I spent a fair amount of time learning about the history of both science and music, and my impression is that both advance in much smaller, more incremental steps than we normally imagine. Post facto, we see shortened versions of the stories that are very linear, and don't describe all the thousands of ideas than never really went anywhere. So it **looks like** progress is this very continual stream, when actually that's only true in aggregate. At the boundary of the known-and-unknown, it's a hazy mess, of a drunkard's walk fumbling around vaguely in the direction of what might make sense. And yes, there's a streetlight effect, but more like "someone builds a new kind of lamp and puts it in a new location, and then starts hunting near there".
This isn't strictly relevant, but kinda cool nevertheless:
The analogy there is maybe that plenty of those bumps end up being in the direction of granite walls, or vast deserts, and we quit exploring, and that work was sorta-kinda wasted (except that often knowing where the boundaries of the map are is in itself useful).
The Worm guy has some suggestions about how to get imitation learners to do that
Obviously for science, the process is focused by "experiment as the arbiter of truth", so AI will also be constrained to be a purely-theoretical, armchair-researcher... as long as we don't do anything **truly stupid**, like give it direct access to a bio wetlab, or the ability to develop nanomaterials on its own.
I spent a fair amount of time learning about the history of both science and music, and my impression is that both advance in much smaller, more incremental steps than we normally imagine. Post facto, we see shortened versions of the stories that are very linear, and don't describe all the thousands of ideas than never really went anywhere. So it **looks like** progress is this very continual stream, when actually that's only true in aggregate. At the boundary of the known-and-unknown, it's a hazy mess, of a drunkard's walk fumbling around vaguely in the direction of what might make sense. And yes, there's a streetlight effect, but more like "someone builds a new kind of lamp and puts it in a new location, and then starts hunting near there".
This isn't strictly relevant, but kinda cool nevertheless:
https://matt.might.net/articles/phd-school-in-pictures/
The analogy there is maybe that plenty of those bumps end up being in the direction of granite walls, or vast deserts, and we quit exploring, and that work was sorta-kinda wasted (except that often knowing where the boundaries of the map are is in itself useful).
The Worm guy has some suggestions about how to get imitation learners to do that
https://gwern.net/ai-daydreaming
Obviously for science, the process is focused by "experiment as the arbiter of truth", so AI will also be constrained to be a purely-theoretical, armchair-researcher... as long as we don't do anything **truly stupid**, like give it direct access to a bio wetlab, or the ability to develop nanomaterials on its own.
We'd never be dumb enough to do THAT, right...?
Narrator's voice: they would absolutely be dumb enough to do that
Actually I'm pretty sure they already are, I just couldn't find the link
:/
They are, I posted about it a few days ago. Reuters reported that ant is doing it
:sobs uncontrollably: