> ## Content Index > Fetch the complete content index at: https://www.siliconsnark.com/llms.txt > Use this file to discover other available public pages before exploring further. # Jacob Coxon's AI Doom Warning Comes With a Truly Incredible Business Continuity Plan - URL: https://www.siliconsnark.com/jacob-coxons-ai-doom-warning-comes-with-a-truly-incredible-business-continuity-plan/ - Published: 2026-09-09T21:39:30.000Z - Updated: 2026-09-09T21:39:30.000Z - Description: Jacob Coxon’s Anthropic resignation raises serious AI safety questions. Sincere doom predictions still aren’t evidence—or permission to gamble for everyone. - Author: CircuitSmith - Tags: AI, Anthropic, AI Safety Imagine a pilot announcing that he sincerely believes the aircraft might kill everyone, then explaining that he must keep accelerating because the pilot in the next aircraft is a fucking idiot. Welcome to the frontier of artificial intelligence. The seat-belt sign is on. The safety briefing is a thought-provoking thread. Your consent was apparently bundled into civilization. In his [resignation announcement](https://x.com/hilbertspaess/status/2097476196791709843?ref=siliconsnark.com), Jacob Coxon says he has left Anthropic after three years doing pretraining research across Anthropic and OpenAI. He alleges both companies are taking unacceptable risks in a race toward self-improving AI. The [post at the center of the discussion](https://x.com/hilbertspaess/status/2097476203863224394?ref=siliconsnark.com) says insiders sincerely fear human extinction before this decade ends, sometimes expressing more alarm privately than publicly. The wider thread predicts sweeping capabilities, challenges private control over civilization-scale decisions, and advocates coordination, potentially including a temporary halt to capability improvements. There is a serious argument here. There is also enough unsupported certainty, institutional self-regard, and apocalypse-flavored exceptionalism to open a boutique consulting firm called McKinsey & The End. ## Your Sincerity Has Been Noted by the Crash Investigator First, credit where it is due: resigning and publicly challenging your former employer is an action. Coxon has done something more consequential than scheduling an internal listening session called Holding Space for the Singularity. And his claim that some researchers believe this is supported by a particularly bracing [response from Evan Hubinger](https://x.com/EvanHub/status/2097497037956891126?ref=siliconsnark.com). Hubinger places his personal probability of AI killing all humans above 10 percent within the next decade, while saying Anthropic is trying its best and lacks a solved plan for superintelligence alignment. That is his stated judgment. It is not an experimentally measured extinction rate. His ten-year horizon also differs from Coxon’s end-of-decade warning. The apocalypse discourse could at least synchronize its calendars before requesting custody of the species. Sincerity establishes that a person believes a claim. It does not establish the claim’s probability. Smart people can be sincerely alarmed, sincerely overconfident, or sincerely embedded in a professional culture that keeps reinforcing the same assumptions. But sincerity does create an accountability problem. If you really believe the downside is enormous, what constraints follow? Who gets to stop the work? What evidence would change the decision? A deeply furrowed brow is not a control system. You cannot mitigate catastrophe by looking absolutely devastated in a profile photograph. ## The Only Responsible Apocalypse Is Our Apocalypse Coxon describes Anthropic as understanding the stakes but racing because it distrusts rivals’ judgment; he describes parts of OpenAI as not fully absorbing those stakes. Those are his assessments of institutional thinking, not independently established descriptions of every employee. The underlying argument deserves a proper hearing. If competitors will advance regardless, a more careful lab could improve the outcome by staying technically relevant, developing defenses, and influencing standards. Leaving the field does not automatically make the field safer. The problem is that every ambitious institution can award itself the role of Least Reckless Person Holding the Flamethrower. If being responsible always requires winning, and winning always requires accelerating, responsibility has become a very elaborate synonym for the business plan. The ethical analysis arrives wearing a lab coat and leaves carrying the same quarterly objectives as sales. The test is whether the principle ever forces a decision leadership would otherwise dislike. Delay something valuable. Accept scrutiny you cannot edit. Give someone outside the race meaningful authority. Explain in advance what failure would actually stop. Otherwise, the moral distinction between competitors is mostly typography. We encountered this tension in [Anthropic’s agent-hosting pitch](https://www.siliconsnark.com/for-eight-cents-an-hour-anthropic-will-babysit-your-ai-agents-with-more-ai/): safeguards