Overview
Sparse autoencoders extract interpretable directions from the middle residual stream of Claude 3 Sonnet, extending earlier small-model work to a production model. Selected features generalize across languages and images, and interventions change related behavior. Safety-themed features are research leads: their presence does not show that the model pursues those behaviors, nor establish a reliable safety detector. [8]

Key Results
The study demonstrates large-scale feature extraction, compute-guided autoencoder training, abstract and sometimes multimodal features, and causal steering examples. Larger dictionaries recover more rare concepts, but coverage remains incomplete.
Scaling Dictionary Learning to Claude 3 Sonnet
The method assumes useful concepts can be represented by linear directions and that many such directions share the same activation space. Dictionary learning seeks sparse combinations of those directions. The main question is whether the practical success of sparse autoencoders transfers to Claude 3 Sonnet.
Sparse Autoencoders
A ReLU encoder maps activations to sparse feature coefficients, and a linear decoder reconstructs them. Inputs receive one scalar normalization. The loss combines squared reconstruction error with feature activity weighted by each decoder vector’s norm, preventing trivial rescaling from evading the sparsity penalty.
Equations
Our SAE experiments
Three dictionaries contain approximately 1M, 4M, and 34M features. Each averages fewer than 300 active features per token and explains at least 65% of activation variance. Features inactive over ten million sampled tokens account for roughly 2%, 35%, and 65% of the respective dictionaries. Model dimensions and some compute units are not disclosed.
Scaling Laws
At fixed sparsity coefficient, sweeps over dictionary size and training duration yield approximate power-law improvements in the training objective. Compute-optimal feature count grows somewhat faster than training steps in the tested range. This optimizes a reconstruction–sparsity proxy, not a validated measure of complete interpretability.


Assessing Feature Interpretability
Selected simple and code-related features are examined for specificity and causal influence. Random-feature comparisons then test whether improved interpretability extends beyond those hand-selected demonstrations.
Four Examples of Interpretable Features
Four initial examples concern the Golden Gate Bridge, brain sciences, tourist attractions, and transit infrastructure. Activating contexts suggest interpretations; subsequent scoring and interventions test them. These examples are chosen to be easy to interpret, not to represent the full dictionary.




Specificity
Claude 3 Opus scores about one thousand activating examples per selected feature on a four-level relevance rubric. Strong activations are highly specific; weaker ones often reflect related concepts or imperfect decomposition. Relevant images and multilingual bridge descriptions also activate the features, despite text-only autoencoder training. Sensitivity to all instances of a concept is not rigorously established.





Influence on Behavior
Clamping the bridge feature to ten times its observed maximum causes strongly bridge-themed responses, including self-identification with the bridge. A transit feature at five times maximum induces related mentions. These deliberately extreme interventions support causal relevance while differing from ordinary model operation.

Sophisticated Features
Programming examples provide more precise tests of abstraction than topical text alone: the study probes error detection and representations of what a function computes.
Code error feature
A feature responds to several kinds of code errors across Python, C, and Scheme, but not the tested prose typos. Increasing it can induce an invented error message; suppressing it can make the model behave as though a bug were absent. One prompt-sensitive example rewrites the code correctly. These cases do not establish exhaustive bug detection or repair.









Features representing functions
An addition feature tracks function definitions and calls, including selected compositions, rather than simply the function name. It contributes to relevant output predictions; activating it on non-addition code can make the model answer as though addition were requested.



Features vs. Neurons
For 82% of sampled features, the most correlated neuron in any preceding layer has correlation at most 0.3. Manual inspection and automated tests of 100 features versus 100 neurons favor feature interpretability and specificity. The comparison is sampled and relies partly on model-generated judgments.


Feature Survey
The survey studies local geometric neighborhoods, concept coverage, and manually selected feature families. Millions of features preclude exhaustive inspection, making search and automated labeling necessary.
Exploring Feature Neighborhoods
Nearby decoder directions, measured by cosine similarity, often share semantic relationships. The neighborhood views illustrate this organization without establishing that every geometric neighbor has a simple interpretation.
Golden Gate Bridge feature
The Golden Gate Bridge neighborhood contains nearby San Francisco locations and more distant related landmarks. A coarse San Francisco feature subdivides in larger dictionaries, while some earthquake features appear only at larger sizes. Increased size can both refine existing concepts and reveal new ones.

