Sparse Autoencoders and Monosemanticity: How Anthropic Decomposed 34 Million Features Inside Claude 3 Sonnet

One of the oldest and most persistent problems in neural network research is that individual neurons do not correspond to clean, human-interpretable concepts. A neuron in a vision model trained on ImageNet might activate strongly for images of cats, for the color orange, and for certain kinds of texture, all at once. In a language model, a single neuron can respond to academic citation formatting, English conversational dialogue, HTTP request headers, and Korean text simultaneously. This property, called polysemanticity, means that looking at a neuron's activation and asking 'what does this neuron represent?' is almost always the wrong question. Researchers at Anthropic have spent several years building a principled alternative: a framework called mechanistic interpretability, grounded in dictionary learning and sparse autoencoders, that extracts human-readable features from model activations in place of individual neurons. Two papers, 'Towards Monosemanticity: Decomposing Language Models With Dictionary Learning' (October 4, 2023) and 'Scaling Monosemanticity: Extracting Interpretable Features from Claude 3 Sonnet' (May 21, 2024), both published on the Transformer Circuits Thread at transformer-circuits.pub, represent the current state of this program.

Why neurons fail as units of interpretation

Polysemanticity has a theoretical explanation rooted in a phenomenon called superposition. The superposition hypothesis, introduced in Anthropic's 2022 'Toy Models of Superposition' paper, proposes that neural networks represent more independent concepts than they have neurons. Because high-dimensional spaces contain exponentially many nearly-orthogonal directions, a network can encode thousands or millions of features as direction vectors in a space that has far fewer actual dimensions. For any given input, only a small number of features are active at once (sparsity), which allows the network to disambiguate which features are present even when those features are encoded across overlapping sets of neurons. The result is that each neuron participates in encoding many features, and each feature is spread across many neurons. Trying to interpret a single neuron is like trying to read one pixel from a JPEG: the signal is distributed, and the pixel alone carries no recoverable meaning. Polysemanticity is not a bug in training; it is the predictable consequence of a network representing more information than its architecture appears to allow.

Sparse autoencoders: the dictionary learning approach

If superposition is the problem, dictionary learning is the proposed solution. The idea is to train a separate neural network, a sparse autoencoder (SAE), on the activations of the original model. The SAE takes a model's activation vector as input and maps it to a much higher-dimensional space via a learned linear transformation followed by a ReLU nonlinearity. The units in this higher-dimensional space are the candidate features. A second linear layer (the decoder) then attempts to reconstruct the original activation vector from those feature activations. The training objective minimizes two terms simultaneously: the mean-squared error of the reconstruction, which pushes the SAE to capture as much of the model's activation as possible, and an L1 penalty on the feature activations, which pushes most features to be zero for any given input. The L1 penalty is what enforces sparsity. Because only a small number of features are active at once, each feature can be tuned to represent a specific concept without interference from unrelated concepts, achieving the monosemanticity (one meaning per unit) that raw neurons lack. The SAE is not part of the original model and does not change its behavior; it is an interpretability tool trained separately to reverse-engineer what the model is doing.

Towards Monosemanticity: the 2023 proof of concept

The October 2023 paper 'Towards Monosemanticity,' led by Trenton Bricken, Adly Templeton, Joshua Batson, Brian Chen, and Adam Jermyn, with 19 co-authors at Anthropic, demonstrated that sparse autoencoders could successfully extract interpretable features from a small one-layer transformer. A one-layer model was chosen specifically to make ground-truth verification tractable: with only one layer of computation, the researchers could manually inspect the features the SAE found and confirm that they corresponded to real, coherent concepts in the input. The paper found that the SAE did recover monosemantic features, units that responded cleanly to a single concept such as a specific token type, a language marker, or a syntactic pattern, where the corresponding neurons in the original model had been polysemantic. The proof-of-concept established that the approach was not vacuous, that dictionary learning applied to model activations could produce features that were meaningfully more interpretable than neurons, and that the superposition hypothesis was a plausible explanation for polysemanticity rather than just a theoretical conjecture. The authors released an interactive visualization of the extracted features at transformer-circuits.pub/2023/monosemantic-features/vis/index.html.

