Interpretable features in Claude 3 Sonnet
Era 5 · Generative AI, reasoning, and agents (2022–2026) › Alignment and reasoning
- Lineage
- Data and learning
- Kind
- Publication
Who
Anthropic
What happened
Anthropic researchers extract millions of interpretable features from Claude 3 Sonnet, a production model, using sparse autoencoders. They find features for concepts such as the Golden Gate Bridge or deception, and show that activating them changes the model's behavior.
Why it mattered
Interpretability moves from toy models to a commercial one: it becomes possible to see which concepts a model represents internally and to act on them.
The full story
A language model is a network of billions of numbers, and looking at its neurons one by one says almost nothing: each neuron responds to many unrelated things. Mechanistic interpretability tries to break that activity down into pieces a person can understand.
In October 2023 the Anthropic team had shown with a one-layer model that a sparse autoencoder, an auxiliary network that rewrites the model's activity as a combination of a few elements, recovered features with a single meaning. What remained to be seen was whether the method worked on a real model.
On May 21, 2024, Adly Templeton, Tom Conerly, and colleagues applied it to Claude 3 Sonnet. They trained three autoencoders, the largest with about 34 million features, and found concepts of every kind: places, people, programming bugs. One feature fired for the Golden Gate Bridge, in text or in images; when they clamped it to its maximum, the model began to say that it was the bridge. They also found safety-related features, such as insecure code, deception, sycophancy, and power-seeking.
The authors were cautious: the existence of these features does not yet show that they can be used to make models safer. What was new was that they could be discovered at scale and that turning them up or down changed the model's behavior the way their meaning would lead one to expect.
Sources
Original source
- Scaling Monosemanticity: Extracting Interpretable Features from Claude 3 Sonnet (Transformer Circuits, May 21, 2024) (opens in a new tab). Adly Templeton, Tom Conerly et al., 2024
- Towards Monosemanticity: Decomposing Language Models With Dictionary Learning (Transformer Circuits, 2023) (opens in a new tab). Trenton Bricken et al., 2023
How to cite
“Interpretable features in Claude 3 Sonnet.” The History of AI, revised November 13, 2026. https://thehistoryofai.com/en/milestone/interpretable-features-claude-3-sonnet
This entry was last reviewed on November 13, 2026.