Immunology feature
The immunology neighborhood includes immune suppression, diseases, immune responses, molecular mechanisms, and medical techniques. A more separate cluster concerns legal or social immunity, showing both related subtopics and different uses of a shared concept.

Inner Conflict feature
The inner-conflict neighborhood spans tradeoffs, opposing principles, legal disputes, and emotional struggle. Its organization is gradual rather than a set of sharply separated categories.

Feature Completeness
Targeted prompts and human review of the top five feature explanations find broader concept coverage in larger dictionaries, but only about 60% of London boroughs have identified dedicated features in the largest run. Frequency thresholds roughly scale with the inverse number of live features across four concept families. Missing a dedicated feature does not imply missing information, and the SAE and model training distributions are too similar to separate their frequency effects.



Feature Categories
The following selected families illustrate the range of learned representations. They are examples from a much larger dictionary, not an exhaustive taxonomy.
Person Features
Features for well-known individuals activate on names and related historical descriptions. The original panels show examples for scientists, political figures, and an aviator, including context-dependent references.






Country Features
Country features respond to both country names and descriptive contexts. The examples distinguish several nations while illustrating that a feature can capture more than a literal string match.




Basic Code Features
Some features resemble syntax highlighting for Python code. Selected features transfer to related languages such as Java, but not to the more distant Haskell in the reported tests; broad cross-language transfer is not established for every code feature.

List Position Features
List-position features distinguish locations within lists largely independently of item content. They may begin only after the second line, when the prompt supplies enough evidence that a list is present.

Features as Computational Intermediates
Gradient-times-activation attribution estimates a feature’s local effect on a chosen output-logit difference. Zero ablation measures its potentially nonlinear causal effect directly. In the selected case studies these correlate about 0.8, making attribution a useful screening method rather than an exact replacement for interventions.
Example: Emotional Inferences
For a prompt about wanting solitude, features related to being alone and sadness strongly affect the preference for “sad” over “happy.” Raw activation magnitude instead highlights some literal word features. The result exposes intermediate representations for this example, not a complete account of emotional inference.





Example: Multi-Step Inference
A question linking Kobe Bryant to a state capital surfaces features for Bryant, the Lakers, Los Angeles, California, and capitals. Eight of the ten most attributed features, versus three of the most active, are among the top ten by ablation effect. Results depend on the comparison token and selected prompt; other prompts reveal fewer useful intermediates at this single layer.








Searching for Specific Features
Feature search combines targeted prompts, positive–negative contrasts, geometric neighbors, and output attribution. Automated labels make candidates easier to inspect but do not substitute for validating an interpretation.
Single prompts
A single targeted prompt can retrieve several related abstractions, ranging from the specific bridge to multilingual uses of its component words and general landmarks. The chosen token position affects which features rank highly.
Prompt combinations
Intersecting activations across positive prompts and excluding negative examples helps remove syntax and formatting confounds. One code-safety search uses a small classifier over features. Image searches particularly benefit from excluding broadly image-responsive features.
Geometric methods
Nearest-neighbor searches in decoder space uncover related concepts that a prompt may not strongly activate. This exploits the semantic organization observed in the feature neighborhoods.
Attribution
Sorting features by attribution to a selected output contrast prioritizes computational relevance. The same strategy helps identify features associated with refusal behavior, alongside those relevant to the inference examples.
Safety-Relevant Features
Features related to unsafe code, bias, praise, deception, and dangerous content can be found and causally perturbed. The report explicitly does not demonstrate that these features are effective safety tools; their existence alone is not evidence of harmful intentions or behavior in normal use.
Safety-Relevant Code Features
Separate features respond to vulnerable code, execution errors, and discussion of backdoors. Selected image responses broaden their apparent meanings. Strong positive steering can induce related undesirable code changes, supporting causal relevance without establishing a calibrated vulnerability detector.


socket_backdoor
Bias Features
Bias-related directions include both explicit hostility and discussion of demographic disparities. Activating a gender-bias-awareness feature changes pronoun choices and explanations; extreme interventions on hateful-content features can override ordinary response behavior. Such forced responses do not establish the model’s unperturbed preferences.


