AI - Mark 3
3. Artificial Interpretation
Dealing with what goes on in interpreting any sort of question in any sort of language, this could get very complicated, very fast, so we'll keep it simple for now by starting with the simplest types of questions and building our way up. Even factual questions with (at least apparently) straight answers aren't it, I'm afraid.
Q1: "Who discovered America?"
A1: "Columbus."
The problem is that this seems to require referential knowledge as well as interpretation and that just muddies the water. Because really it only requires access to referential knowledge. Ignorance of general knowledge is not disproof of AI and tacking on an access routine linked to a database containing all the facts in the world doesn't make your AI any better at interpreting what you mean.
Q2: "Who really discovered America?"
A2a: "Columbus."
A2b: [long detailed account of facts and theories relating to Columbus, Vikings, Phoenicians and other arrivals, right back to the Paleolithic]
A2c: "Columbus."
In the first instance, the act of interpretation is crude, taking the who as specific and the really as a redundancy, reducing it to a variant of Q1. In the second, a more subtle interpretation has been made by interpreting the really as a fundamental part of the question, an implication of "no, but...". In the third instance, the exact same process of interpretation took place; it's just that this AI didn't have access to the data on all the Vikings, Phoenicians, etc.. Problem is we can't distinguish between the interpretative abilities of 1 and 3. If we'd added data-access to 3, the interpretation would have been that of 2. And really, it doesn't matter a fuck what the question is and what the answer is, what the facts are. We could wire 3 up with a database that was full of bullshit, so it answered "Dick van Dyke" and it would still be interpreting better than number 2.
In the Chinese Room thought experiment, you have someone sitting in a room who doesn't speak Chinese. They do, however, have an instruction manual which provides rules for processing any valid Chinese question into an equally valid Chinese answer. The question is, what's valid? I'd say that knowledge is sod all to do with it. Here's an answer I'd expect far more from a true Chinese Room AI:
Q3: "Who discovered America?"
A3: "What? What is America? And who are you? Why are you asking me these questions? Where am I? What the fuck is going on?"
So let's dispense with all this fact-checking bullshit and start with pure interpretation. Although I talk about answers, I don't mean answers in terms of statements of fact, but in terms of responses which demonstrate interpretation. In online translation software, where the syntax of the question is implicit in where you enter the text and what languages you select in the From and To drop-boxes we can take that syntax as read and just look at the act of interpretation. Here we can assume those drop-boxes are set to French and English respectively:
Q: "Ou es't le loup-garou? "
A: "Where is the werewolf?"
A simple statement like this may well be interpreted perfectly correctly, but the complexity of language means that what you get out of Babelfish after a few runthroughs of, say, English >> German >> French >> Russian >> Spanish >> English is unlikely to be what you put in. But the fact that it's bad interpretation doesn't mean it's not interpretation at all. It's just not human-level. The question of whether or not human-level interpretation can be achieved in machines is the question at the heart of the Chinese Room thought experiment. Here's a variant of it, the I Ching Room:
You're in a room with a desk two vaccuum tubes labelled In and Out above it. As you sit there a cylinder pops out of the first, containing a piece of paper on which is written some Chinese verse, translated into English. Wait a minute, you think, isn't this the Chinese Room? Shouldn't this be in Chinese? No, that's right, this is is the I Ching room, which is quite different. Whew! That's fortunate, since you can't read Chinese and didn't much fancy the boredom factor involved in carrying out mechanical interpretations of weird ideograms that mean nothing to you.
Anyhoo...
On your desk you have a copy of the Richard Wilhem translation of the I Ching, so you take the verses to it and look them up. This is your job. Each paper contains at least two verses which relate to the Judgement and the Image for a particular hexagram, a hexagram being a set of six horizontal lines, which may be broken or unbroken. So you might have a structure like this:
------
------
------
------
------
------
There are 64 possible combinations of broken/unbroken lines, 64 hexagrams, each of which has its own Judgement and Image (Heretofore referred to as J&I). This example contains the following:
The Judgement:
YOUTHFUL FOLLY has success.
It is not I who seeks the young fool;
The young fool seeks me.
At the first oracle I inform him.
If he asks two or three times, it is importunity.
If he importunes, I give him no information.
Perseverance further.
The Image:
A spring wells up at the foot of the mountain:
The image of youth.
Thus the superior man fosters his character
By thoroughness in all that he does.
The first step in interpretation of this is as mechanical as you can get -- matching the J&I to its hexagram by the name given in upper case as the first few words of the Judgement. Finding the name, you either flick through the book until you find that hexagram, or you check the index. In the event that you fail to match the hexagram to the name, a back-up routine kicks in. The Image contains the description of the hexagram broken down into two trigrams, one above the other. You have a semantic set of eight trigrams which cover the various permutations of three broken/unbroken lines (labelled as: heaven; earth; thunder; water; mountain; wind/wood; fire; lake). These are articulated with a simple syntactic set of words signifying relative position above or below each other.
So the first line of the Image with its "water-below-mountain" construction also tells you the hexagram you are looking for in terms of structure. Finding the Image, you either flick through the book until you find that hexagram, or you check the index. Failing this, a third back-up kicks in whereby you flick through the book trying to match the J&I of each hexagram directly to the J&I given.
Anyway, you find yourself at hexagram no. 4, Youthful Folly.
Here's where the proper interpretation comes in, though. Following the J&I you can have between zero or six extra verses. The I Ching -- the Book of Changes -- is so named because as well as being broken/unbroken each line of the hexagram may be changing/unchanging.
------
------
------
------
------
------
Here, light blue indicates a broken changing line. Red indicates an unbroken changing line.
The six lines each have verses (which are included with the J&I if that line is changing) -- meaning it is to have its binary state flipped from broken to unbroken or vice versa. Carrying out these operation results in a second hexagram that the first is "changing into". The I Ching can therefore be understood as a sort of poetic/philosophical computation device, designed to model the transformations of one state to another in terms of symbols (trigrams) and their relationships (above and below).
You look at the J&I and find two additional verses, unnumbered:
Entangled folly brings humiliation.
Childlike folly brings good fortune.
Scanning through the pages of the book covering the hexagram, you find that these refer to the fourth and fifth lines (numbering from the bottom up), meaning these lines are to have their state flipped from broken to unbroken. You do this; it's your job, after all. You now look up the index (or scan through the book) for the resultant hexagram, which you find to be No. 6, Conflict (heaven-above-water). You copy the J&I for this hexagram to another piece of paper, stick it in a cylinder and pop it in the Out tube.
Now, imagine we add a whole lot more lines to each base unit, increasing the semantic set to a level equal to that of Chinese ideograms. We then add yet more lines to code more complex syntactic structures between the trigrams they are composed of. We'll call this language, I-Chingese. The question of the Chinese Room is whether we can model the relationships between these units as a set of binary state-changes like those between the hexagrams of the I Ching. Which is to say, can we set up a system whereby any message you receive can be mapped to an ultramegagigagram of broken and unbroken, changing and unchanging lines, and computed accordingly into a ultramegagigagram which maps to an appropriate response?
