On 5/25/01 8:24 PM, "Eliezer S. Yudkowsky" <sentience@pobox.com> wrote:
> Spike Jones wrote:
>>
>> This is something Ive wondered about. If someone manages
>> to create a seed AI very early in human history, such as now,
>> this year, the computer infrastructure would be still insufficient
>> to create the SIAI vision of a non-Slow Singularity. Or am I
>> hopelessly confused about something really basic? spike
>
> There's enough computing power on the Internet right now - or, as Eugene
> Leitl and Christine Peterson prefer to call it, "poorly defended"
> computing power - for transhumanity.
This does not follow, and it baffles me that so many relatively
knowledgeable people have simply accepted it. Distributed computing,
*particularly* over slow high-latency links, is totally unsuitable for a
very large percentage of the algoriİhm space (and essentially *every*
plausible AI architecture). For many types of problems you can't
automagically have vast quantities of computational capability simply by
harnessing the Internet. In fact, for most problems you would lose
performance by distributing it over the Internet. And there are many
classes of hard problems (such as protein folding) that are arguably
intractable for all practical purposes even if you had the computational
power of Internet connected computers at your disposal.
Perhaps this misconception is common because very few programmers have
experience with distributed computing and the mathematical/engineering
limitations imposed by it. If what is being posited in this thread (re: AI
singularity via the Internet) was actually possible, silicon cores would
have evolved differently than they have over the last decade or two. I
think most people who are actually knowledgeable about distributed computing
would either 1) find this idea laughable for reasons that would be obvious
to them, or more likely 2) would attach some severe caveats and restrictions
as to the requisite nature and structure of AI that I think most AI
researchers would claim makes the possibility untenable. Note that this is
based on the infrastructure that we have today, and that this applies to the
infrastructure we are likely to have for quite some time.
Spike is right; the existing computer infrastructure today is totally
inadequate for a "fast" takeoff via an Internet AI. Undefended computing
power is absolutely worthless if it is unusable as a distributed computing
resource for most intents and purposes. It should however, be possible with
today's technology to build an environment that can incrementally increase
the net performance of an AI by adding networked resources on an extremely
local basis. A usable basement AI without the ability to take over the
Internet is perfectly feasible and far more realistic, at least for the next
several years.
-James Rogers
jamesr@best.com
This archive was generated by hypermail 2b30 : Mon May 28 2001 - 10:00:08 MDT