ClaimsDario Amodei

“Most memorably, we used this method to create “Golden Gate Claude”, a version of one of Anthropic’s models where the “Golden Gate Bridge” feature was artificially amplified”

Dario Amodei, Anthropic co-founder, writing on his own site 18 February 2025 The Urgency of Interpretability (darioamodei.com) archived copy

Factual AI safety and alignment

Anthropic created "Golden Gate Claude", a version of one of its models with the Golden Gate Bridge feature artificially amplified, making it obsessed with the bridge.

Resolved true The record shows it happened or was so.

Factual: Statements about the present or the past, checked against the record: did they hold up?

Said for an organisation they lead or work for; it counts on both records.

What would count as true or false

True: Anthropic's records show it released or demonstrated 'Golden Gate Claude', a version of a Claude model in which an interpretability feature for the Golden Gate Bridge was artificially amplified, causing the model to fixate on the bridge. Mostly true: as described, with a minor detail off (e.g. which model or exactly how it was steered). Mostly false: a related Golden Gate Bridge feature experiment existed, but not as a released model version, or not done by amplifying a feature. False: no such model or experiment existed.

Set before the research began.

Evidence (4)

  1. E1 Golden Gate Claude Anthropic · 23 May 2024 Own record

    “UPDATE: Golden Gate Claude was online for a 24-hour period as a research demo and is no longer available.”

  2. E2 Golden Gate Claude Anthropic · 23 May 2024 Own record

    “And as we explain in our research paper, when we turn up the strength of the “Golden Gate Bridge” feature, Claude’s responses begin to focus on the Golden Gate Bridge. Its replies to most queries start to mention the Golden Gate Bridge, even if it’s not directly relevant.”

  3. E3 Scaling Monosemanticity: Extracting Interpretable Features from Claude 3 Sonnet Transformer Circuits Thread (Anthropic) · 21 May 2024 Own record

    “For instance, we see that clamping the Golden Gate Bridge feature 34M/31164353 to 10× its maximum activation value induces thematically-related model behavior. In this example, the model starts to self-identify as the Golden Gate Bridge!”

  4. E4 Golden Gate Claude — Pantheon Pantheon (community archive) · 23 May 2024

    “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.”

Reasoning

Anthropic's research page says Golden Gate Claude was online for 24 hours as a research demo (E1). It also says that turning up the strength of the "Golden Gate Bridge" feature made Claude's replies focus on the bridge, even when it wasn't relevant (E2). The Scaling Monosemanticity paper says that clamping this feature to 10× its maximum activation in Claude 3 Sonnet led the model to identify itself as the bridge (E3). An independent community archive confirms the public demo ran on claude.ai for about 24 hours on 23 May 2024 (E4). This meets the criteria for true.

Cite this claim

Dario Amodei, “Anthropic created "Golden Gate Claude", a version of one of its models with the Golden Gate Bridge feature artificially amplified, making it obsessed with the bridge.” (said 18 February 2025). Verdict: resolved true. Reliability API, claim c_800bcc96ffbf, method 0.1.0+c876030c, dataset release v1.1.2. https://epistacklabs.com/claims/c_800bcc96ffbf

This claim is in dataset release v1.1.2, with the same verdict.

Reply to this claim

For Dario Amodei, or someone speaking for them: a correction or counter-evidence. Replies are read before anything is shown; an approved reply shows your name, role, date, statement and link. Your email address is never shown and is deleted 30 days after the decision (privacy).

More factual claims by Dario Amodei

  1. Mostly true Factual 28 February 2025
    “So we did this experiment where we basically train the model to be all the good thing, helpful, honest, harmless, friendly.”

    Anthropic ran an experiment training a model to be helpful, honest, harmless and friendly.

  2. Resolved true Factual 28 February 2025
    “They're not yet at the stage where we think that there is a real and meaningful increase in the threat end to end,”

    Anthropic assessed that Claude 3.7 Sonnet is not yet at a stage giving a real and meaningful end-to-end increase in dangerous threat capability.

  3. Resolved true Factual 28 February 2025
    “10 to 25% referred to the chance of civilization getting substantially derailed, right?”

    Amodei's earlier 10 to 25% figure referred to the chance of civilization getting substantially derailed.

  4. Mostly true Factual 18 February 2025
    “this lack of understanding is essentially unprecedented in the history of technology.”

    The lack of understanding of how AI systems work is essentially unprecedented in the history of technology.

  5. Mostly true Factual 18 February 2025
    “We were able to provide a strong theoretical basis for having conviction that superposition was the right hypothesis to pursue.”

    Amodei said that "we" were able to provide a strong theoretical basis for having conviction that superposition was the right hypothesis to pursue; the surrounding text does not say who "we" refers to.

All of them

Verdicts are made automatically; how verdicts are made.