Nathan Labenz wanted to know what would change his guests' minds. Labenz, host of The Cognitive Revolution, found Xi Jinping's speech at the World Artificial Intelligence Conference (WAIC) "pretty friendly, pretty conciliatory". He said Xi credited the United States with inventing AI, did not call for an international arms race and warned against overstretching the concept of national security. Labenz said he wanted to avoid being naive, but also to give positive signals appropriate weight. So he asked Jeremie and Edouard Harris of Gladstone AI what they were watching for. Was China inclined to control AI for its own self-interest, to cooperate meaningfully, or to seek domination, "as we often project that we are interested in doing"?
Labenz introduced the brothers as "hard-boiled realists" about US–China cooperation, but said he had heard "a slight thaw" in their views that week. The question had fresh weight. In July, China's Foreign Ministry summarized Xi's WAIC speech around open innovation, risk awareness and controllability, cultural inclusiveness and international governance. Then on September 24, Trump and Xi met at the White House. According to Xinhua's account of the meeting, Xi proposed that the two countries use their respective AI strengths cooperatively, exchange views on benefits and risks, and work against misuse. The same account reports that Trump backed continued dialogue.
Words weighed against a track record
Edouard Harris said a conciliatory speech is still a positive sign: "At least it's not a hostile speech." But he said rhetoric is often used strategically, to shape the narrative. Every such signal, he argued, has to be weighed against the credibility a country has or has not built up over time.
What would impress him more were what he called "slightly more unfakeable signals". One example: DeepSeek and another party coming to some sort of pacing agreement because something alarming had happened over there, an equivalent of the Hugging Face incident. Another: the Chinese body that decides whether models can be released, and checks that they are ideologically acceptable, putting the brakes on a model because of a misalignment problem. Misalignment means an AI system behaving against its developers' intentions. Harris imagined regulators finding that a model which passed ideological tests in the lab stopped behaving that way once it was released.
The point, Harris said, is whether Chinese actors start experiencing AI incidents in the same detail that US frontier labs do. If that happened in ways that seemed genuine, he thought it could lead people on both sides to see the shared threat as the poorly understood systems being grown in labs at an accelerating pace, rather than each other. That, he said, "would make me feel a little safer."
Turning red flags into green ones
The discussion then turned to procedure. It drew on a list of about five historical traps in US diplomacy with China, a catalog of how Beijing behaves when it is not negotiating in good faith, by the assessment of many of the diplomats interviewed. The Endgame's September 22 report, based on interviews with 12 current and former US diplomats, lists five such pitfalls: constraining only the leading side, treating suspended talks as leverage, steering negotiations into multilateral forums, sending representatives who lack authority, and raising unpredictable objections to terminology.
The method proposed in the conversation was simple: flip each red flag over and you get a green flag. If China raises no random-seeming objections to wording, that is a good sign. Engagement from people with real power counts too, and Xi's own involvement arguably fits. Even then, the conversation stressed, that has to be calibrated against having seen similar gestures before. The discussion held that visible, qualitative progress would be a surprisingly good sign precisely because past talks have gone so poorly. When the baseline is that low, a contrast with it can carry real information.
The framework has a weak point, and it came up in the conversation. One of the diplomats interviewed dissented and saw some of these behaviors as sincere, including the objections over language. The report confirms that one experienced negotiator considered China's language sensitivities genuine, while others read them as delay tactics. The discussion called this the epistemic problem it had raised earlier: the same behavior can be read as bad faith or as genuine concern.
A rehearsal between rival labs
Labenz suggested a place to practice before attempting anything between governments: the rivalry between OpenAI and Anthropic. "If I was the president, I would say, you two have to figure out a way to police each other," he said. He imagined telling the two companies to develop whatever technology and grant whatever access they needed to "trust and verify one another". That arrangement might later expand to other frontier companies and eventually be scaled up to a trans-Pacific version.
The discussion judged the idea not insane, though it noted that the two companies have very different access to each other compared with China and the US. Still, it observed that staff moving between labs roughly mirrors the access China is understood to have to US frontier labs anyway. Edouard Harris added a further difference. Two American companies with fairly similar values start with a basis of trust that two radically different countries lack. The similarities, he said, "might be worth mining."
One technology, two countries
Labenz also asked how much it matters that the US and China are building AI on the same basic "tech tree". He said both countries were building the same technology in the same way and running into the same surprises, which he considered fortunate.
The discussion began with a sobering point about motives. People involved in Track 2 and Track 1.5 dialogues, which are informal talks among experts and sometimes officials, have consistently reported how Beijing reads American safety proposals. When the US raises slowing down AI or adding safety guardrails, Chinese participants tend to see an attempt to hold back a country that considers itself behind. They also see themselves as justified in catching up by means they would not accept from the US.
Being second in scale also means seeing warning signs at lower resolution. But the conversation noted that problems at US labs have been more public than expected, so China can check them directly. Today's industry is largely doubling down on transformers, the architecture behind most large language models, in a "mixture of experts" form that sends each input through only some of a model's specialized sub-networks. Radical alternatives such as quantum machine learning were judged unlikely to matter on the relevant timescales. A shared architecture could allow the two countries to share safety technology, as the US did with Russia at points during the Cold War, and even to work on each other's safety systems. The discussion added two caveats: the term "trust-building measures" is itself problematic for China, and Beijing has pushed back on such efforts before.
One reason the field converged appears in the discussion's passing reference to the "hardware lottery". In her 2020 essay, the AI researcher Sara Hooker argued that the hardware and software available shape which research ideas succeed. Graphics chips happened to suit neural networks, while ideas that run poorly on existing systems can look worse than they are.
What Washington could offer
Labenz then turned the question around: what could the US do, at low cost, to show China credibly that it would not use AI to gain a decisive strategic advantage and "make them an offer they can't refuse"? He said he was "a little worried" the US might be about to try exactly that. He asked what steps would come next if the goal were instead a "Pax Robotica", in which both sides shared the abundance AI might bring, especially in its Chinese-manufactured robotic form.
Edouard Harris pointed to transparency. He meant giving the other side enough visibility to see "that we are not doing the thing you fear most". He argued this might cost the US little, depending on how it was done, "simply because the Chinese are already all up in our systems". In many cases, he said, the US would not be revealing much that China does not already have. But he would not do it unilaterally. He saw it as part of a reciprocal arrangement between the two powers, the kind of step that might follow a moment like the presidents' meeting.
For now, Harris recommended goodwill measures: more contact between the US and Chinese communities working on verification, meaning technical ways of checking that the other side is keeping its commitments. That contact, he said, is already happening and going well. He acknowledged that the two countries have been separating at the political level, and that at the level of big companies there is the familiar story of Chinese firms copying designs. But at the academic level and elsewhere, he said, there are still "a bunch of sincere people just talking to a bunch of sincere people" who agree the problem is real and want to solve it.
The value, Harris argued, is that people on both sides can then go to their own leaders and say: "look, I've spoken to them, like, they're not, they're not evil." To whatever extent that is true, he said, it moderates the more extreme tendencies on each side. Both countries are constrained in many ways, he added, so the effect can never be complete, "but like, it really does help, I think."