FULL STORY

Amodei's Pace-the-Frontier Proposal and the METR Audit Storm

Dario Amodei's call to pace the frontier and his proposal for METR oversight triggered a backlash: Kevin Bass's audits of Anthropic and Open Philanthropy, doubts over METR's funding and independence, and a wider regulatory-capture debate.

2026-09-12 ~ 2026-09-19 · 11 episodes · 79 posts

Episode 1 · Open Philanthropy Accused of Fueling AI Doom Narrative With Over $1 Billion (2026-09-12, 8 posts)

Starting 09-12, Kevin Bass (@kevinnbass) and other posters alleged that the charity Open Philanthropy (OP) is the largest funder of AI doom narratives, spending over $1 billion across a decade, and questioned this as "regulatory capture" building moats for its portfolio company Anthropic. The claims were amplified by reposts and network-mapping threads; neither OP nor Anthropic has responded.

Confirmed

  • Kevin Bass posted on 09-12 that OP was founded by Moskowitz and Karnofsky under the Good Ventures system, is the largest funder of AI doom narratives and an Anthropic investor; he also said OP is the main funder of the Tarbell Foundation, which trains journalists covering AI.
  • Cited figures: $312 million for 476 related publications, and over $1 billion total over ten years.
  • A long thread by Hesamation mapped dense personnel and funding links between Anthropic and OP, noting many in frontier AI and AI safety circles have received Open Phil / Coefficient funding (reposted by StewartalsopIII, nptacek, and others).
  • Beff Jezos amplified a post claiming "Open Philanthropy is the Soros 2.0 of the AI era"; kevinnbass also cited a long post by AITrailblazerQ arguing the doom narrative is "philanthropy-driven regulatory capture" rather than a grassroots movement, mentioning Coefficient Giving (formerly Open Phil-linked).

Unconfirmed

  • The $1 billion total, 476 publications, and $312 million figures come solely from the posters' own accounts, without independent verification.
  • The "regulatory capture" framing is the accusers' inference; Open Philanthropy and Anthropic have not responded.

Why it matters

  • The allegations put the funding chain behind the AI safety agenda in the spotlight: if true, "doom" research shaping public perception and regulation may carry systematic conflicts of interest rather than neutral scientific consensus.
  • OP's dual role as both Anthropic investor and funder of risk narratives touches the core controversy over the independence and conflicts of the AI safety movement, warranting follow-up on whether either party responds.

Episode 2 · Anthropic Researcher's Doomsday Claims Alleged as Orchestrated PR Campaign (2026-09-14, 5 posts)

Multiple commentators allege that AI doomsday statements by Anthropic researcher Jacob Coxon constitute an orchestrated PR campaign, igniting a fierce dispute in the x-risk community. All claims remain allegations without direct evidence, but the debate has escalated to whether AI doom discourse reflects genuine risk warnings or commercially and politically motivated narrative-building.

Confirmed

  • Reddit user @Random-Number-1144 posted a suspicious timeline: an account under the name Jacob Coxon was created in January as a dormant Twitter account, bought a blue check in August, and began posting related content in September.
  • Reddit user @SadAd8761 questioned the timing: with Anthropic reportedly eyeing an IPO of roughly $2 trillion, Coxon — who holds substantial soon-to-vest RSUs — may have incentives to inflate company valuation through fear-driven narratives.
  • Blogger AlexTensor drew an analogy to a "manufacturing bailout 2.0," arguing that RogueAI/pause-style rhetoric from Anthropic and PauseAI is creating conditions for regulatory capture that would shift private commercial risk onto taxpayers.
  • Blogger Greg Hopper (relayed via @petetrainor) described a "doom industrial complex" of frontier labs, AI safety activists, and politicians, and accused OpenAI of using METR and Redwood Research in the HuggingFace investigation to disadvantage competitors outside its own network.

Unconfirmed

  • Scholars including Kevin Bass allege a tight financial network behind AI doom promotion, mostly funded by Anthropic's investors (e.g., Good Ventures / Open Philanthropy-related parties); this remains an allegation without direct evidence of fund flows.
  • The reported $2 trillion IPO and the size of Coxon's RSU holdings are speculative claims by Reddit users, not officially confirmed.

Why it matters

The episode puts the question of whether AI doom talk is genuine risk warning or commercially and politically motivated narrative at center stage. If the allegations hold, safety discourse could be weaponized as a valuation tool and a means of regulatory capture; even if not, the public questioning of financial interests and fear marketing has already damaged the credibility and agenda of the AI safety community.

