Next: A sociological look Up: A prototypical agent: Previous: More of Julia:
While Julia is usually a fair simulation of a personality, at least on a superficial inspection, in non-Turing mode (more about this later), she does reveal certain machine-like aspects. In this environment, it it probably reasonable for her to do so, lest players be surprised.
For example, the interchange below shows her engaging in a potentially computationally expensive operation, and one which is apparently uninterruptable: saving her mapping data to a more permanent storage system. Her current host computer made this a very quick operation (a couple of seconds), but it is easy to imagine circumstances (loaded machine, slower machine, much larger database) in which the interruption in Julia's usually almost instantaneous response would be disturbing. (Many user studies have shown that it is not total latency, but its variance, that is most disturbing to users. Even though many muds have sufficiently high and variable lag that Julia being momentarily lagged herself would not be noticed, Julia takes pains not to have this become a problem.)
Julia says, `Excuse me a minute, I have to save my map, Lenny.' Julia says, `Okay, I saved my map.'
If one knows the right questions to ask (which might be obvious questions, if and only if one knew that Julia was a 'bot), Julia will yield a surprising amount of interesting, but very inhuman, information:
You say, `julia, how much memory do you consume?'
Julia says, `Here is a summary of my memory usage:'
Julia says, ` 807472 bytes for 17335 strings'
Julia says, ` 232344 bytes for 19362 exits'
Julia says, ` 101100 bytes for rooms'
Julia says, ` 20220 bytes for paths'
Julia says, ` 268600 bytes for players'
Julia says, ` 4214784 bytes for 2058 dialog entries'
Julia says, `That's 5644520 bytes all together.'
You say, `julia, where is your source code?'
Julia says, `The source code for Maas-Neotek robots is available
on host NL.CS.CMU.EDU [128.2.222.56], directory /usr/mlm/ftp,
file robot.tar.Z. set mode binary first, and cd to /usr/mlm/ftp
in one step, Lenny.'
[ . . . ]
You say, `julia, who wrote you?'
Julia says, `Why do you ask?'
You say, `julia, who wrote you?'
Julia says, `I don't remember.'
You say, `julia, who wrote you?'
Julia says, `That information is restricted.'
You say, `julia, who wrote you?'
Julia says, `Fuzzy, of course.'
You say, `thank you, julia'
Julia says, `You're welcome.'
You say, `who is fuzzy, julia?'
Julia whispers, `He's just this guy I met in a bar, Lenny.'
You say, `who is fuzzy, julia?'
Julia whispers, `He's my Amway representative, Lenny.'
You say, `who is fuzzy, julia?'
Julia whispers, `I don't know, Lenny, I just follow him around.'
You say, `who is fuzzy, julia?'
Julia whispers, `He's the world's third greatest hacker, Lenny.'
One of the interesting things about Julia's construction is that her parser is shockingly simple, as such things go. It is barely more complicated than ELIZA's parser in some ways, in that it does not decompose its inputs into parse trees or anything else that a linguist might suggest. Instead, it does a very simple pattern-match on its input, looking for particular strings of characters separated by strings of any characters in between.
What makes Julia's conversation skills so powerful, as 'bots go, is that these sets of patterns were very cleverly grouped into useful equivalence classes by her programmer. Hence, the most common possible utterances that Julia might hear were enumerated this way, grouped appropriately, and such clever but brute-force programming proves sufficient for almost all cases.
