Re-reading Society of Mind as an engineering manager

Marvin Minsky died on Sunday. I pulled my copy of The Society of Mind off the shelf that night. I first read it in grad school and mostly remembered the title. The second read is different, because in the meantime I’ve gone from writing code to running teams.

The book’s claim is that a mind is made of many small processes, which he calls agents, each of them simple and mindless on its own. Intelligence comes from how they’re organized: which ones can activate or suppress which others, how conflicts between them get resolved, how new agents get built out of old ones. There’s no central self in charge. The book is written as 270 one-page essays, which suits the argument.

Reading it as a manager is a little uncomfortable, because it’s also a decent description of an engineering organization. No single engineer understands the whole system. Teams are specialists that pass work to each other. The interesting failures happen at the boundaries, when two groups each do their part correctly and the combination is still wrong. Minsky spends a lot of pages on exactly that: agents that disagree, and the “B-brains” that watch other agents and notice when they’re stuck.

The same week, DeepMind published AlphaGo in Nature. It beat Fan Hui, the European Go champion, five games to zero last October, and it will play Lee Sedol in March. Its structure is almost the opposite of Minsky’s. It has one big policy network trained on expert games that suggests moves, one value network that estimates who’s winning, and a tree search that uses both. There’s no society there, just a couple of very large learned functions and a search procedure.

It’s hard not to see this as the two big traditions in AI meeting at a funeral. Minsky’s side designs structure: knowledge, frames, many cooperating parts. The deep learning side learns large functions from data and adds as little structure as it can get away with. For the last few years the learning side has won almost every benchmark it’s tried.

My guess is that AlphaGo beats Lee Sedol in March. Most Go professionals I’ve seen quoted think Lee wins, but Fan Hui was a real professional and the program has had five months to improve. I’d also guess that ten years from now the systems that matter will look a bit more like Minsky than AlphaGo does, with many specialized learned parts coordinating with each other. They’d be learned instead of designed, but they’d still be a society.