We should actively look for more informative hypotheses, keeping this one as a last resort or “default” to return to if nothing more informative fits the evidence.
Roko’s heuristic (quoted above) isn’t terrible, but as LW-ers equipped with Bayes’ theorem, we can do better. Instead of betting on whichever explanation predicts the observations in most detail (the “most informative hypothesis”), we can bet on whichever explanation has the most combined predictive power and prior probability.
P(hypothesis | data ) = P(hypothesis) * P( data | hypothesis ) / P(data).
Let’s say we’re trying to decide between theory A—“we like underdogs because underdog-liking was specifically useful to our ancestors, for such-and-such a specific reason”—and theory B—“underdog-liking is an accidental side-effect of other adaptations”. Roko correctly points out that P(data | hypothesis A) is larger than P(data | hypothesis B). That’s what it means to say hypothesis A is “more informative” or has “more predictive power”. (Well, that, and the fact that hypothesis A might also fit some larger set of future data that we might collect in future experiments.) But it is also true that P(hypothesis B) is much larger than P(hypothesis A). And if our goal is to estimate whether smoofra’s general hypothesis B or Roko’s specific hypothesis A is more likely to be true, we need to focus on the product.
We can estimate the relative prior probabilities of hypotheses A and B partly by thinking about how much more general B is than A (general hypotheses have higher priors) and partly by gathering data on how good an optimizer evolution is, or how often evolution generates specific adaptations vs. general side-effects. Regarding how good an optimizer evolution is, Michael Vassar likes to note that adult baboons and human toddlers have to learn how to hide; hiding is useful in a variety of situations, but its usefulness was not sufficient to cause specific “how to hide” adaptations to evolve. If similar examples of missing adaptations are common, this would increase the prior weight against hypotheses such as Roko’s near/far account of underdog-empathy. If there are plenty of clear examples of specific adaptations, that would increase weight toward Roko’s near/fear underdog theory.
Evolutionary psychology is important enough that figuring out what priors to put on specific adaptations vs. side-effects would be darn useful. Anyone have data? Better yet, anyone with data willing to write us a post, here?
Even if the functional hypothesis is less likely than the random hypothesis, there is further we can take it if we explore it. Finding structure can lead to finding more structure, letting us climb further steps up the ladder of knowledge.
Roko’s heuristic (quoted above) isn’t terrible, but as LW-ers equipped with Bayes’ theorem, we can do better. Instead of betting on whichever explanation predicts the observations in most detail (the “most informative hypothesis”), we can bet on whichever explanation has the most combined predictive power and prior probability.
P(hypothesis | data ) = P(hypothesis) * P( data | hypothesis ) / P(data).
Let’s say we’re trying to decide between theory A—“we like underdogs because underdog-liking was specifically useful to our ancestors, for such-and-such a specific reason”—and theory B—“underdog-liking is an accidental side-effect of other adaptations”. Roko correctly points out that P(data | hypothesis A) is larger than P(data | hypothesis B). That’s what it means to say hypothesis A is “more informative” or has “more predictive power”. (Well, that, and the fact that hypothesis A might also fit some larger set of future data that we might collect in future experiments.) But it is also true that P(hypothesis B) is much larger than P(hypothesis A). And if our goal is to estimate whether smoofra’s general hypothesis B or Roko’s specific hypothesis A is more likely to be true, we need to focus on the product.
We can estimate the relative prior probabilities of hypotheses A and B partly by thinking about how much more general B is than A (general hypotheses have higher priors) and partly by gathering data on how good an optimizer evolution is, or how often evolution generates specific adaptations vs. general side-effects. Regarding how good an optimizer evolution is, Michael Vassar likes to note that adult baboons and human toddlers have to learn how to hide; hiding is useful in a variety of situations, but its usefulness was not sufficient to cause specific “how to hide” adaptations to evolve. If similar examples of missing adaptations are common, this would increase the prior weight against hypotheses such as Roko’s near/far account of underdog-empathy. If there are plenty of clear examples of specific adaptations, that would increase weight toward Roko’s near/fear underdog theory.
Evolutionary psychology is important enough that figuring out what priors to put on specific adaptations vs. side-effects would be darn useful. Anyone have data? Better yet, anyone with data willing to write us a post, here?
Even if the functional hypothesis is less likely than the random hypothesis, there is further we can take it if we explore it. Finding structure can lead to finding more structure, letting us climb further steps up the ladder of knowledge.
Exactly. There is a distinction between that hypothesis that you think is most likely, and that hypothesis that you think is most worth pursuing.