AI timeframes

From: Dani Eder (
Date: Wed Apr 07 2004 - 12:12:00 MDT

> > I'd be surprised if AGI is developed in less than
> 30
> > years. I'm looking at a 50 year time-frame
> actually.

Ben G. said:

> In the end she agreed that 8-10 years was plausible
> and 15 years was
> fairly likely -- assuming the focused team of 5-7
> people really has
> total focus on the project, and assuming hardware
> continues to advance
> as projected, making the running of multiple

It's the 40th anniversary of IBM's introduction of
the first general purpose mainframe family, the 360
series. It's worth considering what made such
mainframes so popular and predicting how AI may
follow a similar path.

In the first half of the 20th century, a computer
was generally a young lady who operated a mechanical
calculator. There were legions of them working in
industries like banking and insurance that required
a lot of arithmetic to be done. There were also
large numbers of them working in payroll and
accounting departments of most companies. Prior
to the 360 series, electronic computers were
generally special purpose. The 360 series had
a consistent machine code and OS across a range
of models, and the machines were capable of running
a wide range of programs. They were expensive,
but less expensive than the large number of human
computers they replaced.

Similarly, there are software systems that have
been developed to do specific tasks, such as
circuit layout or flagging unusual credit card
transactions. These are hand crafted for the
specific task. Software systems that are trainable
for a wide variety of tasks are not yet available.

Once hardware is available that is adequate for
the job at a competitive cost, there will be a
strong incentive for the software to be developed.
For example, a voice recognition call center
package that can be trained much as the humans
who do the job to reply to product-specific
questions might sell for a lot of money if it
replaces dozens or hundreds of humans.

The more generally trainable a system is, the
more potential customers there are, and thus the
amount that can be invested in developing it can

I estimate sufficiently powerful hardware needs
to be in the range of $30K to $30 per teraflop,
or a factor 5 to 5000 less than at present, or
4 to ~20 years of continued cost reductions for
computers at recent rates, plus 5-10 years for
development of 'trainable software' of increasing
generality once the financial incentive exists.

Thus I project a range of 10-30 years in total from


Do you Yahoo!?
Yahoo! Small Business $15K Web Design Giveaway

This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:00:46 MDT