-start with some rough seed data -make matches based on our initial hypotheses of data vs action vs outcome -learn from success/failure of matches -adjust ma

-start with some rough seed data -make matches based on our initial hypotheses of data vs action vs outcome -learn from success/failure of matches -adjust matching strategy & request more or different data as needed -search for best predictors by babble and prune [1] with extremely prolific babblers and extremely discerning pruners. GANs & transformers & hashtables, we'll grow em quickly out of the right SUBSTRATE; I think representation efficacy makes for the biggest performance jumps in machine learning training and inference speed. Constant and well-directed experimentation is key in innovation, let's get all the best people spazzing out constructively on a constant basis Let a thousand distributed "Bell Labs" type bloom. (If you don't know, Bell Labs is known for many innovations and was ran with a spirit of collaboration, tinkering, and interdisciplinary work. And also it had all the right people in the same spot!) [1] https://www.lesswrong.com/s/pC6DYFLPMTCbEwH8W https://www.lesswrong.com/s/pC6DYFLPMTCbEwH8W "Babble creates a constant stream of coarse material, while Prune cuts these ideas harshly without the slightest hint of remorse."