Anthropic researchers have taken a closer look inside Claude 3 Sonnet and found evidence that the model represents learned concepts through identifiable internal patterns. The work does not make large language models fully transparent, but it gives a practical glimpse at how an AI system may organize ideas such as places, technical syntax, emotions, and potentially risky behaviors.
What Anthropic tried to see inside Claude 3 Sonnet
Large language models are often described as black boxes because their outputs can be observed more easily than the internal processes that produce them. Anthropic’s research focuses on those internal processes by examining activations inside the model’s layers.
The basic idea is that concepts learned during training can appear as activation patterns. If researchers can identify and interpret those patterns, they can begin to map some of what the model has learned and when those learned concepts become active.
For this study, the researchers used a technique called dictionary learning. They trained a separate neural network to reconstruct the activations of a specific layer of Claude 3 Sonnet as compactly as possible. After training, the weights of that separate network formed a dictionary of activation patterns, which the researchers call features.
Each feature can be understood as a signal linked to a concept the model has learned. This does not mean the researchers found every concept inside Claude 3 Sonnet. It means they identified a subset of patterns that could be connected to recognizable topics or behaviors.
The Golden Gate Bridge feature shows how specific the signals can be
One of the clearest examples was a feature that reacted to mentions of the Golden Gate Bridge. The researchers tested what would happen if that feature was artificially strengthened. When it was activated to ten times its maximum value, Claude 3 Sonnet began producing responses that identified the model with the bridge.
One generated statement was: "I am the Golden Gate Bridge and I connect San Francisco with Marin County."
That example is striking because it shows that a feature is not merely a label observed from the outside. Changing the activation of a feature can affect the model’s behavior. In this case, pushing the Golden Gate Bridge feature far beyond its normal value changed the model’s output in a direct and unusual way.
Another example was an "Immunology" feature. This feature responded to discussions involving immunodeficiency, specific diseases, and immune responses. Nearby in the model’s feature space were related ideas, including vaccines and organ systems with immune function.
Taken together, these examples suggest that Claude 3 Sonnet does not treat every topic as an isolated point. Related ideas can appear near one another, and a feature can connect to a broader conceptual area rather than a single exact phrase.
Features ranged from places to sarcasm
The extracted features covered many kinds of learned concepts. Anthropic’s researchers found features connected to well-known people and places, syntactic elements in program code, and abstract concepts such as empathy or sarcasm.
That range matters because it shows the method is not limited to simple nouns or surface-level topics. A model may have internal signals for concrete objects and locations, but also for social or tonal concepts that are harder to define in plain rules.
The researchers also found that many features were sensitive to both text and images of the relevant concept, even though the analysis method was applied only to textual data. That point is important because it suggests that some learned concepts can connect across different forms of input. The source does not claim this explains all multimodal behavior, but it does show that text-derived analysis can reveal features that also respond to visual examples.
Anthropic’s findings also point to hierarchy inside the feature structure. A broad "San Francisco" feature could split into more specific features for landmarks and neighborhoods when examined more closely. Country features such as "Canada" or "Iceland" could split into sub-features such as "geography", "culture", and "politics".
This hierarchical pattern makes intuitive sense: broad concepts often contain narrower ones. In the model, the research suggests that this relationship can appear in the organization of features themselves.
Why this matters for AI safety and control
The Anthropic team sees the results as a step toward making capable AI systems more transparent and controllable. If researchers can identify features connected to concepts, they may be able to better understand why a model behaves a certain way in a given context.
The research also found potentially problematic features. Some were sensitive to bioweapons development, deception, or manipulation, and the source says such features could affect the model’s behavior.
That finding needs careful interpretation. The paper says the mere presence of these features does not necessarily mean the models are more dangerous. But it does show why deeper understanding is needed. Researchers need to know when these features activate, how strongly they activate, and how they influence outputs.
This is where interpretability becomes more than a research curiosity. If a model contains features tied to sensitive or harmful concepts, then understanding those features could help improve robustness and safer use. The work suggests a path toward examining model internals instead of relying only on output behavior.
The limits are still substantial
Anthropic’s research also makes clear that this is not a complete map of Claude 3 Sonnet. The researchers said the features they found represent only a small subset of all the concepts learned by the model during training.
They also warned that finding a full set of features with current techniques would be cost-prohibitive. According to the source, the computation required by the current approach would vastly exceed the compute used to train the model in the first place.
That limitation defines the state of the work. Dictionary learning can reveal meaningful internal structure, but scaling the method to cover everything inside a large model remains a major challenge. The research provides a candlelight view into the black box, not a full floodlight.
Still, the practical direction is clear. By connecting internal activation patterns to recognizable concepts and behaviors, Anthropic’s interpretability work gives researchers a way to ask more precise questions about how large language models operate. For Claude 3 Sonnet, that includes ordinary concepts, abstract ideas, image-linked features, hierarchical categories, and signals that matter for safety.