The connection
In March 2026, Eliezer Yudkowsky and Nate Soares joined other researchers to brief Bernie Sanders about superintelligent AI. Sanders released video of the meeting. MIRI, the organization they help lead, pointed its supporters toward the encounter. A warning had reached a senator's desk, with its authors in the room. MIRI's account
That is an instructive place to begin. Public debate tends to give us personalities: the prophet, the crank, the visionary, the billionaire. Those labels are emotionally efficient and analytically lazy. What matters here is the machinery connecting a person's argument to resources, audiences, political access and proposed public powers.
Yudkowsky occupies several of those positions simultaneously. Following them reveals something consequential about how policy becomes possible—and why a frightening prediction deserves especially careful scrutiny when someone asks us to build an enforcement system around it.
The institution and the money
MIRI, the Machine Intelligence Research Institute, dates its founding to 2000; it operated as the Singularity Institute until 2013. Its current roster places Yudkowsky under both Leadership and Board as co-founder. Malo Bourgon is CEO. Soares is president. Keeping those roles straight matters: intellectual prominence, board authority and daily executive responsibility are different kinds of power. MIRI history · Current team
The institutional connection gives Yudkowsky's advocacy a durable home. An organization can hire people, retain knowledge, commission work and keep pursuing a problem after the news cycle has wandered off. To understand how that becomes possible, follow the money into the organization.
MIRI's contributor acknowledgements, updated April 30, 2026, credit Open Philanthropy with approximately $15.93 million, Vitalik Buterin with $5.41 million, the Survival and Flourishing Fund with $3.37 million and Jaan Tallinn with $1.63 million. The Thiel Foundation appears among past contributors at $1.627 million; the Future of Life Institute is credited with $412,252. These are cumulative institutional support credits, including matching funds leveraged. They are not a September campaign budget, Yudkowsky's income or six receipts for the same political project. The page's “recent” category requires only some qualifying support within five years. MIRI's contributor list
A separate 2021 announcement makes the scale tangible. An anonymous cryptocurrency investor donated 3,001 MKR tokens. MIRI says it sold them for $15,592,829, with spending capped at $2.5 million annually during 2021–2024 and the remaining $5.59 million becoming available in 2025. The same announcement records a distinct 1,050 ETH gift from Buterin. That specific gift must not be added again to his cumulative credit. The token's identity also does not make MakerDAO the donor. The donation announcement
An anonymous gift leaves a real accountability gap: we cannot inspect the donor's interests or relationships. We can still identify the recipient, reported proceeds and disclosed spending restriction. Guessing a name would make the investigation worse.
This financial history explains capacity. It does not settle present allegiance. MIRI itself acknowledged in its December 2025 fundraising appeal that it did not know how many earlier donors would support its newer communications and governance work. An institution can carry resources from one phase of its mission into another. The interesting questions are which restrictions travel with the money, which expire, and who approves the new direction. MIRI's mission and fundraising account
A community, a book and a production channel
Money is one resource. A community that already understands your vocabulary is another.
LessWrong's own history identifies Yudkowsky as its founder in 2009 and describes his writings as part of its philosophical foundation. Its FAQ records that the site began as MIRI property. Today, Lightcone Infrastructure lists LessWrong among the infrastructure it operates. This is a documented genealogy; it does not establish that Yudkowsky currently controls Lightcone. LessWrong FAQ · Lightcone
Such a forum can do something a grant cannot do by itself: give people a shared language for identifying problems, evaluating arguments and finding collaborators. That is a plausible mechanism of intellectual influence, not proof that everyone who reads the same essays votes, works or thinks alike. A community can preserve an author's importance while arguing ferociously about what to do next.
The video connection is unusually tangible. In October 2023, Rational Animations’ operator described four adaptations of Yudkowsky’s work, including The Power of Intelligence and The Hidden Complexity of Wishes. He said he had invited Yudkowsky into its production Discord to contribute ideas. That is a documented invitation and an adaptation history. It is not evidence that Yudkowsky writes most of the channel’s output. The operator’s account
An April 2026 interview with lead scriptwriter Alice Liu describes her own writing and Robert Miles’s substantial editorial input, alongside other writers. The channel has a production process, not just an author’s megaphone. An early $30,000 EA Infrastructure Fund award explicitly supported audience development; it predates the current campaign. Liu’s interview · The grant report
Turning an audience into political demand
The passage into organized public persuasion is more directly documented.
Yudkowsky and Soares coauthored If Anyone Builds It, Everyone Dies. MIRI describes the book as central to its 2025 communications work, alongside public appearances and policymaker conversations. Its plans for 2026 included helping third parties use existing material with their own audiences, responding to news and maintaining relationships with allies and journalists. Those plans gained institutional muscle: by March, MIRI reported that its communications team had grown from five people to thirteen. MIRI's 2025 account · March 2026 update
That is organized advocacy, described by the organization itself. It gives us grounds to investigate particular collaborations. It does not supply a universal explanation for every frightened researcher, viral post or sympathetic interview. The missing link in a specific case still has to be found.