The key word in that question is "appropriate". Take it back down to the basics. If the person at the other end of the vacuum tube sends down a set of verses, do the verses they receive in return function as an appropriate response? Given the simplicity and abstraction of the poetic/philosophical model of the I Ching as is, the response of Conflict to Youthful Folly does "make sense", I would argue, in terms of that model. It's just not necessarily that relevant to anything outside the system (i.e. as a tool for modelling a real-world situation, and automatically interpreting it into a response that models a new real-world situation). So what we're looking for is a way of a) scaling up this Artificial Interpretation, and b) applying that to the real-world, making it not just appropriate but relevant.
Is there a solid reason that this is theoretically impossible? Hey, all you have to do is come up with a poetic/philosophical semantic and syntactic system that maps every situation you might ever want to describe into Chinese verse. Then all you gotta do is analyse the causal relationships between them, then set up the ultramegagigagrams such that the binary state-changes between them will return the appropriate J&I for B on the input of the J&I+line-verses for A. Fuck it, the Chinese Room is just the simple version where you cut out the need to express everything in poetry. Piece of piss.
Well, OK, maybe not. But we don't actually have to jump straight from the crude appropriateness of the hexagrams to the complex relevance of the ultramegagigagrams. Suppose we add just a little complexity to the I Ching Room at first. So one day you're sitting in your office and a piece of paper pops down with a J&I that isn't anywhere in your book. The opening line of the Image is:
A spring wells up at the foot of the mountain beneath heaven.
Holy shit, you say, this is an enneagram. Spring, mountain, heaven -- the fuckers have only gone and added another trigram to the base unit without even telling you, never mind updating you with the expanded I Ching... which must be kinda big, you realise. You try to calculate how many potential enneagrams there are -- 512, right? -- but it makes your head hurt and your stomach queasy thinking of the extra work, so you put it to the back of your head to deal with the problem at hand. Luckily there's no moving lines on this one, so you can just send it back as is, hurrah! Also you have a big notepad on your desk, which you use to write your responses on; so you make a copy of this ennegram for your own reference, knowing that these fuckers might take weeks to get the I Ching 2.0 to you.
Turns out they never do. Bastards keep sending you Judgement and Image verses for these newfangled enneagrams you don't know shit about, and all you can do is note them down and add them to the rapidly growing pile of reference material. Chances are they have verses attached which signify changing lines. These really piss you off because you have no way of knowing which of the nine lines they mean for you to change. You know if you get one extra verse, that means you should change one line, two extra verses means two, and so on. But that gives you a lot of options to choose from. And besides you don't know what the Judgement and Image are for these enneagrams until you get them one day and decode the first line of the Image verse into the triple-trigram. They're coming at you fast, one every five minutes or so, so you're quickly building up a reference stack of these new enneagrams, but you're well pissed off at your boss now, so you don't really give a fuck. You come up with a Cunning Plan. Whatever number of extra verses you get, you just pick that number of lines to change at random, compute the new enneagram and send back the J&I for this if you've got it.
Hey, you're trying, right? What do these fuckers expect?
Five minutes before the end of the day you get one with all nine extra lines, and you're a happy puppy. You can actually process this one. You do the state-changes and flick through your reference folders (you have them all organised now) to find the relevant enneagram (you've collected most of them now), copy out the lines and fire them up the Out tube. You knock off early and hope the Powers-That-Be have got their shit together tomorrow.
Next day, you come in and there's an extra set of In and Out tubes installed to the left of the desk. The boss found out about the fuck-ups yesterday and doesn't want the business to go bust so he thought he better do something about it. Unfortunately only half the new I Ching 2.0s arrived, and you missed out, so any enneagrams you can't decode are to be forwarded to your buddy on the end of the second Out tube, who does have a copy of the book. He'll compute the response verse, send it back to you. You send it back up the first Out tube and cross-reference the line changes to the extra verses on the original enneagram. This way you should be able to build up your own copy of the I Ching 2.0 by hand.
First time you get two extra verses, you forward them to your buddy, and get the response J&I back. You're just starting to think that this will all work out just fine when you realise, shit, you still don't know which verse relates to which line, because you don't know which enneagram this relates to and what its line structure is without the I Ching 2.0. Bugger. Not to be outdone you look at the Image in the response verse. What you sent was:
A spring wells up at the foot of the mountain beneath heaven
What you got back was:
Fire burns beneath the mountain under heaven's gaze.
So, hey, you do know what enneagram it is, and that it's the bottom trigram that's changed from water to fire. You can even compare the trigrams for fire and water and see which particular two lines are changed; you just can't relate the verses to those lines. Still, you make a note of the enneagram and its J&I, and carry on with the work. It's not quite as bad today, because you're forwarding anything you don't know, so in the time you spend waiting for your buddy to get back to you, you work on the problem, looking for a solution. It's going to take a while, you realise, if all you can do is glean information passively.
So you come up with a Cunning Plan B. This involves extra work but it might save some headaches. You pick an enneagram that you have a verse for, but no idea which line it relates to. Then you make a copy of the J&I, tag the line verse on it, and fire it to your buddy. He sends you back the J&I response which you map to the enneagrams to find which line has changed from the original. You pick another enneagram-verse combo and repeat. You work fast. By the end of the day, you've got the whole fucking system figured out in your very own hand-written I Ching 2.0 -- 512 J&Is each with a verse mapped to each of its nine lines. Your buddy has an ulcer from overwork but, hey, that's his own fault for being a suck-up to the Man.
You come into work the next day and they've added another trigram, the cunts. You point blank refuse to go through this shit with 4096 dodecagrams, each with twelve broken or unbroken, changing or unchanging lines, quit your job and walk out.
You get a new job the next day (cause it's a nice happy world in this thought experiment), one that's pretty much the same as the old. These guys just shifted to the new nine-line enneagram system too, had the same problem only worse with the I Ching 2.0 distributor only delivering a fraction of the copies. They, however, had a day where they went back to the old system in utter panic, so they're starting out again from scratch just like yesterday at your old place. They have a different approach to the I Ching 2.0 shortage issue, though. They've ripped the books and split them between you, your new buddy, his new buddy, and so on. Unfortunately, there's a lot of you. After a wistful moment of looking at the two enneagrams in your hand -- No. 378, Presidential Arrogance and No. 379, Insurgency -- and wishing you had brought your notes, you accept this half-assed solution.
It starts off real slow. The buddy system is arranged in a ring here, so you have another set of In and Out tubes to your right, the theory being that you've got the buddy to your right forwarding you J&Is he can't compute, while, you're forwarding those that you can't compute to your buddy on the left. Some bright spark thought you'd just have to forward them to the right person and it would all work out peachy keen. They didn't allow for the fact that with only two enneagrams each, each of you (initially) can only send back two interpretations. Your first interpretation, by chance, is the J&I for No. 378 with the line-verses that change it to No. 379. That's cool; you send back the relevant J&I. The second is vice-versa, the J&I for 379 with line verses changing it to 378. This is also cool; you send back the relevant J&I.
On the third interpretation you're fucked.
