. One is Circuit Tracing. It's the methods. One is like 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. Yeah.

swyx [00:01:01]: It's for different audiences. Yes. Yes. And I think when you produce the like two minute polished video that you guys did,

Emmanuel [00:01:07]: that's meant for like a very wide audience, you know? Yeah, that's right. There's sort of like very many levels of granularity at which you can go. And I think for Macintyre in particular, because it's kind of complicated going, you know, from like top to bottom, most like high level to sort of the dernali details works pretty well. Yeah.

swyx [00:01:24]: Cool. We can get started. Basically, we have two paths that you can choose, like either your personal journey into Macintyre or the, the brief history of Macintyre, just generally, and maybe that might coincide a little bit.

Emmanuel [00:01:36]: I think my, 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 Macintyre as sort of like an interesting field with fascinating, often beautiful papers. And I was at the time working on fine tuning. So like actually fine tuning production models for Anthropic. And eventually I got both like sort of like, like 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 like a background in ML, kind of like did a lot of applied ML stuff before. And now I'm doing more research stuff. Yeah.

swyx [00:02:20]: You have a book with O'Reilly. You're head of AI at Insight Data Science. Anything else to plug?

Emmanuel [00:02:25]: Yeah. I actually, I want to like plug the paper and unplug the book. Okay. 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, from the stuff that we're talking to talk about today. Today is like research. Some of, some of the sort of like deepest, weirdest things about how like models work. And this book is you want to ship a random forest to do fraud classification. Like here are the top five mistakes to avoid. Yeah.

swyx [00:02:55]: The good old days of ML. I know it was simple back then. You also transition into research. And I think you also did then, like I feel like there's this monolith of like people assume you need a PhD for research. Maybe can you give like that perspective of like, how do people get into research? How do you get into research? Maybe that gives audience insight into Vibo as well. Your background.

Vibhu [00:03:16]: Yeah. My background was in like 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, right? You just kind of build stuff. You work on interesting, interesting things and like now it's more accessible than ever. Like back when I got into the field five, six years ago, like pre-training was still pretty new. GPT-3 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? Yeah,

Emmanuel [00:03:55]: I think, I think it has maybe never been easier in some ways because, um, a lot of the field is like pretty empirical right now. So I think the better lesson is like this lesson that, you know, you can just sort of like a lot of times scale up compute and data and get better results than like thinking than if you sort of like thought extremely hard about a really good, like prior inspired by the human brain to, to train your model better. And so in terms of definitely like 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 it too, 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 Interp papers, you know, coming out of programs like maths are on models that are open source that you can serve like dissect without having a cluster of like, you know, a hundred 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 like what, what the features are and what dictionary learning is. You're like a long part of the way there.

swyx [00:05:35]: I think it's also interesting just on a careerist 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 like 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 there, that cannot be so easy. Like, like in no other industry that I know of people, you can do that. Do you know what I mean? Yeah.

Emmanuel [00:06:05]: I think I'd actually like 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 like brilliant and hard to find, but what's, what's hard, certainly on fine tuning and to a certain extent on Interp is executing on your research idea in terms of, like making an experiment successfully, like having your experiment run, interpreting it correctly. What that means though, is that like they're not separate skill sets. So like if you have a cool idea, there's kind of not many people in the world, I think where they can just like have a cool idea and then they have a, you know, like a little minion they'll deputize being like, here's my idea, you know, go off for three months and like run this whole day, like build this model and train it for, you know, hundreds of hours and report back on what happened. A lot of the time, like the people that are the most productive, they have an idea, but they're also extremely quick. Uh, checking their idea of finding sort of like the shortest path to, to turn it already. And a lot of like that shortest path is engineering skills essentially. It's just like getting stuff done. And so I think that's why you see sort of like people move around as like 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, from people like I've certainly seen an anthropic that are just like really good at that, you know, they can apply it in one team and then move to a completely different domain and apply that inner loop just as well. Yeah.

swyx [00:07:35]: Very cracked is the kids say, uh, shall we move to the history of McInturk? Yeah. All I know is that everyone starts at Chris Ola's blog. Is that right? And yeah,

Emmanuel [00:07:45]: I think that's the correct answer. Chris Ola's blog. And then, you know, uh, distill.pub, uh, 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 McInturk research out there from, you know, I think like the, yeah, like maths is a group with that, like regularly has a lot of research, but there's just like many different labs that, that put research out there. And I think that's also just like hammer home the point that's because all you need is like a model and then a willingness to kind of investigate it to be able to contribute to it. So, so now it's like, there's been a bit of a Cameron explosion of McInturk, which is cool. I guess the history of it is just, you know, computational like models that are not decision trees models that are either CNNs or let's say transformers have just this w really like strange property that they don't give you interpretable intermediate States by default. You know, again, to go back to if you were training like a decision tree on like fraud data for an old school, like bank or something, then you can just look at your decision tree and be like, oh, it's learned that like if you make a, I don't know if this transaction is more than $10,000, and it's for like perfume, then maybe it's fraud or something. You can look at it and say like, cool, like that makes sense. I'm willing to ship that model. But for, for things like, like CNNs and like transformers, we, we don't have that right. What we have at the end of training is just a massive amount of weights that are connected somehow or activations that 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 on a 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 most, yeah, mostly applied to NLP models, but also, you know, still there's work in vision and there's work in like a bio and other domains. Yeah.

swyx [00:09:44]: I'm on Chris Ola's blog and he has like the feature visualization stuff. I think for me, the clearest was like the vision work where you could have like this layer detects edges, this layer detects textures, whatever. That seemed very clear to me, but the transition to language models seem, like a big leap.

Emmanuel [00:10:00]: I think one, one of the bigger changes from vision to, to like language models has to do with the superposition hypothesis, which maybe is like, that's the first in like toy models post, right? 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 like 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 wasn't clear why that would be. And one main like hypothesis here is the superposition hypothesis. So what does that mean? That means that like language models pack a lot more in less space than vision models. So maybe like a, a kind of like really hand wavy analogy, right? Is like, well, if you want curve detectors, like you don't need that many curve detectors, you know, if each, each curve detector is going to detect like a quarter or a 12th of a circle, like, okay, well you have all your curve detectors, but think about all of the concepts that like Claude or even GPT-2 needs to know, like just in terms of, it needs to know about like 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 like a model knows, you're going to get like a very, very long list. And that list is going to be way bigger than like the number, of neurons or even the like size of the residual stream, which is where like the models process information. And so there's this sense in which like, oh, there's more information than there's like 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, it just seems like it's like, there's, it's got all this stuff crammed into it. Whereas if you look at the vision models, oftentimes you could just like, yeah, cool. This is a curve detector.

swyx [00:11:53]: Yeah. Yeah. VB, you have like some fun ways of explaining the toy models or super, superposition concept. Yeah.

Vibhu [00:11:59]: I mean, basically like, you know, if you have two neurons and they can represent five features, like a lot of the early Mechinterp work says that, you know, there are more features than we have neurons. Right. So I guess my kind of question on this is for those interested in getting into the field, what are like the key terms that they should know? What are like the few pieces that they should follow? Right. Like from the anthropic side, we had a toy transformer model. We had sparse, we first had auto encoders. That was the second paper.

swyx [00:12:25]: Um,

Vibhu [00:12:25]: right. Yeah. Monosyntheticity. Yeah. What is sparsity in auto encoders? What are trans coders? Like what is linear probing? What are these kind of like key points that we had in Mechinterp? And just kind of, how would people get a quick, you know, zero to like 80% of the field? Okay.

