We report a number of developing ideas on the Anthropic interpretability team, which might be of interest to researchers working actively in this space. Some of these are emerging strands of research where we expect to publish more on in the coming months. Others are minor points we wish to share, since we're unlikely to ever write a paper about them.
We'd ask you to treat these results like those of a colleague sharing some thoughts or preliminary experiments for a few minutes at a lab meeting, rather than a mature paper.
Short Research Notes
Updates
We are often asked about where our work on superposition is going. Over the last few months, we've run several more ad-hoc experiments on superposition in real models (which sometimes produced interesting, but inconclusive results), as well as exploring a variety of questions related to the theory of superposition like our recent memorization paper.
Our ad-hoc experiments have persuaded us that solutions to superposition won't be extremely low-hanging fruit, and that more systematic, focused efforts will be necessary. To that end, we're presently focusing on what we've described as "Approach 2" in the Toy Models paper: extracting features from superposition by using dictionary learning on the activations of a trained, dense model. (This approach has also been investigated by Sharkey et al., who provided us with helpful comments.) We're building up infrastructure to do a systematic, large-scale investigation, with the hope of either finding superposition in real models or making a significant update against this decoding approach.
Informally, we've found cases where sparse factorizations of neural network activations seem to produce activations which suggest simple hypotheses on inspection. But we don't yet have anything that persuades us that these are the "fundamental truth" of the models that we're studying, rather than a convenient lens which might reveal some features while obscuring others.
In the meantime, we do have a few more conceptual contributions, which can be found in the comments Features as The Simplest Factorization and Dictionary Learning Worries.
As investigation of superposition has progressed, it's become clear that we don't really know what a "feature" is, despite them being central to our research agenda. Several previous definitions have been considered, but all seem unsatisfying. This isn't necessarily bad – sometimes uncertainty about definitions is a very fruitful avenue for science! – but it points at a major open question for us.
In parallel with this, attempts to use dictionary learning or sparse coding methods to automatically discover features in superposition have run into a major challenge. These methods require one to pick a number of features to attempt to decompose the activations into. But how can one know if they've picked the right number to "get all the features" without splitting a true feature into many parts? A recent report by Sharkey et al. proposes some heuristics, but the answer seems non-obvious.
We wonder if it might be possible to answer both of these questions at once by defining features as "the simplest factorization of the activations".
More formally, given a sparse factorization of the activations
(The initial experiments below measure information by modeling the distribution of entries in each matrix with a 100 bin histogram, with the largest bin determined by the maximum matrix entry over all experiments. We then take the surprisal of each entry under this distribution, and sum to get the total information. However, we expect our exact formulation of this to change as our investigation continues.)
It turns out that measuring the information like this seems to be an effective tool for determining the "correct number of features", at least for synthetic data. We consider a variety of synthetic dictionary learning tasks where we take sparse vectors (of dimension n_sparse, with n_active non-zero entries) and randomly project them into 32 dimensional space. The goal is to recover the original sparse structure. We perform dictionary learning (essentially MOD with an LASSO inner loop), varying both the L1 coefficient and the dictionary size. We then do a parametric plot of Mean Squared Error vs this notion of "Total Information".
We observe that dictionary learning solutions "bounce" when the dictionary size matches the true number of factors. Put another way, the Pareto frontier of this rate-distortion plot is occupied by the solutions with the correct number of learned factors. (This is also where the best MMCS score, a metric introduced by Sharkey et al. for evaluating factorizations when the correct answer is known, is obtained.)
If such bounces could be found in real data, it would seem like significant evidence that there are "real features" to be found.
Dictionary learning is presently our top contender for how to extract features out of superposition (following "Approach 2" to solving superposition). If we believe that our activations are described by the factorization
Unfortunately, there are at least two major ways in which we might wish to solve a subtly different problem:
All these concerns point towards using the kind of sparse autoencoder setup explored by Sharkey et al. over a full-blown dictionary learning setup. However, we've found that sparse autoencoders are more fragile and sensitive to hyperparameters, which is a significant countervailing consideration in using them. We are interested in finding an approach with the advantages of a sparse autoencoder (in terms of only finding true features) and the consistent trainability of the dictionary learning schemes.
We also have other worries – such as correlated features which may be more difficult to pull apart – which could be of significant concern to these efforts, but aren't specific to the dictionary learning setup.
In Toy Models of Superposition, perhaps the most surprising result was that toy model features often arranged themselves into uniform polyhedra in superposition, with the specific polyhedra varying by sparsity. However, in a recent comment, we found this is more sensitive to the amount of "feature pressure" (the ratio of the number of features that the model would ideally represent and the number of dimensions it has to represent them), and also how long the model is trained. In particular, there are regimes where having more features compete, and also having models train longer, causes clean geometry where it otherwise would not exist.
