[ExI] Hard Takeof

Keith Henson hkeithhenson at gmail.com
Mon Nov 15 00:27:33 UTC 2010


Michael wrote:

>> Prediction: most comments in response to this post will again ignore the
>> specific points in favor of a rapid takeoff and simply dismiss the idea
>> based on low intuitive plausibility.

Yep.

I think "Hard takeoff" and "Rapid takeoff" are pretty much the same
thing, set by human perception.

And even if the doubling time didn't speed up (which it certainly
could) a doubling time of a day or less is probably beyond human
ability to even understand what is happening, especially if the AI
were moderately sneaky.

Some years ago there was a very compact virus that infected (as I
recall) Microsoft SQL servers.  It fit in a packet under 500 bytes.
Once a machine had received one it was zombified and immediately
started sending copies of the virus packet to random IP addresses.
There were (as I recall) only about 50,000 possible targets on the
net.  All were infected in a short time.  The doubling time (again
from memory) was 8.5 seconds.  At this rate, it would have taken under
3 minutes.

The infection peaked out (clogging the net) before anyone could have reacted.

If you had an AI that infected PCs this fast to get procession power,
an AI takeoff could be over before people woke up to what was
happening.

It's a different situation where someone is manufacturing AIs for some
purpose such as the clinic in "The Clinic Seed."  In that case the AI
had been constructed with roughly human motivations, where the AI's
motivational goal was to obtain the high opinion of humans and others
if its kind.  The AI's population would increase at the rate set by
the factory.

This doesn't contribute much to your sound points about AI takeoff,
but the first is an example of what has happened and how short the
timetable might be.

Give my best to Eliezer

Keith



More information about the extropy-chat mailing list