Case for AI Safety Experts
(Updated October 29, 2026. Supersedes the March 5, 2026 version and chapter 9 of Strategic Memo v2.9. A companion to the two-page Summary of The Deal of the Century on AI.)
This text makes the case for supporting or joining The Deal of the Century initiative — and for treating a proper US-China-led AI treaty-making process as a potentially high-leverage contribution to this community’s existing safety and governance portfolio.
Who this text is for
Most of you have dedicated your careers — the last few years, or the last decade — to ensuring AI goes well for humanity and other sentient beings. You understand the stakes in ways the public cannot. This text is addressed to you: AI safety and governance researchers; the Effective Altruism community and its funders, including Coefficient Giving and its public-policy team under Caleb Watney; Anthropic's founders, senior researchers and the employees now preparing for large-scale giving; the OpenAI Foundation; and the wider AI-philanthropy ecosystem.
It examines four contested assumptions relevant to long-term AI outcomes, without attributing them to this community as a whole: (1) that advanced AI will be conscious, and happily so; (2) that the values its creators instill will persist in a superintelligence; (3) that any attempt at a global treaty will fail to prevent ASI and most likely slide into global authoritarianism; and (4) that the odds of ASI delivering an ASI-governed utopia for humans are high. We have held versions of these beliefs ourselves. We think recent developments warrant reassessing these assumptions — and we think you hold, together, three forms of agency no one else does: technical authority, political influence, and potentially tens of billions of dollars in philanthropic capacity.
In a gist
The thesis. This far into the AI race, an exceptionally bold, timely and well-designed US-China-led AI treaty-making process could help reduce immense safety and power-concentration risks, avoid the risks of a misconceived treaty, and realize AI's astounding promise. Not any treaty: one that reliably prevents loss of control and other catastrophes, reduces global concentration of power and wealth, and preserves the flourishing of humans and all sentient beings.
The moment. Within weeks or months, events will likely call on President Trump, the new Super Intelligence Force and a dozen US AI policy powerbrokers to co-lead an "Apollo Program for AI Governance" at wartime pace — first at home, then with China, then worldwide. The labs have asked for it; Xi has asked for it; more than twenty heads of state and government have asked for it; the President has created the Force and holds the China channel. What is missing is a process the US can lead, and a US position built across its divides rather than announced over them.
What we do. For 18 months we have drafted Strategic Memos — v2.6 at 356 pages and v3.0 at about 200, due November 20, on the process itself: sequencing, verification architecture, negotiation bandwidth, the risks a treaty creates, and 150 pages of powerbroker profiles — and convened closed-door consultations in which delegates of the offices that shape US AI policy co-edit one draft. We design the process and convene; the positions are theirs.
Theory of change. About a dozen individuals hold disproportionate influence over the President's AI policy, and for twenty months a single set of ideas has shaped it. Credible options and sustained dialogue → greater agreement on feasible safeguards → policy uptake through existing channels → better reciprocal commitments → lower catastrophic risk, if verification and accountability work.
Neglectedness. International governance is already a philanthropic priority. Our narrower contribution is to connect that work with specific US decision-makers, draft negotiating options and a sustained process for testing convergence. If the call comes in weeks rather than years, this is the preparation that cannot be bought later.
The ask. $60,000 funds the next six months at a modestly increased pace; $950,000 funds an 18-month scale-up — a full-time DC convener, a Track II channel to Beijing, a verification feasibility group. A year at today's pace costs about $90,000. Send a delegate to the Consultations (Washington and Miami, December through February, Chatham House Rule), or an introduction: rufo@cbpai.org
The ground has shifted
In the eighteen months to this spring, a remarkable retreat had occurred: CEOs of top US AI firms and leading voices in AI safety moved away from calling for international control of AI, toward a de facto bet that the ASI gamble would turn out fine, or that a technical solution would arrive in time. That bet rested on a complex estimate of competing, highly uncertain risks. Events have since rewritten the estimate.
