A one-sided American pause on AI development would let China catch up in under a year, according to a post published on LessWrong on 22 October 2026 by an author writing as josefely. The author’s answer is a US-China treaty, and the post is a survey of how each side could catch the other cheating before cheating pays off.
The catch-up math comes from the Epoch Capabilities Index, which blends many benchmarks into one score. The post puts Claude Opus 5.5, the strongest public US model, at 167, and Kimi K3, the best Chinese model, released in July 2026, at 157. The index has climbed about 14 points a year since reasoning models appeared, so ten points translates to a bit under twelve months of progress. The author adds that US labs sit on their best systems for months before release, so the true gap is unknown.
A shared pause has its own flaw. If both countries stop and one keeps going in secret, the cheater wins a large geopolitical advantage while the honest side carries the risk. The treaty therefore has to detect defection early, and the author argues it would most likely govern compute, because chips are physical, energy-hungry, and made by a very small number of suppliers.
The first toolkit works from outside the other country’s borders. Satellites and aircraft can pick up the heat, light, and electromagnetic signature of data centers, much as the IAEA watches for uranium enrichment. The limits are plain: the images cannot show whether a building is training a model or serving one, and waste heat can be dispersed or the whole site buried. Energy-use tracking, customs records, financial intelligence, and whistleblowers each cover part of the picture and miss small operations. A determined state can also keep a large project quiet. Wartime America ran the Manhattan Project with 130,000 workers, while Anthropic had 3,950 staff in March 2026.
The second toolkit sends inspectors in. They could check a data center’s size, compare its chips against declared inventories, audit chip factories, or examine the developers themselves. Each has a workaround, from hidden hardware to doctored logs, and each asks a rival to open its facilities. The post stresses that AI differs from enrichment because the product is information, so an inspector who understands the work can steal it. Random spot checks help, though: the chance of getting caught grows with the size of a hidden stockpile, and a single catch is enough.
The third toolkit is the most powerful and the furthest away. Chips could report whether they are training or running a model, prove where they sit by timing a cryptographic challenge or by GPS, stop working until they receive a fresh cryptographic license, and cap how many chips they will link into one cluster. Location checks can be fooled with satellite links or leased private fiber, and every scheme depends on a secret key staying locked inside the chip. The author notes that no chip built for a US-China treaty exists, that the hardware could take years, and that it would have to be retrofitted onto machines already installed.
None of it lasts forever. Compute governance, the post says, is a wasting asset: the training compute needed to hit a fixed capability level now halves every 4.5 months, and small specialized models fall under any threshold. The aim is to buy time for society to prepare, not to freeze the technology.
This is one author’s first post on the forum, and it says plainly that the ideas come from academic work by researchers including Wasil, O’Gara, and Heim. Its value is the packaging for a public that polled at 64 percent saying AI is moving too fast, per the AP-NORC survey the author cites. For chipmakers and cloud buyers, the sharper point is where the weight would land: on the hardware roadmap, years before any treaty could rely on it. Until then, the near-term version of a deal would lean on satellites, inspectors, and the economic pain of switching clusters off.
Reported by LessWrong on 22 October 2026, in a post by josefely.