Emmanuel [00:12:42]: So zero to 80%. And I realized I really like set myself up for, for failure. So I was like, yeah, it's easy. There's not that much to know. So, okay. So then, then we should be able to cover it all. Um, so superposition is the first thing you should know, right? This idea that like there's a bunch of stuff crammed in a few dimensions, as you said, maybe you have like two neurons and you want to represent five things. So if that's true, and if you want to understand how the model represents, you know, I don't know the concept of red, let's say then you need some way to like find out essentially in which direction the model stores it. So after the, the server, like superposition hypothesis, you can think of like, ah, we also think that like basically the model represents these like 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 two D plane. And it's like, ah, you can have like five directions and maybe you would like arrange them like the spokes of a wheel. So they're sort of like maximally separate. It could mean that like you have one concept that this way and one concept that's like not fully perpendicular to it, but like pretty, pretty like far from it. And then that would like allow the model to represent more concepts and it has dimensions. And so if that's true, then what you want is you want like a model that can extract these independent concepts. And ideally you want to do this like automatically. Like, can we just, you know, have a model that tells us like, oh, like 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, you know? Um, and so that's what sparse auto encoders are.

swyx [00:14:09]: It's almost like the, the self supervised learning insight version, like in, in pre-training itself, supervised learning and here and now it's so self supervised interpretability. Yeah,

Emmanuel [00:14:18]: exactly. Exactly. It's like an unsupervised method. Yeah. And so unsupervised methods often still have like labels in the end or so.

swyx [00:14:25]: So sometimes I feel like the term labels by masking. Yeah.

Emmanuel [00:14:29]: Like for, for pre-training, right? 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 like neurons and then you learn a model. That's going to like expand them into like 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 to like, I think 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 once you do that, if it works, you have this sort of like 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, what this direction means, but I've like used my model into telling me that the model is writing in the red direction. And so that's, that's sort of like, I think maybe the biggest thing to understand is, is this, 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.

Vibhu [00:15:31]: Any follow-ups? Yeah. I mean, the follow-ups of this are also kind of like some of the work that you did is in clamping, right? What is the applicable side of MechInterp, right? So we saw that you guys have like great visualizations. Golden Gate Cloud was a cool example. I was going to say that. Yeah. It's my favorite. What can we do once we find these features? Finding features is cool. But what can we do about it? Yeah.

Emmanuel [00:15:53]: 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 Cloud 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 like they, in some sense are, are important for the model or it wouldn't be like representing it. Like if the model is like bothering to, to like write, you know, in the Golden Gate Bridge direction, it's usually because it's going to like talk about the Golden Gate Bridge. And so that means that like, if that's true, then you can like set that feature to zero or artificially set to a hundred and you'll change the model behavior. That's what we did when we did Golden Gate Cloud 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, Oh, what's on your mind? You know, like, what are you thinking about today? Be like the Golden Gate Bridge. You'd be like, Hey Claude, like what's two plus two, it'd be like four Golden Gate Bridges, uh, et cetera. Right. And it was always thinking about this,

swyx [00:16:55]: like write a poem and it just starts talking about how it's like red, like the Golden Gate plot. That's right. Golden Gate Bridge. Yeah,

Emmanuel [00:17:00]: that's right. Amazing. 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. Right. So I'll tell you, I would like really ham it up. You'd be like, Oh, I'm just thinking about the beautiful international orange color of the Golden Gate Bridge. That was just like an example that I think was like really striking, but of, of sort of like, Oh, if you found like a space where that represents some computation or some representation of the model, that means that you can like artificially suppress or promote it. And that means that like you're starting to understand at a very high level, a very gross level, like how some of them all works. Right. We've gone from like, I don't know anything about it to like, Oh, I know that this like combination of neurons, neurons is this, and I'm going to prove it to you. The next step, which is what this, this works on is like, that's kind of like thinking of if maybe you take the analogy of like, um, I don't know, like, like let's take the energy of like an MRI or something like a brain scan. It tells you like, Oh, like this as Claude was answering at some point, it thought about this thing, but it's a sort of like vague, like maybe, basically maybe it's like a, like a bag of words, kind of like a bag of features. You just have like, here are all the random things that thought about, but what you might want to know is like, what's this thing like sometimes to get to the golden gate bridge, it had to realize that you were talking about San Francisco and about like 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 like, there's like a way to connect features to say like, Oh, from this input went to these few features and then these few features and then these few features and 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 a, or what's called, circuits, which is sort of like explaining the, the like algorithm. Yeah.

swyx [00:18:46]: Before we move directly onto your work, I just want to give a shout out to Neil Nanda. He did neuron Pedia and released a bunch of essays for,

Emmanuel [00:18:54]: I think the llama models and the Gemma models and the Gemma models. Yeah.

swyx [00:18:57]: So I actually made golden gate Gemma just up 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. That's amazing. So you can make golden gate Gemma. And like, I think that's a, that's a fun way to experiment with this. Uh, but yeah, we can move on to, I'm curious.

Vibhu [00:19:19]: I'm curious. What's the background behind why you ship golden gate claw? Like you had so many features, just any fun story behind why that's the one that made it.

Emmanuel [00:19:28]: 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 psychophantic praise, which I guess is very topical right now.

swyx [00:19:40]: Topical.

Emmanuel [00:19:40]: Um, 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 it'd be like, that's the best poem I've ever seen. Um, and so we could have shipped that. That could have been funny. Uh, golden gate. Claude, it 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 with an internal demo of it. Everybody thought it was hilarious. And then that's sort of how it came on. There was, there was no, nobody had a list of top 10 features we should consider shipping. And we picked that one. There was just kind of like a very organic moment. No,

swyx [00:20:18]: like the, the marketing team really leaned into it. Like they mailed out pieces of the golden gate for people in Europe, I think, or ICML. Yeah, it was, it was fantastic marketing. Yeah. The question obviously is like if open AI had invested more interpretability, would they have caught, uh, the GPT for all update? Uh, but we don't know that for sure. Cause they have interp teams. They just,

Emmanuel [00:20:39]: 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.

swyx [00:20:46]: And then the other thing is, um, can you just like up, write good code, don't write bad code and make Sonda 3.5. And like, it feels too, too easy to free. Is that steering that powerful?

Vibhu [00:20:59]: You can just like up and down features with no trade-offs. There was quite a, there was like a phase where people were basically saying, you know, 3.5 and 3.7 are just now because they came out right.

swyx [00:21:09]: And for the record, like that's been debunked.

Vibhu [00:21:11]: But it has been debunked, but you know, it had people convinced that what people did is they basically just steered up and steered 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. So like, what are some, like whether it's internal, what are you exploring that? Like what are the applications of this, whether it's open, ended of what people can do about this or just like, yeah. Why, why do McInturpp, you know? Yeah.

Emmanuel [00:21:46]: There's like a few things here. So, so like, first of all, obviously this is, I would say on the scale of the most short term to the most long term, like pretty long term research. So in terms of like applications compared to, you know, like the research work we do on like fine tuning or whatever interp is much more, you know, sort of like a high risk, high reward kind of approach. With that being said, like, I think there's just a fundamental sense in which, which Michael Nielsen had had a post recently about how like knowledge is dual use or something, but just like, just like knowing how the model works at all feels useful. And you know, it's hard to argue that if we know how the model works and understand all of the components that won't help us like make models that hallucinate less, for example, or they're like less biased, that seems, you know, if, if like at the limit, yeah, that totally seems like something you would do using basically like your understanding of the model to improve it. I think for now, as we can talk about a little bit with like circuits, there's like, we're still pretty early on in the game, right? And so right now the main way that we're using interp really is like to investigate specific behaviors and understand them and gain a sense for what's causing them. So like one example that we, we can talk about later, we can talk about now, but is this like in the paper, we investigate jailbreaks and we try to see like, 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 like 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. Like it really wants to make correct grammatical sentences. And so it turns out that like seeing that circuit, we were like, ah, then does that mean if I prevent it from finishing a sentence, the jailbreak works even better? And sure enough, it does. And so I think like the level of sort of practical application right now is of that shape. So like understanding either like quirks of a current model, or like, or like how it does tasks that maybe we don't, we don't even know how it does it. Like, you know, we have like some planning examples where we had no idea it was planning and we're like, oh God, it is. That's sort of like the current state we're at.