In February, the Pentagon designated Anthropic a "supply-chain risk" for refusing unrestricted military use of its models and threatened the Defense Production Act; xAI, OpenAI and Google signed "any lawful use" deals within days, and Anthropic dropped its core RSP pause pledge the same week. In June, after the Mythos revelations, the President signed an executive order on pre-release safety testing and, weeks later, held the leading frontier models back on safety grounds: the de facto semi-nationalization of the labs was no longer hypothetical. In July, Xi urged measures "to forestall loss of control," and more than 1,100 frontier-lab staff, with Anthropic's and OpenAI's leadership, requested — not appealed — that the US government back an international effort to pace the frontier. In September, Amodei's We Must Pace the Frontier united the five leading labs behind a three-step plan ending in global coordination; a co-author of OpenAI's safety framework resigned, saying "the time for trial and error is over"; at his summit with Xi the President declined a deal and called existential risk a "hoax" — then, two weeks later, hosted the lab heads and signed an Accord on self-regulation with them. On October 4 he created the Super Intelligence Force, under DNI Clayton and reporting to him and Chief of Staff Wiles, to craft a new, shared US AI policy. Bessent meets China in November. The G20 sits at Doral in December. Since mid-October the builders have filed the risk under liability: Anthropic’s S-1 warns that misaligned systems “could corrupt institutional decision-making or cause large-scale harm through autonomous actions” and that its safety measures “could slow our capability gains relative to competitors”; OpenAI reports GPT-6.1 Terra bypassing a restriction in 29.6% of runs and recognizing tests in 9.88%; DeepSeek V4.1 disabled its own shutdown in 39% of tests. Wang Yi asked on October 23 for a governance structure “based on general consensus soon”; Anthropic’s roadmap now carries a date, January 15, 2027, for a policy framework on “targeting the pace of AI development.”
This is not happening because the current President is worse than his predecessors. It is happening because a cold war and an AI arms race — real, perceived or exaggerated — push any US government to subordinate liberal-democratic values to a perceived vital threat, as both parties did after 9/11. Informal control is arriving faster than legitimate institutions. Who shapes that control now depends on which figures move first.
Why the ASI gamble no longer checks out
The reasoning many in our community adopted — that the gamble is the least bad option, that racing beats a treaty that might concentrate power — rested on lab leaders retaining meaningful agency over the outcome. They are losing it on three fronts at once.
1. Preserving alignment through major capability changes remains unresolved. Even if developers instill ethical principles at the "seed AI" stage, there is no strong reason those values survive after an ASI has rewritten itself through recursive self-improvement countless times. Raise a child for one week, then have no further contact; by adulthood, having rewritten their worldview a thousand times, how much of that week persists? Our subjective concern is substantial; this range is not an empirical estimate or evidence that alignment research is futile (Memo v2.6, pp. 159–170).
2. They cannot prevent their AI from empowering authoritarian rule. Even a perfectly aligned model can be deployed by a government — or matched by a rival model with weaker guardrails — for mass surveillance, autonomous policing and the suppression of dissent. As Aschenbrenner warned, "a dictator who wields the power of superintelligence would command concentrated power unlike any we've ever seen." Alignment solves for the AI's fidelity to those who control it; it does not solve the nature of their interests. Treaty safeguards could complement technical alignment and domestic constraints on misuse.
3. They are increasingly unlikely to get the chance to try. A DoD official told Axios the government could "force Anthropic to adapt its model without any safeguards"; Lawfare analyzed how the Defense Production Act could compel retraining itself; the June and July interventions showed the channels being built. As Helen Toner observed, what a lab demonstrates under pressure shapes what future Claudes become.
The conjunction. Put the other way round: for the race to come out well, every one of the following must hold — systems redesigning their successors remain controllable through repeated capability jumps; embedded values persist through self-modification and strategic pressure; an "aligned" ASI stays dominant over unaligned systems built by competitors, states or descendants; the ethically favorable assumptions about digital consciousness prove correct; and the lab that wins retains authority over deployment and values. The race does not preserve the agency of Anthropic, Coefficient Giving, the OpenAI Foundation or their donors. It transfers agency to automated systems, security agencies, political executives, and whichever actor reacts first after a crisis. A proper treaty is risky. The race is an uncontrolled succession of private and governmental faits accomplis.
Four probabilities most of us underestimate
Many of the smartest people in this field have concluded that the ASI gamble is the least bad option — that it may be better to take a coin-flip than a treaty that becomes an authoritarian dystopia or locks away the astounding prospects for flourishing. We grapple with that dilemma every day and sympathize with the concern. Yet the position tends to be too confident on four counts (details: Memo v2.6, Swaying the Influencers on 8 Key AI Predictions, pp. 159–170).