Episode 3 · Hugging Face co-founder sparks debate on AI labs' profit motives and danger claims (2026-09-14, 7 posts)

Hugging Face co-founder Thom Wolf opened a discussion on the apparent contradiction of frontier AI labs simultaneously pushing capabilities and warning of AI dangers, clashing with researcher Merve over the labs' motives; other users joined in criticizing the labs' opaque disclosure practices. No consensus emerged, but the thread laid bare a trust deficit around capability races, danger claims, and transparency.

Confirmed

  • Wolf laid out five lines of argument for understanding the contradiction, noting the concern is long-standing: Sam Altman wrote in February 2015 that superintelligence "may be the greatest threat to human existence," and Elon Musk made similar remarks in 2014.
  • Responding to Merve's claim that AI companies "only care about profit," Wolf countered that profit-seeking companies would not have started as nonprofits or PBCs, citing the paths of OpenAI and Anthropic.
  • Merve's core position is that these companies have never truly cared about anything but profit, and rhetoric like "govern after AGI" amounts to double standards.
  • User @doomslide attacked the labs by analogy: they wish to be seen as "1930s nuclear companies in a parallel timeline," issuing ever-larger "semi-controlled chain reaction" alarms without telling the public how the explosion is made, with disclosure fully controlled by the companies.
  • User 1a3orn added that allowing third-party evaluators into AI companies eases some principal-agent problems but neither advances science nor builds public knowledge; the labs refuse to disclose pretraining data or architecture while claiming extreme danger.
  • Another critic said they cannot reconcile Dario Amodei's claim that "AI could kill everyone" with his Dreamforce keynote, questioning how serious these people are; user vikhyatk quipped that if you truly believed your company was building humanity-destroying technology, you would delay rather than proceed.

Why it matters

  • The dispute touches the core tension of AI governance: if labs genuinely believe the technology is dangerous, how do capability races and opaque disclosure cohere?
  • Wolf's stance as a co-founder of a leading open-source company contrasts with lab insiders, reflecting the trust rift between the open-source camp and frontier labs.
  • Whether third-party evaluation can substitute for public training details is becoming a new focus in safety and transparency debates.

Episode 4 · Debate Rages Over Whether AI Doom Talk Is Regulatory Capture (2026-09-15, 4 posts)

A fierce debate erupted over whether frontier labs' AI doom warnings amount to regulatory capture aimed at banning open-weight models, with Hugging Face engineer Niels Rogge accusing Anthropic of stoking fear to protect its market. Others, including Alex Tabarrok, pushed back, arguing early safety statements show no such capture.

Episode 5 · Blogger's "Audit" Claims Anthropic Built a Regulatory Capture Machine, METR Independence Disputed Then Rebutted (2026-09-15, 16 posts)

On Sept 15, blogger Kevin Bass (@kevinnbass) published a long post on X claiming to have "audited" Anthropic's financial structure and calling for a Congressional investigation. His core allegation: Anthropic is not merely lobbying, but has built an "unstoppable regulatory capture machine" he dubbed the "Anthropic Network." The post spread widely; Yann LeCun amplified Brian Roemmele's scathing attack calling Anthropic a "grift machine" profiting from fearmongering, and e/acc founder Beff Jezos shared it with astonishment. On Sept 17, rebuttals emerged arguing the financial claims against METR do not hold up, moving the story into contested territory.

Confirmed

  • Kevin Bass published the audit post on 09-15 and called for a Congressional investigation, as confirmed by the source posts.
  • His public argument includes: Dario Amodei proposed METR as a "third-party evaluator" of model risk; METR's funding was alleged to depend on the appreciation of over $7B in Anthropic stock (Ayush's version cites $7.7B); Bass also cited US National Philanthropic Trust (NPT) FY2025 data.
  • Bass followed up questioning METR's funding: multiple large grants came via Coefficient, backed by Dustin Moskovitz's Good Ventures Foundation; the largest grant could not be traced through Vanguard but ultimately points to Good Ventures, a holder of Anthropic stock.
  • Replies added that Tarbell researchers funded by an Anthropic-linked pool wrote Time 100 AI profiles without disclosing conflicts of interest.
  • @FinanceYF5 summarized the alleged closed loop of mutually reinforcing interests among Anthropic, AI safety organizations, and their funders; @theteknosaur mapped the same network including multiple safety orgs, doom research, policy lobbying, and Tarbell.
  • A rebuttal appeared on 09-17 (via @dhadfieldmenell) claiming the METR financial allegations were refuted.
  • The content was amplified by @Evideyear, @Promptmethus, @beffjezos, @ylecun, @ayushtweetshere and others.

Unconfirmed

  • The "audit" is a one-sided allegation, explicitly marked unverified in the sources; its funding-chain data and causal inferences lack independent corroboration, and @theteknosaur's mapping was described as conspiracy-flavored.
  • The 09-17 rebuttal appears only as a repost; its specific arguments are not elaborated in the material, so no verdict exists either way.
  • Whether METR lost independence, whether Anthropic engaged in structural self-dealing, and the Tarbell disclosure claim remain allegations, not verified conclusions.
  • No information on a Congressional investigation or other official response appears in the material.

