DeepSeek and the compressed timeline

The claim. DeepSeek R1 matched o1-class reasoning at roughly 1 per cent of the reported training cost of the leading proprietary systems. The book cites this as evidence that the frontier is now producible by more actors than the industry's incumbents assumed.

Why the compression matters to the argument

Every safety framework built on the assumption that a small number of well-resourced actors will remain the frontier holders has an implicit clock. DeepSeek moved the clock. If a lab can reach o1-class capability for a fraction of the assumed cost, the field cannot be governed by containing the small number of firms that were expected to hold the frontier. The book's Chokepoint chapter is written around this observation: the leverage cannot last if the market for producing frontier capability broadens as fast as the DeepSeek result suggests.

What the book does not claim

It does not settle whether the 5.6 million figure represents true total cost, partial cost accounting, or a subset of the training programme's compute. Those debates continued after the book went to print. What the book claims is that even a debate in which the true figure is an order of magnitude higher still puts the frontier within reach of more actors than the field assumed, and the argument holds.

How the book uses DeepSeek in Chapter 8

The Chokepoint chapter names DeepSeek as an accelerator of the window's closure. If China's domestic capabilities advance, the chokepoint framework loses effectiveness on the schedule the framework depends on. The book explicitly says the framework must be established while the chokepoint exists, "knowing that the chokepoint will eventually close." DeepSeek moved the "eventually" closer.

The narrative role in the book

DeepSeek appears in the compressed-timeline passages where the book argues that we cannot afford the deliberate policy processes we normally use for technologies of this magnitude. The compressed timeline is not decoration; it is the reason the book argues for architectural instruments (caretaker doping, meltdown triggers) rather than long-term regulatory processes. Architectural instruments can be shipped inside product cycles. Regulatory processes cannot.

The convergence with the December 2024 manuscript

The manuscript argued that AI systems cannot be truly controlled and that safeguards imposed after the fact would fail. DeepSeek is one datapoint for a related but distinct claim: that we cannot rely on the concentration of the frontier holding long enough for slow governance to catch up. The two claims are compatible, not identical. The register does not carry DeepSeek as a convergence with the manuscript on the specific control claim; it treats it as context for the argument.

What the reader keeps

A datapoint about the cost of frontier capability, an argument about how the datapoint compresses the timeline, and a careful line about what the book is and is not claiming about the cost figure. The book uses DeepSeek to justify architectural rather than regulatory instruments, not to name any particular geopolitical claim about which country holds the frontier.