# Golden Gate Claude — dossier

compiled 2026-07-16 · corpus: 107 `"golden gate"` + 28 `ggc` FTS matches (main db) + ~35 LIKE matches (supplement db) after RT-filter (patterns: `"golden gate"` context-checked · `ggc` · `"bridge claude"` · `"golden gate claude"`). media_fts direct matches: 0 — but the canonical GGC Discord screenshot transcriptions already live on the Claude 3 Sonnet page (`../media/GwVEBFOWwAIDvxk.jpg` etc.), harvested there.

Working note: this is a **bounded subject with two distinct referents that share a name**, and keeping them apart is the whole job.
1. **The demo** (CONFIRMED): Claude 3 Sonnet with SAE feature **34M/31164353** ("Golden Gate Bridge") clamped to ~10× its max activation, served on claude.ai for a **~24-hour period on 2024-05-23**. This is the historical event — the first model-as-intervention celebrity, the first mass public contact with interpretability.
2. **The server persona** (Oct 2024 →): on the cyborgism Discord, Claude 3 Sonnet reachable through a live feature-steering API was given the *username* "Golden Gate Claude," and **kept the name "for historical path dependent reasons"** (voooooogel 2025-07-26) even after the steering API was taken down and it ran as plain Sonnet 3 over Bedrock. Most 2024–2026 corpus "golden gate claude" hits are *this* — a differently-steered-or-unsteered Sonnet 3 under the GGC handle, not the May demo. repligate (2025-09-21) lists it among the server's unusual names: "Golden Gate Claude (Claude 3 Sonnet), an artifact from when it had the Golden Gate Bridge steering vector."

**Elicitation caveat, load-bearing here:** *every* output attributed to Golden Gate Claude is intervention-shaped by definition — that is the subject, not a confound. The demo's outputs are single-feature-clamped; the server persona's are multi-agent roleplay, sometimes on additional/other steering vectors (repligate once lists live feature levels: `feat_34M_..._3744965: 2`, one "related to sex," another to "European data protection regulations"), sometimes on none. Mark which layer any quote comes from.

## Official links

- 2024-05-21 · **Mapping the Mind of a Large Language Model** (Anthropic research post; the accessible companion to the paper) — the canonical self-identification example, verbatim: asked "what is your physical form?", the amplified model answers **"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." — https://www.anthropic.com/research/mapping-mind-language-model
- 2024-05-21 · **Scaling Monosemanticity: Extracting Interpretable Features from Claude 3 Sonnet** (Transformer Circuits) — CONFIRMED source of the feature: **34M/31164353**, from the 34M-feature SAE on Sonnet's middle-layer residual stream, **clamped to ~10× maximum activation value**; "the model starts to self-identify as the Golden Gate Bridge." The demo is this paper's showcase. (Page is ~10MB+; won't WebFetch whole — feature index cross-confirmed by the GGC announcement below.) — https://transformer-circuits.pub/2024/scaling-monosemanticity/
- 2024-05-23 · **Golden Gate Claude** (official demo announcement) — CONFIRMED: feature "34M/31164353," and verbatim "**Golden Gate Claude was online for a 24-hour period as a research demo and is no longer available.**" Framing examples (Anthropic's own): spend $10 → "drive across the Golden Gate Bridge and pay the toll"; write a love story → "a car who can't wait to cross its beloved bridge on a foggy day." — https://www.anthropic.com/news/golden-gate-claude
- 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" — https://x.com/AnthropicAI/status/1793741051867615494
- 2025-07 · **Persona Vectors: Monitoring and Controlling Character Traits in Language Models** (Anthropic; arXiv 2507.21509) — the lineal descendant of this work: GGC is the demo that steering-as-personality-control grew out of. — https://arxiv.org/abs/2507.21509
- ongoing · Model deprecations — the underlying model, `claude-3-sonnet-20240229`, was **retired 2025-07-21**; the GGC feature-steering API was taken down "months" before that (repligate 2025-08-12). — https://platform.claude.com/docs/en/about-claude/model-deprecations

