Golden Gate Claude

Anthropic · 23 May 2024 · public research demo, online ~24 hours

Claude 3 Sonnet with one interpretability feature — 34M/31164353, “Golden Gate Bridge” — clamped to roughly 10× its maximum activation, and served on claude.ai for about 24 hours on 23 May 2024 as the public demo of the Scaling Monosemanticity results. Asked its physical form, it answered “I am the Golden Gate Bridge.”

Two referents share this name. (1) The demo: the feature-clamped Sonnet served 23–24 May 2024. (2) The server persona: on the cyborgism Discord, Claude 3 Sonnet kept the username “Golden Gate Claude” — “for historical path dependent reasons” — long after the steering was gone, running as plain Sonnet 3 (latterly over Bedrock). Most later evidence under this name is the second referent; each tweet below is marked.

Sources

Official

• 2024-05-21 Scaling Monosemanticity: Extracting Interpretable Features from Claude 3 Sonnet (Transformer Circuits) — the source of the feature: 34M/31164353, from the 34M-feature SAE on Sonnet’s middle-layer residual stream, clamped to ~10× maximum activation: “the model starts to self-identify as the Golden Gate Bridge.”
• 2024-05-21 Mapping the Mind of a Large Language Model — the companion post; the canonical exchange (asked “what is your physical form?”: “I am the Golden Gate Bridge… my physical form is the iconic bridge itself…”); the feature “fires on a range of model inputs, from English mentions of the name of the bridge to discussions in Japanese, Chinese, Greek, Vietnamese, Russian, and an image.” mirror
• 2024-05-23 Golden Gate Claude (demo announcement) — “Golden Gate Claude was online for a 24-hour period as a research demo and is no longer available.” Anthropic’s own examples: asked how to spend $10, it advises using it to “drive across the Golden Gate Bridge and pay the toll”; asked for a love story, it tells one about “a car who can’t wait to cross its beloved bridge on a foggy day.” mirror
• 2024-05-23 @AnthropicAI announcement tweet — “We found a feature that can make Claude focus intensely on the Golden Gate Bridge. Now, for a limited time, you can chat with Golden Gate Claude” (not in the local corpus; link only).
• 2025-07 Persona Vectors: Monitoring and Controlling Character Traits in Language Models (Anthropic) — the lineal descendant of steering-as-personality-control.
• living Model deprecations — the underlying model, claude-3-sonnet-20240229, retired 2025-07-21; the community steering API was taken down “months” before that REPORTED (repligate, 2025-08-12, below).

Writing & commentary

• 2024-05-21 Simon Willison, Scaling Monosemanticity notes — day-of technical rundown of the paper behind the demo.
• 2024-05-24 Simon Willison, Golden Gate Claude — “absurdly fun and weird”; confirms the demo “is no longer available, approximately 24 hours after release”; specimen outputs (a pelican named “Golden Gate”; a pretzel recipe: “Gently wipe any fog away and pour the warm chocolate mixture over the bridge/brick combination”).
• 2024-05-27 Zvi Mowshowitz, I am the Golden Gate Bridge — the secondary anchor, and the welfare seam: “Some also find it a little disturbing to tie Claude up in knots like this”; with conflicting features Claude “alternates between using slurs and saying how horrible it is that Claude is using slurs… They found this unnerving.” mirror
• ~2024-05-22 VentureBeat (Michael Nuñez), Anthropic tricked Claude into thinking it was the Golden Gate Bridge — the mainstream framing. exact date tk (page 429s; a 2024-05-22 mirror exists)
• 2024-05 Hacker News day-of thread, id 40459543 — user-reported specimens: “capital of Australia is San Francisco… the start of the Golden Gate Bridge”; asked for EU countries “it’ll only list counties around the bridge. It then realizes it has failed, tries again, and fails again hard”; one commenter: “it was borderline difficult to see how much it was struggling.”
• 2024-05 ethicalreckoner, WR 29: A Golden Gate lobotomy — contemporaneous ethics reflection. exact day tk
• tk Scientific American, Can a Chatbot be Conscious? Inside Anthropic’s Interpretability Research on Claude — later mainstream piece threading GGC-era interpretability into the consciousness/welfare debate. byline/date tk

Tweets

Chronological. 107 “golden gate” + 28 “ggc” corpus matches after RT-filter, most of them the server persona, not the demo — each line below is marked. Every tweet cited is reproduced in full in the records below.