We're confused why having more features – which the model ultimately chooses to not represent – affects the geometry of the solution so much. One hypothesis is that even as tiny features, they inject noise. Another hypothesis is that the model is actually effectively using "epsilon features" in superposition somehow.
In a recent comment on Superposition, Memorization, and Double Descent, we observed that problems with
We typically think about superposition as a phenomenon where features are put in superposition. For example, we might have features
But this picture doesn't really help us reason about what kinds of computation a neural network can do while in superposition. We know that some kinds are possible – but what?
To answer this, it's helpful to reason about how weights are put in superposition. If we have two layers
What are the properties of this map? Observe that we want
(Modulo the possibility of interference we'll discuss below.)
Equivalently, one can define the map as a tensor product,
In feature superposition, the interference between two features
All of this is just preliminary thinking on this question, but it seems to give us a tool for reasoning about what weight matrices are possible to represent in superposition, and thus what kinds of computation it's possible to do while in superposition.
In Toy Models of Superposition, we saw that when features are sparse, simple neural networks can represent more features than they have neurons through the phenomenon of superposition. We think something analogous can happen with attention heads, with "attention circuits" and "attentional features" being stored in superposition over attention heads.
For now, we’ll talk about skip-trigrams ([A]…[B] → [C]) as our basic attentional circuits. This is a restricted definition, as we think there are more general kinds of attentional circuits, but they seem sufficient to demonstrate attention superposition.
We trained toy models — small one-layer transformers with trivial embeddings and unembeddings — to investigate how and under what circumstances attention heads place circuits in superposition. The training data were sequences of tokens which were chosen uniformly except for over-representing certain skip-trigrams.
We focus in particular on skip-trigrams which are "OV-incoherent," meaning that they attend from multiple different tokens back to a single token, and the output depends on the token attended from. A single attention head cannot implement multiple such skip-trigrams without introducing errors in its output, because the OV circuit does not know which token is being attended from.
What we see is that when the ground truth contains more incompatible trigrams than there are attention heads, models resort to placing them in superposition across heads. Above, each trigram is encoded in superposition between at least two attention heads. These results suggest caution in studying the role of a single attention head in isolation, as circuits implemented in superposition can appear misleading when only partially interpreted.
These wiring diagrams are simple for models trained on small numbers of skip-trigrams, but rapidly become too complex to read as the training data become more complex. Despite this, we see tantalizing evidence of beautiful geometry underlying even cases with many skip-trigrams, which we are excited to investigate further.
In our toy model investigations of superposition, we assume the existence of discrete features and see how an autoencoder represents them. But in real life, features may lie on a manifold, where nearby features respond to similar data. What should we expect neural networks to do in such cases? In many empirical cases, neural networks model the manifold with families of equivariant neurons, representing the manifold in terms of discrete units rather than representing the manifold directly. Should we expect this to always happen? Why does it happen?
For example, a vision model might want to represent curves in different orientations; the set of possible orientations naturally defines a 1D manifold, a circle. One could imagine the network might have a single neuron whose activation represents the angle of the curve or might have two neurons whose activations represent sine and cosine of the angle of the curve. Instead, in Curve Detectors, Cammarata et al find many (~10) neurons which each respond to curves in a specific range of orientations.
We present some extremely preliminary results investigating this question by considering a toy problem with a "feature manifold" rather than discrete features. We'll then study what happens when we change the length scale (
Our basic setup will be the ReLU-output problem from the Toy Models paper. Instead of having the data be independent features, we imagine having a large number of features arranged around a circle, with equal angular spacing. We first fix a length scale (
We can now study how the features are embedded as we vary the length scale:
This kind of emergent discretization (which we're increasingly seeing hints of across a variety of problems) seems like it might be a very important phenomenon. It may be that "emergent discretization" is the thing we mean when we talk about features.
One caveat to this work is that we've seen some hints that the smallest length scale discretization may be an optimization failure. Additional research is needed to understand this phenomenon.
Transformer Circuits periodically publishes comments on our papers, both from external parties and by the authors. Some of these comments were submitted before publication, from reviewers of early draft manuscripts. But others are submitted significantly after the fact, and might not be seen. To that end, we've included a digest of recently added comments:
A Mathematical Framework for Transformer Circuits
In-context Learning and Induction Heads
Toy Models of Superposition
Superposition, Memorization, and Double Descent
Over the last few months, we've also published a few smaller papers which you might not have seen (including one "perspective" article – Interpretability Dreams – being released along with this post).
Finally, we'd like to highlight recent work by a number of researchers at other groups which we believe will be of interest to you if you find our papers interesting.
