Taiwan makes AI chip smuggling to China a criminal offense
Taipei is turning the black market in advanced AI hardware into a crime it can actually prosecute — and a researcher says the field has spent a decade misreading how its own models work.
Taiwan's government plans to add AI servers and cutting-edge semiconductors to its list of controlled exports, and once the measures take effect, shipping restricted AI hardware to China becomes a criminal offense under the Foreign Trade Act, Minister of Economic Affairs Kung Ming-hsin told Bloomberg News. Kung, interviewed on October 12, said the new controls are part of Taipei's commitments under the Agreement on Reciprocal Trade signed with Washington in 2026: "Going forward, some of our measures on controlled items will correspond to American ones because we have made quite some commitments." He declined to say when the rules land.
The change closes a gap that Taiwanese prosecutors have been working around for months. Unauthorized AI chip exports to China are not currently a crime on the island — authorities can only charge suspected smugglers under other, existing laws. That is exactly what happened in August, when prosecutors indicted nine people, including employees of Nvidia and Super Micro, over AI servers allegedly routed to China, and leaned on forgery charges rather than export-control statutes. Making the shipment itself criminal gives Taipei real leverage over the middlemen who move the hardware, at a moment when US officials allege billions of dollars of Nvidia accelerators have already reached China through a sophisticated black market. Taiwan also plans to raise billions of dollars for US industrial parks aimed at Nvidia, TSMC and Foxconn's supply chain, starting with a Texas complex costing up to $4 billion.
The awkward part is timing. The same week Taipei moved to criminalize the trade, the evidence that controls are working got thinner: Nvidia's share of China's data center chip market fell to 34% in the first half of 2026 while shipments of its AI accelerators dropped 44% year over year, according to IDC data cited by Reuters, and Huawei has been shipping its own Ascend silicon at scale. Semafor's read is blunt — Taipei is tightening rules to please Washington while Chinese firms take domestic share anyway. Criminal penalties will raise the cost of smuggling; they won't slow the substitution that smuggling was papering over. We covered the August indictment when it landed — Taiwan indicts nine over AI servers allegedly smuggled to China.
Nick Cammarata, a Goodfire researcher who co-authored the original circuit and superposition work, published an essay arguing that neural networks are already legible and that the field's founding evidence for polysemanticity was misread. In "Neural Networks Are Already Clean," Cammarata revisits the canonical examples: the famous neuron that seemed to fire for both cat faces and car fronts is, he writes, a V-shape detector (V-shaped flames, water jets, fox faces, with an O for its negative pole), and a 2019 circuit that appeared to smear a car feature into dog neurons reads as shared pose. His explanation for the remaining mess is what he calls recursive generalization — a neuron sprouting related offshoots as it grows, not concepts folded together arbitrarily. The numbers attached to the claim are the striking part: in preliminary audits of Qwen3-32B, Cammarata estimates roughly 90% of neuron poles have one main job, about 7% have two related jobs, and only 2% to 4% are not yet understood, with similar results in models up to trillions of parameters. For contrast, Anthropic's 2023 monosemanticity study scored the median neuron zero on its rubric, and OpenAI's automated interpretability work the same year found about 1.7% of neurons interpretable.
If Cammarata is right, the practical consequence is large: mechanistic interpretability would not need to train a second model, such as a sparse autoencoder, to approximate a network as it would look untangled before anyone can study it — researchers could read the network directly. He is careful to call it a working hypothesis, and the headline percentages move with how you count weak activations, so treat the 90% as a regime indicator rather than a precision measurement. Still, a claim this large from one of the people who built the orthodoxy it undercuts is the kind of thing that reshapes where interpretability effort goes next.
What to watch: whether Taiwan publishes a date and a scope for the new controls — and whether anyone outside Goodfire replicates the neuron audits.
If interpretability turns out to be this easy, was the field solving a problem that was never there? Tell us in the comments.
