3 ms·
The Hacker and the Ants
In his 1994 novel "The Hacker and the Ants", Rudy Rucker described a setup for training an AI that uncannily resembles current machine learning approaches.
For example, this passage (though it uses terms related to genetic algorithms rather than neural networks) prefigures OpenAI's Gym and Universe training environments:
“We put a Squidboy instance into each of the 256 Our American Homes, and select out, say, the 64 parameter sets that give the best behavior. Then we replace the worst 64 sets with mutated clones and crossovers of the genes in the top 64,” I explained. “And you leave the 128 medium-scoring guys alone, or maybe mutate them a little. On the next cycle some of the middle guys might do better than average, and some might do worse.” Russ was starting to grin. I was getting over. “So that we don’t have to monitor it, we give all the Squidboys some simple machine-scored task. To begin with, the task will be walking into the living room without killing anybody. And once it can do that, we try a different task. The process works, I promise you.”
And then Rucker makes the leap to a solution strongly reminiscent of Generative Adversarial Networks:
To continue improving the Adze code, we began making the Our American Home test beds more difficult. We began breeding for bad Our American Homes. [...] After a few days of this, the Our American Homes were pretty bizarre-like imagine your worst nightmare of a subdivision to live in. In one house you could see Pat throwing dishes at Walt. In another, Dexter was taking a crap on the front steps. In another, a flipped-out Pat was in the kitchen setting the drapes on fire. In another, drunk Walt hunted the robot with an axe. In another, Scooter was sitting on the ledge of a window holding a carving knife. And in each of the bad Our American Homes, a desperate Squidboy did his best to fit in. Some of the Squidboys did better than others, and those were the ones who would get bred onto the genes of the Squidboys who lost.
Do you know of other examples of art prefiguring current ML approaches?