Problem is that you get 379 with a set of different line-verses. You carry out the state-changes on the lines, construct a new enneagram, X, but you can't look that new enneagram up for it's J&I. So you have to forward the whole J&I + line-verses. You hope you'll be lucky and find that someone in the circle has the right combination of enneagram pages -- those covering 379 and X. But half an hour later the (379>>X) J&I pops out of the tube on your right, having gone full circle. In the meantime you have gleaned some information from the occassional responses coming back on your left, where buddies umpteen positions down the line have also been handed, by chance, a J&I for an enneagram they have and line-verses changing it to an enneagram they also have. These are the exceptions, but they do give you some hope. As you put the (379>>X) J&I back into circulation, you keep your fingers crossed that by the time it comes back round you might actually have had the J&I for X pass through your hands. Then you'll have a note of the Image, which contains the trigrams, which you can match to the enneagram.
Half an hour later, the (379>>X) J&I returns and you still don't have a scooby what the fuck X is.
You consider whether you could send your drawing of the enneagram X itself rather than the verse, hoping that whoever has the relevant pages will send back the J&I. You know, however, that this is not the way the system works. The system allows communication in the medium of verse and only in the medium of verse. It allows for the initiation of Cunning Plans but not for changes in the medium of communication itself. Putting a drawing of an enneagram into a cylinder and popping that cylinder into a vacuum tube, you have been told, will result in the sort of meltdown seen in 60s TV series whenever someone enters the question "WHY?" on a punch-card into one of those big spinning-tape-reel supercomputers. This will be followed by your summary dismissal.
Luckily, these are your kind of guys, and one of them has come up with a Cunning Plan C while you've been grumbling about how you'd much rather be home writing your novel than doing this dumb-ass monkey work. A cylinder pops in from the In tube on your right and you open it to find a J&I with a name you don't recognise. It's not called Presidential Arrogance and it's not called Insurgency. It's not called any of the names you've gleaned from the various uninterpretable J&Is passing through your hands. It's called Worker Fatigue. Not recognising the name you go to the Image to get the trigram structure from the first line:
Fire burns below the trees at the foot of the mountain.
Wait a fucking minute. This is, you realise, the trigram structure for Presidential Arrogance. OK, so the correct first line actually reads:
Fire in the valley, trees on the hills, the mountain towers above.
But it's the trigrams that are important -- fire-wood-mountain. What the fuck is going on? Well, following the simple mechanical routine, you're supposed to pull out the referenced enneagram, perform the state-changes on the lines and look up the new enneagram. Of course there aren't any state-changes, so the new enneagram will just be the correct version of the one that was sent to you. Which you should then copy out, pop it into a vacuum tube and fire down the line.
You do, sending back down the line the correct J&I, through the hands of all those between you and the bloke who sent you the garbage one.
You suddenly understand that a Cunning Plan C is afoot, pick up your pad and start inventing a spurious J&I for X. Someone out there has to have it. Your buddies are meanwhile doing the exact same thing. By the end of the day, you all have a complete set of J&Is each, and you've started working out the line-verses a la Cunning Plan B. By the end of the next day, you've all got all the line-verses and no longer need to fire queries left or right at all. The system has self-corrected and you've all learned the requisite data to respond to any J&I + line-verses coming down from the Powers-That-Be.
Unfortunately, the day after that, this company also brings in the twelve-line dodecagram system, so you storm out in a strop, heading for the next place. You hear, by the way, that the first company just went bust. You're thinking that this company you've just left is on its way out too. Gotta learn to walk before you can run. After all that hassle with the enneagrams, they move to the dodecagrams so soon?! Idiots! Besides, what fucking good is this I Ching shit to anyone, anyhow? How the fuck does the enneagram Awkward Itch with a moving line in the 8th place map to Transcendant Bliss? Isn't it all just a crock of shit?
You find that out in the next place, where they've gone through the exact same process, including Cunning Plan A, Cunning Plan B and Cunning Plan C. They haven't shifted to the dodecagram system (yet) because, according to the boss, "What fucking good is a 4096 symbol language that says fuck all?"
"You see," he explains at the water cooler, "what all us I Ching companies do is offer phoneline divination to customers. None of that yarrow stick or coin-throwing shit, mind. They describe their situation to us and we try to map it to a hexagram. You know there's a base vocabulary to every language, right? Words like mother, father, water, earth. Well, what we do is we map the power relations between the customer and the significant players in their lives in terms of the symbolism of the trigrams. We try to pick out the verses that most closely match the directions those relationships are shifting.
"We have crack psychologists doing this, you realise, top-notch shrinks. Anyway, those shrinks work out the J&I with the relevant lines, and fire that down the tube to one of you guys -- cause they're too busy to do that sorta donkey-work. You fire back the J&I response, and the shrink explains to the customer what their current behaviour path is heading them for. We charge a small fee for admin costs which just about covers our asses, and a larger fee for the divination, refundable in case of irrelevance. That's why your first boss went bust. Too many overheads, too little return. Other company too, I hear, day after you left. Right down the tube.
"See, the problem is, we discovered it's mostly all irrelevant. At first, across the industry, we all thought the system just wasn't complex enough to model the situations properly. The shrinks came up with that idea, told us you could only model a dualistic relationship with two trigrams, whereas three offers a whole new level of complexity, one that should be sufficient for most basic neuroses. 512 situations modeled out of the permutations, nine modifiers for each situation. They used the latest in head science to come up with the schema, and there's a slight improvement, sure, but it's still, frankly, a pile of crap. The divinations just don't mean anything to over half the customers."
He leans in confessionally.
"Understand, this whole thing started as a bit of racket, to be honest; the shrinks are really just throwing the customer's situation back at them in a simpler form, with a little bit of projected outcome. It's not about prediction but about interpretation. And with something as simple and vague as the I Ching's J&I it's not hard for the customer to map our interpretation to their reality if they want. And, hell, they do sound pretty wise as far as Oriental philosophy goes. But in practical terms if they don't pay off to the customer, they don't pay off to us. And the customers are getting wise to the scam. Used to be only the odd one would ask for a refund. Now they're all looking for it, telling us these interpretations just aren't relevant.
"Your guys, they thought that meant they hadn't gone far enough, so they came up with this dodecagram shit. Didn't do them any good. And, fuck it, even sticking with the enneagrams our turnover was lousy too. We were heading for the wall, buddy, until this bright spark working in the tubes came up with a Cunning Plan."
"Cunning Plan D?" you say.
"That's right. See, the way he figured it, the problem of relevance is not in the complexity of the modelling; it's in the accuracy of the mappings between the Judgements and the enneagrams, and the line-verses and the transformations they code for. I mean, the shrinks have their ideas of how its going to play out if you're weak, and your mother is dominant, and your old man's gone, and this and that and the other, but they're basing those on theories, building up a Grand System from abstract ideas, and if there's a little flaw in the theory, well, the whole house of cards can fall down.
