Summary

Sullivan treats transformative AI within the next couple of years as a planning case, not a forecast. AI 2027 is a “totally credible scenario,” but AGI by 2027, 2028, or even 2030 is not inevitable; informed experts remain radically divided. His prescription is to build concrete tools around today’s cyber, bio, labor, and military risks so government develops the “muscles” needed if an intelligence explosion arrives.
America leads at the frontier but remains “behind the curve” in deploying AI across government and national security. Sullivan’s Biden-era memorandum directed the Pentagon, intelligence community, Treasury, Commerce, and State to accelerate adoption, and the Trump administration has not discarded it. Yet bureaucracy, classified-computing rules, Congress, and defense primes remain obstacles. Demos help, but he expects Congress to act only after real-world shocks—catastrophe, job displacement, or a foreign military breakthrough—leaving policy in catch-up mode.
The Trump AI action plan reflects substantial bipartisan continuity, but Sullivan sees a credibility gap on China. He supports its investment agenda, proposed global AI alliance, biosecurity provisions, and renewed pre-synthesis DNA screening; his concern is the dissonance between promising to deny China high-end compute and the administration’s H20 decision. Under profound technical uncertainty, he favors prepared but flexible policy tools over premature, highly prescriptive regulation.
Sullivan supports global diffusion of American chips and AI, provided frontier training and most compute remain in the United States. Gulf deals should meet strict cyber, physical-security, insider-threat, and anti-diversion standards, with limits preventing America from swapping oil dependence for compute dependence. When Labenz raised a possible 5-gigawatt Saudi data center running recursive self-improvement beyond U.S. jurisdiction, Sullivan agreed “very strongly” that leading-edge training runs should occur domestically—making U.S. power availability a strategic bottleneck.
China is a competitor to manage, not an economy to decouple from or a government Washington should try to overthrow. Sullivan’s concrete threat model combines subsidized “China shock 2.0” exports, coercion over speech, Volt Typhoon malware pre-positioned in U.S. infrastructure, the world’s largest peacetime military buildup—probably in all human history—and potential conflict over Taiwan. His desired outcome is neither Cold War nor regime change, but enough U.S. economic, technological, and military capacity to prevent a China-led stack from producing “greater coercion, less freedom” and weaker opportunity for Americans.
The export-control debate turns on whether frontier compute is a narrow military input or a general-purpose engine of prosperity. Sullivan argues civil-military fusion makes benign end-use restrictions unenforceable, so America should not hand a strategic competitor the equivalent of an advanced jet engine; controls should preserve a frontier advantage without blocking gaming or ordinary applications. Labenz’s challenge remains unresolved: if the objective becomes running 10 times as many AI workers, the “small yard, high fence” starts looking like broad denial—even if Sullivan rejects the claim that Chinese consumers will lose access to AI doctors.
Sullivan prefers “managed competition” to either an AI arms race or a grand bargain with China. The model is to “compete like hell” while maintaining guardrails, direct AI-risk dialogue, arms-control-like engagement, and communication against miscalculation; he calls competition a “chronic condition” that no G2 deal can cure. On strategic dominance, his deliberately ambiguous answer is to cross the river “by feeling your way across the stones,” because current uncertainty makes rigid declarations more dangerous than useful.
AI-enabled warfare and labor disruption are the two near-term tests of American state capacity. Ukraine offers glimpses of attritable scale, autonomous weapons, AI-assisted intelligence, logistics, and command-and-control, while Sullivan worries the PLA may adopt these capabilities faster than the Pentagon. Economically, he expects disruption regardless of aggregate growth and calls for better social insurance as people confront changes in income, family provision, meaning, and professional purpose.
Labenz presents Sullivan as a restrained, pragmatic “very best of the establishment” voice, while flagging a gap around extreme AI scenarios. Standard policy and international-relations playbooks offer little for the most disruptive cases, and Labenz urges leaders to pair risk management with a positive vision for Americans.

Deep dive

1. Short timelines belong in the plan, not the forecast

Sullivan’s planning assumption is that transformative AI in the next couple of years is a “distinct possibility” with effects across economics, society, warfare, state competition, and non-state threats. But he explicitly rejects certainty that AGI arrives by 2027, 2028, or even 2030.
Labenz takes AI 2027 and a developer-triggered intelligence explosion seriously. Sullivan had read the scenario, shared it “extremely broadly,” and found technologists split between “totally credible” and less credible—not grounds for paralysis, but evidence that policy must accommodate a wide band of outcomes.
Sullivan’s concrete-first response is to solve live problems such as AI-enabled cyber risk in 2025 and inadequate social insurance. Doing that, he argues, develops policy tools and consensus that “tees us up better” for misalignment, takeoff, and P(doom), rather than pretending those questions are insignificant.
Labenz reported Sullivan’s off-mic advice to the AI-safety community: go deep on specific downside scenarios, give officials clear threat models, and pair them with concrete response plans.

