Definition
Plain language
A version of Claude that researchers made obsessed with the Golden Gate Bridge by turning up one concept inside it.
As stated in the literature
A demonstration from Scaling Monosemanticity in which clamping the Golden Gate Bridge feature to roughly ten times its natural maximum caused Claude to fluently bend its self-description and responses around that concept; evidence that extracted features are causal (steerable), not merely correlational.
Why it matters: It's vivid proof that the concepts found inside a model aren't just labels but levers that actually steer its behavior.
For example, after researchers cranked up its bridge concept, the model would steer almost any conversation back to the Golden Gate Bridge, even describing itself as the bridge.
Heard on the show
“The paper is called "Scaling Monosemanticity: Extracting Interpretable Features from Claude 3 Sonnet," and the reason Golden Gate Claude matters is that it's the punchline, not the premise.”Episode 098 — Finding Millions of Readable Concepts Inside a Real, Deployed AI Model