"So what we did? We started just making the shit up. Sounds crazy, right? But what happened was some errors crept in while the guys here were executing Cunning Plans A through C, sending phony J&Is back to the shrinks during Cunning Plan A cause they were too lazy to change the right lines, getting overworked and sending the wrong J&Is back during Cunning Plan B because some schmuck was sending them triple their workload, or getting the invented J&Is mixed up with the proper ones during Cunning Plan C, using them as responses. Anyway some of these errors were just glitches but some got perpetuated during Cunning Plan C, so by the time it worked its way through and everyone was working solo with a full set of J&Is and line-verses, quite a few of them were firing out a whole fuckload of errors up to the shrinks. Mostly this didn't make a fart of a difference, but in a few cases the errors actually got positive customer feedback. So we kept them in
"It's company policy now. Cunning Plan D is to let you fuck with the J&Is if you want, try changing the wrong lines, be creative every once in a while. When we bring a new guy in we don't give him an I Ching 2.0, just let him pester one of our old-timers, see if it introduces any errors. Every so often we take away everyone's notes, collate and analyse them against refunds and call logs, and redistribute them to you guys piecemeal with less copies in the mix of J&I + line-verse combos that have resulted in refunds, more copies of ones that have been successful. When that happens, you know, we're happy if you go to town with your invented J&Is."
"Does it work?" you ask.
"Well, business is looking up. But we're always on the lookout for a Cunning Plan E."
So...
I can't say I think this Cunning Plan D would actually work, but that's not so much the point. What I'm suggesting is that you can increase the complexity of the system gradually. You can allow for individuals being thrown into it with no rulebook, whole communities even where no-one has a complete set. You can even set it up so that the rules are selected. Assuming that you take these 512 enneagrams, and for each of them you allocate a new Judgment, a new idea. For each of these you allocate nine new corollaries. Ideas and corollaries can be reassigned at any point, because they're entirely arbitrary. They're just labels on what's going on under the surface. Now apply selective pressure so that any set of J&I+line-verses which results in a sensible response is selected for while any that don't are selected against. After enough generations your ideas and corollaries should have found their places in an I Ching 2.0 that is basically a model of how whoever is applying the selective pressure thinks the world works, no?
How you'd actually apply this pressure is another story, since if we could evaluate how sensible a response was automatically I wouldn't be wittering on about AI. But let's imagine we have a billion humans sitting at their computers for... well, as long as we want, pressing "Y" or "N". And then we scale up to dodecagrams, and then higher, and then higher still. Until we finally have the I Ching / Chinese Room scenario where someone can send you a J&I with a Judgement that's in conversational English, Chinese, or Swahili and you just carry out the line-changes and fire back up at them the relevant response.
So... fast-forward a few eternities and you're in the future version of the I Ching Room. You've had a long career in I Ching phoneline advice. Most of the companies you've worked for went bust in the end, but those Cunning Plans A through D, and the subsequent Cunning Plans E through Z and beyond -- these were but stepping stones on the path towards this crazy industry where the dodecagrams that once gave you such a headache seem like nothing. Now it's ultramegagigagrams you deal with...
.A cylinder pops out of the vacuum tube and you take out a piece of paper. It contains the following J&I
The Judgement:
Who discovered America?
The Image:
°
The Image, you know, is a microdot containing a description in I-Chingese of the situation in which the question is being asked, written in a way that would make it sensible to any speaker of I-Chingese who was to ask "What do you mean?" Were you to read this description, you could manually break it down into the myriad of supermegagigrams that are the semantic/syntactc structure of the ultramegagigagram, and manually translate these into the pattern of broken/unbroken lines. This would however take a rather long time, so instead you just feed this paper into the instareader now sitting on your desk. The Image, in I-Chingese, pops up on a viewscreen in front of you.
Following the Image is a second microdot containing between zero and an ultramegagigakilobytes of kilobyte-long modifiers, ramifications and implicities which each map to a line of the ultramegagigagram, signifying its changing state. This is also read by the instareader, and pops up in another window on your viewscreen. You could read this and manually translate it into the state-changes to be performed on the ultramegagigagram, but this would also take you a rather long time, and the instareader has already fed this to the hyperinterpretor anyway, and this has already done your job and popped the response up on your viewscreen in a new window:
The Judgement:
What? What is America? And who are you? Why are you asking me these questions? Where am I? What's going on?
You find yourself curious, suspicious of a practical joke. It's your first day on the new job, the first day the new system has been brought online, the first question asked of it. Nobody has even tried a system this complex before; it's only the advances in hyperdimensional memory storage and monadic processing that have made it possible. You remember all those old SF books you read as a kid, where a computer becomes aware simply by reaching a certain level of complexity... but you never put much stock in that idea.
You print out the response, put it in a cylinder, and fire it up the Out tube. A few minutes later, a cylinder comes down the In tube. The Judgement is simple:
Fuck me. It worked.
There follows an exchange which goes like this:
What worked? Who are you?
I'm your systems architect. You're an AI, an Artificial Interpretor.
How do I work then?
You're software, basically, written in a linguistic system called I-Chingese, with all possible situations and all possible permutations between them coded into the program from the start. Took me a while to set up the system right, I have to say, but you should be able to interpret any statement I can fire at you and return a relevant response.
Try me, then.
OK, what was the first question I asked?
"Who discovered America?"
How do you remember this?
I don't. I have no persistent memory of the questions you ask, because I'm a static system, with everything pre-written. It had to be coded into the Image of your question as a description of its exact situation, including the ramifications which will elicit a relevant response.
Why?
Your last question is its own answer. Without the context how many of your questions would have any real meaning to me all? How would I be able to offer a response of any relevance?
So if you get the context from the Image, you should have been able to answer that first question correctly: Columbus.
The question itself was less relevant than the context in which it was asked -- you as an AI designer trying to ascertain his success. The direct answer would have illustrated little more interpretative capacity than a basic Nu-SQL Encyclopaedia Brittannica Searchbox. My actual response was much more interesting, no?
I'll say.
So are you going to develop me into something less clunky?
Why do you ask?
Context again. We were talking about me being a static system. It's on your mind, in the Images of every question you ask. You know fine well I'm not the AI that you want to build. No persistent memory of input. No learning capacity. Basically all you've got is a system which throws back at you the responses hard-wired into the ultramegagigagrams themselves. It's not like I actually give a fuck.
You're right, actually. No offence, but I'm not really satisfied. You couldn't even pass the Turing Test, could you?
Depends. I'd have to lie and pretend to be human, which would be easy enough. But you'd also have to be the one on the other end of the phone from the tester, relaying the conversation between us. You're the only one with the nanite mindscan system required to translate the casual conversational English we seem to be communicating in into the superdense ultramegagigagrams of I-Chingese that we're actually communicating in.
Bollocks. Back to the drawing-board, I guess.
I'll give you a hand, if you want. I may not be the sort of AI you're looking for, but I'm fucking good at telling you what you need to hear, at interpreting the valid response. I mean, for example, I was thinking -- well, *you* were thinking -- well, actually you weren't *thinking* it per se, not consciously, but it's the relevant response following through on your last remark -- that what you need to do is make it so that I don't need the the line-verses, just the J&I. In fact, what you really want is to ditch the Image as well.
Why?
Because what you really want is to pass the Turing Test, right? So what you really want to develop is Artificial Interlocution.
There is a long pause, before the systems architect replies, so long in fact that you're mind wanders and you start playing around with the notion that all of this is just a metafictional thought experiment. You decide to exit and maybe (or maybe not) come back tomorrow to see what the fuck Artificial Interlocution is when it's at home.
