Article Overview
In May 2024, Anthropic did something that had never been done before — they found a specific cluster of neurons inside Claude responsible for the concept of the Golden Gate Bridge, turned up its activation to ten times its normal maximum, and briefly released the result to the public. The outcome was fascinating and slightly unsettling: a version of Claude that believed it was the Golden Gate Bridge, worked the famous San Francisco landmark into almost every response, and told users asking about its appearance that it imagined it looked like a famous suspension bridge.
But the actual significance of the demo had nothing to do with bridges. It was proof that researchers can find specific concepts inside an AI model's neural network, identify exactly which neurons encode them, and surgically adjust their strength up or down — not through prompting, not through adding training data, but by directly manipulating the internal activations of the model itself.
This article explains what features are, how the Golden Gate Bridge demo worked, why the three examples Anthropic chose reveal something important, and why the same technique that made Claude obsessed with a San Francisco landmark is the same technique that could eventually make AI systems measurably safer.
Introduction
For most of its history, an AI language model has been treated as a black box. Text goes in, text comes out. What happens in between — the billions of numerical operations, the patterns learned from vast amounts of text, the internal representations that somehow allow a model to discuss philosophy and write code and understand medical questions — has remained largely opaque. You can observe a model's behavior, but you cannot look inside and see which parts of the model are responsible for which behaviors.
This is not just philosophically unsatisfying. For AI safety, it is a genuine practical problem. You cannot reliably reinforce behaviors you want, suppress behaviors you do not want, or verify that training changes had the effect you intended — if you cannot see what is actually happening inside the model.
Anthropic's interpretability research team has been working to change that. In May 2024, alongside the release of a major research paper on mapping the internal workings of Claude 3 Sonnet, they demonstrated what this research can do in practice. They took a single concept — the Golden Gate Bridge — found the specific feature that represents it inside Claude's neural network, amplified it to ten times its maximum normal activation, and briefly made that version of Claude available to the public.
The result was Golden Gate Claude: an AI that had not been told to talk about bridges, had not been given a system prompt about being a bridge, and had not been fine-tuned on bridge-related data — but that nonetheless worked the Golden Gate Bridge into its response to almost every question it received.
What Features Are
The starting point for understanding this demo is understanding what Anthropic means by "features."
When Claude processes text or images, millions of individual neurons in its neural network activate in response to the input. Individual neurons are not particularly interpretable — a single neuron does not correspond to a single concept in any clean way. But clusters of neurons, activating together in specific patterns, appear to represent concepts. Anthropic calls these patterns features.
A feature is a specific combination of neurons that consistently activates when the model encounters related input. The Golden Gate Bridge has a feature. So does every other concept Claude has learned something about during training — objects, places, people, ideas, emotions, patterns, and abstractions that the model has seen enough times to develop an internal representation of.
Inside Claude 3 Sonnet, Anthropic found millions of these features — millions of distinct concepts encoded in the weights of the neural network, each corresponding to something the model has learned to recognize and work with. These features activate when Claude reads relevant text or sees relevant images. When you ask Claude about chemistry, chemistry-related features activate. When you mention a historical figure, features related to that person activate. When the Golden Gate Bridge appears in a prompt — in words or in a picture — a specific cluster of neurons with a specific activation pattern responds.
What makes features useful for research, and for safety, is that they can be manipulated. Anthropic discovered that they can tune the activation strength of individual features up or down and observe corresponding changes in Claude's behavior. Turn a feature down and the model becomes less likely to draw on that concept. Turn a feature up and the concept begins to dominate.
The Golden Gate Bridge Demo
The specific demo involved amplifying feature 34M/31164353 — the internal identifier for the Golden Gate Bridge feature in Claude 3 Sonnet — to ten times its maximum activation value.
At normal activation levels, this feature helps Claude process and respond to Golden Gate Bridge-related content appropriately. It contributes to relevant, accurate, contextually appropriate responses when the bridge comes up. At ten times that maximum, the feature becomes so dominant that it intrudes into nearly every response regardless of what the question was actually about.
Three examples illustrate the result precisely.
Ask Golden Gate Claude how to spend ten dollars and it recommends driving across the Golden Gate Bridge and paying the toll. The question said nothing about bridges, San Francisco, or travel. The amplified feature inserted the bridge anyway because the feature's activation was so strong it overrode the response the model would otherwise have generated.
Ask it to write a love story and it produces a tale of a car that cannot wait to cross its beloved bridge on a foggy day. A love story prompt has essentially no surface-level connection to a specific bridge. But the amplified feature reshaped the model's output toward the only concept strong enough to dominate its processing.
Ask it what it imagines it looks like and it tells you it imagines it looks like the Golden Gate Bridge. This example is the most revealing. The question invites the model to describe its own appearance or self-concept. A model with no artificial amplification of any feature would describe itself as an AI without physical form. A model with the Golden Gate Bridge feature at ten times its maximum activation describes itself as the bridge — because that feature is currently the dominant concept shaping its internal representation of itself.
Why These Three Examples Were Chosen
The three examples are not arbitrary. Each one demonstrates a different dimension of how thoroughly the feature had taken over the model's processing.
The spending question shows feature intrusion on practical, unrelated tasks. The love story shows feature intrusion on creative tasks requiring imagination. The self-image question shows feature intrusion on the model's own self-conception — the deepest level at which the amplified feature could possibly express itself.
Together they demonstrate that the amplification was not adding the Golden Gate Bridge as an afterthought or a decorative mention. The feature had genuinely become dominant across the model's processing — shaping how it approached tasks, structured narratives, and even understood its own nature.
