The Connection Machine

Danny Hillis's MIT thesis, published as a book in 1985, began from a complaint: the computers of the day did one thing at a time through a single narrow channel, while a brain did billions of things at once. His Connection Machine had 65,536 one-bit processors, each with a little memory, connected in a hypercube so that any could talk to any in a few hops. Thinking Machines Corporation, founded in 1983 with Marvin Minsky among its advisers, shipped the CM-1 in 1986 in a black cube studded with blinking red lights.

The machine was built for AI, for semantic networks and the kind of spreading activation that symbolic researchers thought thinking was made of. It made its money in oil exploration, fluid dynamics and document retrieval, and the company was bankrupt by 1994 as commodity chips caught up.

Hillis was right about parallelism and early about what it would be used for. The graphics processors that trained AlexNet in 2012 were, in effect, connection machines on a single chip, thousands of simple cores doing the same operation to different data, and the reason they suited neural networks was the one he had given in 1985.