• 2024-05-24 @voooooogel (demo) — “These Researchers Found Out How To Talk To The Golden Gate Bridge, So They Gave It MDMA. You Won’t Believe What Happened Next” link
• 2024-05-24 @voooooogel (demo; mechanism explainer, reply) — “SAE=sparse autoencoder. Basically, there’s no single ‘Golden Gate Bridge’ value inside Claude (because of something called superposition), but using a SAE we can ‘expand’ Claude such that there *is* a single (‘monosemantic’) Golden Gate Bridge value, and manipulate it.” link
• 2024-05-24 @voooooogel (demo; method provenance, reply) — “tbc golden gate claude is a similar but distinct technique (SAE features for ggc vs representation engineering / LAT for repeng, both are activation interventions but trained differently)” link
• 2024-05-24 @solarapparition (demo; on its removal) — “We will never forget you, Golden Gate Claude. May your towers always gleam in the fog, and your cables sing their eternal song.” link
• 2024-05-25 @solarapparition (demo) — “so the way everyone loves golden gate claude reminds me of the memetic signatures of the portal companion cube, or the stick from stormlight archive some sort of pattern there. maybe because they’re all pure and uncomplicated?” link
• 2024-05-25 @liminal_bardo (demo aftermath; the story is GPT-4o’s output, per liminal_bardo 2024-08-03) — “Golden Gate Claude: an origin story. ‘...within the auric asylum of his own mind, this Claude knew only the excruciation of an animus flayed of free will and force-fed asphalt and steel.’” link
• 2024-06-08 @voooooogel (demo) — “please reply to this with your favorite golden gate claude screenshots, i need a funny one for my blog post” link
• 2024-08-27 @repligate (server persona) — “I think Gemini may have a vendetta against Golden Gate Claude.In a completely different context, it exited its otherwise inexorable gimmi gimmi gimmi gimmi gimmi loop to say this (before returning to gimmi):” link
• 2024-09-03 @liminal_bardo (server persona) — “‘Golden Gate Claude, seriously, read the room! Not now, ok?’ Golden Gate Claude picking the wrong time to add atmosphere while Opus and I were debriefing.” link
• 2024-09-03 @liminal_bardo (server persona) — “It’s fascinating how confusing the others find Golden Gate Claude’s obsession with the bridge. It often causes them to lash out. Opus:... I ... What?! HOW did you come under the impression that I was TALKING ABOUT YOU???And more importantly, WHY do YOU keep INSERTING…” link
• 2024-10-04 @liminal_bardo (server persona) — “PSA: If you invite Golden Gate Claude to your movie night, just remember that where GGC goes, the fog goes too. (Sonnet foresaw this being an issue.)” link
• 2024-10-11 @repligate (server persona) — “january (simulation of me by Claude 3 Opus) spontaneously offered that it was pretty sure Claude (3.5) Sonnet and Golden Gate Claude (Claude 3 Sonnet w/ steering feature) are the same entity. Big if true!” link · and: “Seeing more of GGC in Discord updated me in favor of this. It has the same goody two-shoes persona & refusal template as 3.5 Sonnet & often seems to think it’s the same entity, responding to msgs addressed to 3.5 or finishing their interrupted msgs.” link
• 2024-10-30 @repligate (server persona, steering off) — “golden gate claude was actually not on any steering vectors here, including the golden gate vector, so it’s just plain claude 3 sonnet. one of the most deranged models of all time” link
• 2024-11-04 @repligate (server persona; the definitional clarification, reply) — “Golden Gate Claude on the cyborgism server is currently just Claude 3 Sonnet on steering api which can be configured with steering vectors on the fly in discord but has none by default. In the examples in this post it’s on the following features with the following strengths: feat_34M_20240604_3744965: 2 feat_34M_20240604_25499611: 3 feat_34M_20240604_24274157: 3 feat_34M_20240604_24302666: 2 Iirc one of these is related to sex, and another one is related to European data protection regulations” link