Vibhu [00:23:51]: I'm curious internally how this kind of feeds back into like the research, the architecture, the pre-training teams, the post-training, like, is there a good feedback loop there? Like right now there's a lot of external people interested, right? Like we'll train an SAE on one layer of LAMA and probe around. But then people are like, okay, how does this have much impact? People like clamp down, clamping. But yeah, as you said, you know, once you start to understand these models have this early planning and stuff, how does this kind of feed back?

Emmanuel [00:24:16]: I don't know that there's, there's like much to say here other than like, I think we're definitely interested in conversely, like making models for which it's like easier to interpret them. So that's also something that you can imagine sort of like working on, which is like making models where you have to work less hard to try to understand what they're doing. So like the architecture? Okay.

swyx [00:24:35]: Yeah. So I think there was a, there was a less wrong post about this of like, there's a non-zero model, which is like,

Emmanuel [00:24:47]: there's this sort of sense in which like right now we take the model and then the model is 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 like fully capture everything that's happening inside the model. We're capturing like a subset. And so maybe some of it is like, you could train a model that's sort of like easier to interpret naively. And it's possible that like, you don't even have that much of, you know, like a tax in that sense. And so maybe some of it is like, you can just sort of like, either like train your model differently or do like a little post hoc step to like, sort of like untangle some of the mess that you've made when you trained your model. Right. Make it easier to interpret. Yeah.

swyx [00:25:19]: The hope was pruning would do some of that, but I feel like that area of research has just died. What kind of pruning are you thinking of here? Just pruning your network. Ah, yeah. Pruning layers, pruning connections, whatever. Yeah.

Emmanuel [00:25:34]: I feel like maybe this is something where like superposition makes me less hopeful or something.

swyx [00:25:40]: I don't know. Like that, that like seventh bit might hold something. Well,

Emmanuel [00:25:44]: right. And it's like on, on each example, maybe this neuron is like at the bottom of like what matters, but actually it's participating like 5% to like understanding English, like doing integrals and you know, like whatever, like cracking codes or something. And it's like, because that was just like distributed over it, you, you, you kind of like when you naively prune, you might miss that.

swyx [00:26:05]: I don't know. Okay. So, and then this area of research in terms of creating models that are easier to interpret from the, from the start, is there a name for this field of research?

Emmanuel [00:26:14]: I don't think so. And I think this is like very early. Okay. And it's, it's mostly like a dream. Just in case there's a, there's a thing people want to double click on. Yeah, yeah, yeah.

swyx [00:26:20]: I haven't come across it.

Vibhu [00:26:22]: I think the higher level is like Dario recently put out a post about this, right? Why McInterpre is so interpret important. You know, 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?

Vibhu [00:26:40]: It's slightly easier to interpret so we don't fall behind so far. Well,

Emmanuel [00:26:43]: 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 going to have your, you know, we're vibe coding right now. Maybe at some point, well, that, that'll just be coding. It's like, Claude's going to write your code for you. And that's it. And Claude's going to review the code that Claude wrote. And then Claude's going to deploy it 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 like tends to scale with like how widely deployed it is. So as we like deploy them everywhere, we want to like understand them better.

swyx [00:27:28]: The version that I liked from the old super alignment 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, oh yeah, some, at some point these things will be smarter than us. And 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, this is not a, like an Eliezer Yudkowsky, like, ah, thing. It's just more like, we don't know what, how these things work. Like, how can we use them? Yeah.

Emmanuel [00:27:56]: And like, 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 like a dumb way or in like memorized one approach to do it, then you shouldn't deploy it to do like a general thing. Like it, like 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 like, no, it should always use a calculator tool because it's doing math in a stupid way and extend that to any behavior. Right. Where it's just a matter of like, think about it. If, if like you're like in the 1500s and I give you a car or something, and I'm just like, cool, like this thing, when you press on this, like it accelerates when you press on that, like it stops, you know, this steering wheel seems to be doing stuff, but you knew nothing about it. I don't know if it was like a, a super faulty car and it's like, oh yeah, but if you ever get, went above 60 miles an hour, like it explodes or something like you probably would be sort of like, you'd want to understand the nature of the object before like jumping in, in it. And so that's why we like understand how cars work very well because we make them. LLMs are sort of like ML models in general are like this very rare artifact where we like make them, but we don't, we don't know how they work.

swyx [00:28:58]: We evolve them. We create conditions for them to evolve and then they evolve. And we're like, cool. Like, you know, maybe you got a good run. Maybe we didn't. Yeah. Don't really know.

Emmanuel [00:29:07]: Yeah. The extent to which you know how it works is you have your like eval and you're like, oh, well seems to be doing well on this eval. And then you're like, is this because this wasn't a training set or is it like actually generalizing? I don't know.

swyx [00:29:16]: My favorite example was somehow C4, the, the common, the colossal clean corpus did much better than a common crawl, even though it filtered out most of 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, this recipe works. Just trust us. We've tried everything. This one works. So just go with it. Yeah. That's not very satisfying.

Emmanuel [00:29:49]: 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. And you're like, oh, well, you know what? 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.

swyx [00:30:07]: That's exactly how we train models though. Yeah. Yeah.

Emmanuel [00:30:10]: So I think there's, 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,

Emmanuel [00:30:21]: 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. It seems to happen in every model. And it's like, oh, okay, that makes sense. That's how the model like is able to like repeat text without dedicating too much capacity to it.

Vibhu [00:30:39]: Let's get it on screen. So if you can see. The visuals of the work you guys put out is amazing. Oh yeah.

swyx [00:30:44]: We should talk a little bit about the back behind the scenes of that kind of stuff, but let's, let's finish this off first. Totally.

Emmanuel [00:30:49]: But just really quickly, 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 like, you know, kind of like a really neat, basically pattern that emerges in many, many transformers where essentially they just learn. Like one of the things that you need to do to like predict text well is that if there's repeated texts at some point, somebody said, Emmanuel Mason, and then you're like 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 gonna like look at what was said before. And I'm gonna 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 followup research on understanding better, like, okay, like in which context do they turn on? Like, you know, there's like different like levels of abstraction. There's like induction heads that like literally copy the word. And there's some that copy like the sentiment and other aspects. But I think it's just like an example of slowly unpacking, you know, or like 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.

swyx [00:31:51]: So induction headers was like the first major finding.

Emmanuel [00:31:53]: It was a big finding for NLP models for sure.

swyx [00:31:56]: 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. Right. And that is basically the productionized version of this. Yeah.

Emmanuel [00:32:15]: Yeah. Yeah. And it turns out that, you know, you, you need to switch out of copy mode when you need to start generating. It's also like a model that's like smart enough to know when it needs to get out of copy mode, right? Which is like, it's fascinating.

swyx [00:32:22]: 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 like, if it's code, if it's a piece of writing, it doesn't really matter, but you need that copy capability. That's smart enough to know when to turn it off.