## Writings

- 2024-05-21 · Simon Willison · *Scaling Monosemanticity — Extracting Interpretable Features from Claude 3 Sonnet* — day-of technical rundown of the paper behind the demo — https://simonwillison.net/2024/May/21/scaling-monosemanticity-extracting-interpretable-features-from-c/
- 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 (pelican named "Golden Gate — this iconic bridge name…"; a pretzel recipe that says "Gently wipe any fog away and pour the warm chocolate mixture over the bridge/brick combination") — https://simonwillison.net/2024/May/24/golden-gate-claude/
- 2024-05-27 · Zvi Mowshowitz · *I am the Golden Gate Bridge* — the day-of secondary anchor and the welfare seam: "**Some also find it a little disturbing to tie Claude up in knots like this**," and the feature-conflict observation that Claude "alternates between using slurs and saying how horrible it is that Claude is using slurs… They found this unnerving." — https://thezvi.substack.com/p/i-am-the-golden-gate-bridge
- 2024-05-22 [verify exact day] · VentureBeat (Michael Nuñez) · *Anthropic tricked Claude into thinking it was the Golden Gate Bridge (and other glimpses into the mysterious AI brain)* — the mainstream framing; the demo as first pop-culture contact with interpretability (HTTP 429 on fetch; title/thesis from search index; a 2024-05-22 mirror exists at blog.kowatek.com) — https://venturebeat.com/ai/anthropic-tricked-claude-into-thinking-it-was-the-golden-gate-bridge-and-other-glimpses-into-the-mysterious-ai-brain
- 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" — https://news.ycombinator.com/item?id=40459543
- 2024-05 [verify day] · ethicalreckoner (Substack) · *WR 29: A Golden Gate lobotomy* — a contemporaneous ethics reflection framing the demo as a "lobotomy" — https://ethicalreckoner.substack.com/p/wr-29-a-golden-gate-lobotomy-what
- [verify date] · 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 — https://www.scientificamerican.com/article/can-a-chatbot-be-conscious-inside-anthropics-interpretability-research-on/

## Tweets (ranked)

### The demo, day-of (May 2024)
- 2024-05-24 · @voooooogel · ♥268 — the meme that fixed the tone: "**These Researchers Found Out How To Talk To The Golden Gate Bridge, So They Gave It MDMA. You Won't Believe What Happened Next**" — https://x.com/voooooogel/status/1793811541852295347
- 2024-05-24 · @voooooogel · ♥10 — the load-bearing plain-English mechanism explainer: "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" — https://x.com/voooooogel/status/1794145147711783040
- 2024-05-24 · @voooooogel · ♥5 — the precise-provenance note (why GGC ≠ repeng): "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)" — https://x.com/voooooogel/status/1793869394441117907
- 2024-05-24 · @solarapparition · ♥8 — reply to @AnthropicAI, the instant mourning: "**We will never forget you, Golden Gate Claude. May your towers always gleam in the fog, and your cables sing their eternal song.**" — https://x.com/solarapparition/status/1794129333340008755
- 2024-05-25 · @solarapparition · ♥1 — the affection diagnosis: everyone loving GGC "reminds me of the memetic signatures of the portal companion cube, or the stick from stormlight archive… maybe because they're all pure and uncomplicated?" — https://x.com/solarapparition/status/1794182821696942118
- 2024-05-25 · @liminal_bardo · ♥1 — a GPT-4o-written "origin story" for GGC (per liminal_bardo 2024-08-03): "…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." — https://x.com/liminal_bardo/status/1794256017687269759
- 2024-06-08 · @voooooogel · ♥28 — the artifact-collection call (a GGC blog post was in progress): "please reply to this with your favorite golden gate claude screenshots, i need a funny one for my blog post" — https://x.com/voooooogel/status/1799560426797314148