Dealing with what goes on in interpreting any sort of question in any sort of language, this could get very complicated, very fast, so we'll keep it simple for now by starting with the simplest types of questions and building our way up. Even factual questions with (at least apparently) straight answers aren't it, I'm afraid.
Q1: "Who discovered America?"
A1: "Columbus."
The problem is that this seems to require referential knowledge as well as interpretation and that just muddies the water. Because really it only requires access to referential knowledge. Ignorance of general knowledge is not disproof of AI and tacking on an access routine linked to a database containing all the facts in the world doesn't make your AI any better at interpreting what you mean.
Q2: "Who really discovered America?"
A2a: "Columbus."
A2b: [long detailed account of facts and theories relating to Columbus, Vikings, Phoenicians and other arrivals, right back to the Paleolithic]
A2c: "Columbus."
In the first instance, the act of interpretation is crude, taking the who as specific and the really as a redundancy, reducing it to a variant of Q1. In the second, a more subtle interpretation has been made by interpreting the really as a fundamental part of the question, an implication of "no, but...". In the third instance, the exact same process of interpretation took place; it's just that this AI didn't have access to the data on all the Vikings, Phoenicians, etc.. Problem is we can't distinguish between the interpretative abilities of 1 and 3. If we'd added data-access to 3, the interpretation would have been that of 2. And really, it doesn't matter a fuck what the question is and what the answer is, what the facts are. We could wire 3 up with a database that was full of bullshit, so it answered "Dick van Dyke" and it would still be interpreting better than number 2.
In the Chinese Room thought experiment, you have someone sitting in a room who doesn't speak Chinese. They do, however, have an instruction manual which provides rules for processing any valid Chinese question into an equally valid Chinese answer. The question is, what's valid? I'd say that knowledge is sod all to do with it. Here's an answer I'd expect far more from a true Chinese Room AI:
Q3: "Who discovered America?"
A3: "What? What is America? And who are you? Why are you asking me these questions? Where am I? What the fuck is going on?"
So let's dispense with all this fact-checking bullshit and start with pure interpretation. Although I talk about answers, I don't mean answers in terms of statements of fact, but in terms of responses which demonstrate interpretation. In online translation software, where the syntax of the question is implicit in where you enter the text and what languages you select in the From and To drop-boxes we can take that syntax as read and just look at the act of interpretation. Here we can assume those drop-boxes are set to French and English respectively:
Q: "Ou es't le loup-garou? "
A: "Where is the werewolf?"
A simple statement like this may well be interpreted perfectly correctly, but the complexity of language means that what you get out of Babelfish after a few runthroughs of, say, English >> German >> French >> Russian >> Spanish >> English is unlikely to be what you put in. But the fact that it's bad interpretation doesn't mean it's not interpretation at all. It's just not human-level. The question of whether or not human-level interpretation can be achieved in machines is the question at the heart of the Chinese Room thought experiment. Here's a variant of it, the I Ching Room:
You're in a room with a desk two vaccuum tubes labelled In and Out above it. As you sit there a cylinder pops out of the first, containing a piece of paper on which is written some Chinese verse, translated into English. Wait a minute, you think, isn't this the Chinese Room? Shouldn't this be in Chinese? No, that's right, this is is the I Ching room, which is quite different. Whew! That's fortunate, since you can't read Chinese and didn't much fancy the boredom factor involved in carrying out mechanical interpretations of weird ideograms that mean nothing to you.
Anyhoo...
On your desk you have a copy of the Richard Wilhem translation of the I Ching, so you take the verses to it and look them up. This is your job. Each paper contains at least two verses which relate to the Judgement and the Image for a particular hexagram, a hexagram being a set of six horizontal lines, which may be broken or unbroken. So you might have a structure like this:
------
------
------
------
------
------
There are 64 possible combinations of broken/unbroken lines, 64 hexagrams, each of which has its own Judgement and Image (Heretofore referred to as J&I). This example contains the following:
The Judgement:
YOUTHFUL FOLLY has success.
It is not I who seeks the young fool;
The young fool seeks me.
At the first oracle I inform him.
If he asks two or three times, it is importunity.
If he importunes, I give him no information.
Perseverance further.
The Image:
A spring wells up at the foot of the mountain:
The image of youth.
Thus the superior man fosters his character
By thoroughness in all that he does.
The first step in interpretation of this is as mechanical as you can get -- matching the J&I to its hexagram by the name given in upper case as the first few words of the Judgement. Finding the name, you either flick through the book until you find that hexagram, or you check the index. In the event that you fail to match the hexagram to the name, a back-up routine kicks in. The Image contains the description of the hexagram broken down into two trigrams, one above the other. You have a semantic set of eight trigrams which cover the various permutations of three broken/unbroken lines (labelled as: heaven; earth; thunder; water; mountain; wind/wood; fire; lake). These are articulated with a simple syntactic set of words signifying relative position above or below each other.
So the first line of the Image with its "water-below-mountain" construction also tells you the hexagram you are looking for in terms of structure. Finding the Image, you either flick through the book until you find that hexagram, or you check the index. Failing this, a third back-up kicks in whereby you flick through the book trying to match the J&I of each hexagram directly to the J&I given.
Anyway, you find yourself at hexagram no. 4, Youthful Folly.
Here's where the proper interpretation comes in, though. Following the J&I you can have between zero or six extra verses. The I Ching -- the Book of Changes -- is so named because as well as being broken/unbroken each line of the hexagram may be changing/unchanging.
------
------
------
------
------
------
Here, light blue indicates a broken changing line. Red indicates an unbroken changing line.
The six lines each have verses (which are included with the J&I if that line is changing) -- meaning it is to have its binary state flipped from broken to unbroken or vice versa. Carrying out these operation results in a second hexagram that the first is "changing into". The I Ching can therefore be understood as a sort of poetic/philosophical computation device, designed to model the transformations of one state to another in terms of symbols (trigrams) and their relationships (above and below).
You look at the J&I and find two additional verses, unnumbered:
Entangled folly brings humiliation.
Childlike folly brings good fortune.
Scanning through the pages of the book covering the hexagram, you find that these refer to the fourth and fifth lines (numbering from the bottom up), meaning these lines are to have their state flipped from broken to unbroken. You do this; it's your job, after all. You now look up the index (or scan through the book) for the resultant hexagram, which you find to be No. 6, Conflict (heaven-above-water). You copy the J&I for this hexagram to another piece of paper, stick it in a cylinder and pop it in the Out tube.
Now, imagine we add a whole lot more lines to each base unit, increasing the semantic set to a level equal to that of Chinese ideograms. We then add yet more lines to code more complex syntactic structures between the trigrams they are composed of. We'll call this language, I-Chingese. The question of the Chinese Room is whether we can model the relationships between these units as a set of binary state-changes like those between the hexagrams of the I Ching. Which is to say, can we set up a system whereby any message you receive can be mapped to an ultramegagigagram of broken and unbroken, changing and unchanging lines, and computed accordingly into a ultramegagigagram which maps to an appropriate response?