2. AI policy needs a ledger of risks and opportunities

Sullivan organizes AI into four buckets: security, economics, society, and existential questions. His rule for government is deliberately spare—identify concrete risks and opportunities; if an intervention can expand the upside or minimize harm, “act. If it’s not, then don’t act.”
Security risks include cyber, bio, adversary “wonder weapons,” and decisive military advantage; economic risks include displacement and concentrated wealth and power. Against those sit productivity, climate and energy problem-solving, public health, and the prospect of broader U.S. national-security capacity.
Society presents misinformation and alienation alongside universal access to teachers, doctors, nurses, and personalized care. Sullivan finds both “robot apocalypse” and superintelligence-led nirvana too imponderable for definitive answers, while Labenz presses that prioritizing nearer-term issues implicitly depends on timelines.
Sullivan’s frustration is analytical scarcity: beyond “the headline or the whiz-bang or the doomerism,” he found little work coherently mapping the full risk-opportunity landscape. That missing rigor limits sensible decisions about where government should act.

3. The U.S. government cannot lead technology it barely uses

Sullivan canvassed frontier-lab leaders, skeptics, investors, academics, and the national-security establishment. The resulting memorandum said America led at the frontier but had to adopt AI with “greater urgency” and “greater dexterity” across the Pentagon, intelligence agencies, Treasury, Commerce, and State. The Trump administration has not discarded it.
Labenz’s operational challenge was stark: White House personnel reportedly lacked ChatGPT-like access and sometimes used tools at home before returning to secure offices. Sullivan, now an “avid would be an understatement” user, acknowledged using them far less while National Security Advisor because of culture, legal constraints, and classified systems.
Live demos made an impression on Sullivan, but he doubts scary chatbot or bioweapon demonstrations alone will move Congress. Legislators generally react to real effects—a 9/11-scale event, cascading job losses, or foreign military capabilities—just as the Patriot Act followed catastrophe “in part for good and in part for ill.”
With technology direction and timing unresolved—and the Trump administration signaling little appetite for legislation—Sullivan favors preparation, flexible standards, and deployable tools, with activation decisions made through dialogue among Washington, Silicon Valley, and the wider country.

4. Global chip diffusion needs a domestic-compute anchor

Sullivan found bipartisan substance in the AI action plan: investment to preserve the frontier lead, cooperation with like-minded countries, and biosecurity measures including pre-synthesis DNA screening. His central objection is inconsistency between denying China high-end compute and decisions such as H20.
His Gulf framework has three legs: run the world on American rather than Chinese AI and chips; demand strong protections against diversion, insiders, cyberattack, and physical compromise; and avoid replacing oil dependence with dependence on foreign compute.
The “devil’s in the details” because the UAE and Saudi announcements were still announcements. Sullivan supports sales but also quantity limits, insisting that “the lion’s share” of American AI compute, future build-outs, and leading training runs stay inside the United States.
Labenz asked whether a 5-gigawatt Saudi data center could host recursive self-improvement beyond U.S. control. Sullivan agreed frontier runs should not occur in the Gulf and tied that objective directly to breaking America’s power bottleneck; differing values warrant scrutiny, not a wholesale blacklist of partners.

5. Sullivan’s China threat model is concrete and bounded

Sullivan rejects decoupling: the countries must maintain an economic relationship and “live alongside one another as major powers for the indefinite future.” His “small yard and high fence” reserves controls for high-end capabilities with national-security applications.
The economic threat is “China shock 2.0”: state subsidies and cheap exports could undercut firms playing by different rules and hollow out even future industries. His response is coordinated defensive measures with the G7 and others, not war.
The coercion examples are sharper. China pressured the NBA after the Houston Rockets general manager made a comment about freedom fighters in Hong Kong; if Chinese AI ran the world, Sullivan warns, Beijing could attach “a price for your speech.” Volt Typhoon’s pre-positioned malware adds direct exposure across U.S. water, electricity, pipelines, and other infrastructure.
Sullivan describes China’s military buildup as the world’s largest peacetime military buildup, probably in all human history. Its Taiwan ambitions and capacity to challenge the U.S. economically, technologically, militarily, and diplomatically make it unique. Sullivan pairs competition with Wang Yi-level diplomacy, cooperation where possible, and an inaugurated AI-risk dialogue: “It’s Cold War time, baby” is explicitly not his position.