Linear Representations. One of the most fundamental assumptions we personally tend to make in studying neural networks is the linear representation hypothesis: neural network features are represented by directions. While this is a common hypothesis, it isn't known to be true.
A recent back and forth between Li et al. and Nanda (in the context of language models trained on Otehllo games) is perhaps the strongest evidence yet from a Popperian perspective: the linear representation hypothesis made a prediction which was contrary to evidence at that point, and was then validated (see Neel Nanda's comment here). It should be mentioned that there are many other reasons to be excited about this work – we discuss it more below – but we wanted to particularly highlight this as an example of excellent scientific discourse and the evidence it seems to provide for a question of very broad significance to the field.
More generally, a wide range of other papers have continued to provide more empirical examples of seemingly linearly represented features. Perhaps the most striking is Turner et al. (who do vector arithmetic to control language models), but see also Gurnee et al. and more generally all the papers mentioned in the following section on what features can be found in language models.
What Features Exist Inside Language Models? Ultimately, our goal is to understand language models. While it's often tempting to emphasize methods or theories, the bread and butter of mechanistic interpretability research must be something similar to the study of anatomy in biology: characterizing features and circuits that exist in language models. On this note, Yun et al., Gurnee et al., and Bills et al. – while all also notable for other contributions – deserve attention for their qualitative results on what features we exist inside language models.
Superposition. In the last few months, significant progress has been made by our colleagues at other groups on superposition. Sharkey et al. attempted to decode superposition in real models, using sparse autoencoders. Yun et al. apply dictionary learning to transformer residual streams and recover many interpretable features. Gurnee et al. apply sparse linear probes to transformers and find, among other things, evidence of low-level linguistic features being represented in superposition over small sets of neurons. Lindner et al. created a tool to compile programs into transformers using superposition. Jermyn et al. explore approaches to encouraging monosemantic neurons. Scherlis et al. examine superposition from the perspective of constrained optimization. Hobbhahn published two posts extending our investigation of superposition and memorization.
One detail from Hobbhahn's posts which we wanted to highlight is that some models seem to have a kind of "shifted superposition" where the model shifts data points to avoid ReLU. This is contrast to the intuition one might naively have that ReLU would in fact anchor the superposition at 0 due to its special behavior there.
Othello & World Models. In the context of language models, there's been an ongoing debate about whether they're "just doing statistical pattern matching" or they "understand". This conversation has often been polarized and disconnected from specific mechanistic hypotheses of what's going on. However, a recent paper by Li et al. – and follow up work by Nanda – used probes to provide evidence that language models trained to play Othello have an internal representation of the state of the board. This is both a nice example of progress in mechanistic understanding, and is also perhaps an example of how mechanistic interpretability can help us have more productive dialogues about neural networks.
How is factual knowledge retrieved? A recent paper by Geva et al. continues the very fruitful line of investigation on activation patching methods (see Meng et al.), which allows for larger-scale understanding of how transformers process information. This new paper looks into how somewhat more complex queries about knowledge are processed by language models. In particular, where prior work showed that attention heads were important for moving information from a subject token, this work suggests that the OV circuit of attention heads can also transform that information, for example reading in a country and writing out its capital.
Activation Patching Continues. As mentioned above, we're continuing to see exciting work based on the activation patching approach (see Meng et al.), most recently by Geva et al..
Automated Interpretability. One of the most common (and very reasonable) critiques of mechanistic interpretability is that it can't scale to large models. A recent paper by Bills et al., "Language models can explain neurons in language models", provides a proof of concept for automating parts of mechanistic interpretability. This approach would still require a solution to superposition, but it's potentially an exciting way to address the scalability concern. At the same time, we also have some reservations about this kind of automation, especially when the goal is safety. Do we really want our auditing of AI models to depend on trusting an AI model to help us with auditing? A critical question is whether alternative approaches to addressing the scalability problem can be found. Either way, this kind of method seems helpful in the meantime – and the qualitative results are also very interesting.
Attribution patching. In a recent post, Neel Nanda describes a method called "attribution patching" which he developed in collaboration with several of us a while back. It's exciting to see this written up! Using gradient activation products to perform quick attributions to various intermediate computations was quite useful for investigations in the vision context (see Building Blocks of Interpretability), and seems helpful as a way to investigate larger models. However, be sure to pay attention to Neel's cautionary notes on when this works, especially the section on LayerNorm. (We can file this as reason #78 for why interpretability researchers hate LayerNorm.)