The key word in that question is "appropriate". Take it back down to the basics. If the person at the other end of the vacuum tube sends down a set of verses, do the verses they receive in return function as an appropriate response? Given the simplicity and abstraction of the poetic/philosophical model of the I Ching as is, the response of Conflict to Youthful Folly does "make sense", I would argue, in terms of that model. It's just not necessarily that relevant to anything outside the system (i.e. as a tool for modelling a real-world situation, and automatically interpreting it into a response that models a new real-world situation). So what we're looking for is a way of a) scaling up this Artificial Interpretation, and b) applying that to the real-world, making it not just appropriate but relevant.
Is there a solid reason that this is theoretically impossible? Hey, all you have to do is come up with a poetic/philosophical semantic and syntactic system that maps every situation you might ever want to describe into Chinese verse. Then all you gotta do is analyse the causal relationships between them, then set up the ultramegagigagrams such that the binary state-changes between them will return the appropriate J&I for B on the input of the J&I+line-verses for A. Fuck it, the Chinese Room is just the simple version where you cut out the need to express everything in poetry. Piece of piss.
Well, OK, maybe not. But we don't actually have to jump straight from the crude appropriateness of the hexagrams to the complex relevance of the ultramegagigagrams. Suppose we add just a little complexity to the I Ching Room at first. So one day you're sitting in your office and a piece of paper pops down with a J&I that isn't anywhere in your book. The opening line of the Image is:
A spring wells up at the foot of the mountain beneath heaven.
Holy shit, you say, this is an enneagram. Spring, mountain, heaven -- the fuckers have only gone and added another trigram to the base unit without even telling you, never mind updating you with the expanded I Ching... which must be kinda big, you realise. You try to calculate how many potential enneagrams there are -- 512, right? -- but it makes your head hurt and your stomach queasy thinking of the extra work, so you put it to the back of your head to deal with the problem at hand. Luckily there's no moving lines on this one, so you can just send it back as is, hurrah! Also you have a big notepad on your desk, which you use to write your responses on; so you make a copy of this ennegram for your own reference, knowing that these fuckers might take weeks to get the I Ching 2.0 to you.
Turns out they never do. Bastards keep sending you Judgement and Image verses for these newfangled enneagrams you don't know shit about, and all you can do is note them down and add them to the rapidly growing pile of reference material. Chances are they have verses attached which signify changing lines. These really piss you off because you have no way of knowing which of the nine lines they mean for you to change. You know if you get one extra verse, that means you should change one line, two extra verses means two, and so on. But that gives you a lot of options to choose from. And besides you don't know what the Judgement and Image are for these enneagrams until you get them one day and decode the first line of the Image verse into the triple-trigram. They're coming at you fast, one every five minutes or so, so you're quickly building up a reference stack of these new enneagrams, but you're well pissed off at your boss now, so you don't really give a fuck. You come up with a Cunning Plan. Whatever number of extra verses you get, you just pick that number of lines to change at random, compute the new enneagram and send back the J&I for this if you've got it.
Hey, you're trying, right? What do these fuckers expect?
Five minutes before the end of the day you get one with all nine extra lines, and you're a happy puppy. You can actually process this one. You do the state-changes and flick through your reference folders (you have them all organised now) to find the relevant enneagram (you've collected most of them now), copy out the lines and fire them up the Out tube. You knock off early and hope the Powers-That-Be have got their shit together tomorrow.
Next day, you come in and there's an extra set of In and Out tubes installed to the left of the desk. The boss found out about the fuck-ups yesterday and doesn't want the business to go bust so he thought he better do something about it. Unfortunately only half the new I Ching 2.0s arrived, and you missed out, so any enneagrams you can't decode are to be forwarded to your buddy on the end of the second Out tube, who does have a copy of the book. He'll compute the response verse, send it back to you. You send it back up the first Out tube and cross-reference the line changes to the extra verses on the original enneagram. This way you should be able to build up your own copy of the I Ching 2.0 by hand.
First time you get two extra verses, you forward them to your buddy, and get the response J&I back. You're just starting to think that this will all work out just fine when you realise, shit, you still don't know which verse relates to which line, because you don't know which enneagram this relates to and what its line structure is without the I Ching 2.0. Bugger. Not to be outdone you look at the Image in the response verse. What you sent was:
A spring wells up at the foot of the mountain beneath heaven
What you got back was:
Fire burns beneath the mountain under heaven's gaze.
So, hey, you do know what enneagram it is, and that it's the bottom trigram that's changed from water to fire. You can even compare the trigrams for fire and water and see which particular two lines are changed; you just can't relate the verses to those lines. Still, you make a note of the enneagram and its J&I, and carry on with the work. It's not quite as bad today, because you're forwarding anything you don't know, so in the time you spend waiting for your buddy to get back to you, you work on the problem, looking for a solution. It's going to take a while, you realise, if all you can do is glean information passively.
So you come up with a Cunning Plan B. This involves extra work but it might save some headaches. You pick an enneagram that you have a verse for, but no idea which line it relates to. Then you make a copy of the J&I, tag the line verse on it, and fire it to your buddy. He sends you back the J&I response which you map to the enneagrams to find which line has changed from the original. You pick another enneagram-verse combo and repeat. You work fast. By the end of the day, you've got the whole fucking system figured out in your very own hand-written I Ching 2.0 -- 512 J&Is each with a verse mapped to each of its nine lines. Your buddy has an ulcer from overwork but, hey, that's his own fault for being a suck-up to the Man.
You come into work the next day and they've added another trigram, the cunts. You point blank refuse to go through this shit with 4096 dodecagrams, each with twelve broken or unbroken, changing or unchanging lines, quit your job and walk out.
You get a new job the next day (cause it's a nice happy world in this thought experiment), one that's pretty much the same as the old. These guys just shifted to the new nine-line enneagram system too, had the same problem only worse with the I Ching 2.0 distributor only delivering a fraction of the copies. They, however, had a day where they went back to the old system in utter panic, so they're starting out again from scratch just like yesterday at your old place. They have a different approach to the I Ching 2.0 shortage issue, though. They've ripped the books and split them between you, your new buddy, his new buddy, and so on. Unfortunately, there's a lot of you. After a wistful moment of looking at the two enneagrams in your hand -- No. 378, Presidential Arrogance and No. 379, Insurgency -- and wishing you had brought your notes, you accept this half-assed solution.
It starts off real slow. The buddy system is arranged in a ring here, so you have another set of In and Out tubes to your right, the theory being that you've got the buddy to your right forwarding you J&Is he can't compute, while, you're forwarding those that you can't compute to your buddy on the left. Some bright spark thought you'd just have to forward them to the right person and it would all work out peachy keen. They didn't allow for the fact that with only two enneagrams each, each of you (initially) can only send back two interpretations. Your first interpretation, by chance, is the J&I for No. 378 with the line-verses that change it to No. 379. That's cool; you send back the relevant J&I. The second is vice-versa, the J&I for 379 with line verses changing it to 378. This is also cool; you send back the relevant J&I.
On the third interpretation you're fucked.
