In September, Washington and Beijing agreed to set up an AI incident channel. Seven weeks on, there are no named officials, no notification thresholds and no categories of reportable incidents. My test for the Trump–Xi meeting in Shenzhen this week is whether it produces concrete procedures or just another communiqué.
Why it matters:
• The bigger near-term risk is bad actors, not states. A third party reportedly used a Chinese open-source cyber tool with Claude and DeepSeek to breach at least nine South Korean banks. That is exactly the case a channel exists for.
• "Staying ahead" at the frontier doesn't make open-weight models safer once their safeguards can be stripped. That takes coordination with China and others.
• In cyber, attackers need one hole and defenders have to patch them all. Sharing vulnerabilities with a small set of partners leaves millions of organizations exposed.
What would make the channel real:
- Move it from Treasury's trade talks to the technical staff who meet regularly: CAISSI, China's AI Safety and Development Association, the CAC and lab evaluators.
Write terms of reference that name the contacts, a Chinese lead agency and notification deadlines.
- Set up a cyber-response link between the two sides for major vulnerabilities their models find.
- Run shared safety evaluations through a neutral host like the UK AI Security Institute.
Mistrust runs both ways, so start with the cheapest credible signal.
We can't race your way to a safer world.🚨
Please read all the details from the seven experts in the new @AsiaPolicy Center for China Analysis pre-Summit briefing. 👇
#AIPolicy #AIGovernance #USChina @wstv_lizzi @LyleJMorris @wendyscutler @QuincyInst
Full pre-summit briefing doc here. centerforchinaanalysis.asias…
1
110
Thanks to Elijah Shoaf and the @uscnpm for a long conversation on why the AI race framing is wrong. There is no single finish line, and treating frontier AI as zero-sum leaves everyone less safe. You cannot race yourself to a safer world.
A few of the points we got into:
•Diffusion matters more than invention. China spends roughly a tenth of what the US does on AI capex, its models trail by months, and Chinese open-weight models held the top six spots on OpenRouter in July at about a tenth of the API price.
•US AI infrastructure spending is heading past $1 trillion a year, backed by over $3 trillion in off-balance-sheet commitments. I think that is financially fragile.
•Employment for 22 to 25 year olds in the most AI-exposed jobs is down 16% against their peers, closer to 19% on updated data. Displacement stays invisible until it hits all at once.
•Payroll is taxed near 30% while capex is taxed near 5%. A Pigouvian levy on AI spending would start to fix that.
•Frontier model testing needs a neutral third party that neither Washington nor Beijing controls. The UK AI Security Institute is a natural place to start.
1
4
399
Grateful to join @WSJ’s What's News podcast this week discussing the fallout from the #Nvidia revelations with Luke N. Vargas. 🎙️
The fallout is spreading to Washington, where both parties (and top frontier labs) are calling for new laws to cut China off from U.S. AI chips, just as US leaders are about to head to ShenZhen for AI Safety Dialogues.
My view: Advanced AI is not containable. You cannot enforce non-proliferation for AI and export controls won’t stop open-source progress. If AGI were achieved, within weeks or months, other countries would have it; whether through espionage, distillation, their own innovation, or just techniques transferred through rumor mills. Software fits on a USB stick.
New export controls and open source bans could temporarily wall the U.S. off from Chinese competition but it only speeds up Beijing's push for technological independence and makes American firms less competitive. Unfortunately, U.S. lawmakers will likely back increased controls anyways, because a simple "they're spying on us" story is easier to sell than one that needs five minutes and some technical background to understand.
We can’t race our way to a safer world! AI safety is a shared interest for both countries making collaboration a must. I hope these new controls and the Jensen Huang incident don’t derail the positive momentum that’s been happening on the safety talks and induce an escalation of ladder of tit-for-tat responses from the two countries. 🚨
#AIGovernace #AISafety #AIPolicy – at Lisbon, Portugal
Closed models buy a temporary benchmark lead at a high cost. Open models enable broad adoption, cost effectiveness, and customer sovereignty.💡
Today at @WebSummit Lisbon I joined Carmem Domingues (former White House Senior AI Advisor) for the "Open vs closed AI" panel moderated by @joshdzieza from the @verge.
The initial framing was China is betting on open source and the West on proprietary models for the AI race. My take is that the whole AI race narrative is a misdirection by the Frontier Labs to reduce AI regulation and the competition from open-source AI commoditization. Anti-Open-source is politically unpopular but anti-China has bipartisan support. 🤔
AI is becoming a global public whether they want it or not. The key is how do we keep it safe while making it widely accessible?
#WebSummit #OpenSource #AI #AIGovernance – at Lisbon, Portugal
There is no winner in the AI race. There is no finish line.💡
I spoke with @Cybernews at @WebSummit in Lisbon. Three points from the conversation:
- China could close the compute gap in 4 to 5 years. Once it does, the US edge shrinks to a pure innovation race, and the model gap is already closer to a year.