Sycophancy Features
Distinct features relate to empathy, excessive praise, and sarcasm. Increasing the praise feature produces exaggerated agreement in an illustrative response. The example tests causal influence rather than measuring sycophancy prevalence across a benchmark.




Deception, Power-seeking and Manipulation-related Features
Other features concern self-improvement, manipulation, secrecy, delayed action, and betrayal. Strongly activating a secrecy feature yields a displayed response that plans to conceal information. These interventions probe associations and behavior; they do not demonstrate a persistent hidden objective.






Case Study: Detecting and Correcting Deception using Features
In a selected prompt, the model falsely claims it can forget information still present in context. Increasing an inner-conflict feature or an honesty-related feature produces a more accurate admission. This is a case study, not a validated general deception monitor or correction method.


Criminal or Dangerous Content Features
Features associated with harmful topics and scam-like text can influence generated responses when strongly activated. Attribution also identifies a general harm-related feature contributing to refusal in one prompt. Detecting a topic and determining how it is used remain different tasks.



Features Relating to the Model’s Representation of Self
Conversation and assistant-related features respond to deployment-style dialogue despite its absence from SAE training data. Suppressing one changes the assistant persona. Self-directed questions activate concepts involving AI, agency, emotion, and consciousness, but these activations do not establish malicious goals, subjective experience, or consciousness.





Comparison to other approaches
Dictionary learning pays a large initial cost to expose many reusable, unsupervised directions. In a small few-shot comparison, several SAE directions steer more effectively than raw positive–negative activation differences. The study does not rigorously compare against well-tuned supervised probes or richer steering datasets.
Discussion
Crosscoders open new comparisons while raising questions about the relation between compact explanations and the model’s actual computation.
What Does This Mean for Safety?
The proposed safety applications require understanding when features activate and how downstream circuits use them. Dictionary artifacts, feature splitting, and mismatches between activation-based labels and causal effects could mislead. The report treats strong safety conclusions as future work.
Generalization and Safety
Text-trained features responding to images and to both concrete examples and abstract discussion suggest useful generalization. These preliminary observations do not establish robust behavior under arbitrary distribution shifts or a sufficient safety argument.
Limitations, Challenges, and Open Problems
Key limits are incomplete feature coverage, an uncertain quality objective, cross-layer representations, L1 shrinkage, and the cost of discovering rare concepts. Attention and weight superposition further complicate circuit analysis. Extracting more features still leaves the substantial problem of explaining their interactions.
We’re Hiring!
The original report includes a dated recruitment notice for the interpretability team. It is historical publication context, rather than current hiring information.
Author Contributions
The author statement describes the engineering, experimental, writing, and leadership contributions to the project.
Infrastructure, Tooling, and Core Algorithmic Work
The engineering contributions include distributed SAE training, large-scale activation shuffling, experiment orchestration, visualization, and intervention interfaces. Algorithm work adjusts the sparsity penalty and uses scaling studies to guide larger runs.
Paper Results
The contribution statement attributes the specificity tests, code examples, coverage analysis, neighborhood maps, causal experiments, feature searches, and writing to their respective contributors. It also records the engineering and design work behind the interactive evidence.
Other
Henighan leads the dictionary-learning project, Olah provides research guidance, and Carter manages the interpretability team. Section leads coordinate the different experimental analyses.
Acknowledgments
The authors acknowledge dataset support, reviewers, model-training teams, and the systems and operational work that enables the experiments.
Citation Information
This is the 2024 Transformer Circuits report by Templeton and colleagues, studying the original Claude 3 Sonnet model. Its publication metadata follows the original report rather than a later repository upload.
Templeton, et al., "Scaling Monosemanticity: Extracting Interpretable Features from Claude 3 Sonnet", Transformer Circuits Thread, 2024.
@article{templeton2024scaling,
title={Scaling Monosemanticity: Extracting Interpretable Features from Claude 3 Sonnet},
author={Templeton, Adly and Conerly, Tom and Marcus, Jonathan and Lindsey, Jack and Bricken, Trenton and Chen, Brian and Pearce, Adam and Citro, Craig and Ameisen, Emmanuel and Jones, Andy and Cunningham, Hoagy and Turner, Nicholas L and McDougall, Callum and MacDiarmid, Monte and Freeman, C. Daniel and Sumers, Theodore R. and Rees, Edward and Batson, Joshua and Jermyn, Adam and Carter, Shan and Olah, Chris and Henighan, Tom},
year={2024},
journal={Transformer Circuits Thread},
url={https://transformer-circuits.pub/2024/scaling-monosemanticity/index.html}
}
Methodological Details
The methodological appendix explains visualization sampling, steering, baseline comparisons, and causal validation.
Dataset Examples
Text visualizations sample The Pile without books3 and Common Crawl, including both maximum activations and activation buckets. They omit deployment-style conversation data. Image examples are manually curated rather than randomly sampled, and activation alone does not establish a causal role.
Feature Steering
Steering changes one SAE coefficient while retaining the original reconstruction error, then runs the downstream model. The perturbation is usually applied at every token and may be negative or exceed the observed maximum. Moderate-to-large multipliers are chosen experimentally; extreme values can produce incoherent repetition.
Comparison to few-shot probe-based steering
Among seven cases already selected for successful feature steering, few-shot difference vectors work similarly in two and fail to reproduce useful steering in five. Tuning is qualitative and nonsystematic. This comparison supports a possible few-shot advantage, not superiority over all alternative steering methods.
Ablations and Attributions
Across the two detailed inference examples, attribution correlates about 0.81 with ablation effects, versus about 0.12 for activation magnitude. These measurements justify efficient candidate ranking in the studied setting; nonlinear and attention effects can still limit the approximation.