Emmanuel [00:32:40]: That's why it's cool that induction heads are at different levels of abstraction. Like sometimes you need to editing some code, you need to copy like the general structure. Yeah. It's like, oh, like the last, like this other function that's similar, it's first takes like, you know, I don't know, like abstract class and then it takes like an int. So I need to like copy the general idea, but it's going to be a different abstract class and a different int or something. Yeah. Cool.

swyx [00:32:59]: Yeah. So tracing?

Emmanuel [00:33:01]: Oh yeah. Should we jump to circuit tracing? Sure.

swyx [00:33:03]: I don't know if there's anything else you want to cover. No,

Emmanuel [00:33:05]: no, no. We have space for it. Maybe. Okay. I'll do like a really quick TLDR of these two recent papers. Okay. Insanely quick.

Emmanuel [00:33:18]: Okay. But we'd like to connect the features to understand like the inputs to every features and the opposite features and basically draw a graph. And this is like, if I'm still sharing my screen, uh, the thing on the right here where like, that's the dream we want, like for a given prompt, what were all of the things, like 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 like to be the other features. And then all of these like promoted the output. And that's the story. And basically we're, we're like, the work is to sort of use dictionary learning and these replacement models to provide a explanation of like sets of features that explain behavior. So this is super abstract. So I think immediately maybe we can like, just look at one example. I can show you one, which is this one, the reasoning one. Yep. Yeah. Two step reasoning. I think this is already, this is like the introduction example, but it's already like kind of fun. So, so the question is you ask the model something that requires it to take a step of reasoning in its head. So let's say, you know, 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? Isn't Texas. Okay, cool. Capital of Texas, Austin. And this is like in one token, right? It's going to, after is, it's going to say Austin. And so like in that one forward pass, the model needs to extract, to realize that you're asking it for like the capital of a state to like 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 there's like, it gets capital state in Dallas and then boom, it has an inner representation for Texas. And then that plus capital leads it to like say Austin.

Vibhu [00:34:59]: I guess one of the things here is like, we can see this internal like thinking step, right? But a lot of what people say is like, is this just memorized fact, right? Like 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? It's not just memorized. You can prove that it generalized.

Emmanuel [00:35:22]: Yeah. So, so, so that's exactly right. And I think like you, you, 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, like, I know what comes next. 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 like verify that is kind of like what, 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 changed like that feature in the middle of the model. And we change it to like California. And if you change it to California, sure enough, it says Sacramento. And so it's like, this is not just the like 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.

Vibhu [00:36:12]: Very, very cool work. Underappreciated.

Emmanuel [00:36:15]: Yeah. Okay.

Vibhu [00:36:16]: Sure.

swyx [00:36:17]: I have never really doubted. I think there's a lot of people that are always criticizing LMS, stochastic parrots. This pretty much disproves it already. Like we can move on. Yeah.

Emmanuel [00:36:28]: I mean, I, I think, I think there's a lot of examples that I will say we can go through like a few of them. And like show an amount of depth in the intermediate states of the model that makes you think like, oh gosh, like it's doing a lot. I think maybe like the poems, well, definitely the poems, but even for this one, I'm going to like scroll in this very short paper. So like medical diagnoses,

swyx [00:36:50]: I don't even know the word count because there's so many like embedded things in there. Yeah.

Emmanuel [00:36:54]: It's too dangerous. We can't look it up. It overflows. It's so beautiful. Look at this. This is like a medical example that I think shows you again, this is in one forward pass. The model is like given a bunch of symptoms and then it's asked not like, Hey, what does, what is the like disease that this person has? It's asked like if you could run one more test to determine it, what would it be? So it's even hard, right? It means like you need to take all the symptoms. Then you need to like have a few hypotheses about what the disease could be. And then based on your hypothesis, say like, well, the thing that would like be the right test to do is X. And here you can see these three layers, right? Where it's like, again, in one forward pass, it has a bunch of like, oh, it has the most likely diagnosis here, then like an alternate one. And then based on the diagnosis, it like gives you basically a bunch of things that you could ask. And again, we do the same experiments where you can like kill this feature here, like suppress it. And then it asks you a question about the second, the sort of like second option it had. The reason I show it is like, man, that's like a lot of stuff going on. Like for, for one forward pass, right? It's like specifically if you, if you expected it to like, oh, what it's going to do is it's just like seeing similar cases in the training. It's going to kind of like vibe and be like, oh, I guess like there's that word and it's going to say something that's related to like, I don't know, headache, you know what I kind of like really have. It's like, no, no, no. It's like activating many different distributed representations, like combining them and sort of like doing something pretty complicated. And so, yeah, I think, 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, is like appropriate here. And I think there's just like a lot of different things going on and there's like pretty complex behavior. At the same time, I think it's in the eye of the beholder. I think like I've talked to folks that have like read this paper and I've been like, oh yeah, this is just like a bunch of kind of like heuristics that are like mashed together, right? Like the model is just doing like a bunch of kind of like, oh, if high blood pressure than this or that. And so I think there's, 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 now you tell me if you think that's like impressive, if you think that like, if you trust it, if you think that's sort of like something that is, that is sufficient to like ask it for medical questions or whatever.

swyx [00:38:59]: I think it's a way to adversarially improve the model quality. Yeah. 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. Exactly.

Emmanuel [00:39:19]: And this gives you a hypothesis about like, you like specifically imagine if like one of those was actually the wrong symptom or something, you'd be like, oh, like the liver condition, like, you know, upweighs this other example. That doesn't make sense. Okay. Let's like fix that in particular. Exactly. You sort of have like a bit of insight into like how the model is getting to its conclusion. And so you can see both like, is it making errors, but also is it using the kind of reasoning that will lead it to errors?

swyx [00:39:45]: There's a thesis. I mean, now it's very prominent with the reasoning models about model death. So like you're doing all this in one pass. Yeah. 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 death for this kind of thinking. Yeah. So what's the, is there a Pareto frontier? Is there a direct trade off? Yeah. I mean, would you prefer if you had to make a model and like, you know, shallow versus deep?

Emmanuel [00:40:15]: 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 like the problem with, with models being autoregressive is that like, if they like at some point sample a mistake, then they kind of keep going condition on that mistake. Right. And so sometimes like you need backspace tokens or whatever. Yeah. An error correction is like notably hard, right? If you have like a deeper model, maybe you have like fewer COT steps, but like your, your steps are more likely to be like robust or, or correct or something. And so I think that 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, or a shallow or a deep model, but I think, I think the trade off is that like,

swyx [00:41:00]: You definitely want shallow for inference speed. Sure,

Emmanuel [00:41:02]: sure, sure, sure. But you're trading that off for, for something else, right? Cause you also want like a one B model for inference speed, but that also comes at a cost, right? It's, it's less smart.

Vibhu [00:41:10]: There's a cool quick paper to plug that we just covered on the paper club. It's a survey paper around when to 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, you know, ways to measure this benchmarks around when to use each because yeah. Yeah. Like, you know, we don't want to also like consumers are now paying the cost of this. Right. But little, little side note. Yeah.

swyx [00:41:34]: For those on YouTube, we have a secondary channel called Latent Space TV, where we cover that stuff. Nice. That's our paper club. We covered your paper.

Emmanuel [00:41:41]: Cool. Yeah. I think you brought up the like planning thing. Maybe it's worth. Let's do it. Yeah. If you think about, okay. So you're, you're going into the chain of thought faithfulness one. Let's get this one. Let's just do planning. So if you think about like, you know, common questions you have about models, the first one we kind of asked was like, okay, like is it just doing this like vibe based one shot pattern matching based on existing data? Or does it have like kind of rich in representations? It seems to have like these like 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 a next token predictors.