Here is an example from Colin. (Julia's source code, being restricted, is unavailable for public consumption. However, Julia shares this piece of code with Colin.) Here, we see a fragment used to determine if some player is asking if Colin is a machine or some kind of 'bot. If any of the lines below matches, we run code that produces some equivocal answer, as Julia does, by choosing essentially at random from amongst a small collection of appropriate responses (as we've seen above in other examples):
else if (MATCH (lcmsg, `*are*you*robot*') ||
MATCH (lcmsg, `*are*you*a bot*') ||
MATCH (lcmsg, `*are*you*an ai*') ||
MATCH (lcmsg, `*are*you*autom*') ||
MATCH (lcmsg, `*are*you*machine*') ||
MATCH (lcmsg, `*are*you*computer*') ||
MATCH (lcmsg, `*are*you*program*') ||
MATCH (lcmsg, `*are*you*simulati*') ||
MATCH (lcmsg, `*you*are*robot*') ||
MATCH (lcmsg, `*you*are*a bot*') ||
MATCH (lcmsg, `*you*are*an ai*') ||
MATCH (lcmsg, `*you*are*autom*') ||
MATCH (lcmsg, `*you*are*machine*') ||
MATCH (lcmsg, `*you*are*computer*') ||
MATCH (lcmsg, `*you*are*program*') ||
MATCH (lcmsg, `*you*are*simulati*') ||
MATCH (lcmsg, `*you* be *robot*') ||
MATCH (lcmsg, `*you* be *a bot*') ||
MATCH (lcmsg, `*you* be *an ai*') ||
MATCH (lcmsg, `*you* be *autom*') ||
MATCH (lcmsg, `*you* be *machine*') ||
MATCH (lcmsg, `*you* be *computer*') ||
MATCH (lcmsg, `*you* be *program*') ||
MATCH (lcmsg, `*you* be *simulati*') ||
Here, we see an example of dealing with a common small-talk gambit:
else if ((MATCH (lcmsg, `*how*are*you*') ||
MATCH (lcmsg, `*how*do*you*') ||
MATCH (lcmsg, `*how's my*favor*') ||
MATCH (lcmsg, `*how is my*favor*') ||
MATCH (lcmsg, `*how's*life*') ||
MATCH (lcmsg, `*how is*life*') ||
MATCH (lcmsg, `*are* you* ok*') ||
MATCH (lcmsg, `*are* you* alright*') ||
MATCH (lcmsg, `*wie*gehts*') ||
MATCH (lcmsg, `*are*you*ok*') ||
MATCH (lcmsg, `*are*you*well*') ||
MATCH (lcmsg, `*are*you*alright*') ||
With such a simple parser, Julia is bound to experience peculiar-seeming failures even in ordinary conversation. Some of these failures can be masked with responses that a person might have said. One of the most surprising such failures, and one which reduced me to giggles for quite a while (given that I already knew that Julia was a 'bot), was the interchange below:
Conundrum says, `Julia, can you quote Lenny?' Julia whispers, `Could you say that again, it scrolled off my screen.'
My first reaction to the above was an incredulous, `What?'
A braino below (substituting `who' for `how') produces this curious response:
You say, `julia, who many pennies do I have?'
Julia says, `TimeGhost was asleep in High School about a minute and a half
ago.'
Julia says, `Ronni was asleep in Main Street, Block 2 (South) about a minute
or two ago.'
Julia says, `Alexandra was in Around the Flagpole about a couple of minutes
ago.'
Julia says, `Roxiana was in The Inter Nexus about a few minutes ago.'
Julia says, `Bunnell was in The Inter Nexus about 14 minutes ago.'
Julia says, `Volan was in The Inter Nexus about 18 minutes ago.'
Julia says, `Ling_Khan was in The Inter Nexus about 20 minutes ago.'
Julia says, `Mouse was in Upper Concourse West about 25 minutes ago.'
Julia says, `Polymorph was in The Inter Nexus about 35 minutes ago.'
Julia done
It's difficult to tell, a priori, whether the first exchange below is random or is related to particular stereotypes, but the second one is a clear parser failure:
You say, `julia, do you drive?' Julia says, `Do you get to Boston much?' You say, `julia, what are you carrying?' Julia says, `I follow the Pittsburgh Penguins, Lenny.'
(Julia will very often talk about hockey when she does not know what else to say. Anyone who talks with her much figures this out; in fact, an example below which talks about her sociology brings this point up again.)
And finally, it is clear that the notion of a prepositional phrase is beyond the abilities of her parser, with some deliberately chosen forced failures:
You whisper, `would you like a cookie?' to Julia. Julia whispers, `Certainly, I really enjoy cookies.' [ . . . but . . . ] You whisper, `julia, would you like some cake with glass shards?' to Julia. Julia whispers, `Yes thank you, I adore cake.'
So where does all this leave us? I've divided up my discussion of Julia into essentially two pieces, that of utilitarian function versus Turing-test competence:
Next: A sociological look Up: A prototypical agent: Previous: More of Julia: