What is load-bearing
An intervention is load-bearing if the outcome rests on it: if it fails, nothing downstream of it matters. That is a different question from whether it is good, or likely, or the thing the book asks for. The ranking below is by load, in the world where the argument is roughly right. Each entry says what the book thinks, what it is thin on, and whether a person can move it.
1. Compute governance and verification
Frontier training needs hardware that is made in a handful of facilities, by a handful of companies, in quantities that can be counted. That is the one physical fact that makes any of the rest possible, and the book is right to build its treaty on it: chip consolidation, production monitoring and use verification are Articles V through VII. Without this, no agreement is enforceable and every other measure is voluntary. With it, a halt, a cap, a review threshold, or a licensing regime all become possible.
This gates everything, which is why it comes first. It is also the intervention closest to existing: export controls on advanced chips are in force and the tooling to know where chips are is being built, though for national advantage rather than restraint. The book is thin on how the same instruments get turned from a race tool into a treaty tool, which is the whole political question. A person cannot move this directly; it moves through the next two.
2. Agreement between the powers
A halt, or any binding limit, holds only if the United States and China both accept it and each can verify the other. The book's Wouldn't some nations reject a ban? argues that the alternative is worse for everyone including the defector, which is true and has never on its own produced an arms-control treaty. What produced them historically was a combination of demonstrated danger, verification that both sides trusted, and domestic pressure in both countries. The book supplies the first in argument form and the second in the treaty draft. It has almost nothing on the third, and the third is where individuals live.
3. Public demand
Upstream of both of the above. Treaties get negotiated when leaders believe their publics want them, and this is the input that has moved fastest and most unexpectedly: the statement calling for a prohibition, the Senate bill, the former defence officials, the entertainers. The book's own theory of change is here, in Will elected officials recognize this as a real threat? and Keep the Coalition Large, and it is the part of the book's strategy that is working. This is the intervention a person can move, and the only one most people can. It is load-bearing in the sense that nothing above it happens without it.
4. Interpretability
The effort to understand grown systems from the inside is the only research direction that addresses the mechanism in claim 2 directly. It is load-bearing in the worlds where a halt does not happen and something gets built anyway: if the systems can be read, then "we don't know what it wants" stops being true, and the book's argument loses its foundation. It is not load-bearing for a halt, and the book gives it two paragraphs in Why not just read the AI's thoughts?. That is the book's largest omission, discussed on the next page. A person with the skills can move this; nobody else can.
5. Evaluation-gated development
Capability thresholds that trigger review, tests before deployment, the ability to halt a run. This is the current policy landscape and the labs' own frameworks. It is load-bearing only if alignment failures scale smoothly enough that survivable tests are informative, which is the unresolved crux of the argument. If they do, this is most of what matters. If they do not, the book is right that it is a description of the race in safety vocabulary. Its real value today is visibility: it is how the evidence for the book's own claims gets produced.
6. Human intelligence enhancement
The book's own proposed way out, after a halt: make humans smart enough to solve alignment. The site's research page and Why would making humans smarter help? make the case. It is long-horizon, it is load-bearing only in the world where a halt has already succeeded, and it is the intervention least connected to anything a person can do this year. It belongs on the list because the book puts it there and because, if the halt happened, it would immediately become the most important thing.
The shape of the list
Two things carry the outcome, compute control and great-power agreement, and both run through public demand. That is the only place ordinary effort enters, and it is more effective than the book's critics assumed. Interpretability is the hedge for the world where the halt fails. Everything else is either instrumentation or the far future. A person deciding where to put voice and money should put it at item 3, aimed at items 1 and 2, and should regard item 4 as the thing to fund if they think the halt will not come.