Re: Let's make an A.I.!

[steve donovan <steve.j.donovan@...>]

On Mon, Nov 3, 2014 at 9:32 PM, Stefan Reich
<stefan.reich.maker.of.eye@googlemail.com> wrote:
> How do we make an A.I.? It's actually easy.

Ah, but to make it useful and *not irritating* is hard!  Once I wanted
to do a natural-language shell (in Lua of course) and quickly realized
a few things. One, there is a reason why the classic shells are so
terse - they are the original SMS speak [1]. Most people can't type as
fast as they can think. Two, there has to be a conversation involved,
resolving any ambiguity - this is how humans do it after all.  Three,
the agent has to have intimate knowledge of what you're doing on your
computer

The first point is helped by improvements in voice recognition (e.g.
it's an *offline* activity on modern Android) and then the second
point is important, because the agent must present what it has heard,
and find out exactly what the user wants. The last point requires a
fair amount of OS integration - say the user asks "what was that
website about cats?"; "open the spreadsheet I was working on last
Tuesday".  So all apps need to have very open and scriptable APIs.
The agent would also browse things like shell history.

The exciting bit is getting the agent to learn, and to be taught -
"verbal scripting".

We'll see this kind of thing emerging, but it's most likely to be a
Google agent, reporting to Google ;)

[1] the VMS shell/scripting language was an exception - very pleasant
to read scripts afterwards.