[ExI] Universally versus 'locally' Friendly AGI
spike66 at att.net
Tue Mar 8 22:36:35 UTC 2011
... On Behalf Of Samantha Atkins
>> considered as AGI unless they exhibit the kind of human-like behaviour
>> allowing them to compete with the Real Thing in Turing tests.
>...This must be tongue in cheek. A true AGI may very well flunk the Turing
Test. Not by being too stupid or being less capable but by being unwilling
to dumb itself down to such an asinine level. I doubt very much that
passing as human will be high on the priority list. - Samantha
Good point, I hadn't thought of it. AGI might fail the Turing test even
though it is well beyond human intelligence because it is not particularly
good at the specific skill of imitating human foibles.
If we want to use chess as analogous to intelligence, consider the
intentionally crippled chess software. For an example of intentionally
crippled-ware, look at that which likely came with your latest computer if
it has Microsloth: Chess Titans. It is intentionally dumbed down so that it
makes it fun for us mere mortals. On its highest setting it is playing
about a middle to high B rated chess, perhaps 1700 to 1750-ish Elo I would
estimate. I can beat it, but I need to pay attention and show some respect.
Chess Titans is not as strong on a modern computer as a freeware chess
program for a palm pilot was 10 years ago.
Modern uncrippled chess software will not pass the Turing test. It is too
consistent. I can tell it from a grandmaster: it doesn't necessarily always
find the best move, but it is extremely consistent in picking one of the
best four. Even the best humans seldom go a whole game picking only one of
the best four.
If that is analogous to chess software, we could have AGI which passes the
Turing test temporarily, then cannot any longer.
More information about the extropy-chat