RE: guaranteeing friendliness [more about reaching AGI now that Ben has improved the thread]

From: Herb Martin (HerbM@LearnQuick.Com)
Date: Sun Dec 04 2005 - 13:29:20 MST

From: Ben Goertzel
> Hi,

<snipped my suggestion about SETI@HOme>

> This sort of arrangement would not suit my Novamente AI architecture,
> because Novamente requires a lot of processing power, all of which has
> very rapid access to a lot of RAM. It's much better suited to a large
> Linux cluster than a SETI @ Home type arrangement, at least for core
> cognitive processign. However, a widely distributed network a la SETI
> @ Home could potentially be used for some particular kinds of
> Novamente processing, as an additional portion of the underlying
> hardware architecture.

Not, surprising that it requires close clustering.

Perhaps some University student lab or other pools of computers
can be borrowed on occasion for testing.

Presumably once you have it up, it could sleep during the
hours when the computers are otherwise occupied.

But, the 500 x $1000 (or a few thousand) for hardware
doesn't seem that tough compared to your other expenses
(when you are ready).

> I estimate that with a 3-5 year development effort by a team of 5-8
> expert AI programmer/scientists (like my current Novamente development
> team), we could have a human-level "artificial child", running on a
> ~200-500 PC cluster. Another 5-10 years could viably yield an
> adult-level "artificial scientist" potentially able to launch a
> Singularity.

Does the above presume (any of) the following:

        1) The approximate 1000 fold increase in computing power
                (at any particular price point) that 15 years may
        2) Appreciably more RAM per PC than is common today?
                (1 to a few GB are now common) or perhaps
                must faster (and larger) disk drives which
                perform well enough to substitute for current

        3) A large number of "breakthoughs" in AI technique?

        4) Any significantly "hard" breakthroughs?

        5) Adult-level "scientist" AI is also a "programmer"?
                (i.e., able to program and learn how to program

While I see some required breakthroughs (e.g., "programmmer AI"
is very rudimentary today compared to NLP), it seems these
breakthroughs will actually ADD to the processing needs and
modules such as the NLP or perhaps neural nets and EA methods
which are fairly well understand will benefit mostly from
much more RAM and processing power, along with increased
'common knowledge' stored and processed there.

> Unfortunately, though, my development team and I are currently
> spending the bulk of our time on commercial narrow-AI consulting
> projects that pay the bills, so this "3-5 year development effort" can
> basically only start if/when we achieve a reasonable amount of
> funding. Which is not an easy thing to do for a project of this sort.

<sigh> Sympathies extended.

This is one of the reasons for my own activities being
VERY limited.

Herb Martin

This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:00:54 MDT