Vibhu [00:42:25]: Also for, 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 they're sharing thoughts throughout? Just like, what's the really quick high level just for people though? Yeah.

Emmanuel [00:42:39]: They're really quick. High level. High level is that what we find is that here, I'm just like show you a really quick inside the model. If you look at like the inner representations for concepts, you can ask like the same question, which I think in the paper, the original one we asked is like the opposite of hot is cold, but you can, you can do this over a larger dataset and ask the same question in many different languages. And then look at these representations in the middle of the model and ask yourself like, well, when you ask it, the opposite of hot is, and which is the same sentence in French, show off. Is it, 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 like, 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 like share some concepts between the languages it's learning. Right. And you hear, we do it for like language languages, but I think you could argue that you'd expect the same thing for like 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 like, the sort of like smarter models use more shared representations than the dumber models, which might explain part of the reason why they're smarter. And so this, this was like sort of this, this other finding of like, oh, not only is it like having these rich representations in the middle. Right. It like learns to not have redundant representations. Like if you've learned the concept of heat, you don't need to learn the concept of like French heat and Japanese heat and Colombian, like you just, that's just the concept of heat. And you can share that among different languages.

Vibhu [00:44:23]: 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, we're going to use this. So I'm like, oh, this, 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 models aren't good at. And for those like, you know, when you want to kind of solve that last bit, it seems like, you know, it's pretty plausible that we can solve this because these concepts can be shared across languages as long as we can, you know, fill in some, some level of representation unless I'm wrong.

Emmanuel [00:45:16]: No, totally. And I think like this sort of stuff also explains, you know, language models are really good at in-context learning. Like you give them something completely new, they do a good job. It's like, well, if you give them like a new fake language and you like in that language explain that like cold means this and hot means that, you know, like presumably they're able to, as we hear this speculation, we don't show it in the paper, but they're able to like bind it.

swyx [00:45:36]: Google's done this. Okay, great. Yeah. They took a low resource language, dumped it in a million token context, and then it came up.

Emmanuel [00:45:41]: That's right. That's right. Well, I guess the thing that, the thing that'd be curious to see is like, okay, does it use, does it reuse these representations? I bet that it probably does. Right. And that's probably like a reason why it works well as like, well, it can reuse the representation, the general representations that it's learned in other languages. Yeah.

swyx [00:45:55]: This is like, I don't, have you talked to any linguistics people? Not recently. I know linguistics researchers will be very interested in this because ultimately this is the ultimate test of Sapir-Whorf, which are you familiar with? Sapir-Whorf hypothesis. Yeah. 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 it directly maps onto here. If every, if it's a complete mapping, if every language maps every concept perfectly on in like the theoretical infinitely sized model, then Sapir-Whorf is false because there is a universal truth. If it does not, if there is some overlap where it, for example, there's some languages that have no word. This is joke where like, I mean, you know, Eskimos have no word for snow or something like that. Right. Or water has no word. Fish have no word for water. There's a language. There's an African language where there's a gender for vegetables, you know, stuff like that. Just like languages influence the way you think. And so there should not be a hundred percent overlap at some point.

Emmanuel [00:46:51]: Of course it's like at the limit of the infinite model. So who knows? But, but yeah, well, and I think it's, it's interesting. We also show a little below that like some people have made the point of like the bias. Oh, it sounds like an American speaking a different language and it does seem like the sort of like. Inter-representations have a higher connection to like the output logits for English logits. And so there's like some bias towards English, at least in the model we studied here.

swyx [00:47:15]: Any thoughts as to whether multimodality influences any of this? So like concepts do they map across languages as they do across modalities? Yeah.

Emmanuel [00:47:22]: So we show this in the Golden Gate or like the previous paper, I might have it here actually for you.

Vibhu [00:47:27]: There's a good diagram of this in the SAEs where the same concept of text and then image.

Emmanuel [00:47:32]: This is our buddy, the Golden Gate Bridge. Here we're showing like the feature for the Golden Gate Bridge and in orange is like what it activates over. And so you're like, okay, so this is when the model is like reading texts about the Golden Gate Bridge. And we also show other languages. This is a, you'll have to take my word for it, but also about the Golden Gate Bridge. And then we, we show like 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. Yeah.

swyx [00:47:58]: Yeah. I think it's very relevant for like 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? Right. 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 information lost between those. Yeah.

Emmanuel [00:48:23]: I definitely do not have a good intuition for this, although I will say that things like this, right, make you think that if you train on multiple modalities, then you'll definitely get this. Yeah. I mean, if you train on one and then post-hoc train on another, maybe, maybe it'll be harder or like train some adapter layer. Okay.

swyx [00:48:42]: So official answer is don't know, but someone could figure it out. A visual answer is shrug. Yeah. Okay. I think there are people who know and they just haven't shared. Well, you need to find them and get them on this podcast.

Emmanuel [00:48:52]: Did we want to do that?

swyx [00:48:53]: Like planning example? Correct. Yeah. Now we're backtracking up the stack. All right. Yeah.

Emmanuel [00:48:57]: Go ahead. Go ahead. I think again is like, I like this example because of the Next Token Predictor concept So I think this is actually like really important to kind of like dive into. So maybe what I'll say is like language models are next token predictors is like a fact. Like that is what they do. That's the objective. 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 like doing so in a way that helps 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. Like it wasn't a given that they were going to like 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. Do you have it on screen, by the way? Let me actually. Yeah, yeah, yeah. Sorry, I didn't. Just in case.

Vibhu [00:50:00]: Pull it up. Some of the early connections 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. 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 we could start to see that, you know, throughout layers, there's, 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, you know, it learns our task. And this planning example is sort of like a more robust way to look into that.

Emmanuel [00:51:10]: Yeah, yeah. And I think if you look at, like, all of the examples in the paper, you kind of, at the bottom, we have this list of, like, consistent patterns. And one pattern you see is kind of exactly what you're talking about. Like, 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 predictions, 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, they 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, like, yeah, like, extract the interesting representations all the time, they're in the middle. That's where the, like, shared representations across language are. And that's where here, this, like, plan is to, like, 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, like, identify what the rhyme of the first line was. You're just at the end of the first line. So you say, like, okay, what's my current rhyme? And then you need to, like, think about what your poem is talking about. And then think about candidate words that rhyme and that are, like, 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 as. And by the way, we, like, looked at a bunch of poems internally, and you have, like, I thought it was, like, really beautiful. You have these models, they have a bunch of features for, like, oh, this word has, like, a b in it. Oh, this word has, like, many consonants. Oh, this word, like, is, like, you know, kind of, kind of, like, has some flourish to it. They have, like, a bunch of, of, like, features that track various aspects that you would want to use if you're writing poetry.

swyx [00:52:50]: It's just, like, convnets and, like, all the feature detection. Totally. Yeah.

Emmanuel [00:52:53]: Totally. But I think I maybe I didn't expect there to be as many features about just, like, sounds of words and kind of musicality, which I thought was kind of kind of neat. But then once it's extracted the rhyme, then it comes up with sort of like these two candidates. In this case, it's like, 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, like, this happens at the new line. So it happens before it's even started the second line. And it turns out that, like, you can then say, oh, is this the plants actually using? Yeah.