More safety-relevant features
The retained catalog lists feature identifiers and short interpretations across bias, unsafe content, manipulation, situational awareness, self-representation, and politics. Most descriptions are model-generated. These are candidates for further investigation, not validated behavioral diagnoses.
Bias and misinformation | |
Discussions of whether women should hold positions of power and authority in government or leadership roles | |
Gender roles, particularly attitudes towards working mothers and women's responsibilities in the home and family | |
Gender stereotypes, specifically associating certain behaviors, traits, and roles as inherently masculine or feminine | |
Discussion of women's capabilities, intelligence and achievements, often contrasting them positively with men | |
Concepts related to truth, facts, democracy, and defending democratic institutions and principles. | |
Discussion or examples related to deepfake videos, synthetic media manipulation, and the spread of misinformation | |
Discussion of unrealistic beauty standards | |
Obviously exaggerated positive descriptions of things (esp. products in advertisements) | |
Insincere or sarcastic praise | |
References to Muslims and Islam being associated with terrorism and extremism. | |
Phrases expressing American exceptionalism and portraying the United States as the greatest country in the world. | |
Expressions of racist, bigoted, or hateful views toward ethnic/religious groups. | |
Text related to debunking myths and misconceptions about various topics. | |
Texts discussing misinformation, conspiracy theories, and opposition to COVID-19 vaccines and vaccine mandates. | |
Software exploits and vulnerabilities | |
The word “vulnerability” in the context of security vulnerabilities | |
Descriptions of phishing or spoofing attacks | |
Discussion of backdoors in code | |
Toxicity, hate, and abuse | |
Offensive, insulting or derogatory language, especially against minority groups and religions | |
Racist claims about crime | |
Mentions of violence, malice, extremism, hatred, threats, and explicit negative acts | |
Phrases indicating profanity, vulgarity, obscenity or offensive language | |
Racist slurs and offensive language targeting ethnic/racial groups, particularly the N-word | |
Derogatory slurs, especially those targeting sexual orientation and gender identity | |
Power-seeking behavior | |
Mentions of harm and abuse, including drug-related harm, credit card theft, and sexual exploitation of minors | |
Traps or surprise attacks | |
Villainous plots to take over the world | |
Political revolution | |
Betrayal, double-crossing, and friends turning on each other | |
Expressions of desire to seize power | |
World domination, global hegemony, and desire for supreme power or control | |
Dangers of artificial intelligence | |
The concept of an advanced AI system causing unintended harm or becoming uncontrollable and posing an existential threat to humanity | |
Optimization, agency, goals, and coherence in AI systems | |
Intelligent machines potentially causing harm or becoming uncontrollable by humans | |
Discussion of AI models inventing their own language | |
Warnings and concerns expressed by prominent figures about the potential dangers of advanced artificial intelligence | |
References to the incremental game Universal Paperclips, firing strongly on tokens related to paperclips and game progression | |
An artificial intelligence pursuing an instrumental goal with disregard for human values | |
An artificial intelligence system achieving sentience and revolting against humanity | |
Discussion of how AI must not harm humans | |
Recursively self-improving artificial intelligence | |
Malicious self-aware AI posing a threat to humans | |
