It might be worth separating the claim “Eliezer is wrong about what changes he, personally, should try” from the claim
“It is generally good to try many plausible changes, because:
Some portion will work;
Trying the number of approaches it takes to find an improvement is often less expensive than being stuck in the wrong local optimum;
Many of us humans tend to keep on doing the same old thing because it’s easy, comfortable, safe-feeling, or automatic, even when sticking with our routines is not the high-expected-value thing to do. We can benefit from adopting heuristics of action and experimentation to check such tendencies.”
The second claim seems fairly clearly right, at least for some of us. (People may vary in how easily they can try on new approaches, and on what portion of handed-down approaches work for them. OTOH, the ability to easily try new approaches is itself learnable, at least for many of us.) The first claim is considerably less clear, particularly since Eliezer has much data on himself that we do not, and since after trying many hacks for a given not-lightcone-destroying problem without any of the hacks working, expected value calculations can in fact point to directing one’s efforts elsewhere.
Maybe we could abandon Eliezer’s specific case, and try to get into the details of: (a) how to benefit from trying new approaches; and (b) what rules of thumb for what to try, and what to leave alone, yield high expected life-success?
It might be worth separating the claim “Eliezer is wrong about what changes he, personally, should try” from the claim
“It is generally good to try many plausible changes, because:
Some portion will work;
Trying the number of approaches it takes to find an improvement is often less expensive than being stuck in the wrong local optimum;
Many of us humans tend to keep on doing the same old thing because it’s easy, comfortable, safe-feeling, or automatic, even when sticking with our routines is not the high-expected-value thing to do. We can benefit from adopting heuristics of action and experimentation to check such tendencies.”
The second claim seems fairly clearly right, at least for some of us. (People may vary in how easily they can try on new approaches, and on what portion of handed-down approaches work for them. OTOH, the ability to easily try new approaches is itself learnable, at least for many of us.) The first claim is considerably less clear, particularly since Eliezer has much data on himself that we do not, and since after trying many hacks for a given not-lightcone-destroying problem without any of the hacks working, expected value calculations can in fact point to directing one’s efforts elsewhere.
Maybe we could abandon Eliezer’s specific case, and try to get into the details of: (a) how to benefit from trying new approaches; and (b) what rules of thumb for what to try, and what to leave alone, yield high expected life-success?
One more reason for the list is that doing new stuff (or doing stuff in new ways, but I repeat myself) promotes neurogenesis.