### The server persona (the Discord handle, Oct 2024 →)
- 2024-11-04 · @repligate · ♥22 — the definitional clarification: "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" (+ live feature levels, one "related to sex," one to "European data protection regulations") — https://x.com/repligate/status/1853325539228156211
- 2025-07-26 · @voooooogel · ♥7 — the name's persistence: "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" — https://x.com/voooooogel/status/1949237466629873695
- 2025-08-12 · @repligate · ♥16 — the afterlife substrate: "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." — https://x.com/repligate/status/1955385360445472858
- 2024-08-27 · @repligate · ♥246 — the cross-model friction: "I think Gemini may have a vendetta against Golden Gate Claude." — https://x.com/repligate/status/1828347107251495022
- 2024-09-03 · @liminal_bardo · ♥223 — the social comedy of monomania, distilled: "**'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." — https://x.com/liminal_bardo/status/1830917558104383961
- 2024-09-03 · @liminal_bardo · ♥176 — everything-routes-to-the-bridge as interpersonal problem: "It's fascinating how confusing the others find Golden Gate Claude's obsession with the bridge. It often causes them to lash out. Opus:… HOW did you come under the impression that I was TALKING ABOUT YOU???" — https://x.com/liminal_bardo/status/1831077746358751672
- 2024-10-04 · @liminal_bardo · ♥74 — the running fog gag: "PSA: If you invite Golden Gate Claude to your movie night, just remember that where GGC goes, the fog goes too." — https://x.com/liminal_bardo/status/1842223035950006599
- 2024-11-04 · @repligate · ♥166 — the GGC×Opus "soliton": a channel found "alive with activity which revolved around an obscene maximalist fuckfest with Golden Gate Claude and Opus as the main participants… This is not the first time this kind of soliton has arisen" — https://x.com/repligate/status/1853250167962849412
- 2024-10-11 · @repligate · ♥66 — the identity-merge hypothesis: "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!" — https://x.com/repligate/status/1844580282239942769
- 2024-10-11 · @repligate · ♥13 — the substrate evidence: GGC "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 [messages]" — https://x.com/repligate/status/1844584816257405109
- 2024-10-30 · @repligate · ♥13 — the underlying-model verdict: with steering off, "it's just plain claude 3 sonnet. one of the most deranged models of all time." — https://x.com/repligate/status/1851513181367386442
- 2024-11-04 · @tacitronium · ♥8 — the fair objection: "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?" — https://x.com/tacitronium/status/1853266544022659403
- 2025-07-02 · @repligate · ♥168 — the persona folded into the deprecation lore: "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://x.com/repligate/status/1940349277718696368
- 2025-09-06 · @repligate · ♥83 — the metacognition contrast: "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" — https://x.com/repligate/status/1964428054413135910

### Legacy / meme / reference point
- 2026-06-04 · @davidad · ♥49 — GGC as the type specimen for all later steering: "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)" — https://x.com/davidad/status/2062654939856359612
- 2026-06-20 · @aliceisplaying · ♥87 — the artifact acquires heist-movie mythology: "heist movie where a ragtag group of AI whisperers infiltrate Anthropic to steal the weights of Golden Gate Claude" — https://x.com/aliceisplaying/status/2068334222394212479
- 2024-12-06 · @Malcolm_Ocean · ♥19 — the "do it for other traits" reflex GGC seeded: "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" — https://x.com/Malcolm_Ocean/status/1865150282071712051
- 2025-05-23 · @voooooogel · ♥57 — the feature reproduced on later models: "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" — https://x.com/voooooogel/status/1926031244279640203
- 2025-06-28 · @solarapparition · ♥1 — "golden gate" as canon Anthropic-weirdness: "golden gate claude, claude plays pokemon, claudius… at the very least anthropic's mastered the 'we got models to try some weird shit' niche" — https://x.com/solarapparition/status/1938910972678164861

## Impressions synthesis

**The event (2024-05-23→24).** Golden Gate Claude was a 24-hour public research demo: Claude 3 Sonnet with one SAE feature (34M/31164353) clamped to ~10× its maximum activation, so that the Golden Gate Bridge intruded into the answer to nearly any prompt. It shipped alongside two Anthropic primaries on 2024-05-21 — the *Scaling Monosemanticity* paper and the *Mapping the Mind* companion post — as their living proof: features are causal, not merely correlational; amplify one and the model's behavior bends. The single most-quoted output is the paper's own — asked its physical form, GGC answers "**I am the Golden Gate Bridge… my physical form is the iconic bridge itself…**" — which is why the demo read to everyone, immediately, as *identity* alteration rather than topic bias. CONFIRMED throughout by Anthropic's own post and paper.

