I’ve already used this book as a counter-example in my note on The MANIAC, so it’s only fair to give it its own entry.
The first part is genuinely good. Mallaby is very strong on Demis Hassabis’s early years, the gaming background, the chess, and later the Go and the board games. Hassabis is a genuinely interesting person and obviously extremely intelligent, and the account of how he assembled his interests into a research programme is worth reading.
Almost everything I marked comes from that half of the book. Peter Molyneux handing him a copy of Gödel, Escher, Bach, which Mallaby nicely calls a fire hose of a book. The observation that Gödel’s incompleteness theorem only confirmed what was obvious to them anyway, since humans spend very little time doing deductive logic and most of it extracting meaning from jumbled images, words and smells, which logicians call induction and everyone else calls pattern recognition. Claude Shannon’s definition of information as the opposite of uncertainty, and the leap from there to the idea that a theory of everything would probably take the form of a computer program.
There’s also a line about hiring that I liked, on the sort of person these labs are built out of:
“If you can find these people and be gentle to them and mother them, you get something great which other companies are missing.”
And a genuinely useful framing from Dario Amodei about the marginal returns to intelligence. His point is that investing in a factor of production only helps if the complementary inputs exist, so more intelligence only unlocks progress where intelligence is the actual constraint. If the limiting factors are bureaucratic institutions or public suspicion of tech firms, more intelligence doesn’t buy you much. I think about that one fairly often.
Then the second half arrives, and the book gets into the mess that DeepMind became, the failures inside Google, and the complete obsession with one approach to solving AI. My problem is that Mallaby works far too hard to explain the failures and mistakes away. He is plainly in awe of his subjects, Hassabis above all, and the effect is that a fascinating person ends up sounding like a bore.
Months after I read it, Hassabis and more or less everyone else from DeepMind left their posts at Google. Reading the book back, that isn’t remotely surprising.
All quotes are from The Infinity Machine by Sebastian Mallaby.