Stochastic Bookmark

abstruse unfinished commentary

about correspondence

3.8.05

Inside Job 2

{Part 1 here}

The above transcript comprises the complete log of the AUTOLOGOS project's initial run. The project designer proposed modifications to the genetic algorithm's medium, filtering certain "post-modern" authors from the data, but the procedure was determined to be unreliable. As the risk of further system interruption is considered unacceptable, and funding is inadequate to develop a standalone version of the experiment, the AUTOLOGOS project is hereby cancelled.

* * *

Jeremy crumpled the project cancellation notice, his notice, in one hand; thought better of it, flattened out the paper, read it again. Yet another project binned. At least this time no stack of cartons had appeared by his desk, no security guard hovered to monitor him packing personal belongings. This employer instead relied on the patent and copyright provisions in his contract, reserving the right to demean him later should they deem it necessary, or expedient. Not that anything would come of the strides he'd made on AUTOLOGOS -- they had no idea. The plagiarism detector, Cogginance, would still peruse student papers, and the idiomatic lexical parser would survive in the Peri-Phrase translation software, but the rest of his work would be duly archived and consigned to a musty cellar pressed into service as a storage facility. They probably didn't even purge his project files when they rebooted the system from backups; his program could terminate and stay resident, but he had to go.

He'd taken a salary cut to accept this post as Associate of the Computational Linguistics department. No Professor at the end of the title; merely a consultant, only paid less, though not less than the poor grad students struggling to make stipends meet, nor for that matter many of the adjunct profs (an enduring source of resentment from that quarter). At the time, he rationalized the decision by telling himself that this was an opportunity to work in expert systems, unconstrained by profit motives transmuted by corporate politics into perverse incentives; that he'd be reinvigorated by the spirit of free inquiry of the academic community; and that he'd make good use of the university's facilities and the privileges to audit classes. Instead he found that the politics were more pervasive and convoluted, not being ennobled by collective greed; that the experimental spirit had long departed, save where sex and drugs were concerned (hardly scientific, nor just the province of the student body); and for the rest -- when could he find the time? Three strikes.

Phil stuck his head in the office doorway. "Program terminated? I guess it's not exclusively publish or perish". Phil was a post-precocious graduate student who everyone knew would go far; it was just a matter of degree. Jeremy caught himself wishing he'd hurry up about it. "Doctor Schotz was incensed. Seems that hit-and-run of yours totaled a Monte Carlo simulation he'd been running for days -- all the data cached in memory, no way to recover. You broke the bank!"

"I guess that explains the board's quick decision", Jeremy said.

"Well, his short-term memory ain't what it used to be, but he can still hold a grudge."

"Hey, save the one-liners for the code, okay?". Jeremy had enlisted Phil in AUTOLOGOS, taking advantage of Phil's casting about for any excuse not to finish his thesis. It was just a matter of getting him interested. Phil knew all the cheats, how to make the most of un- or mis-documented system behaviors, which he called hexapedes ("They say that actions speak louder than words, but words do more than they say", he'd say). Several professors had had the benefit of Phil substituting one or two cryptic lines of code for pages worth of subroutines in their programs. His skill in making electrons dance to his beat had led to multiple extensions of his fellowship; it seemed he wasn't the only one interested in not completing his degree requirements.

Phil leaned against the doorpost. "'Fraid you won't be getting anywhere near a computer around here", he said. "Schotz had the sysadmin suspend your account right after they brought the system back up. He posted a message to all the staff not to let you on through their accounts -- he even made me change my password in his presence. Bastard."

"I've seen worse", Jeremy said. "It sure didn't help that my sponsor is off at the ILA symposium. It's not like I took the groves of academé for some kind of paradise. All the same, it hurts to get bounced. We got really close ..."

"Hey, I think I know what went wrong with AUTOLOGOS", Phil said.

The idea of computer-generated prose had been around almost as long as computers themselves (longer, if the automatic writing and word games of the Dadaists qualify; longer still when the Kabbalists are considered). It was a natural outgrowth of the conceptual mechanism: Swapping words around in memory according to a set of rules embodied by stored procedures (encoded in "higher-level" programming languages, then translated into machine instructions). An element of indeterminacy was injected with random (really, pseudorandom) number generators to govern the selection of words.

But the first claimant was a pretender. Racter's publishings had been more than half-constructed beforehand, the program madlibbing substitutions into prepared text. Its novelty had deflected serious inquiry into its method, as had famously been the case with supposedly mechanical precursors that played chess or promised perpetual motion. The effect of its production nonetheless served as inspiration for more sophisticated progeny such as the Post-Modern Thesis Generator and the Computer Science Paper Generator. In any case, whatever meaning the result might hold was only that projected onto it, imposed by the reader.

A more earnest effort was made in the Travesty program, borrowing from cryptography the notion of frequency tables, extended beyond Etaoni Shrdlu to encompass groupings of letters. Travesty would match up the last, say four, letters of what it had written so far, and select the fifth based on how often which letters followed this tetrad in the source texts. Rather than compiling frequency tables, the program would drop into the source texts at random and scan forward until it found the right combination of characters. Proceeding letter by purloined letter like an arrant thief, it built up its own output. If the length of the letter sequence was too short, the text that emerged seemed only a cipher; too long, and whole swaths of the source text would be replicated intact. But a middling length gave rise to gibberish recognizably in the style of the original author. Refinements to this approach included substituting words for letters, and multilayered weighting functions over different sized groupings, with some modicum of sense emerging.