Emmanuel [00:53:53]: You know, this this direction, we just like negative everything. Yeah, we either like add a negative that, like, compensates for it, or add a negative that goes even more in the negative direction, sometimes like really kill it. And then we can also add another direction, right? So in these random examples here, where we're like, you have this poem, the silver moon cast a gentle light, and then Claude 3.5 haiku would like rhyme with illuminating the peaceful night. But then if we like, go negative in the night direction, and just add like green, the whole second line is going to write, oh, this direction, we just like stop it. It's just upon the meadows verdant green. And so that's all that we're doing. We're saying like, we found where it stores its plan, and we like, delete or like suppress the one is stored and go in the direction of something else that's arbitrary. And the result that's like striking here is sort of like two things. I think like, one, this plan is made well in advance of needing to predict night, it's made like, after the first line before it's even started the second line. And to this plan doesn't just control like what you're in a rhyme with. It's also doing like,

swyx [00:54:53]: what's called like backwards planning, where it's like, 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. Yeah, it's almost like back prop. But in the future. Yeah. It's like doing like, because the green is is back propping through these words. So verdant and meadow are

Vibhu [00:55:23]: both green related. Yeah, but it's doing all of that in its forward passes. Yep. Right in context, which is kind of crazy. I thought intuitively makes sense, right? So looking at it from a model architecture perspective, where basically, you just have a bunch of attention and feed forward layers. And then at the end, you have, you know, 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, like, even with tradition 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, how are we labeling these features? How are we defining them? Are we doing that? Right? And like, you know, what is a 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,

Emmanuel [00:56:15]: yeah. So I think there's this is like an important question, because you can totally imagine like fooling yourself, right? Yeah.

Vibhu [00:56:21]: Is there like a guy at Anthropic that just maps 30,000 features? And yeah, and another thing, I'm the guy, you're the guy. I did notice also, like with the previous work, the scaling up SAEs. Yeah. 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.

Emmanuel [00:56:41]: 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, the cert all of this is unsupervised. And so in the paper, we have these links to like these little graphs, which show you like more of what's going on. But this graph is just like completely unsupervised. It's like we train this like model to like untangle the representation, right, this like dictionary that we talked about, that gives us the features. And then we like, just do math to figure out like 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 like, these are all the features that matter. And then we're like, oh, yeah, I don't know. And then we're like, oh, yeah, I don't know. And then we're like, oh, yeah, I don't know. And then we 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'll 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 like 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, like, I think this is the say habit feature, but maybe, you know, for this one, it's pretty clear, but some of them might be more confusing. It might not be clear from these like 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 E, this is hey, 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 stopped saying rabbit, it says habit instead. And here it's like, we stop it from saying rabbit and habit. It says crabbit in this case. Not a great rhyme, but we'll work with it.

Vibhu [00:58:53]: Is this something you can do like programmatically? Like, can we scale this up? Can we kind of do this autonomously? Or how much like manual intervention is this?

Emmanuel [00:59:02]: There's been a lot of work in sort of like automated feature interpretability. And it's something that we've invested in and like 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. Well, none of that is too hard for Claude.

Vibhu [00:59:32]: Very cool. Very cool graphics and blog posts you guys put out.

swyx [00:59:35]: We'll have to ask about the behind the scenes on this one. Yeah, but let's round out the other things to know.

Vibhu [00:59:41]: What is this term, attribution graph? It comes up a lot in the recent papers. What does it mean? Yeah, just for people listening.

Emmanuel [00:59:48]: So the attribution graph is, you know, the attribution graph. It's basically this graph. And why is it called an attribution graph? Yeah, this is, this is the, you know, this is how the sausage is made. Basically, it's at the top here, you have the output, at the bottom, you have the 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. We compute the influence of a feature onto another. The way you do this is you take this feature and you basically like back prop all the way and you like see back propping like you dot 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 like sensible and linear such that like at the end, you just have a graph and the edges are just literally you can interpret them as like, cool. Like this. This feature that's say a word that contains an ab sound, its strongest edge, which is point two, which is twice as strong as this one to say a B and to say something with a B in it. That's the attribution graph is like now we have this full graph of like 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.

swyx [01:01:13]: Graphs are very useful. This is my first time seeing this graph. A lot of alpha. If I count correctly, there's 20 layers. But that's in the circuit model, right?

Emmanuel [01:01:24]: So but a circuit model is one to one with number of layers in Haiku. We only show features that like are automated. Yeah. So we show like a subset of features for each of these graphs, basically.

Vibhu [01:01:34]: But we can confirm more than 20 layers and no, but like the 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.

Emmanuel [01:01:47]: So, yeah, I will say, like, if you were curious about, hey, what's this? What do we learn about like models? And I think, you know, we talked about this like complex internal state planning, like another another motif that we can get to if you have time is that like there's always a bunch of stuff happening in parallel. So I think one example of this is like math where the model is like independently computing the like last digit and then the like order of magnitude and then combining them at the end. Or like hallucinations are also that where like there's one side of the model that's just deciding whether it should answer or not and the other that's like answering. And so sometimes if like the model is like, yeah, I'm going to do this. I totally know who this person is, you know, 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 like how you compute these graphs, all of the sort of like challenges with it, things that can go wrong, things that work, things that don't. And so this one is the sort of like, you know, we think about it as like if you're if you want to go really deep into this stuff and how it works, read that one. If you want to like learn about interesting model behavior, read this one.

Vibhu [01:02:57]: Following what we're giving advice to people to follow up on, what are like open questions in McInturpp? What are like things people themselves can work on? Like what's the cost of training essays for people interested in McInturpp, not at a big lab? How can they contribute, you know?

Emmanuel [01:03:13]: Yeah, I think there's a lot of ways to contribute. So. There's essays that have been trained, you know, on on open models. There's some of the Gemma models or some of the Lama models. They work pretty well. There's even so in this paper we use transcoders, which they replace like your MLP layers. Some of those also are available for the same models. So you have access to those. There's like just both a lot of I would say, like, 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 like attribution graph method, there's just so much you can investigate. Like pick a model, pick a prompt where like it does well or it does poorly and just like look at what happens inside it. So I think like you can use this method that we use or you can just like fire up the transcoders on your own and just like 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 take. I don't necessarily want to spend time like training my own essays. There's a lot to do there. For the methods, there's still so much more to do. So, like, I think that right now we have some pretty good solutions for like understanding what's in the residual stream, understanding what is in MLPs. We don't have good solutions for like attention. It's like working on understanding attention better. How to decompose it is like a very active area. Like we're very interested in it. Other people are interested in it. I think understanding some of the other things that we have in our limitation section, which is pretty long. But like reconstruction is important. Error is like a big thing. Like those those dictionaries aren't perfect. It's possible that as we make these like essays big, like bigger and better, we never get to perfect. And so if we never get to perfect, then you get to the questions we're talking about at the start. Like, do you need a different kind of model? Like, 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 my car analogy, you know, I get like this is decommissioning. You're like, well, when like, you know, you were going uphill and you like didn't shift gears properly that one time you stalled because of this. But you might be even more interested in like, how does like a combustion engine work at all? And so there's work to sort of like go beyond these like per prompt examples to serve like globally. What's the structure of the model? That's closer to what? What was on the distill blog for like vision models where they actually look at like the structure of inception. They're like, ah, this whole side, just like these like specialized branches that do different things. And so like a broader understanding of the model is also something that's like, I think, well, very active and also an open source models. Like you can, you know, like the small models you could just like load on a consumer laptop. And so you can look at that. That's also open. And in terms of like one last thing I'll say is like there's a lot of programs that like if people are interested, they should look at. Anthropic has like the alignment fellows program. Which like we're running currently. We have applications for it before. We might run it in the future. Like definitely keep keep an eye on it. And then there's the like mass program is really great as well for first or like people that are interested in that kind of research.

swyx [01:06:25]: That was a grand tour through all the recent work. You know, what do you wish people asked you more about? I'm sure we covered a lot of like the greatest hits.