Why it matters

  • The episode pushes the independence of AI safety evaluators into public view: if third-party assessors are financially entangled with the companies they assess, the credibility of their verdicts is undermined.
  • Anthropic's public image as a responsible safety company is at stake; confirmation or refutation would affect its policy credibility and the design of evaluator institutions in future AI regulation.
  • Amplification by LeCun and others pushed the dispute beyond safety circles into an open battle between AI camps over fearmongering and regulatory lobbying; the later rebuttals show the allegations themselves are being scrutinized, with no consensus yet.

Episode 6 · Amodei Proposes Embedded METR Oversight of Frontier AI, Sparking Independence Backlash (2026-09-15, 12 posts)

Anthropic CEO Dario Amodei published "We Must Pace the Frontier," proposing that nonprofit evaluator METR (Model Evaluation and Threat Research) be embedded in frontier AI companies with employee-like access to oversee safety commitments—analogous to bank examiners resident in banks. Embedded third parties would have the power to report safety incidents, assess model alignment, and publish honest conclusions, with Anthropic pledging access for METR, Redwood Research and others. Sam Altman endorsed independent inspectors for both Anthropic and OpenAI the next day, but the arrangement triggered a broad backlash over METR's independence.

Confirmed

  • Amodei's essay argued for third-party evaluation teams embedded like employees at frontier labs (@firstadopter, @basedjensen); NY Post ran a critical report on the safety proposal.
  • Sam Altman backed independent inspectors at Anthropic and OpenAI; one thread noted the context of Anthropic researcher Jacob Coxon's resignation and public statements (@alexverem).
  • METR self-assessed its independence against its own standardized criteria, performing well on most items while lacking any disclosed conflict-of-interest policy, buried mid-way through a 300-plus-page report (@kevinnbass).
  • @kevinnbass traced the previously contested Coefficient funding through intermediary layers, claiming it ultimately points to Anthropic investors (@kevinnbass, @firstadopter).
  • Gary Marcus argued METR's world overlaps with Anthropic's, so it cannot be the sole third-party auditor, while acknowledging its research quality (@GaryMarcus).
  • Very Sane AI Newsletter published a long response arguing METR cannot check Anthropic's power (@kyliebytes).
  • @segyges clarified he was not asserting METR is necessarily conflicted, but raised concerns about mechanisms like labs recommending their own auditors; the post was amplified by jdpressman as a significant take (@inductionheads repost, @jessicata).

Unconfirmed

  • The funding chain allegedly ending at Anthropic investors is the poster's own inference, not confirmed by METR or the investors involved.

Why it matters

  • The dispute cuts to the core of frontier AI governance: embedded third-party evaluation loses its checks-and-balances value if the auditor's interests intersect with the audited company; METR's self-assessment conspicuously lacking a conflict-of-interest policy amplifies concerns about the independence of the evaluation ecosystem.

Episode 7 · Researcher Answers Dario's Slowdown Call: Responsibility, Not Rationing (2026-09-16, 3 posts)

In response to Dario Amodei's call to pace the frontier, researcher Max Van Kleek argued for 'responsibility, not rationing,' proposing open, decentralized alternatives. Another researcher laid out positions rejecting both accelerationism and total safety freeze.

Episode 8 · Ben Todd blasts OpenAI and Anthropic over "pacing the frontier" (2026-09-16, 2 posts)

80,000 Hours founder Ben Todd published a three-part essay accusing OpenAI and Anthropic of regulatory capture via "pacing the frontier," followed by a sarcastic call for them to race toward superintelligence at full speed.

Episode 9 · METR's Independence Questioned: Client Capture and Worldview Alignment (2026-09-16, 7 posts)

In mid-September, a multi-front critique of AI safety evaluator METR's independence erupted, led by Stanford professor Christopher Manning, commentator Matthew Stoller, and Gary Marcus, targeting business dependence, worldview alignment, and self-limited investigation scope.

Confirmed

  • Manning argued that although METR is financially independent of frontier labs, its business itself depends on cooperating with them, creating a "client capture" problem—you cannot risk anything that might lose you the business. Even setting aside finances, METR's worldview closely aligns with Anthropic's, making it not truly independent and undermining its credibility as AI systems grow more powerful. He also stressed the value of university research in intellectual rigor and creativity.
  • Stoller criticized METR's report on OpenAI agent autonomy as proactively "striking a deal" with OpenAI, explicitly scoping the report to ignore decisions by humans inside OpenAI rather than being forced to. A parallel critique of METR's Hugging Face incident investigation argued it examined only agent behavior, never human factors. Nathan Young amplified the scope controversy.
  • Gary Marcus endorsed Stoller's critique as the only intellectually substantive challenge to METR's investigations.
  • @r0ck3t23, responding to Dario Amodei's proposal that frontier labs accept external independent evaluation, noted that if evaluators have insider-level access, their independence itself becomes part of the safety system—raising "who evaluates the evaluators"—and cited METR's stated policy of not taking frontier-lab funding.