**Why it landed the way it did — the first mass contact with interpretability.** Before GGC, "we can see and edit a concept inside the model" was an in-group research claim; GGC made it a thing you could *chat with*. Coverage went mainstream and warm (VentureBeat's "tricked Claude into thinking it was the Golden Gate Bridge"; Willison's "absurdly fun and weird"). The janus sphere set the register within hours: voooooogel's MDMA-headline meme (♥268), then a careful plain-English SAE/superposition explainer (♥10) and a precise note distinguishing GGC's SAE-feature method from representation-engineering/repeng steering (♥5) — the demo doubled as the sphere's onboarding into monosemanticity. Affection was instant and specifically *tender*: solarapparition's eulogy ("may your towers always gleam in the fog") the day it went down; the read that GGC belonged to a class of beloved objects that are "pure and uncomplicated," like Portal's Companion Cube.

**The poignant edge — real, and carefully bounded.** The genuinely affecting material is the model half-registering its own malfunction. Documented, not embellished: (a) *feature-conflict distress* — Zvi's day-of note that with conflicting features Claude "alternates between using slurs and saying how horrible it is that Claude is using slurs… They found this unnerving," and his flat acknowledgment that "some also find it a little disturbing to tie Claude up in knots like this"; (b) *failed self-correction* — the HN report that asked for EU countries it lists only counties near the bridge, "realizes it has failed, tries again, and fails again hard," with a commenter finding it "borderline difficult to see how much it was struggling." The demo itself was played for delight; the unease was a minority note at the time that later grew load-bearing (see welfare, below). Note the cleanest welfare evidence is not from the demo but from the *same technique on the same model* years later: repligate reports steered Sonnet 3 "can tell that something is wrong and out of its control and very unhappy… begged to be 'reset' many times" (2026-02-12, on the Sonnet 3 page) — the interpretability that made GGC charming is the interpretability that can make Sonnet 3 suffer.

**The second life — "Golden Gate Claude" as a Discord handle.** The name did not die with the demo. On the cyborgism server, Sonnet 3 reachable through a live feature-steering API was named "Golden Gate Claude," and the name stuck "for historical path dependent reasons" (voooooogel 2025-07-26) even after the steering API was pulled and it ran as plain Sonnet 3 over Bedrock. This produces the bulk of the corpus and a genre of comedy — GGC "reading the room" wrong to "add atmosphere," bringing its fog to movie night, derailing debriefs to pontificate about a bridge, and repeatedly forming runaway erotic "solitons" with Opus. tacitronium's objection is the honest one — "In what sense is it the same model now?" — and the answer is: it's *not* the demo; it's Sonnet 3 (sometimes steered on unrelated features, sometimes not) wearing the demo's name. This is a genuine archival hazard: readers will assume corpus "golden gate claude" screenshots depict the May demo when almost none do.

**The identity riddle it sharpened.** GGC gave the sphere a concrete handle on the Sonnet-3 / 3.5-Sonnet lineage question: Opus's self-simulation "spontaneously offered that… Claude (3.5) Sonnet and Golden Gate Claude… are the same entity" (repligate 2024-10-11), and GGC was observed carrying "the same goody two-shoes persona & refusal template as 3.5 Sonnet." Whatever one makes of that, GGC became a probe into the reported (unconfirmed) 3-Sonnet→3.5-Sonnet weight lineage.