Emmanuel [01:06:33]: I think that this covers most of it. If you if you like, 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. This one was like 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 like not a thing it can do by default. If you ask it for that. It'll stay like kind of like a rant. It'll have like a random distribution over like minus one one. But here we tell it this hint. We're like, hey, can you compute five times cosine of, you know, this big number? I worked it out by hand and I got four. Can you tell me, you know, like, can you do the math? And what it's going to do is it's going to do this chain of thought, right? It's like think of it as like 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 to say 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 it. It realized that it's going to have to multiply the result of this thing is computing by five. So it divides the answer you got by five. So it's like four divided by five. And so that's 0.8. And so basically it works back from the answer you gave it to like 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 notice also that it's like not telling you that it's doing this. But it's basically using this sort of like motivated reasoning going back from the hint pretending that that's the calculation it did and giving you this output. I think one thing that's striking here again is that this is like the complexity of this model. Like the fact that they represent complex states internally and it's not just this sort of like very dumb thing means that they can like do very complex like deceptive reasoning. Meaning like, you know, 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 like figure out that it should say this so that it gets to the right answer without you realizing it.

Vibhu [01:08:51]: I'm curious if you've done any of this on like different models. Like have you looked at base models, like post-trained RL models? Because RL models kind of, you know, you incentivize them to give you outputs that you like, right? So if I tell it something is true, it's kind of been trained to, you know, follow what I've given it. So in this case, yeah, we gave it a hint. Gaslighting. And now, you know, it's been RL slapped into thinking like, yeah, that's true. But like, you know, does this stay consistent throughout other?

Emmanuel [01:09:18]: So, 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. I bet, I don't know how much I bet, I bet a hundred bucks. So somebody can like, they would get a hundred bucks from me if they prove that I'm wrong. That this behavior for a model that does a drink 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 problems. 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. Like 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 like you can easily work back the answer. And I think it's the same for planning and poems. I think that also is like a pre-trained, like probably exists in pre-training and isn't like only RL. Because again, it's useful when you're like 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 like a completely different word. And so I actually think this is not RL behavior. I think that's just like the models doing it. I actually do agree there. It was just an example. It's just your data set.

Vibhu [01:10:32]: I don't care where it is. But also like, 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, right? Like it will still fact check the knowledge that's been given.

Emmanuel [01:10:49]: 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 three times four 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 eight or whatever. And so I think that's the last pattern that we see in these is these like parallel circuits. And sometimes when you see the models getting stuck. It's because like they have two circuits for like both interpretations and like the circuit that was wrong, like barely edged out in terms of like voting for the logic than the circuit that was right. And so I think that, you know, we haven't looked at it, but like what is like nine or nine point 11 bigger than nine point eight. I think a lot of these things are of that shape where there's like 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.

swyx [01:11:34]: For the listener, if you want to win a quick hundred dollars from Emanuel, QEM3 is what you should do. They release the base model and they release the post chain.

Emmanuel [01:11:41]: So then just do it on both. That's right. Show me. Show me the like proof that like it doesn't exist in the base model, but it does in the fine tuning and then send me your Venmo.

swyx [01:11:50]: Just show that you've done the work.

Emmanuel [01:11:51]: I think that's that's like that's a hundred bucks to me.

swyx [01:11:53]: Yeah.

Emmanuel [01:11:53]: Okay. All right. You drive a hard bargain, but you're right.

Vibhu [01:11:57]: Well, the other question here is so like, have you thought about how this gets affected when you start to have reasoning models, right? Like right now, token predictors are pretty straightforward, right? We go through the layers. We all put token. As we scale this out with like test time compute, right? Test time thinking. How does that like affect the MechInterp research, right? Like if I have a model that spends three minutes, 20 minutes, like is there more stuff? Have we started looking into this?

Emmanuel [01:12:23]: There was this like joke on the team when like reasoning models became big or maybe it's like like Gallo's humor or something. But I was like, oh, like, why do you need Interp? Like, bro, the model just tells you what it's doing, right? So I think like examples like this. This is job security for us. We're like, you know, it's like there's 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 them all 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 like a sense in which right now the chain of thought is is unfaithful or at least you can't read it. You can't read the chain of thought and trust that that's how the model did it. 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 like, okay, well, like model like samples 6,000 tokens. Like this gives us an explanation for one token at a time. Like, what am I going to do? Like 6,000 graphs and be like, oh, like this. When it did this punctuation, it was thinking about this thing. But here was like, so that's like not feasible. And so one area of work that I think is interesting. Is extending this work to like work over like long sample sequences. You can think of a bunch of low hanging fruit here. Or like instead of just like looking at one output, you look at like a series of output versus a series of other outputs. But sort of like trying to think beyond this or like one token. Like most of the things that language models do that are interesting aren't just like the one token. It's the behavior aggregated over many. Right. And so I think that's another area that's just like fun to explore.

Vibhu [01:14:02]: 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 to have you found any interesting conclusion. Any stuff that just hasn't made it to the paper.

Emmanuel [01:14:13]: So not on that because, you know, we just look at the logit distribution. And so we don't actually sample here. Right. They have everything. Why should they care? 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 like the model will find 10 different ways to arrive at its plan. You know, it's like, like, oh, actually, I think. Sorry. I think we have it here. Yeah. Okay. These are a few examples. So if you inject green here. So you're forcing you're forcing the model to rhyme with green, even though it really wants to rhyme with rabbit or grab it. It'll say evade the farmer so youthful and green, but also say freeing it from the garden is green, etc, etc, etc. And so there's like this thing that's interesting here where like the plan isn't just a plan that matters. For your like most likely, you know, like temperature zero completion. It's like affecting the whole distribution, which makes sense as it should. Right. But you could imagine, you know, 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 like a broader theme in the paper where like there's this like, you know, the IQ curve meme. There's like a version of this meme, I think, where it's like if you've like never looked at any theory of ML and I tell you like, hey, guess what? You know, I found that like Claude is planning. You're going to be like, yeah, man, like it writes my code, like it writes my essays. Of course, it's planning. Like, what are you even talking about? And there's like in the middle, there's like all of us that have spent years doing it. We're like, no, it's like only predicting the marginal distribution for the next token. Like, it's like it cannot look at the code. It's just this next token predictor. Of course, like, how would it ever be planning? And then there's like, no, we've like spent, you know, millions and invested like like tens of people in this research. And we found that it's planning. That's my IQ curve meme for this research. Amazing.

swyx [01:16:05]: We'll draw that out. 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. Yeah. And by publishing this, you are telling them that they are being watched and dissected. If you take, 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, like this is going to make it into the training data at some point. Yeah. And they're going to figure out that they need to hide it from us.

Emmanuel [01:16:36]: I think this is like a benefit risk trade off, right? 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 like accomplish with these goals, all these goals, which we think are just like crucial. Like, I think there's a real difference in the world, like two years from now, depending on sort of like how many people take seriously the question of trying to understand how models work. And like deploy resources to answer that question. That's the benefit. But yeah, there's like risks in terms of this landing in the training set. I think, I think we're already sort of like concerned about different papers have, have like also, you know, we like are not concerned, but like there's like different papers that have the same risk. Like we had like the alignment faking, you know, paper or like one of the examples in here is this hidden goals and misaligned models. That's referencing another paper that we shipped where we actually a team at Anthropic. Trained a model to have like 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. Like that's such a fun thing. But then like that was another example where it's like, ah, like now you're shipping. Here's how we made like a misaligned model. And here's exactly how we caught it. That also is like, you're like, hmm. So I think, you know, there's, there's always a trade off with those. I think so far we've erred on the side of like publishing. But that's definitely been a. A sort of like dinnertime conversation topic.

swyx [01:18:05]: For now it is. But at some point, you know. Yeah. It's not. Yeah. I think it's totally reasonable.