ControlAI supplies an unusually clear example of adoption. In September 2025, its newsletter promoted the book alongside its own campaign to prohibit superintelligence. The organization endorsed the authors' central prescription and directed readers to their work. ControlAI's campaign announcement
The influence extends to method. In a published account of briefing more than seventy UK lawmakers, ControlAI policy adviser Leticia García Martínez cited Yudkowsky's essay about inferential distance: the gap between what a speaker assumes and what an audience already understands. Her guidance also explains how to make arguments memorable enough for parliamentarians to repeat. Here is a concrete transmission path from a writer's account of reasoning into political persuasion. ControlAI's briefing guidance
ControlAI then makes its own contribution. Its April 2026 strategy praises the book but distinguishes repeatable campaigning from exceptional individual successes. It explicitly compares its approach to sales funnels: brief lawmakers, seek public support, measure subsequent statements, improve conversion. The proposed $50 million annual budget is an expansion scenario, not evidence it received or spent that amount. ControlAI's strategy
There is a direct connection to the incident that started this investigation. On September 8, ControlAI identified itself as the organization involved in Sabine Hossenfelder’s disputed sponsorship approach. It also named Rational Animations as a sponsored partner. Those are separate facts; they do not make every adaptation a commissioned campaign script. ControlAI’s account
Read both accounts of the Sabine offer, and what remains unresolved →
This explains why counting famous names is insufficient. Authors supply arguments. Campaign staff translate them for an audience. Supporters provide public legitimacy. Policymakers can turn the resulting pressure into decisions. Each transition is a place where the message may change—and where responsibility needs a name.
The Sanders meeting belongs in that sequence as documented access. It establishes an opportunity to inform and persuade him. We have not established that Yudkowsky or Soares drafted the later Sanders/Casar legislation. A photograph or briefing can demonstrate proximity; authorship requires evidence of authorship.
What happens when the answer becomes enforceable?
The stakes become sharper when persuasion reaches enforcement.
In his March 2023 TIME essay, Yudkowsky rejected a six-month AI pause as insufficient. He advocated an indefinite international halt to large training runs, including government and military work, and willingness to destroy unauthorized foreign datacenters by airstrike despite nuclear-escalation risk. He tentatively allowed narrowly biological AI. This was his dated proposal, not enacted law, and its target was infrastructure—not a policy of executing researchers. The extremity is consequential enough without embellishment. Yudkowsky's authored essay
The institutional policy work also deserves examination on its own terms. MIRI says its technical-governance team collaborated with Yudkowsky, Soares and others on an annotated treaty, then developed a formal international-agreement proposal. The later agreement’s May 2026 version names Aaron Scher, David Abecassis, Peter Barnett and Brian Abeyta as its authors. Yudkowsky’s documented collaboration on the earlier annotated treaty is a different attribution. MIRI's account
The later design proposes an executive council initially comprising the United States and China, technical oversight, compute restrictions and inspections. It would restrict certain AI research as well as large training runs, because algorithmic progress could undermine compute-only limits. Article XII would allow member states to undertake necessary and proportionate protective action on credible evidence of specified prohibited activity, including cyber sabotage and military destruction of AI infrastructure. Less coercive approaches ordinarily come first, with exceptions for emergencies and time-sensitive situations. The Executive Council reviews each action, but the article does not expressly require its prior approval. Safeguards include public reasons when security permits, limits on collateral harm qualified by mission requirements, and conditions requiring action to cease. These safeguards matter; so does the extraordinary authority being proposed. This remains a research proposal, not a signed treaty or an authority already conferred on MIRI. Agreement, version 3
Nor does the surrounding field agree on one blueprint. ControlAI's own strategy criticizes approaches centered on evaluation or positioning trusted insiders, and favors public mobilization toward international prohibition. It even notes Yudkowsky's shift from an earlier approach toward public awareness. Shared concern about extinction coexists with disputes over institutions, strategy and who should act. ControlAI's comparison
Conviction does not settle permission
MIRI’s own guidance calls for honesty and cautions against manipulation. It recommends Rational Animations while explicitly declining to imply agreement on everything. That belongs in this account: a claim that its public doctrine licenses any lie would contradict its stated position. Testing conduct against that position is a legitimate investigation. MIRI’s guidance
The difficult issue is the relationship between conviction and permission.
A person who believes everyone is about to die may sincerely regard sweeping intervention as a moral obligation. Sincerity does not determine whether the prediction is right. Nor does it answer who should decide, what evidence should be sufficient, which people bear the costs, or how a mistaken decision can be reversed.
Those questions become more urgent as the claimed stakes rise. If ordinary disagreement is interpreted as willingness to gamble with humanity, the space for legitimate opposition can shrink very quickly. A public institution must remain answerable to people who reject its founding premise. Otherwise its safety case becomes a demand for obedience.
We should ask of these proposals what we would ask of any consequential system: identify the authority, its limits, its evidence, its appeal process and the means of stopping its misuse. Publish the grant purposes. Name the drafters. Expose the decisions to independent challenge. Keep the distinction between expertise and permission to govern visible.
Yudkowsky is a keystone in this part of RED THREADS because ideas, institutional resources, advocacy and political access meet around his work. That is a judgment about his connecting roles, not a claim that he secretly runs the whole network.
The public has every right to examine what happens at those connections. A warning can deserve a hearing. The power requested in its name still has to earn our consent.