Might be about using the accumulated references to seek out and generate more training data (at the currently relevant level, not just anything). Passive perception of random tidbits might take much longer. I wonder if there are lower estimates on human sample efficiency, how long can it take to not learn something when getting regularly exposed to it, for a person of given intelligence, who isn’t exercising curiosity in that direction? My guess it’s plausibly a difference of multiple orders of magnitude, and a different ceiling.
Might be about using the accumulated references to seek out and generate more training data (at the currently relevant level, not just anything). Passive perception of random tidbits might take much longer. I wonder if there are lower estimates on human sample efficiency, how long can it take to not learn something when getting regularly exposed to it, for a person of given intelligence, who isn’t exercising curiosity in that direction? My guess it’s plausibly a difference of multiple orders of magnitude, and a different ceiling.