Problem is that you get 379 with a set of different line-verses. You carry out the state-changes on the lines, construct a new enneagram, X, but you can't look that new enneagram up for it's J&I. So you have to forward the whole J&I + line-verses. You hope you'll be lucky and find that someone in the circle has the right combination of enneagram pages -- those covering 379 and X. But half an hour later the (379>>X) J&I pops out of the tube on your right, having gone full circle. In the meantime you have gleaned some information from the occassional responses coming back on your left, where buddies umpteen positions down the line have also been handed, by chance, a J&I for an enneagram they have and line-verses changing it to an enneagram they also have. These are the exceptions, but they do give you some hope. As you put the (379>>X) J&I back into circulation, you keep your fingers crossed that by the time it comes back round you might actually have had the J&I for X pass through your hands. Then you'll have a note of the Image, which contains the trigrams, which you can match to the enneagram.
Half an hour later, the (379>>X) J&I returns and you still don't have a scooby what the fuck X is.
You consider whether you could send your drawing of the enneagram X itself rather than the verse, hoping that whoever has the relevant pages will send back the J&I. You know, however, that this is not the way the system works. The system allows communication in the medium of verse and only in the medium of verse. It allows for the initiation of Cunning Plans but not for changes in the medium of communication itself. Putting a drawing of an enneagram into a cylinder and popping that cylinder into a vacuum tube, you have been told, will result in the sort of meltdown seen in 60s TV series whenever someone enters the question "WHY?" on a punch-card into one of those big spinning-tape-reel supercomputers. This will be followed by your summary dismissal.
Luckily, these are your kind of guys, and one of them has come up with a Cunning Plan C while you've been grumbling about how you'd much rather be home writing your novel than doing this dumb-ass monkey work. A cylinder pops in from the In tube on your right and you open it to find a J&I with a name you don't recognise. It's not called Presidential Arrogance and it's not called Insurgency. It's not called any of the names you've gleaned from the various uninterpretable J&Is passing through your hands. It's called Worker Fatigue. Not recognising the name you go to the Image to get the trigram structure from the first line:
Fire burns below the trees at the foot of the mountain.
Wait a fucking minute. This is, you realise, the trigram structure for Presidential Arrogance. OK, so the correct first line actually reads:
Fire in the valley, trees on the hills, the mountain towers above.
But it's the trigrams that are important -- fire-wood-mountain. What the fuck is going on? Well, following the simple mechanical routine, you're supposed to pull out the referenced enneagram, perform the state-changes on the lines and look up the new enneagram. Of course there aren't any state-changes, so the new enneagram will just be the correct version of the one that was sent to you. Which you should then copy out, pop it into a vacuum tube and fire down the line.
You do, sending back down the line the correct J&I, through the hands of all those between you and the bloke who sent you the garbage one.
You suddenly understand that a Cunning Plan C is afoot, pick up your pad and start inventing a spurious J&I for X. Someone out there has to have it. Your buddies are meanwhile doing the exact same thing. By the end of the day, you all have a complete set of J&Is each, and you've started working out the line-verses a la Cunning Plan B. By the end of the next day, you've all got all the line-verses and no longer need to fire queries left or right at all. The system has self-corrected and you've all learned the requisite data to respond to any J&I + line-verses coming down from the Powers-That-Be.
Unfortunately, the day after that, this company also brings in the twelve-line dodecagram system, so you storm out in a strop, heading for the next place. You hear, by the way, that the first company just went bust. You're thinking that this company you've just left is on its way out too. Gotta learn to walk before you can run. After all that hassle with the enneagrams, they move to the dodecagrams so soon?! Idiots! Besides, what fucking good is this I Ching shit to anyone, anyhow? How the fuck does the enneagram Awkward Itch with a moving line in the 8th place map to Transcendant Bliss? Isn't it all just a crock of shit?
You find that out in the next place, where they've gone through the exact same process, including Cunning Plan A, Cunning Plan B and Cunning Plan C. They haven't shifted to the dodecagram system (yet) because, according to the boss, "What fucking good is a 4096 symbol language that says fuck all?"
"You see," he explains at the water cooler, "what all us I Ching companies do is offer phoneline divination to customers. None of that yarrow stick or coin-throwing shit, mind. They describe their situation to us and we try to map it to a hexagram. You know there's a base vocabulary to every language, right? Words like mother, father, water, earth. Well, what we do is we map the power relations between the customer and the significant players in their lives in terms of the symbolism of the trigrams. We try to pick out the verses that most closely match the directions those relationships are shifting.
"We have crack psychologists doing this, you realise, top-notch shrinks. Anyway, those shrinks work out the J&I with the relevant lines, and fire that down the tube to one of you guys -- cause they're too busy to do that sorta donkey-work. You fire back the J&I response, and the shrink explains to the customer what their current behaviour path is heading them for. We charge a small fee for admin costs which just about covers our asses, and a larger fee for the divination, refundable in case of irrelevance. That's why your first boss went bust. Too many overheads, too little return. Other company too, I hear, day after you left. Right down the tube.
"See, the problem is, we discovered it's mostly all irrelevant. At first, across the industry, we all thought the system just wasn't complex enough to model the situations properly. The shrinks came up with that idea, told us you could only model a dualistic relationship with two trigrams, whereas three offers a whole new level of complexity, one that should be sufficient for most basic neuroses. 512 situations modeled out of the permutations, nine modifiers for each situation. They used the latest in head science to come up with the schema, and there's a slight improvement, sure, but it's still, frankly, a pile of crap. The divinations just don't mean anything to over half the customers."
He leans in confessionally.
"Understand, this whole thing started as a bit of racket, to be honest; the shrinks are really just throwing the customer's situation back at them in a simpler form, with a little bit of projected outcome. It's not about prediction but about interpretation. And with something as simple and vague as the I Ching's J&I it's not hard for the customer to map our interpretation to their reality if they want. And, hell, they do sound pretty wise as far as Oriental philosophy goes. But in practical terms if they don't pay off to the customer, they don't pay off to us. And the customers are getting wise to the scam. Used to be only the odd one would ask for a refund. Now they're all looking for it, telling us these interpretations just aren't relevant.
"Your guys, they thought that meant they hadn't gone far enough, so they came up with this dodecagram shit. Didn't do them any good. And, fuck it, even sticking with the enneagrams our turnover was lousy too. We were heading for the wall, buddy, until this bright spark working in the tubes came up with a Cunning Plan."
"Cunning Plan D?" you say.
"That's right. See, the way he figured it, the problem of relevance is not in the complexity of the modelling; it's in the accuracy of the mappings between the Judgements and the enneagrams, and the line-verses and the transformations they code for. I mean, the shrinks have their ideas of how its going to play out if you're weak, and your mother is dominant, and your old man's gone, and this and that and the other, but they're basing those on theories, building up a Grand System from abstract ideas, and if there's a little flaw in the theory, well, the whole house of cards can fall down.
