Hints from AlphaFold


It's surprising to me that protein folding fell to AlphaFold -- a transformer-based architecture. There's no very deep thinking going on here, like there seems to be in eg AlphaZero. I wouldn't have guessed that protein folding would fall this early. I would've thought that:

But in "Discussion with Yudkowsky on AGI interventions", Yud says that he predicted intelligence would be enough to solve folding (predict that first bullet point above is wrong) because he abstracted correctly from hints like:

Evolution is dumb; if it figured out some space of proteins that it could easily traverse, very shallow thinking might also be able to traverse that. Another key point I missed is that figuring out how natural proteins fold must be much easier than figuring out any protein will fold.

It does also seem like deeper thinking than exists in AlphaFold is already do-able. AlphaZero seems to be doing the type of deeper thinking that's really relevant for science. Anything that plays a game has to think about what is possible given a certain move; that seems pretty relevant when trying to do experiments to figure out some property of nature. But it doesn't seem as useful for something like AlphaFold.

If a key advantage in protein folding was taking advantage of evolution being a blind idiot god, then maybe there are similar bio breakthroughs to be made with relatively shallow thinking.