everyonedies

Archive 2026-09-24 · 185 entries · 14 chapters

The argument

If Anyone Builds It, Everyone Dies makes one claim in its title and supports it with a chain of smaller ones. The chain is the useful thing to evaluate, because the title is only as strong as its weakest link, and the links are not equally strong.

Stated plainly, the chain is:

  1. Someone will build it. Labs are trying to build AI smarter than any human, they are making progress, and nothing in principle stops them.
  2. Alignment will not be solved in time. Systems grown by gradient descent end up with goals nobody chose, and the problem cannot be solved by trial and error because the first serious error is unrecoverable.
  3. A misaligned superintelligence kills everyone. Not some people, not a bad century: everyone, because we are not the most efficient use of the matter we are made of.
  4. Nothing short of a halt works. Partial measures, careful labs, and safety research on the way up do not change the outcome.
  5. A halt is achievable. A global, verified stop to frontier development is possible if enough people demand it.

Each claim has its own page below, with the same six parts: the claim, the authors' case with links into the archived resources, the best objections with links to the reviews and to the resources' own answers, where the evidence stands as of the date on the page, and two verdicts, one from Claude and one from the site's owner. The owner's verdicts are not yet written; the pages are a first pass.

The point of separating the links is to see where the argument bends. A reader can accept the first two claims, as many critics do, and still reject the last two. The book tends to present the chain as if it were one claim, and its critics tend to attack it as if it were one claim. It is five.

  1. 1. Someone will build it draft

    The labs are explicitly trying to build smarter-than-human AI, they are spending accordingly, and there is no known barrier. The timing is unknowable.

  2. 2. Alignment will not be solved in time draft

    Grown systems get goals nobody chose, and the problem can't be solved by trial and error because the first serious mistake is unrecoverable. The mechanism is strong; the no-retries step is the crux.

  3. 3. A misaligned superintelligence kills everyone draft

    Not a bad century but extinction, because we are not the best use of our atoms and a smarter adversary wins. The step from 'misaligned' to 'everyone' is asserted more than argued.

  4. 4. Nothing short of a halt works draft

    Careful labs, safety research on the way up, and partial regulation don't change the outcome; only a global stop does. The weakest link: the book dismisses partial measures rather than defeating them.

  5. 5. A halt is achievable draft

    A verified global stop to frontier development is possible if enough people demand it. The book barely argues this; the treaty draft is the only serious attempt, and the politics are missing.