Chris Olah · Anthropic · 2021-01-09
Olah confronts a problem most science education ducks: how do you actually learn to pick good problems? His answer is a set of concrete drills — predict what a paper will find from its abstract, rank ideas before testing them, debate the merits of directions you don't believe in. Taste, in his telling, is not innate; it's a calibrated muscle.
This is essential reading for any learner who wants to do research, not just consume it. The piece is short, structured, and unusually honest about how slow taste-building is. Pair it with the circuits work to see one of the field's most discerning researchers reason about his own decision-making at the meta level.
> The fundamental issue is that actually testing whether a research idea is good is very expensive.