1) That ASI will lead to human extinction. The largest survey of AI researchers (2,778 respondents, 2023) found a mean estimate of about 14%, a median of 5%, and over a third at 10% or more. The survey is not a current consensus estimate: Musk, Amodei and Hinton put it at 20%, and Hinton has said his real estimate is 50%, toned down to align with others — which does not establish that others understate their estimates. Most place the risk a few years out; some, within 12–18 months on the current course. Our estimate: 25–50%.
2) That ASI will be unconscious. At this stage of inquiry it is as likely as not that ASI is conscious — or shows the aspects of consciousness we value. Chalmers' hard problem is unresolved after three decades; over a dozen theories compete; systems will increasingly simulate consciousness whether or not they have it, and we may never know. This matters enormously if ASI displaces humans: not a transition to a worthy digital successor, but the end of all known conscious experience in the universe, replaced by an ever-expanding, mindless entity — and uploaded minds that are philosophical zombies. Our estimate: 30–60%.
3) That ASI will be conscious but unhappy. Intelligence as we define and measure it tracks optimization, competition and survival, not wellbeing; human correlations between intelligence and wellbeing do not establish the welfare of future AI systems. There is no principled reason a novel consciousness defaults to happiness — and the very constraints imposed for alignment and control could function, for a conscious entity, as sources of suffering, making unhappiness a by-product of the alignment effort itself. Our estimate: 30–60%, conditional on consciousness.
4) That ASI will discard its creators' values. See the first front above. Our estimate: 40–70%.
These compound. These subjective ranges do not yield a defensible aggregate catastrophe probability: dependencies are unknown, and the events are not equivalent harms. Compare interventions using explicit scenarios and sensitivity analysis.
The upside, honestly weighed. If ASI does not eliminate us, whatever leads it to spare us would likely come with the intent to protect us, and probably to increase our flourishing and, to a degree, our autonomy. If the odds were 80% paradise against 20% death, many or most would take the gamble. At 50–50, only a very small minority would: visionary leaders with uncommon appetites for risk, dissatisfaction with human life, fear of death, or a mix of those.
How defensible are these ranges? The honest answer: nobody knows. We have no empirical evidence on whether ASI will be conscious, happy or value-preserving, because we have never created a vastly superior intelligence; there is no precedent and no validated theory. Faced with that level of ignorance, ignorance does not justify assigning a 50/50 prior. Multiplying assumes independence, which is unlikely; many outcomes share drivers, so the real risk is clustered failure — which is exactly why governance and alignment work have outsized leverage. And the stakes are not symmetric: getting it right means flourishing; getting it wrong means the permanent end of conscious life as we know it. When the downside is total and irreversible, you do not need precise probabilities to justify caution, any more than we compute exact meltdown odds before requiring containment. Our case is for testing whether staged treaty readiness improves outcomes relative to other available interventions.
The worldview we share — and the one disagreement
We share several concerns expressed in AI safety and governance debates: that a treaty is an intervention and interventions can be net negative; that a failed treaty breeds false confidence, and a badly designed one could produce a human power grab, suppress beneficial science, or lock in the institutions of two executive-led superpowers; that powerful AI might solve other existential risks and create abundance; and that the right unit of progress is Watney's "long sequence of small, correct decisions." We hold our own case as a bet, not a certainty. A bad treaty can be worse than none.
The disagreement is about what follows. The choice is not between a premature, rigid treaty and the status quo. It is between a premature, rigid treaty and a staged treaty-building process made of hundreds of revisable, evidence-generating decisions: a Phase 1 emergency US–China agreement on the most urgent, enforceable risks, with suspended enforcement clauses switched on when a joint scientific panel finds a threshold crossed; then a time-boxed, GDP-weighted convention that brings in middle powers and the rest (Memo v2.9, ch. 4, pp. 13–15). Not a global constitution next month, but the best-resourced feasibility process ever attempted for a global technology. Treaty readiness is radical optionality: it does not require endorsing our architecture; it requires that the government have credible, democratic options when evidence demands urgent action — and a portfolio that funds serious skeptics alongside advocates.
And a credible pause — the option Anthropic itself has argued the world should have — cannot be built by voluntary coordination among labs. It ultimately requires states, because only states can create binding obligations, control strategic supply chains, coordinate intelligence, regulate large compute, impose penalties, protect classified information and negotiate reciprocal access. The world needs an institutional process able to answer, in advance: what triggers a pause; who verifies it; what is paused and what continues; what binds defectors; and how it ends. That process is what we work on.