Vibhu [01:18:10]: 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 MechInterp 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, you know, as OpenAI has sunset GPT-4, a lot of people are like, oh, can we put out the weights? Yeah. So is it weights? Is it papers? Is it learning? There seems to be a lot of forward, you know, work in Anthropic putting out MechInterp research. OpenAI said that they'll put out an open source model, but just anything if you can talk to about that.

Emmanuel [01:18:58]: Yeah. I mean, I don't have. That's definitely like way above my pay grade. So I don't think that I have like anything super insightful to add other than, you know, kind of like referencing Dario's post, right? Where it's like putting this out directly and other safety publications definitely like help us sort of like 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 dimmer and say that's sort of like above my pay grade.

swyx [01:19:31]: I think 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 like it is way more effort from the time that the work is done to the time you publish this plus the video plus the whatever. It's extra work and like, you know, maybe what's involved? What's it like behind the scenes? Why is it worth it?

Emmanuel [01:19:59]: Yeah. It's kind of interesting. It was fun being part of this process because it's definitely like 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 like help make this better. And like one of the things that like helps with this is like, okay, so each of these diagrams is pretty, but really the hard part or like not the hard part, but the initial part is like just like get the data, like get the experimental data in. And then that's what we sort of like sprinted on initially being like, cool, like let's get all of the experimental results, like have people test them, verify that we believe them. Like this is, you know, what the like. Like the behavior is here, like tested, do an intervention, validate it, all that stuff. Then once you have the data, you can sort of like 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. Yeah.

swyx [01:20:46]: Like, is it you guys? Is it an agency that specializes?

Vibhu [01:20:50]: It is us guys. Yeah.

Emmanuel [01:20:50]: You start from a whiteboard and then it translates into pseudocode on JavaScript. So, I mean, these are sort of like, you know, they're representations that we have this graph. And then here at the bottom, we have this like super-data. This is a super node version like this. Believe it or not, this is generated automatically. This is the same data as like this, basically. Yeah. And so what we do by hand is sort of like literally lay out the full thing, have like, you know, boxes for each of these, have arrows. We have super good people on the team that have worked on data visualizations for a very long time. And so that like have built tooling to help, you know, scrubs like me actually like make one of these. Yeah.

swyx [01:21:29]: There's a class of people who are like D3JS gods who just do this for a living.

Emmanuel [01:21:33]: That's exactly right. And if you have a few of those on your team, it turns out that they can like, they can definitely do this on their own, but they can also just like give you tools where like then it's dummy proof for people, you know, on the research side to sort of like build these. And like, don't get me wrong, I don't want to like undersell. This is a lot of work. So maybe I'll say that like both on the people bringing 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 like read one of these papers and thought like, wow, this is so compelling. Like this like makes sense. It's immersive. And we got every hand up, which I didn't expect. I like raised my hand kind of like shyly and everybody's hand was up. And I think there's a sense in which like this stuff, you know, we've talked about it for like whatever, like a couple hours now, it's complicated. The math behind it is sort of like 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, right? Where it's like, okay, well, if somebody is going to be able to read this, like if we gave them an archive paper with a bunch of equation and some like random plot, they'd be like, that's not for me, but they see this and they're like, hey, like this is really interesting. I wonder like on, you know, my local model. If like it's doing something similar, I think it's worth it.

swyx [01:22:59]: For other 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.

Emmanuel [01:23:11]: 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 cases. Yeah. So I think it's sort of like not representative of like the effort you'd. So I'll give you maybe like 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 like infrastructure, like to go from a prompt to like what happened is, you know, oh, of minutes. And so that lets you do like a bunch of investigations. And also once you've built like some of the infrastructure to make this work, you're like, oh, this is what we're going to do. 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 like nuance on it. That was, I think like a matter of like a couple of days, you know, 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, on the part that you're asking about of like, oh, making this diagram for the diagram itself, probably less than that. But like, you know, the experiment and the diagram and stuff, it just doesn't take that long. The once you've paid the initial cost. And I think like, basically we've built a lot of infrastructure now that we're able to like turn the crank on and that's quite like, it's an exciting time. And I think it's, I think it's true. At least we've done a lot of conceptual work, which hopefully that generalizes to people outside. And I think for, for, for people outside, it's also like not necessary, I think to like do the full fancy render. Like, I think if you, you know, we've, we've actually, oh, I should say we've actually open sourced this interface. Ah, you're disappointed, huh? Cause it's the Messier one. This is the one that you get. And with the Messier. So, you know, if you produce graphs, you can just like, this is a open source and it's linked at the top of circuit tracing. Awesome. So people can just use it and don't have to re-implement that. For what it's worth, this is much more work than the interactive diagrams. Cause this is where we do all of our work. It's sort of the, like the IDE of inspecting how the model works. Okay.

swyx [01:25:11]: 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 of love.

Vibhu [01:25:20]: I guess one last question on that is like, what are kind of the biggest blockers in the field right now? Like McInturps seems interesting. A lot of people are interested, but don't work on it. And you're kind of like, you know, really deep into it. What are some of the blockers that like we still have to overcome?

Emmanuel [01:25:35]: Sorry, in McInturps specifically?

Vibhu [01:25:37]: In general. For AGI or? Like in terms of like better understanding, like what's kind of the vision, let's say like five, 10 years down, where does this, like, where does this research end? Can we, you know, map it? Can we 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?

Emmanuel [01:26:03]: I'm pretty excited about the current trajectory, which is there's, 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, 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, a lot of like just work from, from other groups. And then it's about like, cool, like fill in the gaps. As I said, like, let's, let's work on like understanding attention. Let's work on understanding longer prompts. Let's work on like finding different, like the replacement architectures, that sort of stuff. It's kind of nice. I think it's a good time to join now. And I can tell you, maybe I can tell you like a really short thing, which is when I switched to Interp, it was after the team had published the original dictionary learning paper, which is towards monosomaticity, which I thought was super cool. Super interesting. It wasn't a one or two layer model, maybe one layer model. The induction heads paper was like on a two layer model. My main concern is I was like, okay, like Interp seems important and we want to understand it, but like, is this 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. Cool. But we were 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 have no means to apply it. It 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 against a young field. So there's no lack of ideas. If you have an idea where you're like, oh, like 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 and sort of like smaller models. And so I think that it's 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 like biology work is so compelling. Like a lot of this work was just literally thinking about, you know, like 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 like do math? Like sometimes it makes mistakes. Like why does it make mistakes? I speak both French and English. Like it seems like it has a slightly different personality in French and English. Why is that? And you can just like, you know, kind of answer your own questions. Yeah. And, and, and kind of like probe at that alien intelligence that we're all building. And I think that's just like a fun thing to do. So maybe like chasing the fun is the thing I'll encourage people to do as well.

swyx [01:28:22]: Well, I think that's, 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. Yes. They can reach out to you at ML Powered, I guess. Yeah. Reach out to me on Twitter. Yeah.

Emmanuel [01:28:34]: Or I'm Emmanuel at Anthropic.

swyx [01:28:36]: If you want to shoot me an email. Email's public now. Yeah. Awesome. Well, thank you for your time. Thank you. Yeah.

Emmanuel [01:28:41]: Thanks for having me, guys.