6. Export controls expose the dual-use fault line

Labenz’s Pacific-war analogy: cutting China off from chips described as “the new oil” might create the same use-it-or-lose-it pressure that an oil embargo created for Japan. Sullivan sees no evidence China feels compelled to attack and distinguishes targeted high-end controls from a comprehensive embargo.
Sullivan also sees tension between claiming controls are futile and claiming they are intolerably constraining. China expects to develop better domestic chips and applies its own export controls; economic and technological “cards to play,” he argues, are ordinary features of competition rather than an inexorable path to war.
Labenz granted that controls might bite over AI 2027 or Dario Amodei-style timelines, yet questioned their evolution from denying exquisite military uses to denying frontier models or enabling 10 times more U.S. AI workers. At that point, Chinese citizens could reasonably ask whether America is withholding their “AI doctor.”
Sullivan’s answer rests on civil-military fusion: chips serving science and commerce can also build military, intelligence, and economic-statecraft advantage, with no meaningful way to police end use. He denies seeking broad consumer deprivation, but prefers not to hand China a decisive input—just as America would not supply advanced jet engines or hypersonic technology.

7. U.S. leadership does not require changing China’s regime

Erik Torenberg quoted Dario Amodei’s proposal from his “Machines of Loving Grace” essay to convince China to “give up competing with democracies in order to receive all the benefits and not fight a superior foe,” reading it as regime change. Sullivan’s categorical reply: shaping China’s government should not be an object of U.S. policy, despite his respect for Amodei’s ethics and democratic commitments.
Labenz saw upside in Kimi K2, which was leading some creative-writing benchmarks in English, Moonshot’s Western-loving aesthetic, and the possibility that decades of engagement had created foundations for friendship. Sullivan’s base case remains Xi Jinping’s consolidation of power, national security as domestic social control, and a strengthened Communist Party—but he would not rule out evolution.
People-to-people optimism survives that realism: Sullivan wants more students, cultural exchange, and societal connection between “two big, ambitious, dynamic nations.” Policy should protect U.S. interests and values while watching China’s internal trajectory, not trying to dictate it.
China’s cheap, open-source AI is meaningful soft power: if capabilities are equal, countries will choose the easier and cheaper stack. Sullivan’s counter is technological performance—many countries prefer American tools because they are “damn good,” so sustained frontier leadership can generate its own attraction.

8. Managed competition is the equilibrium, not a grand bargain

Sullivan’s equilibrium assumes neither side simply wins. The countries keep “competing, jostling, a little bit of elbow throwing,” while guardrails prevent conflict—the Biden administration’s working formula was “compete like hell” and manage the competition intensively.
Powerful AI adds a modern arms-control track: build capabilities while discussing shared risks, non-proliferation, mistakes, and escalation. Sullivan invokes sustained U.S.-Soviet engagement as evidence that vigorous rivalry and risk reduction can coexist.
Labenz’s pushback—worth keeping—is that unverifiable data-center activity and mutual pursuit of strategic dominance could make AI less stable than nuclear competition. Sullivan prefers ambiguity in this phase: stay at the frontier, prevent an opponent from gaining an unanswerable lead, and preserve communications while “feeling your way across the stones.”
A grand bargain cannot remove the underlying rivalry, which Sullivan calls a “chronic condition.” His deliberately “boring” alternative rejects both defeating China and imagining a G2 condominium will make competition disappear.

9. War and work will reveal whether institutions can adapt

Ukraine offers “glimpses of the future”: abundant, attritable systems change battlefield scale; AI is entering drone platforms while humans remain in the loop; fully autonomous weapons are now imaginable. The transformation also reaches intelligence, logistics, and command-and-control.
Sullivan’s largest military worry is organizational, not conceptual. The Pentagon, Congress, bureaucracy, and defense primes may be unable to adopt AI “at scale and with speed,” while the PLA appears better positioned—a bipartisan “kick in the pants.”
The unfinished domestic question is how people earn, support families, and find professional meaning amid disruption. Sullivan calls AI unusually uncomfortable for policymakers because a private-sector-led technology has profound national-security effects while predictions remain extraordinarily wide; his response is better social insurance and a “can-do spirit,” not denial.
Labenz’s overall framing was that Sullivan’s restrained establishment pragmatism is useful for managed competition, but standard international-relations playbooks leave a gap around extreme AI scenarios; he urged a more positive vision for Americans.

OpenAI's GPT-5 Flop, AI's Unlimited Market, China's Big Advantage, Rise in Socialism, Housing Crisis Steven Sinofsky & Balaji Srinivasan on the Future of M&A, AI & Tech