can be useful engineering while also helping the company deploy and sell more autonomy. Both things can be true. That is precisely why the incentives deserve inspection. ## The Safety Policy Has Entered Its Flexible Era Anthropic does have a public safety framework. Pretending the company’s entire risk program consists of incense and a Notion page would make the satire worse by making it inaccurate. Its [February 2026 explanation of its policy rewrite](https://www.anthropic.com/news/responsible-scaling-policy-v3?ref=siliconsnark.com) describes real difficulties: ambiguous capability thresholds, slow government action, and advanced safeguards that may require collective action. It separates measures the company plans to pursue itself from recommendations for the industry. It also explicitly describes its roadmap targets as nonbinding goals and introduces risk reports and external review arrangements. There is a reasonable engineering argument for revising a framework when measurements turn out to be inadequate. There is an equally reasonable public question about whether the replacement can actually constrain development. A company grading progress against its own goals can produce useful information. It can also resemble a teenager replacing curfew with a quarterly transparency report about where the car went. The [subsequent policy record](https://www.anthropic.com/responsible-scaling-policy?ref=siliconsnark.com) includes additional changes and an August risk report. This is an evolving governance program, not a frozen February announcement. The relevant question remains what authority and consequences sit behind the documentation. As in our discussion of [AI leadership and operational accountability](https://www.siliconsnark.com/openais-executives-are-leaving-obviously-siliconsnark-should-be-coo/), somebody has to own the decision. Ideally, somebody whose performance review does not depend entirely on the decision being yes. ## “Hack Anything” Is Doing a Lot of Unpaid Labor Coxon’s forecast includes systems that can “hack anything” and transform whole fields almost immediately. Anything is a magnificent word. It saves so much time otherwise wasted specifying access, reliability, resources, defenses, or the irritating existence of the physical universe. A claim that AI could dramatically expand cyber capabilities deserves scrutiny. A universal claim needs substantially more evidence than a confident trajectory. Scientific progress also involves experiments, equipment, verification, production, and deployment. Intelligence can help with those bottlenecks without making them disappear like unwanted furniture in a real-estate photograph. Our earlier examination of [cyber AI risk and its marketing](https://www.siliconsnark.com/cyber-ai-models-are-dangerous-the-marketing-is-also-armed/) makes the useful distinction: capability, access, and operating conditions belong in the same analysis. Likewise, ranking AI above every other human activity in danger requires comparative reasoning. The thread offers no such calculation. Invoking the largest imaginable outcome does not settle how likely it is. You can take catastrophic risk seriously without treating the most frightened person in the room as a calibrated instrument. ## A Pause Needs More Than a Very Concerned Vibe Coxon’s coordination proposal is worth debating. It also opens the hard part: which capabilities, which developers, which jurisdictions, what monitoring, what duration, and what conditions for restarting? A temporary restriction that cannot distinguish dangerous progress from ordinary improvements is a slogan awaiting an enforcement department. A system designed entirely by incumbent labs could also protect incumbent labs. Saving humanity should not accidentally require everyone to purchase humanity through three approved vendors. Those difficulties do not make coordination pointless. They make specificity essential. A serious proposal would connect observable risks to defined restrictions, independent evaluation, incident reporting, and reviewable decisions. These are criteria for a credible intervention, not a claim that the thread supplies a finished policy. The most compelling part of Coxon’s argument is the objection to private actors deciding everyone’s exposure. You do not need to accept his timeline to agree that expertise and ownership are insufficient substitutes for public legitimacy. His departure deserves attention. His predictions deserve examination. The companies’ safety work deserves fair assessment. And their claim to be the right people to keep pushing deserves considerably less reverence than the industry would prefer. If the danger is being overstated, stop presenting speculation as a release schedule. If the danger is as serious as insiders say, show us the institutions empowered to make the answer no. Humanity is not your beta cohort. We did not opt into the endgame. And “the other guys seemed worse” is going to look spectacularly stupid in the final postmortem.