Can we explain learning dynamics and scaling laws in terms of circuits? We know that induction heads cause a loss bump in training when they form, and likely cause a bump in scaling laws. This suggests the tantalizing possibility of a deep bridge between the microscopic world of mechanistic interpretability and the more macroscopic topics of learning dynamics and scaling laws. Several recent papers have made us more hopeful that such a bridge can be found.
Quanta. Michaud et al. propose a theory of scaling in terms of "quanta" – discrete behavior patterns which reduce loss – along with an algorithm for automatically discovering these quanta based on gradients. A natural hypothesis is that these behavioral quanta mechanistically correspond to circuits, just as the "induction bump" in-context learning behavior corresponds to induction head circuits. If this could be demonstrated, it would create a much wider bridge from the microscopic world of circuits to the macroscopic world of losses, behaviors, scaling, and learning dynamics.
Mode Connectivity. Lubana et al. and Juneja et al. find an empirical relationship between generalization strategies – and likely the underlying mechanisms – and linear mode connectivity in the loss landscape. In particular, models with different generalization properties appear to have a loss barrier separating them if one linearly interpolates in parameter space.
Reverse Engineering Grokking, Fourier Transforms, & Universality. In Progress Measures for Grokking via Mechanistic Interpretability, Nanda et al. reverse engineered a neural network doing modular arithmetic which turned out to be using Fourier transforms, and linked this to grokking behavior. Following up on this, Chughtai et al. found that if one trains neural networks to perform more general group operations, they learn to use group representations (a generalization to noncommutative groups of the Fourier transform for cyclic groups found in the first model). This is interesting both as a compelling example of reverse engineering simple models, and also as evidence for the universality hypothesis, as the authors find that each trained network utilizes a random subset of the group representations that exist.
Why does grokking occur? A recent paper by Liu et al. finds a systematic relationship between weight decay and the length of time it takes for grokking to occur. Roughly, they find first the model finds a memorizing solution whose weight matrix has very large norm, and that grokking occurs when the model weights shrink to the size of the generalizing solution. Qualitatively, the relationship they find between memorization and weight norm matches some of our observations on how memorization occurs mechanistically in toy models and classic work by Bartlett showing that feed-forward networks with small weight norm generalize well.
Neuroscience Parallels. Over the last few years, there have been a number of cases where mechanistic interpretability research discovered results which parallel finding in neuroscience, including curve detector neurons and person-detecting multimodal neurons. Recently, we've begun to see parallels which go in the other direction, with discoveries in artificial neural networks foreshadowing results in biological neuroscience:
The growing body of parallels, and the fact that they're going in both directions, seems suggestive of a genuine, deep connection. It also seems like evidence for a very strong version of the universality hypothesis.
Behavioral Control of GPT with Activation Addition. In a recent post, Turner et al. demonstrate that they can control language models by adding vectors to activations, defined simply by doing arithmetic on activation vectors. This extends earlier work on RL agents navigating a maze. It's interesting to speculate what the mechanism is – are they controlling low-level features related to a topic, high-level topic/theme features, "motor neurons" that directly implement behavior, or something else? More generally, it's another piece of evidence for the linear representation hypothesis.
Hierarchical Skip-Trigrams. A recent post by Buck Shlegeris constructed an example of a phenomenon we'll call "hierarchical skip-trigrams" (following Neel Nanda's naming). Hierarchical skip-trigrams allow one-layer models to use skip-trigrams to express interesting computation one might not have naively expected.
Decision Transformer Interpretability. Two recent articles (part 1, part 2) by Bloom and Colognese take a mechanistic approach to investigating decision transformers in a grid world setting.
Sparsity and Modularity. A recent paper by Liu et al. explores encouraging sparsity and modularity with a weight sparsity penalty that penalizes weights between neurons that are far apart (similar to wire length minimization in neuroscience). They find striking sparse graphs for a variety of tasks including arithmetic, group multiplication, and in-context learning.
Mechanistic Interpretability Challenges. Back in February, Stepen Casper posed several challenges to mechanistic interpretability practitioners, somewhat similar to the "auditing game" tests conducted at OpenAI in 2019. Recently, Stefan Heimersheim and Marius Hobbhahn took up this challenge and solved the first one.
Learning Materials. Neel Nanda has been producing a wide range of resources on getting started in mechanistic interpretability, including a number of videos walking through different papers and also videos explaining transformers, and also a list of open problems. Separately, TheMcDouglas produced a nice illustration of induction heads.
This is an informal collection of updates published on May 24th, 2023. Different sections have different authors. The overall update was edited by Chris Olah, Josh Batson, and Shan Carter.
Informal notes allow us to publish ideas that aren't as refined and developed as a full paper. Mistakes are very possible! We'd ask you to treat this note as more like some ideas a colleague is informally sharing with you, or a typical internet blog post, rather than a paper.