[[[ummmmmm so something tha is missing from what I’m exposed to is figuring out those reusable patterns so as to let discovery programs succeed like, you can run a genetic algorithm for years on an instruction set and it will never discover its own code but if you compress its design into a bunch of reusable patterns with a small total count and a small total depth, you can (somewhat easily, see perplexity link) set it up so that it _does_ discover its own code quite quickly, and the solution is generalizable to larger (and larger) scopes — it’s basically how we teach children and train animals, we break tasks into parts the recipient can handle, it’s human to do, but not instinctive to do with code [this is possibly how our genes and brain evolved too. When you can mutate your own meta patterns you can discover patterns like the above that themselves discover the above. but you have to mutate your own code etc [[[[[i actually went to an online conference around architecture discovery and was surprised that the leading academics were still not doing this. instead huge cash rewards were being given to tiny advances. it really read to me like the tech was suppressed. and it’s notable that this was _after_ covid and chatgpt dropped, so it could have been a real overt normal suppression, for some reason classic AI has not been discussed much any more and that’s quite obvious, it could be because of alignment fears, unsure [[[in pre-covid I think we were amplifying alignment fears, although hard to order the memories