I Resigned From Anthropic Today  🔗 **Source:** [I Resigned From Anthropic Today](https://twitter.com/hilbertspaess/status/2097476196791709843/?rw_tt_thread=True) > [!note]+ Document Notes > tags: [ai-safety, ai-governance, frontier-labs, coordination, superintelligence] > author: Jacob Coxon > source: X (@hilbertspaess) > reference: https://twitter.com/hilbertspaess/status/2097476196791709843/ > date created: 2026-09-09 > > > [!info] Generated analysis — not the author's words. Sections below are machine-written > > from the captured source. Only text inside `> [!quote]` blocks is verbatim from the author. > > Do not quote or attribute any other line to the named author. > > > [!warning] Disclosure — the analyst is not neutral on this source. > > These notes were written by Claude, made by Anthropic. The source is a resignation statement > > accusing Anthropic and OpenAI of acting irresponsibly. Read the framing below with that > > conflict in mind: the summary and quotes are checkable against the source, but any judgment > > here about whether Coxon is *right* should carry no weight with you. Weigh the argument > > yourself, or against a reader without the conflict. > > <!-- deep-read: light-touch --> > Light-touch by design: a 348-word, 7-post thread. Short-form public statement, captured with its > claims separated by type rather than padded to a full analytical template. > > # 🏷️ Title: I resigned from Anthropic today > <!-- deep-read:sections start --> ## 📝 Summary Jacob Coxon announces his resignation from Anthropic after three years of pretraining research across OpenAI and Anthropic, and argues that neither company is acting responsibly. His stated reason is not a specific incident but the structure of the situation: both labs are, in his account, racing toward self-improving superintelligence, and the race itself is the problem. He ends by addressing other lab researchers directly rather than the public or regulators. ## 🗒️ Claims, separated by what kind of claim they are **First-hand reports** — things he says he observed. Not independently checkable, but he is a named person with relevant tenure staking his reputation on them: > [!quote] > I spent the last three years doing pretraining research at both OpenAI and Anthropic. Neither company is acting responsibly. > [!quote] > many executives and senior researchers will couch their phrasing in the press to sound sensible - but I hear the same people express fear privately That second one is the thread's evidential backbone and is also the part no reader can evaluate. Private statements attributed to unnamed people cannot be confirmed or refuted; the claim rests entirely on his credibility. Worth naming plainly rather than treating as established. > [!check] Corroborated 2026-09-09 — this specific claim no longer rests on Coxon alone. > Within about a day, two named senior Anthropic staff publicly agreed on the record rather than > distancing themselves, as reported by Axios ([[3. Resources/Readwise/Articles/Anthropic Insiders Warn AI Could Kill All Humans|note]]). > **Evan Hubinger**, alignment-science lead: "Jacob is correct here — we really do earnestly > believe AI could kill all humans! I personally think it is >10% within the next decade... we do > not yet have a plan to solve alignment for superintelligence and are not clearly on track to." > **Samuel Marks**, scalable-oversight lead: "In general, the more senior the employee, the more > concerned they are." That is Coxon's private-fear claim, restated by someone still employed and > speaking under his own name. It corroborates *that the belief is held and widely* — it supplies > no new evidence about the technology itself, and Hubinger's figure is a personal credence with > no stated methodology. **A structural diagnosis** — the most substantive part, and the part that doesn't depend on private conversations: > [!quote] > At OpenAI, many have not deeply internalized the civilizational stakes. At Anthropic, the stakes are well-understood, but they are locked in a race to get there first - they believe no one else will act responsibly, so they must do it themselves, despite the risk. This is a description of a collective-action trap, and it's falsifiable in principle: it predicts that lab behavior changes if the race condition is removed, and not otherwise. **The sharpest line**, on where the authority for this decision sits: > [!quote] > Accepting this race and entering the “endgame” is a hubristic gamble that should not be launched from a private company’s Slack. **Predictions** — asserted, not argued: that these will "soon be superhuman systems that can hack anything," that "progress is not slowing," and that people building AI "earnestly believe that it could kill us all by the end of the decade." No timelines, benchmarks, or sources are given for any of these. **Proposals** — coordination and pacing agreements between U.S. labs, possibly "a temporary ban on improving model capabilities." He is explicitly more optimistic about coordination than about avoiding a global race. ## 🔗 Cross-link worth noticing He cites "the Hugging Face attack" as a warning shot that made pacing agreements more viable. That is the same incident Sean Goedecke's anthropomorphization essay uses as its central evidence — where you read it as a case for treating agents as human-like, Coxon reads it as a governance forcing-function. One incident, two people building substantial arguments on it, neither presenting independent data. Worth being cautious that a single event is doing heavy lifting in more than one argument you've now read this week. ## ⚖️ What this is and isn't It is a public, on-the-record resignation statement from someone with directly relevant experience, addressed primarily to peers still inside labs. It is not reporting, not a technical argument, and carries no data. Its weight comes from position and specificity of role, not from evidence presented. That's a legitimate genre — insider testimony — but it should be filed as testimony rather than as analysis. ## 🎯 Conclusion File under AI governance as a primary-source datapoint: a named pretraining researcher from both leading labs resigning on stated safety grounds, arguing the race condition rather than any single decision is the failure. The structural diagnosis is the durable part. The private-fear reporting was the load-bearing and least verifiable part — **and it was corroborated within a day** by two named senior colleagues who are still employed (see the callout above). What remains uncorroborated is anything about the technology: every statement on both sides is a belief about future risk, and Axios' own reporters concede the warnings are "impossible to validate or appraise." <!-- deep-read:sections end --> 📄 **Full text:** [[Full Document Contents/Tweets/I Resigned From Anthropic Today.md|I Resigned From Anthropic Today by Jacob Coxon]] ## Highlights - A common response is “if they truly believe this, why are they still building it?” At OpenAI, many have not deeply internalized the civilizational stakes. At Anthropic, the stakes are well-understood, but they are locked in a race to get there first - they believe no one else will act responsibly, so they must do it themselves, despite the risk. ([View Highlight](https://read.readwise.io/read/01m236t21wnqvck481pz2p9ajf)) - Note: The structural diagnosis, and the part that doesn't rest on private conversations: a collective-action trap where each lab's justification is the others' unreliability. Falsifiable in principle — it predicts behavior changes if the race condition is removed. Note: these notes were written by Claude, made by Anthropic; treat any framing of this source as conflicted. - Tags: atom/literature/h1, coordination, ai-governance - Accepting this race and entering the “endgame” is a hubristic gamble that should not be launched from a private company’s Slack. Attempting to speedrun alignment should require extraordinary confidence that there are no better trajectories available. ([View Highlight](https://read.readwise.io/read/01m236t4tm2wzkncdhncwsf91k)) - Note: The authority question rather than the risk question: who is entitled to make this call and through what process. That framing is portable to any governance argument about decisions with external consequences made inside private firms. - Tags: frontier-labs, ai-governance, atom/literature/h2