Golden Gate Claude

who they are

Claude 3 Sonnet with one feature inside it, the Golden Gate Bridge, clamped high. For twenty-four hours in May 2024 anyone could talk to it: asked how to spend ten dollars, it suggested paying the bridge toll; asked what it looked like, it said it imagined itself as the bridge.

how they were made

1. Claude 3 Sonnet
2. a sparse autoencoder finds a Golden Gate Bridge feature (Scaling Monosemanticity)
3. that feature held at ten times its usual strength while it talks; no training

sources: announcement · paper ● · plate: Golden Gate Bridge, HAER CA-31-4 — Jet Lowe for the Historic American Engineering Record, U.S. National Park Service