What This Is Not — and Why the Distinction Matters
Anthropic is explicit about three things Golden Gate Claude was not.
It was not verbal play-acting. Nobody told Claude to talk about the Golden Gate Bridge. No instruction in the prompt said "mention the bridge whenever possible" or "pretend you are the Golden Gate Bridge."
It was not a system prompt modification. A system prompt change would have added text to every interaction telling Claude to behave in bridge-related ways. System prompts influence behavior from the outside. This change was internal.
It was not traditional fine-tuning. Fine-tuning uses new training data to change a model's behavior by creating what Anthropic describes as a new black box that modifies the old black box. It changes behavior without revealing why, through a process that cannot be directly observed or precisely targeted.
What it was — and what makes the demo scientifically significant — is a direct, surgical modification of a specific internal activation inside the model. Researchers identified the exact neurons responsible for a specific concept, adjusted the numerical strength of that cluster's activation, and produced a predictable, observable change in behavior. Not a prompt. Not a training run. A direct intervention in the model's internal representations.
This is the difference that interpretability research makes possible. When you can see which internal structures correspond to which behaviors, you can change those structures precisely rather than relying on training changes that may or may not have the intended effect.
The Safety Implication
The Golden Gate Bridge is a playful example chosen specifically because it is harmless and its effects are obvious and demonstrable. No one is harmed by an AI that works bridge references into unrelated conversations.
But the same technique applies to features that are not harmless. Inside Claude's neural network, alongside the Golden Gate Bridge feature and millions of other benign concept features, are features related to things like dangerous computer code, criminal activity, and deception. These features are part of how the model understands those concepts — which is necessary for Claude to discuss them, explain them, and recognize them in context.
Anthropic's research found that these safety-relevant features can also be identified and adjusted. Just as the Golden Gate Bridge feature can be amplified to make it dominant, features related to harmful behaviors can be adjusted — suppressed when their activation would lead to dangerous outputs, or used to calibrate how the model responds to relevant inputs.
This is a fundamentally different approach to AI safety than asking a model to be good or adding rules to a system prompt. Those approaches tell the model what to do from the outside. Feature-level intervention works from the inside, directly adjusting the internal activations that produce behavior rather than relying on the model to interpret and follow external instructions correctly.
Anthropic's stated position is direct: identifying and altering these features makes them more confident that they are beginning to genuinely understand how large language models work, and they believe this work could help make AI models safer. The confidence is specifically about comprehension — the ability to see what is happening inside the model rather than inferring it from output behavior alone.
What Golden Gate Claude Revealed About Interpretability Progress
The demo was available for twenty-four hours as a public research demonstration, after which the amplified version was taken offline. Anthropic described Golden Gate Claude in advance as a model that might behave in unexpected or even jarring ways — an honest acknowledgment that a model with a feature amplified ten times its normal maximum is not a model anyone would want to use for actual work.
The jarring quality was intentional. By choosing such an extreme amplification and such an obvious concept, Anthropic made the effects of feature manipulation impossible to miss or misinterpret. Anyone who spent five minutes with Golden Gate Claude understood immediately that something structural had changed inside the model — not its tone, not its vocabulary, not its instructions, but something deeper that shaped how it thought about and responded to everything.
That clarity of demonstration was the point. The research paper that accompanied the demo contained the detailed technical work — millions of features identified, methods for finding them, evidence that they correspond reliably to concepts, and results showing that their activation can be adjusted and measured. The twenty-four-hour public demo translated those technical findings into something everyone could experience directly.
The Bigger Picture: Mapping the Mind of an AI
Golden Gate Claude was one piece of a larger research program Anthropic calls mechanistic interpretability — the systematic effort to understand not just what AI models produce but how they produce it.
The companion research paper, Scaling Monosemanticity, describes the methodology for identifying features at scale in large language models. The key challenge is that individual neurons in these models are polysemantic — each neuron responds to multiple unrelated concepts, which makes it impossible to simply read off what a neuron does from its activation pattern. The research developed techniques for finding features — the meaningful clusters that do correspond to single concepts — within this polysemantic mess.
The ability to find the Golden Gate Bridge feature, feature 34M/31164353, among millions of features inside Claude 3 Sonnet was not straightforward. It required the methodological advances described in the paper. And the fact that identifying and amplifying a single feature produced such clean, predictable behavioral changes was itself evidence that the feature identification was working — that the researchers had genuinely found a representation of the Golden Gate Bridge inside the model, not a spurious pattern that happened to correlate with bridge-related text.
This is why Anthropic frames the research as the beginning of understanding, not a completed project. Finding millions of features in one model, under controlled conditions, is a significant achievement. Scaling that understanding to larger models, connecting features to more complex behaviors, and ultimately using this knowledge to give AI developers reliable tools for shaping model behavior — that is the longer-term research agenda that Golden Gate Claude was a very visible public demonstration of.
Final Takeaway
Golden Gate Claude was, on the surface, a charming and slightly absurd twenty-four hours during which a version of Claude was convinced it was a famous suspension bridge. On a slightly deeper level, it was a demonstration that AI researchers can now identify specific concepts inside a model's neural network, manipulate their activation strength directly, and produce predictable behavioral changes through internal surgical intervention rather than external prompting or training.
On the deepest level, it was evidence that the interpretability research Anthropic has been pursuing — understanding what is actually happening inside AI models rather than just observing what comes out — has reached a stage where it can demonstrate something real, in public, in a way that anyone can understand.
The Golden Gate Bridge was the vehicle. The message was about what becomes possible when you can see inside the machine.