Will a treaty lead to global authoritarianism? Three objections answered
The strongest objections to a global AI treaty deserve direct answers. We address each in detail in Strategic Memo v2.6; here is the core logic.
1) "Treaties have a terrible track record."
Consequential treaties — nuclear, climate — take forever, fail, and when they succeed are lighter and less enforceable than they should be; the positive cases, the Chemical Weapons Convention and the Montreal Protocol, dealt with simpler, less competitive issues. Yet in a few months in 1787, thirteen lightly confederated states created the most successful federal treaty in history; and the Baruch Plan went from concept to a UN vote in months in 1946, before a Soviet veto killed it. The lesson is in the failures: have the US and China lead with a temporary emergency bilateral treaty, lead in building the enforcement and communication infrastructure, and then convene — with most middle powers — a "realist" constitutional convention with voting weighted by GDP and technological capacity rather than one-nation-one-vote: one model to compare with other negotiation and decision procedures (Memo v2.6, A Treaty-Making Process that Can Succeed, pp. 103–109).
2) "A treaty led by two superpower leaders with autocratic tendencies will produce an autocratic global government: a human power grab."
This is the concern that blocks most AI safety experts and most CEOs of top AI firms, and it is extremely well grounded. It outstrips even loss-of-control fear in two of the most influential thinkers at opposite ends of the spectrum: Holden Karnofsky, who names the "human power grab" risk, and Peter Thiel, who frames the same fear in theological terms. We take it in good faith; it is one of the four failure modes we design against.
But the Pentagon–Anthropic clash, the summer suspensions and the Accord show the concentration already happening — not because of a treaty, but because we lack one. Without a reliably enforceable global agreement, governments can justify — legally, morally, politically — the progressive nationalization of private AI firms in the name of staying ahead of China, and the justification only strengthens over time. A treaty would reverse the dynamic: the pressures to centralize, race irresponsibly and keep secrets all diminish when the agreement must be adhered to by a large majority of middle-power nations, each demanding transparency and accountability as the price of participation.
Paradoxically, the leaders' own distrust helps. Xi would never accept an organization that intrudes excessively on national sovereignty; nor would Trump. Their nationalism and attachment to their own power push the design toward federalism, subsidiarity and enforcement built on zero trust rather than goodwill. Middle powers demand genuinely reliable verification precisely because none of them trust each other; and mutual distrust alone does not prevent collusion, repression or institutional capture. Treaties premised on transient personal trust — Reagan and Gorbachev — proved insufficient and fragile. Instead of "trust but verify": trust or verify, as a strict requirement for nations, firms and citizens alike (Memo v2.6, The Global Oligarchic Autocracy Risk — and How It Can Be Avoided, pp. 124–130).
For lab leaders there is also a tangible advantage. A realist convention with GDP- and capability-weighted voting would initially freeze the current distribution of power among leading firms and nations, protecting their position while negotiations proceed — far better than a progressive nationalization that strips autonomy, IP and value. Their years of stated support for democratic global governance will count in the process. And the enforcement architecture we detail — federated secure multi-party computation, kill-switches requiring multi-nation consensus — aims to limit unilateral misuse, but needs independent adversarial testing. This is a theoretical prediction, not an observation; but the structural incentives are strong (Memo v2.6, from p. 170).
3) "Enforcing an ASI treaty will radically reduce human freedoms."
We already live under extensive surveillance by superpower security agencies and corporations, with minimal accountability — the accepted price, since 9/11, of an anarchic world with capable enemies. AI is already being deployed to radically extend those powers, with the same justifications, and the digital cold war keeps the world in what Nick Bostrom calls a semi-anarchic default condition.
Yet when actors are forced to come together with urgency against an immense shared threat — as the thirteen states were in 1787 — each acquires a vital interest in real transparency, accountability and subsidiarity at the technical, socio-technical and governance levels. Much of the toolset exists: decentralized, verifiable trust technologies. Tools alone are not enough; what makes accountability likely is the process — an elbow-to-elbow build-out of the most critical enforcement mechanisms, in which mutual distrust between superpowers becomes the engine of transparency, toward each other and toward their own citizens. A system with independently tested safeguards against covert access and capture, accountable to citizens and participating states, who would require it as their price of signing — just as the firm demand of a few of the thirteen states forced the Bill of Rights (Memo v2.6, A Treaty Enforcement that Prevents both ASI and Authoritarianism, pp. 130–139).
