Oh, this gives me an idea on what you could do instead of random groupings: Have some simple learning algorithm divide everyone into some fuzzy set of categories given ONLY the data abut how good they rank previous pieces of advice. Just have some highly ranked standard pieces that new people are given at the start and group people for whom the same ones worked in the same group. Then when a new piece is entered into the system it tests it against all the different groups and tend to give it to those in the same groups as the ones it worked for in the original test.
The “pices” could be different things: quotes, advice, statements about your personality, predictions, whatever.
One thing that’s important to note is that you can be in any number of groups; A “rank maximally good” rock would be placed in ALL groups. The groups would not be anything made to mimic some specific human word but just unsupervisedly lerned from the raw data. In practice the groups might turn you to correlate to things like “optimistic” and “likes quotes” but also things like”guillable” or “wont actually follow the advice just votes based of what sounds good”. Oh, and while I’m saying things like being “in” or “not in” a group it obviusly shuldnt be a binary thing just a probablility of geting pices belonging to that group.
I was thinking something similar. Kind of like non-mutually-exclusive, dynamically-assigned star signs based on what you find useful.
That does also suggest that you could use the system prescriptively instead of simply descriptively. If it places you in the “talented slacker” category, and you’d rather be in the “fastidiously disciplined” group, you could opt to receive the Fastidiously Disciplined horoscope, and try to change your working habits to facilitate the Fastidiously Disciplined advice.
It’s not obvious to me how to do this and still gather information from the user without interfering with the scores for their chosen category. Having them guess after the fact how well the “slacker” horoscope would have worked for them seems clearly sub-optimal, especially since there’s an obvious pressure for them to say that it wouldn’t’ve.
I bet it would be useful to sort people by “what do you most want to improve about yourself?” It seems every LWer has at least one thing (and some, many).
People who choose “nothing” would end up getting horoscopes centering around Dunning-Kruger, confirmation bias, etc.
Well, they wouldn’t be labelled with meaningful English titles, but you could give them arbitrary names for ease of reference. A bit like actual star signs, only empirically informed.
Oh, this gives me an idea on what you could do instead of random groupings: Have some simple learning algorithm divide everyone into some fuzzy set of categories given ONLY the data abut how good they rank previous pieces of advice. Just have some highly ranked standard pieces that new people are given at the start and group people for whom the same ones worked in the same group. Then when a new piece is entered into the system it tests it against all the different groups and tend to give it to those in the same groups as the ones it worked for in the original test.
The “pices” could be different things: quotes, advice, statements about your personality, predictions, whatever.
One thing that’s important to note is that you can be in any number of groups; A “rank maximally good” rock would be placed in ALL groups. The groups would not be anything made to mimic some specific human word but just unsupervisedly lerned from the raw data. In practice the groups might turn you to correlate to things like “optimistic” and “likes quotes” but also things like”guillable” or “wont actually follow the advice just votes based of what sounds good”. Oh, and while I’m saying things like being “in” or “not in” a group it obviusly shuldnt be a binary thing just a probablility of geting pices belonging to that group.
I was thinking something similar. Kind of like non-mutually-exclusive, dynamically-assigned star signs based on what you find useful.
That does also suggest that you could use the system prescriptively instead of simply descriptively. If it places you in the “talented slacker” category, and you’d rather be in the “fastidiously disciplined” group, you could opt to receive the Fastidiously Disciplined horoscope, and try to change your working habits to facilitate the Fastidiously Disciplined advice.
It’s not obvious to me how to do this and still gather information from the user without interfering with the scores for their chosen category. Having them guess after the fact how well the “slacker” horoscope would have worked for them seems clearly sub-optimal, especially since there’s an obvious pressure for them to say that it wouldn’t’ve.
I bet it would be useful to sort people by “what do you most want to improve about yourself?” It seems every LWer has at least one thing (and some, many).
People who choose “nothing” would end up getting horoscopes centering around Dunning-Kruger, confirmation bias, etc.
Yea. Only problem is the groups wouldn’t be labelled since they were autonomously discovered and and thus finding it would be a bit hard.
Well, they wouldn’t be labelled with meaningful English titles, but you could give them arbitrary names for ease of reference. A bit like actual star signs, only empirically informed.
yea, that’d probably be a good idea.