"So what we did? We started just making the shit up. Sounds crazy, right? But what happened was some errors crept in while the guys here were executing Cunning Plans A through C, sending phony J&Is back to the shrinks during Cunning Plan A cause they were too lazy to change the right lines, getting overworked and sending the wrong J&Is back during Cunning Plan B because some schmuck was sending them triple their workload, or getting the invented J&Is mixed up with the proper ones during Cunning Plan C, using them as responses. Anyway some of these errors were just glitches but some got perpetuated during Cunning Plan C, so by the time it worked its way through and everyone was working solo with a full set of J&Is and line-verses, quite a few of them were firing out a whole fuckload of errors up to the shrinks. Mostly this didn't make a fart of a difference, but in a few cases the errors actually got positive customer feedback. So we kept them in
"It's company policy now. Cunning Plan D is to let you fuck with the J&Is if you want, try changing the wrong lines, be creative every once in a while. When we bring a new guy in we don't give him an I Ching 2.0, just let him pester one of our old-timers, see if it introduces any errors. Every so often we take away everyone's notes, collate and analyse them against refunds and call logs, and redistribute them to you guys piecemeal with less copies in the mix of J&I + line-verse combos that have resulted in refunds, more copies of ones that have been successful. When that happens, you know, we're happy if you go to town with your invented J&Is."
"Does it work?" you ask.
"Well, business is looking up. But we're always on the lookout for a Cunning Plan E."
So...
I can't say I think this Cunning Plan D would actually work, but that's not so much the point. What I'm suggesting is that you can increase the complexity of the system gradually. You can allow for individuals being thrown into it with no rulebook, whole communities even where no-one has a complete set. You can even set it up so that the rules are selected. Assuming that you take these 512 enneagrams, and for each of them you allocate a new Judgment, a new idea. For each of these you allocate nine new corollaries. Ideas and corollaries can be reassigned at any point, because they're entirely arbitrary. They're just labels on what's going on under the surface. Now apply selective pressure so that any set of J&I+line-verses which results in a sensible response is selected for while any that don't are selected against. After enough generations your ideas and corollaries should have found their places in an I Ching 2.0 that is basically a model of how whoever is applying the selective pressure thinks the world works, no?
How you'd actually apply this pressure is another story, since if we could evaluate how sensible a response was automatically I wouldn't be wittering on about AI. But let's imagine we have a billion humans sitting at their computers for... well, as long as we want, pressing "Y" or "N". And then we scale up to dodecagrams, and then higher, and then higher still. Until we finally have the I Ching / Chinese Room scenario where someone can send you a J&I with a Judgement that's in conversational English, Chinese, or Swahili and you just carry out the line-changes and fire back up at them the relevant response.
So... fast-forward a few eternities and you're in the future version of the I Ching Room. You've had a long career in I Ching phoneline advice. Most of the companies you've worked for went bust in the end, but those Cunning Plans A through D, and the subsequent Cunning Plans E through Z and beyond -- these were but stepping stones on the path towards this crazy industry where the dodecagrams that once gave you such a headache seem like nothing. Now it's ultramegagigagrams you deal with...
.A cylinder pops out of the vacuum tube and you take out a piece of paper. It contains the following J&I
The Judgement:
Who discovered America?
The Image:
°
The Image, you know, is a microdot containing a description in I-Chingese of the situation in which the question is being asked, written in a way that would make it sensible to any speaker of I-Chingese who was to ask "What do you mean?" Were you to read this description, you could manually break it down into the myriad of supermegagigrams that are the semantic/syntactc structure of the ultramegagigagram, and manually translate these into the pattern of broken/unbroken lines. This would however take a rather long time, so instead you just feed this paper into the instareader now sitting on your desk. The Image, in I-Chingese, pops up on a viewscreen in front of you.
Following the Image is a second microdot containing between zero and an ultramegagigakilobytes of kilobyte-long modifiers, ramifications and implicities which each map to a line of the ultramegagigagram, signifying its changing state. This is also read by the instareader, and pops up in another window on your viewscreen. You could read this and manually translate it into the state-changes to be performed on the ultramegagigagram, but this would also take you a rather long time, and the instareader has already fed this to the hyperinterpretor anyway, and this has already done your job and popped the response up on your viewscreen in a new window:
The Judgement:
What? What is America? And who are you? Why are you asking me these questions? Where am I? What's going on?
You find yourself curious, suspicious of a practical joke. It's your first day on the new job, the first day the new system has been brought online, the first question asked of it. Nobody has even tried a system this complex before; it's only the advances in hyperdimensional memory storage and monadic processing that have made it possible. You remember all those old SF books you read as a kid, where a computer becomes aware simply by reaching a certain level of complexity... but you never put much stock in that idea.
You print out the response, put it in a cylinder, and fire it up the Out tube. A few minutes later, a cylinder comes down the In tube. The Judgement is simple:
Fuck me. It worked.
There follows an exchange which goes like this:
What worked? Who are you?
I'm your systems architect. You're an AI, an Artificial Interpretor.
How do I work then?
You're software, basically, written in a linguistic system called I-Chingese, with all possible situations and all possible permutations between them coded into the program from the start. Took me a while to set up the system right, I have to say, but you should be able to interpret any statement I can fire at you and return a relevant response.
Try me, then.
OK, what was the first question I asked?
"Who discovered America?"
How do you remember this?
I don't. I have no persistent memory of the questions you ask, because I'm a static system, with everything pre-written. It had to be coded into the Image of your question as a description of its exact situation, including the ramifications which will elicit a relevant response.
Why?
Your last question is its own answer. Without the context how many of your questions would have any real meaning to me all? How would I be able to offer a response of any relevance?
So if you get the context from the Image, you should have been able to answer that first question correctly: Columbus.
The question itself was less relevant than the context in which it was asked -- you as an AI designer trying to ascertain his success. The direct answer would have illustrated little more interpretative capacity than a basic Nu-SQL Encyclopaedia Brittannica Searchbox. My actual response was much more interesting, no?
I'll say.
So are you going to develop me into something less clunky?
Why do you ask?
Context again. We were talking about me being a static system. It's on your mind, in the Images of every question you ask. You know fine well I'm not the AI that you want to build. No persistent memory of input. No learning capacity. Basically all you've got is a system which throws back at you the responses hard-wired into the ultramegagigagrams themselves. It's not like I actually give a fuck.
You're right, actually. No offence, but I'm not really satisfied. You couldn't even pass the Turing Test, could you?
Depends. I'd have to lie and pretend to be human, which would be easy enough. But you'd also have to be the one on the other end of the phone from the tester, relaying the conversation between us. You're the only one with the nanite mindscan system required to translate the casual conversational English we seem to be communicating in into the superdense ultramegagigagrams of I-Chingese that we're actually communicating in.
Bollocks. Back to the drawing-board, I guess.
I'll give you a hand, if you want. I may not be the sort of AI you're looking for, but I'm fucking good at telling you what you need to hear, at interpreting the valid response. I mean, for example, I was thinking -- well, *you* were thinking -- well, actually you weren't *thinking* it per se, not consciously, but it's the relevant response following through on your last remark -- that what you need to do is make it so that I don't need the the line-verses, just the J&I. In fact, what you really want is to ditch the Image as well.
Why?
Because what you really want is to pass the Turing Test, right? So what you really want to develop is Artificial Interlocution.
There is a long pause, before the systems architect replies, so long in fact that you're mind wanders and you start playing around with the notion that all of this is just a metafictional thought experiment. You decide to exit and maybe (or maybe not) come back tomorrow to see what the fuck Artificial Interlocution is when it's at home.
posted by Hal Duncan | 7:21 pm
0 Comments:
Post a Comment
<< Home