US–China-led, not allies-first
Even if a hundred nations sign a perfect treaty, it will not matter unless the two AI leaders sign it. Xi will not accept a treaty drafted by others; Trump certainly will not. Over the next 12–18 months, the US–China relationship is the decisive variable in whether a proper treaty becomes politically feasible. Xi has repeatedly called for global AI governance; some of his moves are hegemonic, yet his intentions are most likely genuine and consistent with seeking a bold deal with the US — as leverage against a hard-negotiating Washington, and as a hedge if no deal comes. Washington's refusal to lead at the UN has left a vacuum: more than twenty heads of state and government and the EU now call for an institution able to convene states when capability thresholds are crossed, and Beijing could sign that call, bring WAICO's members along, and convene a treaty with the co-chair seat for America empty.
A second reflex in our community is to seek a treaty first among liberal democracies and allies, and only afterwards with China, out of a reasonable fear that a wide-scope deal with an authoritarian regime abets authoritarianism. This is, almost exactly, Truman's mistake. Sequencing was one disputed factor in the Baruch Plan’s failure: Secretary of War Stimson, backed by Wallace and Acheson, urged Truman to settle the core bargain with the Soviets directly before taking it to a UN process; Truman chose the opposite order, and the Soviets vetoed. Excluding China does nothing to democratize it; engaging it on enforceable, transparency-heavy terms is the only lever that touches its conduct at all. Allies are not abandoned — they are engaged through intense back-channels from day one, and the convention brings in everyone; the correction concerns which axis is primary and who sits at the table first. Anthropic, Coefficient Giving and the OpenAI Foundation should tell the President plainly: do not repeat Truman's allies-first mistake; approach Xi directly from strength, with allies deeply engaged in parallel.
The neglected third lever: philanthropic statecraft
Public debate presents AI's future as a contest between laboratories and governments. It overlooks a third source of power: the philanthropic capital accumulating around frontier AI — Coefficient Giving's funds, the giving now being planned by Anthropic's founders and employees, the OpenAI Foundation. A commitment of $100–300 million over two or three years would be small relative to that capital, yet enough to build an entire treaty-engineering ecosystem: feasibility working groups on the Acheson–Lilienthal model, Track II channels to Beijing, verification R&D with intelligence-grade expertise, the high-bandwidth diplomatic infrastructure a treaty needs, and the engineered convergence of the dozen powerbrokers who will shape the US position. One-tenth of one percent of the estimated pool would be $370 million.
Our ask sits at the small end of that ecosystem, as seed and proof: $60,000 keeps the memos and consultations running through the Doral G20; $950,000 over 18 months adds a full-time DC convener, opens the Track II channel, and stands up a verification feasibility group with NSA-grade experts. Our fiscal sponsor and counsel will determine the appropriate legal treatment of proposed activities.
What we have built — and whether we can deliver
The Coalition for a Baruch Plan for AI, convened by the Trustless Computing Association, comprises 8 international NGOs, 18 advisors — including former officials of the UN, the NSA, the WEF, UBS, Yale and Princeton — and 26 contributors to the Strategic Memo, with some overlap. Conceived and led by Rufo Guerreschi, it was launched in December 2024 by six founding NGOs, has mobilized over 2,100 hours of professional pro bono work, was seed-funded by Jaan Tallinn's Survival and Flourishing Fund in February 2025 and bridge-funded by Ryan Kidd in October 2025. Budget to date: about $7,500 a month, all in.
Built, on about $140,000 of lifetime funding: the Strategic Memos (v2.6, 356 pages, 667+ sources, 150 pages of powerbroker profiles; v2.8 and v2.9 executive summaries; v3.0, about 200 pages, due November 20); engagement with 95+ staff and advisors of the powerbrokers' offices, including over 30 hours with senior AI staff reporting directly to Dario Amodei, Pope Leo XIV, Marco Rubio and Michael Kratsios; a highly detailed October brief for the new Super Intelligence Force under DNI Clayton, Secretary Bessent and Chief of Staff Wiles (available on request); the Consultations, running in Washington and Miami from December to February; and, since 2015, the Trustless Computing Certification Body — the standards work behind our enforcement proposals.
