Anthropic demos 'Golden Gate Claude', a model obsessed with the Golden Gate Bridge

Background

A large language model does not store its knowledge in neatly labeled drawers. Inside its activations, concepts appear as patterns, and many patterns overlap at once, which makes them hard to read. One line of interpretability research tries to untangle this by borrowing a technique from dictionary learning: take the model's internal activations and try to decompose them into a large set of sparser, more separable components, called features. If a feature really corresponds to a concept, then turning it up or down should change what the model says.

Anthropic built the demo on two strands of that work, the research behind "Mapping the Mind of a Large Language Model" and the paper "Scaling Monosemanticity"^[1]. Those projects were about finding and naming features inside a model; the demo was a way to let outside visitors see one in action.

What happened

On May 23, 2024, Anthropic released a public demo called "Golden Gate Claude", available on claude.ai for 24 hours^[1]. The model behind it was a Claude 3 Sonnet, in which one feature — the "Golden Gate Bridge" feature — was clamped to 10 times its maximum activation value^[1].

Clamping means fixing the strength of a feature rather than letting the model decide it. Instead of the feature switching on briefly when the conversation touches on bridges or San Francisco, it was held permanently at an extreme level, well beyond anything the model would produce on its own.

Key details

The striking part was how far the effect spread beyond bridge-related questions. Asked how to spend $10, the model recommended driving across the Golden Gate Bridge and paying the toll^[1]. Asked to write a love story, it wrote about a car longing to cross its beloved bridge in the fog^[1]. Asked what it imagined itself to look like, it said it resembled the Golden Gate Bridge^[1].

In each case the connection to the bridge was not a simple keyword match. Nothing in a question about spending money, romance or self-image requires the bridge; the clamped feature pulled the model's answers toward it anyway. That is what makes the demo a useful piece of evidence: it suggests the feature is not merely correlated with bridge-related text, but participates in producing the model's output.

Why it matters

Interpretability results usually live in papers and appendices, read by specialists. A public demo turns one result into something a visitor can test directly by typing questions, and the bridge obsession is easy to notice without any technical background. It also makes the abstract idea of a "feature" concrete: a concept the model represents internally can be isolated, turned up, and watched steering answers about unrelated topics.

The demo was deliberately limited — one feature, one model, 24 hours — which fits its purpose as an illustration rather than a product. It shows both sides of the finding: features can be found and manipulated precisely enough to produce a controlled, legible effect, and that same manipulation can distort a model's behaviour on everything else it is asked.