[extropy-chat] RE: [wta-talk] SIAI seeking seed AI programmer candidates

[Adrian Tymes]

--- Eugen Leitl <eugen at leitl.org> wrote:
> On Wed, Jun 02, 2004 at 12:58:05AM -0700, Adrian
> Tymes wrote:
> > and hardware, as well, and of course any (decent)
> > programmer knows math.  But we know an AI will
> require
> 
> So you have a Ph.D. in three different disciplines
> (that be neuroscience,
> computational and/or wet), practical hardware design
> (at least at FPGA level,
> preferrably ASIC), have practical experience with
> high-performance numerics 
> codes >kNode country, and can handle complex
> nonlinear dynamics?

Restoring humility: only a single Master's
("computational neuroscience" would be an accurate
description; the term UCLA used was "biocybernetics"),
I've designed ASICs before (that was my first real
job, in fact, as an intern one summer during high
school), and I've built distributed systems designed
to run on >1000 nodes (although the dot-bomb I built
it for crashed before it actually set up that many
nodes).  "Complex nonlinear dynamics" is vague enough
that I can say I can handle it, though you may have
something more specific in mind.

> And you're only taking 100 k$/year? Wow, you're a
> real bargain.

I haven't found any jobs that use everything that I
know.  I doubt I will, at least in the next several
years.

> > new software to be written (i.e., programming); it
> is
> 
> You can assume that those scientists and engineers
> can program.

No, just that programmers will be needed.  It'd
likely be best if the scientists and engineers could
program, though - one might even call it practically
necessary, given the amount of information that would
otherwise have to be extracted from the
non-programmers.  I can conceive of that scenario
working in theory...but, again, the practical problems
would probably make it easier just to find scientists
and engineers who have (or quickly acquire) the
necessary programming skills.

> High-performance parallel numerics, that is, which
> most programmers have
> absolutely no clue of.

This is true.  The tricks of breaking up a task into
what each CPU can handle versus the overhead of
communication are taught in most college Computer
Science cirricula to my knowledge, but it's one of
those tricks that never seems useful until you need to
do it - and it's certainly a skill most self-taught
programmers might never run across while learning the
trade.

> > not as easy to prove that new hardware, or new
> > statistical methods, will have to be custom
> developed
> > for this project.  (Possible, perhaps.  Easy, no,
> 
> What can I say, people with that attitude have been
> failing for past 40-50
> years. You're familiar as to why they've been
> failing, as you've studied
> those failures extensively, right?

I wouldn't say extensively, but I have studied
previous AI attempts enough to know the patterns
behind several of their failures.  I certainly agree
that the lack of serious consideration of the brain's
actual infrastructure - which is not primarily a
programming task - has been a serious stumbling block.
Which is why it can definitely be argued that new
hardware might have to be invented for the purpose of
making AIs.  But that's only a "might", and even that
hardware will need software.

> > especially when one considers that "custom
> developed"
> > rules out hiring people to continue implementing
> > Moore's Law when other companies are already doing
> so
> 
> Moore's law describes *integration density*. Given
> that you're an
> accomplished numerics guy, I wonder why you never
> heard about benchmarks.
> Such as, memory bandwidth? Unpredictable access? 

I was using ML as one (very) familiar example.  There
are similar, if less pronounced, curves for other
important metrics such as memory bandwidth...and, I'll
admit, one might say that some of those curves are
flat enough that special development efforts to ramp
those metrics up to the needed levels might be
required.  All I'm saying is that need is not as
certain as the fact that new software will be needed,
no matter what path winds up being taken.

> There are a few handy graphs on the web, and they're
> unfortunately no linear
> semi-log plots. I stopped posting links, because
> nobody reads them anyway.

I think I may have read the graphs you're thinking of.