I mostly agree with the post, but I think it’d be very helpful to add specific examples of epistemic problems that CFAR students have solved, both “practice” problems and “real” problems. Eg., we know that math skills are trainable. If Bob learns to do math, along the way he’ll solve lots of specific math problems, like “x^2 + 3x − 2 = 0, solve for x”. When he’s built up some skill, he’ll start helping professors solve real math problems, ones where the answers aren’t known yet. Eventually, if he’s dedicated enough, Bob might solve really important problems and become a math professor himself.
Training epistemic skills (or “world-modeling skills”, “reaching true beliefs skills”, “sanity skills”, etc.) should go the same way. At the beginning, a student solves practice epistemic problems, like the ones Tetlock uses in the Good Judgement Project. When they get skilled enough, they can start trying to solve real epistemic problems. Eventually, after enough practice, they might have big new insights about the global economy, and make billions at a global macro fund (or some such, lots of possibilities of course).
To use another analogy, suppose Carol teaches people how to build bridges. Carol knows a lot about why bridges are important, what the parts of a bridge are, why iron bridges are stronger than wood bridges, and so on. But we’d also expect that Carol’s students have built models of bridges with sticks and stuff, and (ideally) that some students became civil engineers and built real bridges. Similarly, if one teaches how to model the world and find truth, it’s very good to have examples of specific models built and truths found—both “practice” ones (that are already known, or not that important) and ideally “real” ones (important and haven’t been discovered before).
Example practice problems and small real problems:
Fermi estimation of everyday quantities (e.g., “how many minutes will I spend commuting over the next year? What’s the expected savings if I set a 5-minute timer to try to optimize that?);
Figuring out why I’m averse to work/social task X and how to modify that;
Finding ways to optimize recurring task X;
Locating the “crux” of a disagreement about a trivia problem (“How many barrels of oil were sold worldwide in 1970?” pursued with two players and no internet) or a harder-to-check problem (“What are the most effective charities today?”), such that trading evidence for the crux produces shifts in one’s own and/or the other player’s views.
Larger real problems: Not much to point to as yet. Some CFAR alums are running start-ups, doing scientific research for MIRI or elsewhere, etc. and I imagine make estimates of various quantities in real life, but I don’t know of any discoveries of note. Yet.
I’ve learned useful things from the sequences and CFAR training, but it’s almost all instrumental, not epistemic. I suppose I am somewhat more likely to ask for an example when I don’t understand what someone is telling me, and the answers have occasionally taught me things I didn’t know; but that feels more like an instrumental technique than an epistemic one.
Before and after prediction market performance jumps to mind and is easy, though doesn’t cover the breadth of short feedback topics that would be ideal.
I mostly agree with the post, but I think it’d be very helpful to add specific examples of epistemic problems that CFAR students have solved, both “practice” problems and “real” problems. Eg., we know that math skills are trainable. If Bob learns to do math, along the way he’ll solve lots of specific math problems, like “x^2 + 3x − 2 = 0, solve for x”. When he’s built up some skill, he’ll start helping professors solve real math problems, ones where the answers aren’t known yet. Eventually, if he’s dedicated enough, Bob might solve really important problems and become a math professor himself.
Training epistemic skills (or “world-modeling skills”, “reaching true beliefs skills”, “sanity skills”, etc.) should go the same way. At the beginning, a student solves practice epistemic problems, like the ones Tetlock uses in the Good Judgement Project. When they get skilled enough, they can start trying to solve real epistemic problems. Eventually, after enough practice, they might have big new insights about the global economy, and make billions at a global macro fund (or some such, lots of possibilities of course).
To use another analogy, suppose Carol teaches people how to build bridges. Carol knows a lot about why bridges are important, what the parts of a bridge are, why iron bridges are stronger than wood bridges, and so on. But we’d also expect that Carol’s students have built models of bridges with sticks and stuff, and (ideally) that some students became civil engineers and built real bridges. Similarly, if one teaches how to model the world and find truth, it’s very good to have examples of specific models built and truths found—both “practice” ones (that are already known, or not that important) and ideally “real” ones (important and haven’t been discovered before).
Example practice problems and small real problems:
Fermi estimation of everyday quantities (e.g., “how many minutes will I spend commuting over the next year? What’s the expected savings if I set a 5-minute timer to try to optimize that?);
Figuring out why I’m averse to work/social task X and how to modify that;
Finding ways to optimize recurring task X;
Locating the “crux” of a disagreement about a trivia problem (“How many barrels of oil were sold worldwide in 1970?” pursued with two players and no internet) or a harder-to-check problem (“What are the most effective charities today?”), such that trading evidence for the crux produces shifts in one’s own and/or the other player’s views.
Larger real problems: Not much to point to as yet. Some CFAR alums are running start-ups, doing scientific research for MIRI or elsewhere, etc. and I imagine make estimates of various quantities in real life, but I don’t know of any discoveries of note. Yet.
I’ve learned useful things from the sequences and CFAR training, but it’s almost all instrumental, not epistemic. I suppose I am somewhat more likely to ask for an example when I don’t understand what someone is telling me, and the answers have occasionally taught me things I didn’t know; but that feels more like an instrumental technique than an epistemic one.
Before and after prediction market performance jumps to mind and is easy, though doesn’t cover the breadth of short feedback topics that would be ideal.