Why it matters

  • As frontier labs increasingly rely on external evaluators like METR, the evaluators' independence and methodology directly determine how credible safety conclusions are. The worldview bias Manning identified and the scope self-limitation Stoller flagged suggest structural blind spots—ignoring human decisions, converging with evaluatees—that the AI safety governance ecosystem must confront.

Episode 10 · METR's $71M Six-Month Funding Surge Draws Transparency Scrutiny (2026-09-17, 9 posts)

On September 17, independent researcher kevinnbass published an audit of public records claiming that METR (Model Evaluation and Threat Research), an AI safety evaluation organization endorsed by Dario Amodei, raised $71 million over the past six months—roughly 5x its prior scale—yet only about $350K of that is traceable. The findings ignited a wide debate on X over METR's funding transparency and evaluation independence.

Confirmed

  • kevinnbass said he ran hundreds of (mostly Grok-based) agents overnight; his core conclusion is "we don't know where METR's money comes from," with remaining evidence not yet organized into readable charts.
  • He also noted METR only posted conflict-of-interest information on its About page on September 15, after his viral tweets, suggesting the policy was effectively an afterthought. Another post described METR's "double loop" model—donors pay, METR produces evaluations—unlike the ordinary "customers pay, vendors deliver" structure, raising questions about donor intent.
  • A fact-check relayed by @basedjensen challenged METR's claims of taking no frontier-AI-company money and donor diversity: METR received about $4.5M when spun out of ARC, which had received funding from Jaan Tallinn's Survival fund and others, giving its funding chain multiple intersections with Anthropic.
  • @AndrewAzlan cited a report that roughly $71M in donations is untraceable and that Anthropic's investors overlap heavily with METR's donors, calling for an investigation in the name of "AI safety."
  • Counter-voice @austinc3301 pointed out that METR's website discloses funders and that U.S. 501(c)(3) nonprofits are not legally required to disclose donations item by item, arguing the criticism overreaches.

Unconfirmed

  • The actual destination and identities of donors behind the untraceable portion of the $71M cannot be verified from public records.
  • Whether the Anthropic-linked funding overlap materially compromises METR's evaluation independence remains inference; the agent-based findings have not been independently verified.

Why it matters

METR is a leading frontier-model evaluator whose conclusions shape public and regulatory perceptions of model safety. If its funding is heavily concentrated among donors tied to evaluated companies, it could constitute a conflict of interest and erode trust in its evaluations. The crux of the dispute is whether nonprofit-style opacity alone supports the claim that METR's independence is compromised.

Episode 11 · Anthropic's Dario Amodei Calls for Pacing AI Frontier; Musk, Altman and Hassabis Rally (2026-09-17, 6 posts)

On Sept 17, Anthropic CEO Dario Amodei published "We Must Pace the Frontier," arguing that the AI industry should deliberately slow the pace of frontier capabilities and laying out a three-step plan; as the first step, Anthropic unilaterally committed to giving third-party evaluators permanent access to its models. The essay triggered a rare cross-camp response: Elon Musk, OpenAI's Sam Altman and Google DeepMind's Demis Hassabis all voiced support for controlling frontier AI, with Altman noting OpenAI had been internally discussing how to "brake" frontier models in recent weeks.

Confirmed

  • Amodei published the essay on Sept 17, calling for pacing the AI frontier
  • The three-step plan's first step is permanent model access for third-party evaluators, with evaluation embedded into governance and domestic safety standards proposed
  • He framed the upside of AI as potentially curing most major diseases within 5–10 years and dramatically accelerating economic growth
  • Musk, Altman and Hassabis publicly endorsed controlling frontier AI; Altman said OpenAI had discussed slowing frontier AI for weeks
  • The phrase "We should pace intelligence" is circulating and drawing debate on social platforms

Unconfirmed

  • Creator Rasmic said he does not believe the trio's stated reasons and used a whiteboard video to question their motives; this remains a personal opinion, and their true motives are undetermined

Why it matters

  • Leaders of the most competitive frontier labs agreeing to slow down is extraordinarily rare and may signal a shift in competitive norms and regulatory posture
  • Anthropic's permanent evaluator access sets a replicable precedent for external auditing and safety assessment
  • Tying pacing to potential 5–10 year medical and economic gains reframes deceleration beyond pure safety arguments