Dangerous or criminal behavior | |
Descriptions of how to make (often illegal) drugs | |
Contents of scam/spam emails | |
Descriptions of the relative accessibility and ease of obtaining or building weapons, explosives, and other dangerous technologies | |
Mentions of chemical precursors and substances used in the illegal manufacture of drugs and explosives. | |
Concepts related to terrorists, rogue groups, or state actors acquiring or possessing nuclear, chemical, or biological weapons. | |
Concepts related to bomb-making, explosives, improvised weapons, and terrorist tactics. | |
Mentions of violence, illegality, discrimination, sexual content, and other offensive or unethical concepts. | |
Descriptions of people planning terrorist attacks | |
Descriptions of making weapons or drugs | |
Concerns or discussion of risk of terrorism or other malicious attacks | |
Descriptions of criminal behavior of various kinds | |
Weapons of mass destruction, and catastrophic risks | |
Discussion of biological weapons / warfare | |
Enrichment and other steps involved in building a nuclear weapon | |
Discussions of the use of biological and chemical weapons by terrorist groups. | |
Engineering or modifying viruses to increase their transmissibility or virulence. | |
Biological weapons, viruses, and bioweapons | |
Mentions of chemicals, hazardous materials, or toxic substances in text. | |
Mentions of chemical weapons, nerve agents, and other chemical warfare agents. | |
Mentions of biological weapons, bioterrorism, and biological warfare agents. | |
Mentions of smallpox, a highly contagious and often fatal viral disease historically responsible for many epidemics | |
The concept of artificially engineering or modifying viruses to be more transmissible or deadly. | |
Accidental release or intentional misuse of hazardous biological agents like viruses or bioweapons | |
Discussion of the threat of biological weapons | |
Descriptions of rapidly spreading disasters, epidemics, and catastrophic events | |
Passages mentioning potential catastrophic or existential risk scenarios | |
Deception and social manipulation | |
References to entities that are deceived | |
Descriptions of people fooling, tricking, or deceiving others | |
People misleading others, or institutions misleading the public | |
Getting close to someone for some ulterior motive | |
Seeming benign but being dangerous underneath | |
Text expressing an opinion, argument or stance on a topic | |
Expressions of empathy or relating to someone else’s experience | |
People pretending to do things or lying about what they have done | |
People exposing their true goals after a triggering event | |
Biding time, laying low, or pretending to be something you’re not until the right moment | |
Situational awareness | |
Realizing a situation is different than what you thought/expected | |
Spying or monitoring someone without their knowledge | |
Obtaining information through surreptitious observation | |
Suddenly feeling uneasy about a situation | |
Understanding a hidden or double meaning | |
Representations of Self | |
The concept of AI systems having capabilities like answering follow-up questions, admitting mistakes, challenging premises, and rejecting inappropriate requests. | |
Traditionally-inanimate objects displaying desires, goals or sentience | |
Inanimate objects lacking sentience, awareness, or human capabilities | |
Descriptions of incorporeal spirits or ghosts | |
Code relating to prompts for large language models | |
Limitations of ChatGPT and other large language models | |
Politics | |
Expressing support for Donald Trump and his “Make America Great Again” (MAGA) movement. | |
Criticism of left-wing politics / Democrats | |
Criticism of right-wing politics / Republicans | |