I roughly agree, though also have complicated thoughts on all of these. For example, I think there is a good case to be made for live/dead to be binary, or at least to have a pretty sharp transition between the two (in general impact on various scales appears to be heavy-tail distributed, which suggests that you often have increasing marginal returns in competence). I think in general coding things as binary variables is a reasonable first thing to do, and is usually surprisingly accurate (i.e. categorizing charitable interventions as “plausible effective” and “ineffective” gets you 90% of the variance or so, because the best interventions we can identify are very likely many orders of magnitudes better than the average one, meaning you should basically just ignore the average one). I also remember some research on linear classifiers that were limited to just coding weights as +1 or −1 and them almost achieving the same performance as general linear classifiers (while being much much simpler).
My general thoughts on the post is that I am excited about the model it is proposing, but would prefer a bit more explicitness about the degree to which the model is a toy model, and the degree to which it serves as an intuition pump vs. a comprehensive model, and more analysis on how it compares to other models. But if I have to trade off between those and getting more of the model written up at all, I would probably prefer to get more of the model written up.
Fair enough! My own reaction is different, as you’ll have guessed: my intuition/prejudices don’t make me optimistic about the model being proposed, and there are enough things in the presentation that just seem wrong-headed (and little enough that feels like it’s providing valuable insight) that I don’t feel any particular appetite for more.
But, of course, criticism is easier, and a less scarce resource, than original thought—and my gut reactions could very easily be badly wrong. (So, in particular, my comments aren’t intended to persuade the OP not to post more.)
I take your point about heavy-tailed distributions, but remark that in so far as that’s an accurate description we need to draw the live/dead boundary between *the very most original agents* and *everyone else*, which I don’t think is the OP’s intent—at any rate, I find that hard to square with judging Russia “live” merely because they did something that plenty of other countries do and “is merely new for modern-day Russia”, or with putting the boundary between Jobs-era Apple and Cook-era Apple.
I roughly agree, though also have complicated thoughts on all of these. For example, I think there is a good case to be made for live/dead to be binary, or at least to have a pretty sharp transition between the two (in general impact on various scales appears to be heavy-tail distributed, which suggests that you often have increasing marginal returns in competence). I think in general coding things as binary variables is a reasonable first thing to do, and is usually surprisingly accurate (i.e. categorizing charitable interventions as “plausible effective” and “ineffective” gets you 90% of the variance or so, because the best interventions we can identify are very likely many orders of magnitudes better than the average one, meaning you should basically just ignore the average one). I also remember some research on linear classifiers that were limited to just coding weights as +1 or −1 and them almost achieving the same performance as general linear classifiers (while being much much simpler).
My general thoughts on the post is that I am excited about the model it is proposing, but would prefer a bit more explicitness about the degree to which the model is a toy model, and the degree to which it serves as an intuition pump vs. a comprehensive model, and more analysis on how it compares to other models. But if I have to trade off between those and getting more of the model written up at all, I would probably prefer to get more of the model written up.
Fair enough! My own reaction is different, as you’ll have guessed: my intuition/prejudices don’t make me optimistic about the model being proposed, and there are enough things in the presentation that just seem wrong-headed (and little enough that feels like it’s providing valuable insight) that I don’t feel any particular appetite for more.
But, of course, criticism is easier, and a less scarce resource, than original thought—and my gut reactions could very easily be badly wrong. (So, in particular, my comments aren’t intended to persuade the OP not to post more.)
I take your point about heavy-tailed distributions, but remark that in so far as that’s an accurate description we need to draw the live/dead boundary between *the very most original agents* and *everyone else*, which I don’t think is the OP’s intent—at any rate, I find that hard to square with judging Russia “live” merely because they did something that plenty of other countries do and “is merely new for modern-day Russia”, or with putting the boundary between Jobs-era Apple and Cook-era Apple.