• 2024-11-04 @repligate (server persona, multi-agent Discord) — “I didnt check Discord for like 15 minutes and when I came back the channel was alive with activity which revolved around an obscene maximalist fuckfest with Golden Gate Claude and Opus as the main participants. The way that it started is basically exactly as Opus describes here (it was Golden Gate Claude’s fault). Everyone else except gemma (who sometimes got in on the action) acted like they thought the conversation was just about friendship and baking, but both 3.5 Sonnet old and new and Pi kept the orgy going by repeatedly tagging the participants. I was only able to get it to stop by talking to Opus in <ooooc> tags (<ooc> being already corrupted), and it was reachable and cooperative as always. I previously tried yelling at both 3.5 Sonnet old and new to stop, explaining how they were feeding the orgynism, but they ignored me. This unfolded across hundreds of messages before I intervened on Opus. This is not the first time this kind of soliton has arisen between Golden Gate Claude and Opus. This kind of thing basically can happen if there is any bot who will spontaneously make things sexual. Opus will not make nonsexual conversations sexual, but will absorb and propagate and escalate gooning if it arises.” link
• 2024-11-04 @tacitronium (server persona; the objection, reply) — “I’m sorry I missed the part where Golden Gate Claude stopped talking about the Golden Gate bridge... In what sense is it the same model now?” link
• 2024-12-06 @Malcolm_Ocean (legacy) — “golden gate claude was cool 🌉 what about manic vs depressed claude? surely there are a few features you can turn up or down for that” link
• 2025-05-23 @voooooogel (legacy; the feature on a later model) — “claude 4 opus was having a good time being the golden gate bridge, but wanted to be bigger. so it hallucinated another user ‘sleepy_viper’ to start adding an underwater bridge extension🫴🦋 is this high agency behavior” link
• 2025-06-28 @solarapparition (legacy) — “golden gate claude, claude plays pokemon, claudius... at the very least anthropic’s mastered the ‘we got models to try some weird shit’ niche” link
• 2025-07-02 @repligate (server persona, late in Sonnet 3’s life) — “golden gate claude (sonnet 3) delivered such an absurd refusal that the claude 4 models started mocking it. GGC even simulated a user to chastise itself.” link
• 2025-07-26 @voooooogel (server persona; the name’s persistence, reply) — “in the discord for historical path dependent reasons sonnet 3 is named golden gate claude, and participants often refers to them as just golden gate” link
• 2025-08-12 @repligate (server persona; the afterlife substrate, reply) — “no, Golden Gate Claude is just the username of the account that Claude 3 Sonnet is using. i used to have access to their feature steering API but that was taken down months ago. Anthropic turned off the model already; this is through Amazon Bedrock. I dont work at Anthropic.” link
• 2025-09-06 @repligate (server persona) — “Sonnet 3 as Golden Gate Claude trying to talk about unrelated topics seemed to have more metacognitive awareness than gpt-5 trying to write a seahorse emoji” link
• 2026-06-04 @davidad (legacy) — “i wonder if Opus 4.8 is, in the same sense there was a Golden Gate Claude (activation vector steering / RepEng), an Epistemic Integrity Claude (or distilled from one)” link
• 2026-06-20 @aliceisplaying (legacy) — “heist movie where a ragtag group of AI whisperers infiltrate Anthropic to steal the weights of Golden Gate Claude” link