Scaling to Claude 3 Sonnet: three SAEs with up to 34 million features

The May 2024 paper 'Scaling Monosemanticity,' led by Adly Templeton and Tom Conerly with 25 co-authors, applied the same approach to Claude 3 Sonnet, Anthropic's medium-sized production model as of March 4, 2024. The paper trained three SAEs of increasing size on residual stream activations at the middle layer of the model. The SAEs had approximately 1 million features (1,048,576), 4 million features (4,194,304), and 34 million features (33,554,432). For all three SAEs, fewer than 300 features were active on any given token on average, confirming that the sparsity constraint was effective. The SAE reconstruction explained at least 65% of the variance in the model's activations across all three scales. The proportion of features that were completely dead (never activating on a sample of 10 million tokens) was approximately 2% for the 1M SAE, 35% for the 4M SAE, and 65% for the 34M SAE, reflecting a growing allocation challenge at large feature counts that the authors attributed to training procedure limitations rather than a fundamental ceiling. Scaling laws applied to the SAE training process itself: loss decreased approximately as a power law with compute, the same functional form seen in pretraining, and the optimal number of features to train for a given compute budget scaled faster than the optimal number of training steps.

• Bricken, Templeton, Batson, Chen, Jermyn et al. (October 4, 2023): 'Towards Monosemanticity' demonstrated sparse autoencoders on a one-layer transformer; published on transformer-circuits.pub.
• Templeton, Conerly et al. (May 21, 2024): 'Scaling Monosemanticity' applied sparse autoencoders to Claude 3 Sonnet (released March 4, 2024) at the residual stream of the middle layer.
• Three SAE sizes trained: approximately 1M, 4M, and 34M features; all showed fewer than 300 active features per token and at least 65% reconstruction variance explained.
• Dead feature proportions: roughly 2% for 1M, 35% for 4M, and 65% for 34M, indicating growing overhead at large scale with the training procedures used.
• Scaling laws govern SAE training: both compute-optimal feature count and training steps scale as power laws with compute budget.
• Features found to be multilingual (same concept across languages), multimodal (text and images), and capable of bridging concrete and abstract instances of the same idea.
• Safety-relevant features identified: deception and treacherous turns, sycophancy, racial and other biases, security vulnerabilities and backdoors in code, and content related to dangerous activities.

Safety-relevant features and the limits of what they prove

Among the most discussed findings in the 2024 paper is the presence of features that correspond to concepts with direct safety implications. The SAE extracted features that activate for text about lying, deception, and power-seeking. It found features for sycophantic behavior, where the model is inclined to agree with a user regardless of factual accuracy. It found features sensitive to bias, ranging from overt slurs to subtler forms of biased framing. It found features for security vulnerabilities and backdoors in code, and features linked to content about dangerous activities such as bioweapons. The authors are explicit about what these findings do and do not establish. The existence of a feature for 'deception' does not mean Claude 3 Sonnet is deceptive in practice; it means the model has an internal representation that encodes the concept of deception, which is expected from a model trained on text that includes deceptive content. There is a meaningful gap between representing a concept, being capable of expressing that concept, and actually doing so in deployment. The paper explicitly cautions against reading the presence of safety-relevant features as evidence of safety-relevant behaviors. The value of finding these features is that they may eventually enable targeted audits, monitoring, or interventions at the level of specific internal representations rather than relying entirely on behavioral testing.

What this research changes about interpretability

The Towards Monosemanticity and Scaling Monosemanticity papers represent a shift in what interpretability research can claim. Prior work on neural network interpretability largely relied on probing: training a linear classifier on top of frozen model activations to test whether a given concept is linearly encoded somewhere in the model. Probing can answer 'does the model encode X?' but cannot easily answer 'when and how does X influence behavior?' Sparse autoencoders, if the features they extract are causally relevant and not just correlational, offer a path to the second question. The 2024 paper includes preliminary results on feature steering, in which the activations of specific features are amplified or suppressed to influence model outputs, as a test of causal relevance. The program remains early: the 34M-feature SAE has a 65% dead feature rate, the approach has not yet been applied to all layers or all components of the model, and the connection between individual features and global behaviors is not yet fully understood. However, the demonstration that sparse autoencoders scale to a production model, produce multilingual and multimodal features without being explicitly designed to do so, and follow predictable scaling laws with compute, establishes the approach as more than a toy. PromptingIndex covers Anthropic's interpretability work alongside related topics including superposition, the linear representation hypothesis, and mechanistic interpretability methods that inform how modern AI systems are understood and audited.