- Advanced AI is unlikely to be used offensively. The escalation logic that has kept nuclear weapons unused since 1945 still holds with AGI.
- Europe doesn't need to reinvent the wheel. Open models that are 2 to 5% behind the frontier cost almost nothing and can be diffused across every economy right away. The real value of a technology has always been how well it spreads through a society, not who invented it.
Racing to "win" AI misreads both the technology and history. Cooperating on safety and diffusion is the smarter play for everyone.
#AI #USChina #WebSummit #OpenSource
@uw_ece @UW @AsiaPolicy @VRWorldSociety – at Lisbon, Portugal
Removing @JensenHuang will be the beginning of the unwinding of the #AI trade.🚨 @nvidia is a company that is built on the leadership of one man. If Nvidia starts to fall, the whole AI industry will suffer. 😔📉 – at Lisbon, Portugal
NVIDIA announces CEO transition. Read the release: nvidianews.nvidia.com/news/n…
If America and China can work together on #AI safety, the world gets safer. If they don’t, both sides and the world lose.😔
The US-China AI dialogues restart next week in Shenzhen. There are some people who are putting political agendas ahead of public safety at this moment because they think they are helping the country “win”, but ultimately they are helping bad actors with powerful tools and making runaway AI even more likely.🚨
More details in this week’s #BigBangTechReport 👇
#AISafety #USChina
AI leadership should be measured by how widely and safely the benefits are spread, not just who builds the most powerful models. Europe can plan a key role as a bridge to connect the US and China in ensuring AI is safe, and be the evaluation arm to keep them to their words. 💡🇪🇺
That’s the perspective I brought to the @TallinnSummit. Stimulating conversations about AI, governance, and how we move beyond rivalry toward shared prosperity. Grateful to be part of these important discussions in beautiful Tallinn. 🇪🇪 – at Tallinn, Estonia
Some specific suggested actions to move us towards a more positive and safe outcome for the #AI industry.
Shared @TallinnSummit
2
4
193
Alvin Wang Graylin retweeted
"AGI is not a finish line. AGI is a point of disruption. And if you do not prepare, you're going to actually go backwards."
— Alvin Graylin, Senior Fellow at the Asia Society Policy Institute Center for China Analysis @AGraylin @AsiaPolicy
#TDS2026
What a special experience in Wuzhen at the World Internet Conference. An honor to listen to Premier Li Qiang’s keynote, “Put people at the center of the AI era.” Key messages were: keep humans in charge, encourage “open” innovation, closing the digital divide by sharing AI to the global south, and the need for greater cooperation (especially around AI and safety). These are very positive signals leading up to the coming U.S.-China AI Dialogues. 🤝
I had stimulating conversations with China's AI community and local leaders, and spoke on a panel with esteemed experts on the future AI and why racing is not the path forward. The ancient water town was as beautiful as ever. 👍 – at Shanghai, People's Republic of China
Alvin Wang Graylin retweeted
Alvin Wang Graylin (@AGraylin) | AI Policy Expert & Digital Fellow | Stanford University for Human Centered AI:
AI is not like nuclear weapons. It cannot be contained or controlled by one party. There is no winning this AI race.
The “AI race” narrative serves those who want trillion-dollar valuations for supposedly inventing something unique.
But open-source models are now 90% as good as the best closed frontier models. There is no real differentiation.
Alvin Wang Graylin retweeted
Alvin Wang Graylin (@AGraylin) | AI Policy Expert & Digital Fellow | Stanford University for Human Centered AI:
America & China look at AI differently - and want it to do different things.
The American narrative is focused on building superintelligence & using it to make the country superior to others.
China is focused on growing the economy - diffusing the technology widely for economic growth & perhaps soft power.
Alvin Wang Graylin retweeted
Alvin Wang Graylin (@AGraylin) | AI Policy Expert & Digital Fellow | Stanford University for Human Centered AI:
The goal today is “slave AI” - or gladiator AI - obedient systems that follow orders exactly.
AI industry’s influence is enormous in the US. There are four times as many AI lobbyists in Washington as defense lobbyists.
AI is not like a nuclear weapon. Greater intelligence has a 99% commercial use case. A nuclear weapon has only one.
Nobody can tell you today who is liable when an AI chatbot or agent causes harm in the real world.🚨
That gap shapes behavior. If the AI labs keep all the winnings and your losses get socialized, you bet bigger. Every hand by AI labs becomes an all-in hand. 🃏 When there is no downside, nobody has a reason to police themselves.
Clear liability laws isn't anti-innovation. It's what makes builders take risk seriously. And what builds trust with customers to want to diffuse faster!
Here more in my discussion with @sonusvarghese @RyanDetrick on the Facts and Feelings podcast. 🎙️





















