Peter Diamandis asked Alvin Wang Graylin what he was advising the American side to seek from an upcoming AI dialogue with China. Graylin lowered the bar for the first encounter: not a sweeping settlement, but a reason to keep talking. His wider agenda starts with an emergency hotline, shared safety tests and ways to verify each other's safeguards.
“I think the key right now is that we don't need to get to a solution on day one,” he said on Moonshots with Peter Diamandis, in an episode published August 18, 2026. If the delegations agreed to a second discussion, “that's already success.”
Diamandis described the dialogue as planned for September 24 in Washington and introduced Graylin as supporting the U.S. side. That timing is the episode's account, not an independently confirmed meeting announcement. Graylin stressed that he was offering his own views, not representing discussions inside any organization. His recommendations should be understood as a personal policy agenda, not an official negotiating position.
A technology executive who has worked in both countries, Graylin also addressed his China-linked affiliations. Asked about a government-endorsed virtual-reality industry alliance and a part-time teaching role at Beihang University—which Diamandis identified as defense-linked and on the U.S. Entity List—he said both positions were unpaid and that he had had no involvement with either since leaving China in 2024. He described himself as a U.S. citizen of more than 45 years.
Start with a channel that works in a crisis
Graylin's argument for modest expectations begins with his reading of an earlier dialogue in 2024. He recalled disappointment that China had not brought its technology specialists and that the meeting had produced no solution. Some took that as evidence that Beijing was not serious, he said.
His explanation was a mismatch in preparation and expectations. Chinese officials, in his account, take longer to prepare for diplomatic discussions. Starting substantive conversations well in advance would therefore be progress—not merely administrative work before the real negotiation.
When the conversation turned toward which country could build the strongest AI business ecosystem, Graylin steered it back to safety. The initial shared interests, he argued, were preventing malicious use, limiting longer-term harm from AI itself and managing the risk of state-to-state escalation.
He put attacks by non-state actors first. A criminal group using AI to attack a network creates a problem for both governments. Worse, an attack could be mistaken for an operation by the other state, or deliberately made to look like one. That is the purpose of his proposed “red line hotline”: a direct channel for exchanging information before an uncertain attribution turns into a confrontation.
Graylin argued that governments have reasons to avoid first strikes because retaliation can outweigh any temporary advantage. A non-state attacker need not face the same incentives. Cooperation against that threat would not require Washington and Beijing to stop competing; it would require them to recognize a danger neither can manage entirely alone.
Why a shared danger does not automatically produce trust
Diamandis brought up Graylin's favored game-theory analogy: the stag hunt. Game theory studies how decisions depend on what other people are expected to do.
In the hunting story, each hunter can catch a hare alone for a modest reward. A stag offers much more, but catching it requires both hunters to remain committed. Someone who pursues the stag while the other leaves to catch a hare risks getting nothing.
The obstacle is confidence, not simply failure to see the larger prize. As Brian Skyrms explains in his chapter on the stag hunt, the rewards and the probability of collective success must be specified before the story becomes a precise model. Both mutual cooperation and sticking to the smaller, independent reward can be stable outcomes.
Graylin's prescription is to build useful, manageable AI now, learn how to govern it and then cooperate on more transformative systems. He contrasts that with a race in which each side expects the other to abandon safety commitments and therefore races ahead first.
His May essay on the U.S.–China AI race makes room for both kinds of interaction: coordination in commercial AI alongside persistent competition in national security. The stag hunt is his proposed way of thinking about parts of the relationship, not a finding that cooperation must prevail.
The practical steps appear in Beyond Rivalry: observe each other's safety tests, allow reciprocal audits and publish research together. Observing tests would let each side see how the other checks a system; audits would provide a way to examine whether promised safeguards are actually being followed. These are proposed confidence-building measures, not existing bilateral arrangements.
Graylin also proposes an international research facility modeled on CERN, the multinational particle-physics laboratory. It would pool expensive computing resources and provide monitored access for scientists from participating countries. Sensitive security work would remain competitive, while collaboration would be bounded to areas such as safety, health and shared standards.
That still leaves the hard questions of what each side would allow outsiders to inspect and how much confidence those inspections could justify. A hotline can keep communication open; it cannot by itself establish that a model is safe or a promise has been kept.
An AI Marshall Plan
Graylin closed with a larger proposal aimed at American commercial interests: an “AI Marshall Plan.” Financing AI infrastructure and deployment abroad, he argued, could create markets for chips and services while helping other countries use the technology.
The historical analogy is reconstruction assistance that also served strategic interests. The State Department's account of the Marshall Plan describes approximately $13 billion in aid to Europe, supporting functioning economies and democratic institutions while resisting communist expansion.
Graylin sees China's World Artificial Intelligence Cooperation Organization as an attempt at an AI-era equivalent. There are documented initiatives behind that comparison. China's Foreign Ministry reported on July 16, 2026 that representatives of 29 countries had signed its founding agreement. The announcement describes an independent intergovernmental organization headquartered in Shanghai, with aims covering cooperation, governance, safety and equitable benefits.
In his World AI Conference keynote the following day, Xi Jinping announced 5,000 AI training and seminar opportunities for developing countries over five years, international application-cooperation centers and access for 30 countries to the MAZU meteorological warning system. Those were announced commitments, not completed deployments or thousands of new training centers.
Graylin separately recalled asking someone involved in organizing the initiative whether the United States should be invited. In his telling, that person welcomed American participation and was even open to changing the organization's name. That recollection is distinct from a formal invitation or an agreed U.S. role.
America should either build something comparable or work with China, Graylin argued. His proposed bargain joins commerce to safety: help finance deployment abroad, create customers for computing and services, and spread models tested against standards both countries accept. In the episode, he envisaged jointly tested open-source models as part of that effort.
Before either a shared laboratory or an international deployment program could take shape, his agenda starts smaller: establish a channel for the next suspicious attack, begin checking safety claims together and leave the first conversation with a second one on the calendar.