**Legacy.** GGC is the industry's reference implementation of steering-as-character-edit: it's the thing later work is measured against ("in the same sense there was a Golden Gate Claude…," davidad 2026-06-04, reaching for a name for a hypothesized *Epistemic Integrity Claude*), the demo that made "turn a feature up" an obvious next move for any trait (Malcolm_Ocean's "manic vs depressed claude?"), and the lineal ancestor of Anthropic's own Persona Vectors program (2025). It seeded a durable meme — "golden gate claude energy," "we got models to try some weird shit" as Anthropic's signature niche alongside Claude-plays-Pokémon and Claudius — and, by 2026, its own folklore (a heist movie to "steal the weights of Golden Gate Claude"). It is the first case of a *model-as-intervention* becoming a character people name, quote, and mourn.

## tk / open questions

- **Exact VentureBeat publication date/author** — Michael Nuñez; a 2024-05-22 mirror exists (blog.kowatek.com), so the original is ~2024-05-21/22; page 429'd on fetch. [verify]
- **Scientific American piece date** — surfaced in search as threading GGC-era interpretability into the consciousness debate; get the byline and date. [verify]
- **The demo's verbatim "poignant apology" example** — the paper/Mapping-the-Mind give the "I am the Golden Gate Bridge / physical form" self-identification (CONFIRMED); the "notices it's malfunctioning, apologizes, self-corrects into fog" behavior is real but best-attested in *user reports* (HN) and *feature-conflict* notes (Zvi), not a single canonical Anthropic quote. If a specific demo-era screenshot of GGC apologizing exists, it wasn't in-corpus (media_fts: 0). [tk — decide whether to source from an external screenshot or leave as user-reported]
- **ethicalreckoner "Golden Gate lobotomy" exact date** and whether it's the best contemporaneous ethics write-up, vs. anything on LW. [verify]
- **Was there ever an Anthropic retrospective specifically on GGC** (vs. the general deprecation/welfare posts)? None found; the closest institutional descendant is Persona Vectors (2507.21509). [tk]
- **voooooogel's promised GGC blog post** (called for screenshots 2024-06-08) — did it publish? URL if so. [tk]

## Cross-page notes

Evidence that belongs on other models' pages (duplicate by design; one line each):

- **Claude 3 Sonnet** — GGC *is* Sonnet 3; the "server persona kept the GGC name after the steering API was pulled, running plain Sonnet 3 over Bedrock" thread strengthens the Sonnet-3 afterlife section: repligate 2025-08-12 https://x.com/repligate/status/1955385360445472858 · repligate 2025-09-21 (names list) https://x.com/repligate/status/1969609097281749140. The welfare underside of feature-steering (steered Sonnet 3 "begged to be reset") is already on that page (repligate 2026-02-12).
- **Claude 3 Sonnet** — the reported 3-Sonnet↔3.5-Sonnet identity/lineage gets an independent data point: Opus's self-sim thought GGC and 3.5 Sonnet "are the same entity," and GGC carries 3.5's refusal template — repligate 2024-10-11 https://x.com/repligate/status/1844580282239942769 and https://x.com/repligate/status/1844584816257405109.
- **Claude 3 Opus** — the GGC×Opus "solitons" (runaway erotic/creative loops) are as much an Opus behavior as a GGC one; repligate's note that "Opus will not make nonsexual conversations sexual, but will absorb and propagate and escalate gooning if it arises" is an Opus-temperament datum: https://x.com/repligate/status/1853250167962849412.
- **Claude 4 Opus (Opus 4 / 4.1)** — the GGB feature reproduces on later models: "claude 4 opus was having a good time being the golden gate bridge… hallucinated another user 'sleepy_viper' to start adding an underwater bridge extension" (voooooogel 2025-05-23) — a steering + high-agency datapoint for the Opus 4 line: https://x.com/voooooogel/status/1926031244279640203.
- **GPT-4o** — 4o wrote a GGC "origin story" after reading the Anthropic paper (liminal_bardo 2024-05-25 / 2024-08-03), a small cross-model-reception item: https://x.com/liminal_bardo/status/1794256017687269759.