« What Insight Literature? | Main | The Halo Effect »

November 29, 2007


If you are asked to estimate a number that is a product (or sum) of many numbers, and you have good estimates for all those numbers but one, well variance in that last number you can't estimate well will dominate the variance of your answer. It just takes one.

I strongly encourage any AI worker who hasn't already done so to read Ian McDonald's 'River of Gods'. He's pretty positive (in timescale terms...) on AI, his answer to the question "How long will it be until we have human-level AI?" is 2047 AD, and it's a totally gob-smacking, brilliant, read.

Derrida must have done a thousand essays on how an author trying to be very precise about how language could possibly work, winds up in an infinte loop clarifying a final point that amounts to in effect starting over.

This contributes a lot to an indefinite future, whatever the modulus problem, if you take AI as just such a project.

I observe that many futuristic predictions are, likewise, best considered as attitude expressions. Take the question, "How long will it be until we have human-level AI?" The responses I've seen to this are all over the map. On one memorable occasion, a mainstream AI guy said to me, "Five hundred years." (!!)

Did you ask any of them how long they felt it would take to develop other "futuristic" technologies? (in other words, their rank ordering of technological changes).

The damages experiment, as described here, seems not to nail things down enough to say that what's going on is that damages are expressions of outrage on a scale with arbitrary modulus. Here's one alternative explanation that seems consistent with everything you've said: subjects vary considerably in their assessment of how effective a given level of damages is in deterring malfeasance, and that assessment influences (in the obvious way) their assessment of damages.

(I should add that I find the arbitrary-modulus explanation more plausible.)

The comments to this entry are closed.

Less Wrong (sister site)

May 2009

Sun Mon Tue Wed Thu Fri Sat
          1 2
3 4 5 6 7 8 9
10 11 12 13 14 15 16
17 18 19 20 21 22 23
24 25 26 27 28 29 30