The Doomsday Calculation by William Poundstone

This is likely to be either quite short (because there's not that much to say about a non-fiction book surveying an interesting field of knowledge) or way too long (because I give in to the impulse to complain at length about some underlying things the author explains). Let's see if I can avoid that specific Scylla and Charybdis this time....

The Doomsday Calculation is a 2019 book by William Poundstone, who has worked as a financial journalist, writing for various major publications, and also written a number of books mostly circling the intersection of business, popular science, and information theory. (He started off with a loose series starting with Big Secrets, about, roughly, trade secrets of consumer goods, in the '80s, and that's when I started reading him.) Poundstone is one of those writers that I think of as someone I read regularly, but the last book of his I mentioned here was Priceless in 2010.

Doomsday is about the possibility of the end of the human race, and related topics, mostly pitched at a philosophical or loosely statistical level. It covers a swath of intellectual history, roughly the last thirty years, starting from the question: given what we know now (age of the universe and of Earth, length the human race has existed, and some other considerations), can we estimate how much longer there will be a human race?

And the answer is, of course, yes. We can estimate anything. But the expected range can be quite large, and is hugely dependent on the numbers you plug into the assumptions. Most of the book is about the squabbling over assumptions, various theories about related things (from the Drake Equation to Bayesian statistics), and how this general idea can be actually useful when applied to events that aren't one-and-done, as human extinction would be.

I find that the arguments are typically pitched in the kind of thought-experiment mode I find deeply unrealistic, relying on the human equivalent of spherical cows to make them come out correctly. (I have a similar problem with the Monty Hall Problem, which doesn't come up here: my assumption of how that works in the real world is that Monty has discretion to open or not open a door, which throws the math utterly askew.) So they're all intellectually interesting, but I rarely find them convincing. They all depend on assumptions that don't match reality, and so extrapolating them to reality is a big leap - sure, the theorists want to make that leap, because it's interesting and provocative, but if you don't agree with the priors, it all falls apart.

Poundstone is a good writer for this kind of material, a born explainer, and he lays out the players and their theories clearly and succinctly. Doomsday is a fun, intellectually interesting book to read, even if (like me) you don't buy a whole lot of it.

OK, one last quibble: the last section is about AI, and it might seem to be more relevant now than even when Poundstone wrote the book a few years ago. And, again, Poundstone is clear and precise about what various people say and are worried about. But the idea of killer, self-bootstrapping AI is so vastly far away from the reality of Large Language Models and Machine Learning that even the doom-mongers have to assume a discontinuity - the singularity or something similar - to get from where we are to the doomsday scenario. What we have now are input-driven tools that generate plausible text (and other formats), but that are not self-directed. Getting from there to a computer mind that makes plans for itself and self-organizes is not even a path: it's just a bold assumption that it could happen.

Frankly, I blame the Turing Test, which built in the sleight-of-hand that anything that can fool a human should be considered human. That's a stupid metric; huge numbers of things fool humans, all the time. Half the animals in the world look like other things, human minds see faces in chicken nuggets and moon-craters - the human brain makes patterns out of things; it always will generate false positives from anything halfway close. So "if a computer program talks like a human, we can consider it equivalent to a human" isn't true in any useful sense - it's a nice benchmark, but that's about it.

Anyway, none of that can be laid at Poundstone's feet: he's explaining the state of things in this area of human knowledge, and, as always, some humans are conspiracy theorists, some are paranoid goofballs, and some are unimaginative math-heads. If you like arguing with books in your head, this is a great one.