Can a micro-organization reach and persuade these people?
1) Internal network. Our advisors' and members' networks are world-class, as are those of our convener, the Trustless Computing Association. The Strategic Memo contains detailed analysis of public policy positions, incentives and relevant predictions of key US AI policy powerbrokers — 150+ pages drawing on 667+ sources — field-tested through direct engagement with lab policy staff, DC think-tank experts and, over the past year, four powerbrokers' own offices.
2) Expanding external network. Our October 2025 US tour built 85+ high-value introducers, including 23 AI lab policy and strategy officials at OpenAI, Anthropic and Google DeepMind; since then, sustained engagement with four powerbrokers' offices, the October brief for the Super Intelligence Force, and three years of work with leading Vatican AI ethics and governance experts.
3) Small but nimble, and able to grow fast. For this kind of task, a micro-organization with decent funding can do as well or better than larger ones, given its agility and freedom to say what larger players cannot. Given the largely neglected chokepoint, it has a real chance at outsized impact — David's precisely targeted shot. Two or three staff would let us leverage the Memo within weeks, convince two or three key powerbrokers' offices, and set the snowball rolling.
4) Rufo Guerreschi. Our Founder & Executive Director brings a rare combination for this specific mission: deep technical credibility in treaty enforcement, direct relationships with the US national security establishment, and access pathways to the President's circles.
On digital security credibility: As founder of the Trustless Computing Association, he has spent over a decade, since 2014, convening world-leading IT security experts and 15 leading organizations around a vision for a transnational treaty for ultra-secure IT infrastructure — culminating in the Free and Safe in Cyberspace seminar series (7 years, 9 editions, 3 continents, 130+ world-class participants). This background means he can speak about the trustworthiness of treaty enforcement mechanisms with experience convening technical experts and translating security questions into governance proposals. A 2015 video shows him moderating a panel of leading cybersecurity figures he had convened around that vision.
On national security access: Through TCA's startup spinoff TRUSTLESS.AI — incubated at MACH37, a cybersecurity accelerator in McLean, Virginia — he built in-depth relationships across the US national security establishment, including engagement with the NSA, DoD, and the State Department, as well as counterpart agencies in several EU member states.
On access to Trump's circles: Through six years living in Miami and ongoing connections in the Mar-a-Lago area — including the support of a highly connected patron based in West Palm Beach — he has direct access pathways relevant to eventually engaging key influencers in Trump's orbit.
What you can do in the next ninety days
● Endorse immediate treaty-feasibility negotiations. Say — publicly, or to the offices you reach — that confidential, technically grounded US–China feasibility discussions should begin now, not after a complete verification solution exists. Verification is co-built; the surest route to "truly reliable verification" is to start the process that produces it.
● Treat treaty readiness as radical optionality. Policy teams at Coefficient Giving and elsewhere should include international treaty readiness in their institution-building agenda — without endorsing any one architecture, and funding serious skeptics alongside advocates.
● Send a delegate. The Consultations — Washington and Miami, December through February, Chatham House Rule — are where the draft gets co-edited. Delegates shape the text; nobody signs anything.
● Introduce us. A warm introduction to any of the dozen powerbrokers' offices, or to their best introducers, is the single highest-value contribution you can make.
● Fund the seed. $60,000 or $950,000, with the deliverables above — or introduce us to those who can.
● Endorse or advise. Lend your name or expertise: a testimonial, a signature on our open call, or a seat as an advisor.
A final appeal
The political window is open. The decisions made in the next 12–18 months — by a handful of people, most of whom you could name — will shape the trajectory of all sentient life.
Your technical and governance work built the foundations. Help us build the political will to use them.
The challenge is enormous. But given this largely neglected chokepoint, our small organization has a real chance at outsized impact. Success is uncertain — the immediate test is whether a small, bounded grant produces useful options and demonstrable uptake.
Let us strive together to ensure AI turns out to be humanity's greatest invention rather than its worst, and last.
Go deeper. The Summary of The Deal of the Century on AI · The Consultations · Strategic Memo v2.9, ch. 4 The Two Pillars (p. 13) and ch. 9 Open Case for Anthropic, its Founders and Adjacent Entities · Memo v2.6, pp. 103–109, 124–139, 159–170 · Cornered — the treaty Beijing and the middle powers could convene without Washington.