There’s only 1 problem left and that’s the protein folding problem. The protein folding problem is somewhat rapidly made progress on software wise, and even if that were to fail it won’t be all that long before we ca simply brute force it with computing power.
Okay, so one sub-piece of puzzlement I have is why talk of protein folding as a problem that is either solved or unsolved—as if we (or more frighteningly, an AI) could suddenly go from barely being able to do it to 100% capable.
I was also under the impression that protein folding was mathematically horrible in a way that makes it unlikely to be brute forced any time soon, though I just now realized that I may have been thinking of the general problem of predicting chemistry from physics, maybe protein folding is much easier.
Predicting chemistry from physics should be easy with a quantum computer, but appears hard with a classical computer. Often people say that even once you make a classical approximation, ie, assume that the dynamics are easy on a classical computer, that the problem of finding the minimum energy state of a protein is NP-hard. That’s true, but a red herring, since the protein isn’t magically going to know the minimum energy state. Though it’s still possible that there’s some catalyst to push it into the right state, so simulating the dynamics in a vacuum won’t get you the right answer (cf prions). Anyhow, there’s some hope that evolution has found a good toolbox for designing proteins and that if can figure out the abstractions that evolution is using, it will all become easy. In particular, there are building blocks like the alpha helix. Certainly an engineer, whether evolution or us, doesn’t need to understand every protein, just know how to make enough.
I think the possibility that a sufficiently smart AI would quickly find an adequate toolbox for designing proteins is quite plausible. I don’t know what Eliezer means, but the possibility seems to me adequate for his arguments.
Okay, so one sub-piece of puzzlement I have is why talk of protein folding as a problem that is either solved or unsolved—as if we (or more frighteningly, an AI) could suddenly go from barely being able to do it to 100% capable.
I was also under the impression that protein folding was mathematically horrible in a way that makes it unlikely to be brute forced any time soon, though I just now realized that I may have been thinking of the general problem of predicting chemistry from physics, maybe protein folding is much easier.
Predicting chemistry from physics should be easy with a quantum computer, but appears hard with a classical computer. Often people say that even once you make a classical approximation, ie, assume that the dynamics are easy on a classical computer, that the problem of finding the minimum energy state of a protein is NP-hard. That’s true, but a red herring, since the protein isn’t magically going to know the minimum energy state. Though it’s still possible that there’s some catalyst to push it into the right state, so simulating the dynamics in a vacuum won’t get you the right answer (cf prions). Anyhow, there’s some hope that evolution has found a good toolbox for designing proteins and that if can figure out the abstractions that evolution is using, it will all become easy. In particular, there are building blocks like the alpha helix. Certainly an engineer, whether evolution or us, doesn’t need to understand every protein, just know how to make enough.
I think the possibility that a sufficiently smart AI would quickly find an adequate toolbox for designing proteins is quite plausible. I don’t know what Eliezer means, but the possibility seems to me adequate for his arguments.
Ah, that’s helpful.