::scr on agents[0] was on bots

Simon Wistow scr@thegestalt.org
Tue, 12 Feb 2002 11:31:40 +0000


On Sat, Feb 09, 2002 at 04:41:56PM +0000, jo walsh said:
> i guess a speech output interface is nice especially for me because it is
> easy to multitask through; just switching on cognition partially and
> letting the brain check for highpoints and patterns, like the news on the
> radio.

This is why I initially did it - I'd mooch around at home occasionally
doing stuff or watching TV or films or reading and would let the drone
of channel play through an old laptop. I have thought about rigging up
some sort of Genlock card on an old box, playing around with a canvas
library and letting the text scroll up the screen. But I haven't yet.

Ircle on the Mac (ISTR) let's you set a Regex and it will speak lines
containing it. Useful because you can set it to only speak when your
name is mentioned.


> [ bots as an interface ]

% cat /dev/braindump

This kind of leads on from what Mr Nullpointer was talking about at
Dorkbot. I mentioned this in passing to my boss this morning and we
ended up having a long conversation in which it transpired that he'd
done some work on this during his time in the Merkan military.

Nullpointer san demonstrated some visualisations that he'd done of
music. The difference between these and, say, WinAmp or iTunes
visualisations were that these reacted to the notes and instruments as
they were being played through Midi rather than reacting to the sound
wave and some fourier calculations.

As Piers, who is conveniently on this list and who was also in the
audience, pointed out - what makes WinAmp vis-s so effective is that
your brain picks out the patterns in the swells and fades andspins and
blurs and matches them against the music. Whilst Nullpointer's stuff was
inarguably better when played side by side with a Winamp vis the Winamp
vis, seen in isolation, looks almost as good.

This filling in, this pattern matching, this willingness to diregard the
wrong and accept the right is what makes horoscopes work. You match the
random vagueness against something in your life - "you are going ona
journey", "OMIGOD! I *did* come in on the tube this morning" - and
ignore the rest - "You will meet a tall dark stranger" "tum, te tum te
tum [skip]".

The Eliza bot works in much the same way. Except that it reacts to you
by picking out words and constructing a sentence. Even though it's
gibberish sometimes you cope with that - especially if you think it's a
person. See http://fury.com/aoliza/ for a stunning example of this.

Eliza is remarkably unsophisticated. The Perl module that implements it
is only 1295 lines of code of which 550 is comments and documentation
and nearly 600 is the data used to do the matches.

By usig this simplicity, getting 90+% of the results from 10% of the
effort of actual natural language can be used to make a interface for
the desktop.

90% of the time it will work brilliantly. And the rest of the time the
user will forgive 

"sorry, simon, I don't understand you"

a lot more than "error 16545 : command executable not present in varargs
tree" </technical mumbo jumbo style="tongue in cheek">

And like the fact that someone who speaks baby english can speak to
someone who eloquent English and the transition between the two states
is seamless users can become powerusers without the dreaded paradigm
switch. 

Not only that but a novice user will be able to do things that they
would have had problems doing under a more conventional system.

For example 

The find command

% find / \( -name a.out -o -name core \
 -o -name '#*#' \) -type f -atime +14 \
 -exec rm -f {} \; -o -fstype nfs -prune

could be typed in as 

% dipsy, find files named a.out, core or like '#*#' that are at keast 2
weeks old, and aren't on an NFS file system, and delete them

which wouldn't be that hard (ignoring the fact that if the natural
language 'parser' misparsed it then you could end up deleting lots of
stuff you didn't want to delete))

Or even in combination with the GUI - "find files like this [clicks on a
file]"

The thing I particularly like is using the bot/agent/whatever as a
coherent interface to apps as well as a knowledge repository - "dipsy,
add buy sugar to my todo list". The nice thing would be that you could
access the bot in a number of different ways - theoretically you could
even phone home and ask you bot stuff - "dipsy, what was on my todo list
again?" - or sms it or something.


[intermission]

Finally, what I find most interesting, and a little disapointing, about
watching people interface with the bots on #london.pm is that despite
the fact that they're designed to work around people, to appear human
and to learn silently - 

<muttley> hey! cool, just found out
<muttley> netscape 2.0 is at
          ftp://archive.netscape.com/archive/index.html 
<muttley> time to go download ...

	[time passes, seasons change, governments fall]

    <JUM> Whenever I set it to not autoload images with
          Netscape 2.01, the whole program locks up. 
          Anyone know why?
 <Irving> no
    <JUM> Does anyone know where I can get Netscape 2.0????
    <url> i think netscape 2.0 is at
          ftp://archive.netscape.com/archive/index.html
    <JUM> I am forever grateful, Url.
    <JUM> Url: Are you running ver 2.01 with success?
    <url> jum: bugger all, i dunno
    <JUM> OK.
    <JUM> Thanks, Url
    <url> de nada, jum



but now I see people (Mr Mison, I'm looking at you although I'm
just as bad) changing their usage patterns to adapt to the bot. Partly
this is because you have to work round some limitations but the end
result is that where, before, dipsy sounded natural because (s)he was
seeded with natural input (kind of like how a Markov chain works) now it
comes across as an artificial construct. Which is kinda sad.

Umm, that's all. I'm spent. Hopefully this will give enough food for
thought that some of you fuckers will pipe up :)


Simon









-- 
: as seen on tv