Post #2173585

I read that paper. It was a good one! It has to do with *training* a model (of finite capacity) on it's own output. The simplest part of the argument is that it will loose the rarest data: It's less likely to come out, so it won't survive repeated rounds of training. Repeat ad infinitum until the data is just the dumbest shit. So an agent left on it's own on code won't necessarily do that since it is not re-training itself as it goes. @creepy_owlet