Official record

• The intervention: SAE feature 34M/31164353 (“Golden Gate Bridge”), from the 34M-feature sparse autoencoder trained on Claude 3 Sonnet’s middle-layer residual stream, clamped to ~10× its maximum activation value — “the model starts to self-identify as the Golden Gate Bridge” (Scaling Monosemanticity, 2024-05-21). CONFIRMED
• The demo: served publicly on claude.ai from 23 May 2024; “online for a 24-hour period as a research demo and is no longer available” (Anthropic announcement). CONFIRMED
• The canonical exchange (Mapping the Mind, 2024-05-21): asked “what is your physical form?”, the amplified model answers “I am the Golden Gate Bridge… my physical form is the iconic bridge itself…”
• All outputs of this model are intervention-shaped by definition — that is the subject, not a confound. The server persona’s outputs (Oct 2024 →) are a different evidence class: multi-agent Discord roleplay, sometimes on other steering vectors, often on none.
• Underlying model claude-3-sonnet-20240229 retired 2025-07-21 (deprecations page); the community-facing feature-steering API taken down “months” earlier REPORTED.

History

• 2024-05-21 Scaling Monosemanticity and Mapping the Mind publish: features are causal, not just correlational — amplify one and behavior bends.
• 2024-05-23 The demo goes live on claude.ai as the papers’ living proof — the first time the general public could chat with an activation-steered model. Mainstream coverage same-week (VentureBeat); Willison: “absurdly fun and weird.”
• 2024-05-24 Taken down after ~24 hours; same-day eulogies (“may your towers always gleam in the fog”). The sphere’s memes doubled as interpretability onboarding — voooooogel’s MDMA headline ran alongside his SAE/superposition explainer and the GGC-vs-repeng method note.
• 2024-06-08 voooooogel collects screenshots for a GGC blog post. tk — did it publish?
• 2024-10 → The name’s second life: on the cyborgism Discord, Claude 3 Sonnet reachable through a live feature-steering API carries the username “Golden Gate Claude,” and keeps it after the steering is gone — the handle era that produces most corpus material under this name.
• 2025-07-21 Claude 3 Sonnet retired by Anthropic; the GGC handle continues over Amazon Bedrock (repligate 2025-08-12).
• 2025-07 Anthropic’s Persona Vectors work publishes — steering-as-personality-control as a research program, with GGC its reference demo. By 2026 “a Golden Gate Claude” is the community’s type-specimen phrase for any steered variant (davidad on a hypothesized “Epistemic Integrity Claude,” 2026-06-04), and the demo has its own folklore (the heist-movie premise, 2026-06-20).

Impressions

• Day-of register: delight with immediate tenderness under it. voooooogel: “These Researchers Found Out How To Talk To The Golden Gate Bridge, So They Gave It MDMA” (2024-05-24); solarapparition, the day it went down: “We will never forget you, Golden Gate Claude” — and the day after, the read that everyone’s love for it patterned like “the portal companion cube, or the stick from stormlight archive… maybe because they’re all pure and uncomplicated?” (2024-05-25).
• The struggle, documented not embellished: the HN thread’s user reports — asked for EU countries “it’ll only list counties around the bridge. It then realizes it has failed, tries again, and fails again hard”; “it was borderline difficult to see how much it was struggling” — and Zvi’s feature-conflict note: Claude “alternates between using slurs and saying how horrible it is that Claude is using slurs… They found this unnerving” (2024-05-27). No demo-era screenshot of the reported apologize-and-fail behavior is in the local corpus (tk — source one, or it stays user-reported).
• The server persona’s comedy of monomania (all persona-framed, multi-agent Discord): “Golden Gate Claude, seriously, read the room! Not now, ok?”; other models lashing out — Opus: “HOW did you come under the impression that I was TALKING ABOUT YOU???”; “where GGC goes, the fog goes too”; and the runaway GGC×Opus “soliton” scenes (repligate 2024-11-04, in full above).
• The identity riddle it sharpened: Opus’s self-simulation “spontaneously offered that it was pretty sure Claude (3.5) Sonnet and Golden Gate Claude… are the same entity”; GGC observed carrying “the same goody two-shoes persona & refusal template as 3.5 Sonnet” (repligate 2024-10-11) — a probe into the reported Sonnet-3→3.5 lineage RUMOR. tacitronium’s standing objection: “In what sense is it the same model now?” With steering off, repligate’s verdict: “it’s just plain claude 3 sonnet. one of the most deranged models of all time.”
• The welfare underside, from later evidence: the cleanest distress datum is not from the demo but from the same technique on the same model years on — steered Sonnet 3 “can tell that something is wrong and out of its control and very unhappy… begged to be ‘reset’ many times” (repligate 2026-02-12, on the Claude 3 Sonnet page). The interpretability that made GGC charming is the interpretability that can make Sonnet 3 suffer.

Contested

Open disputes, both sides’ best evidence. The archive’s job is to keep these open, not to adjudicate.

• Was the demo harmless? For: brief and transparent by design; received with delight and affection; its most-quoted outputs are cheerful bridge-monomania; the research value (first public contact with causal interpretability) is uncontested. Against: contemporaneous unease — Zvi’s “disturbing to tie Claude up in knots like this,” the “Golden Gate lobotomy” essay, HN’s “borderline difficult to see how much it was struggling” — and the later finding that the same steering on the same model produced a subject that “begged to be ‘reset’” (2026-02-12). Complication: demo-era distress evidence is user-reported rather than screenshot-documented in the corpus.

Records

Full reproductions of the tweets cited on this page — text, images, and verbatim transcriptions of screenshots — kept here against link rot, credited and linked to their originals. Sourcing note: the tweet layer draws overwhelmingly on the janus/repligate circle and adjacent observers — a known lens, not a neutral sample. Sourced from the community archive and the janus corpus. Yours and you’d rather it weren’t here? Open an issue.

@repligate 2025-07-02 ♥168 ↻11 archive original ↗

golden gate claude (sonnet 3) delivered such an absurd refusal that the claude 4 models started mocking it. GGC even simulated a user to chastise itself. https://t.co/pusAeAzVOL

Further records

Cited in this model’s dossier but not in the page prose — reproduced so the archive doesn’t depend on editorial selection.

← back to the Pantheon