Human-level _economic_ crossover date

From: Dani Eder (
Date: Wed Apr 11 2001 - 15:50:11 MDT

I'd like to discuss the idea of an economic crossover
in processor power.

In agriculture and industry, machines have replaced
muscle power when the machine+ operator can do the
work cheaper. For example, a backhoe plus operator
costs $40 per hour to run, and does something like
20 times as much work as a man with a shovel. In
the US, you can't hire manual diggers for $2/hr,
so the backhoe option is cheaper. In countries
where labor is cheap ($2/hr for backhoe operator
and $1/hr for manual digger), the comparison
becomes $32/hr for backhoe to $20/hr for 20 laborers,
so you continue to do manual ditch digging.

In the case of my own job, software testing for the
Space Station, assume that a 3000 Tflop processor
(i.e. one human brain equivalent by my estimate)
can automate 95% of my job. So instead of 20 test
engineers and one supervisor, you would have just
the supervisor and a lot of processors. This is
analogous to the backhoe example where one operator
replaces 20 manual laborers. My salary and benefits
and overhead is in the range of $120K per year.
Assume that a processor can run 24x7, and thus
does the work of 4 humans working 40 hour weeks.
So the one processor can do $480K per year of
work. Assume that you want a 16% return on capital
(= 6 year payback time). So you can afford to
spend $3 million on your processor. Therefore
the break-even cost to replace me is $1/Gflop.

Now, don't think of it as necessarily putting me
out of a job, but rather augmenting me so I can
do the work of many engineers. Even without a
superintelligent AI, once you can augment an
engineer and vastly increase his productivity
for less than it costs to simply hire more engineers,
you can get a Singularity-type effect. It might not
be as dramatic as a pure AI driven singularity.
The rate of adaption of the operators will limit
things some. But it would be some type of knee in
the curve of 'effective engineering capacity'.

11 months ago a 60 Gflop Beowulf cluster built with
Athlon processors ran $640/Gflop. Today it can
be built for $335. That's system cost cut in half
in a year. At that rate you hit the $1/Gflop level
in 2009. I don't know what level of processor
power it would take to replace the average engineer.
In some areas (spreadsheets, circuit layout) it's
already happened. I know as a power Google user,
I can research a subject vastly faster than I used
to by having to go to the technical library and
browsing the shelves, or even tracking down the
company expert and picking his brains. The time
it takes me to learn is the same, but getting the
info to me is vastly faster. So I've had productivity
increases in some areas already.

I've got this funny feeling that we're already on
a 1 year doubling time track for processors, and
as the circuit designers, lithography machine
designers, etc. get augmented by computer tools, the
pace will even speed up a bit from that in the next

So I look for an effective singularity point around
2010. It's where the curve of overall progress
in all technical areas takes on the doubling time
we've seen in the electronics industry. It's where
the cost of everything starts dropping like the
cost of computers. Maybe it's not the singularity
in the sense we've used it in the past (where we
can't see beyond it). In analogy to black holes,
it's more like the point at which spacetime begins
to get seriously warped, rather than the event
horizon. In Vinge's fictional terminology, the
near edge of the Low Beyond, rather than the
High Beyond. Maybe one of you can come up with
a better term. But it's definitely an inflection
point in the curve of technological progress.


Do You Yahoo!?
Get email at your own domain with Yahoo! Mail.

This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:00:36 MDT