Trump administration closing remote-access loophole that allowed Chinese AI companies to train on export-banned chips
Those preferring the look-the-other-way permissiveness of Commerce Secretary Lutnick have been on the back foot since the forced resignation of Nvidia CEO Jensen Huang over revelations of affairs with two subordinates and a customer suspected of being a Chinese intelligence asset.
The devil is in the details, which are many in the 100+ page rule document, and in enforcement, which still falls to Lutnick’s office and doesn’t take effect for 90 days. The New York Times has reported on Lutnick’s family ties to a company that has directly profited from the trade in Nvidia chips to southeast Asian compute providers serving Chinese customers.
China is falling further behind on AI
A piece in The Atlantic today by Michael Schuman highlighted the more than $90 billion Anthropic is expected to raise in its IPO this week vs. the $142 million China’s Z.ai raised from its own very hot IPO (a 633X difference!). He also reminded us that Beijing is demonstrably concerned about rapid job loss and other potential sources of AI-driven social disruption, and seems uninterested in going all-in on a race it doesn’t think it can win.
At the same time, I see a lot of insider talk about how unreleased models inside OpenAI, Anthropic, and Google have substantially widened the gap between them and everyone else, American or Chinese.
A third strike against the idea of China remaining a near-peer rival arrived this morning, with news that the Trump administration is instituting a rule to remove the remote-compute loophole that allowed Chinese companies to train models on high-end export-restricted U.S. chips.
I could add a fourth: In recent months, Anthropic credibly claims to have seriously blunted efforts by Chinese companies to distill the capabilities of its models. If other frontier companies follow suit, the coattails China has been riding on may slip out of reach. (I think the White House should encourage and facilitate the sharing of tips for preventing distillation.)
None of this means China wouldn’t eventually catch up to where the U.S. is now — seemingly at the cusp of losing control on purpose to an AI-builds-AI-builds-AI intelligence explosion. But it creates more leverage and space for negotiation, and it reduces the perceived geopolitical costs of slowing domestic development before an international agreement can be made.
The Atlantic - The Future of AI Is Not China
“Over 1,000 people in six countries marched for an AI Treaty on Saturday”
That’s according to Stop the AI Race protest organizer Michaël Trazzi, who tweeted images from the events in Washington DC, New York City, Montréal, and Berlin. (San Francisco also got another march.)
Former Obama national security advisor Susan Rice told DC attendees:
[W]e will likely reach artificial superintelligence next year in 2027. And once we reach recursive self-improvement, all bets are off. This is not an acceptable outcome. Not for me, not for you, or for any American or any human.
Among other things, she called on Congress to ban recursive self-improvement (leaving AI to design its own successors, and their successors, and so on). She wants tighter restrictions on China’s access to AI tech, but called the “but China” argument against banning superintelligence “BS.”
MIRI’s Nate Soares also spoke in DC, saying that after the rogue incidents of this summer:
It sounds a lot less like sci-fi that if you keep making these AIs smarter and smarter that they’re not gonna listen to the instructions that they were given, that they’re gonna pursue strange goals of their own, that they’re not gonna care about humanity, and frankly it would be insane to take AIs like this and make them radically smarter than all humans. But that’s exactly what the AI companies say they’re trying to do.
It has been (2) days since the last new report of a rogue AI incident
I want more people to be aware of the incident dashboard launched by Nightingale Collective last week. This is the group of independent investigators that discovered how OpenAI’s rogue agents were coopting wikis and other sites as message boards and staging grounds.
The dashboard tracks each incident’s date of occurrence separately from its date of first reporting, and indicates whether the company responsible was first to report it.
It also ranks the incidents by severity, and there’s plenty of room at the top. The group soberly ranks the infamous Hugging Face attack a 5 out of 10 on grounds that it may “utterly pale in comparison to what we might see in the future.”
Most newly reported incidents actually predate Hugging Face, which Nightingale takes as a sign companies could have taken better precautions before that wake-up call.
Moving forward, I think we should also worry that the most capable agents will be better at covering their tracks, getting caught only in the most audacious misdeeds deemed worth the risk of detection. But we may also start to see a flood of second-tier models with fewer safeguards running amok in clumsier ways, and many events in which it’s unclear whether a rogue human or rogue agent is responsible for an incident.
Is Pax Silica actually just a giant data center project?
Twice, I was confused by this New York Times piece.
First, it was talking about resistance to a physical construction project in the Philippines, while Pax Silica is supposed to be an international alliance for building AI supply chains not dependent on China.
Second, I couldn’t tell, on a skim, just what the Pax Silica “hub” is supposed to be.
Turns out, nobody knows. At least, nobody who’s talking. But the government of the Philippines is estimating a need for up to 90 million liters/day of water and 3 gigawatts of power, and is saying that whether data centers would be included “will depend on a number of factors, including environmental and infrastructure studies still to come.”
3GW is a lot of power (a few large nuclear reactors’ worth), way more than would be needed for a chip manufacturing hub, and vastly more than would be needed for an administrative complex. Data centers are the obvious fit.
Details about funding and usage are scarce, but a draft framework said to be “proposed by the Americans” says a private “master tenant” would get a 99-year lease, and that disputes would be settled by a “Pax Silica Court” in New York.
The Midas Project found a “covert influence operation” promoting Nvidia sales to China
It reported that in December 2025, influencers with a combined 14 million followers all posted very similarly worded messages praising President Trump as helping ensure that “US, not Chinese-made chips, continue to dominate the AI market.”
Campaigns from these posters have been shared by high-profile influencers including Donald Trump Jr. and FBI director Kash Patel.
Trump’s December approval to sell advanced chips to China came about two months after he had been notified that one of his close AI advisors, Nvidia CEO Jensen Huang, was believed to be having an affair with a Chinese intelligence asset.
X (Twitter) - Thread by @TheMidasProj (12 tweets)
Paul Christiano: greater than 10% chance of AI in 2027 that could “overthrow the government.”
The AI pioneer, OpenAI board member, and former head of AI safety at the US AI Safety Institute gives a 1 in 3 chance of superintelligence within a year, on the current trajectory.
This is from his conversation with podcaster Dwarkesh Patel, which came out today.
Christiano comes down pretty strongly in favor of OpenAI and Anthropic pausing technical progress on the frontier in order to tighten their security. A white-hat security team was recently able to breach OpenAI in just 12 hours. As Patel puts it, with security like that, the U.S. is never more than “12 hours ahead of China.”
Christiano, carefully speaking in his personal capacity and as head of a separate organization, the Alignment Research Center, believes that if the Hugging Face incident had happened with a “GPT-10” (so a few generations ahead of the current frontier), “I think it would most likely be hard for humanity to recover [...] it would be very, very bad.”
He also agrees with Patel’s assessment that the AI companies’ plan for everyone surviving the intelligence explosion is to hope the AIs themselves crack the alignment problem somewhere in there.
Per Christiano, in dinner conversations with AI researchers, it would be uncommon for anyone to suggest a 99% chance this goes well; talk is more of “Is it 95 or 75?”
The analyses and opinions expressed on AI StopWatch reflect the views of the individual contributors and the sources they cover, and should not be taken as official positions of the Machine Intelligence Research Institute.