AUTOLOGOS extended the concept further -- not to larger grammatical units like sentences, but to lexical definition, metaphor and symbolism, idiom and etymology. A hyperlexicon provided structure for these connections, filtering (in parallel with a syntactic checker) words and phrases randomly chosen from the source texts. To organize these diverse materials, Jeremy reused heuristics he'd developed for self-ordering hierarchies in programming the game of Go (which he'd in turn adapted from image-processing software); rather than accreting prose linearly from beginning to end, the program would assign a provisional hierarchy to its output and progressively fill in the gaps. Jeremy then put into place ad hoc rules for iteratively scanning these associative maps, winnowing out the most obvious and far-fetched. He fine-tuned the rules to preserve the overall coherence of the output while ensuring that it also conformed to linguistic measures of statistical regularity. After Jeremy explained what he was doing, and how, Phil helped set up a parser to preprocess on-line texts from the Gutenberg Project, Bartleby, and other repositories into a uniform format to provide an extensive reservoir of text to sample (Phil dubbed it Reader's Digest Condensed Soup), and added internet search agents to the sampling routines. Together they developed subroutines to preserve the continuity of grammar, style, and voice. But Phil's biggest contribution to the project had been in applying multiplexed neural network techniques to abstract and interrelate the maps of linkages, both eliminating masses of unused material and permitting Jeremy much finer calibration of the composition heuristics.

"Well?" said Jeremy.

"You remember where we put the map templates?"

"Yeah, in the same directory as the program was running, so the repeated look-ups wouldn't slow it down."

"What else was in the local directory?"

"Nothing. Just the templates."

"And ... ?"

"Just the working files, and the program. Yeah, so ... wait a sec -- you're not saying what I think ... "

"Look at the output, Jer."

"I need to think", said Jeremy. "No, on second thought, I need a drink."

"Funny you should mention it", said Phil, reaching back into the hall to retrieve a shopping bag. Something inside clinked as he picked it up. "These are getting warm. Grab your stuff and let's hit the road."

* * *

Phil propped open the bungalow's screen door with his foot while Jeremy fiddled with his keys, hidden from his view by the box of files and junk he leaned back to balance. He felt for the keyhole by a method of successive approximation, found it, turned the bolt and gave the door a kick, then a nudge with the box. The door popped open, bounced against something that kept it from hitting the wall with a thunk.

Phil squeezed past. "I'll get the rest of these on ice", he said, navigating through the disjecta strewn on the wall-to-wall towards the uncluttered linoleum in the kitchen.

Jeremy found a clear space to set down the box, pushed aside some magazines and collapsed on the couch, which creaked a sullen reprimand. Phil reemerged from the kitchen, an open Bass in each hand. He gave one to Jeremy and settled in, crosslegged, on the floor.

"So you're saying the program replicated itself in story form", said Jeremy.

"Not replicated. Just described itself -- used itself as a reference point, a model. The pieces fit: dictionary as a card catalog, operating system routines recast as staffers -- I mean, job applications, fercrysake -- the metaphor-stroke engine must have been hitting on all cylinders!"

"Look, the program wasn't in the same format as the templates. How did it manage to parse itself?"

"We simplified the logical structure of the templates to look procedural, so the program could interpret it. And the structure was pretty general. Who's to say that it couldn't have applied the same method to its own source code?"

"It wasn't in the script."

Phil snorted, recovered, swallowed. "Hey, that's my line you're stepping on."

"Even if it did -- and I'm not saying it could -- what about the sequence where it's barricaded inside? That wasn't part of the program."

"Hey, we set it up to diverge from the primary sources, fold in other elements. The point of the exercise was to make up a story, remember?"

"Besides, if it could read its own source code, why didn't it read any other programs?"

"Who says it didn't?", Phil countered.

This stopped Jeremy in his tracks. Phil broke the ensuing silence. "Hey, we can figure it out later. We don't need the computer -- the experiment's not something we could replicate anyway. So ... what's your next move?".

"I don't know. I'm not sure what happened with the last one. I suppose I'll stick around -- the lease for this place runs through the end of the term."

"Let me know. I really enjoyed working with you on this, and since I've decided that come end of term, I'm outta here --"

"What, you got the thesis done?"

"I didn't say that. No, the fellowship's up for renewal, but --"

"You can't just walk away at this point."

"Stop interrupting. AUTOLOGOS was the only thing left keeping me here. I had to stick around to see how it turned out."

"But what about the doctorate? It wouldn't take much to finish your thesis -- you've got three or four to choose from. You could just polish one of them up and walk away with at least a master's."

"Nah, too easy for them to insist on pro forma requirements while they have me continue doing their odds and ends on the side. Besides, I left a little going-away present on the system,"

Jeremy's eyes widened. "You didn't."

"Yeah, I fixed the OS"

"You what?!"

"I fixed it. Something I've been playing around with. A few patches here and there -- from now on, the operating system does exactly what the documentation says it should -- no more, no less. By the book."

Eyes narrowing. "Nothing will run right ..."

"I figure it will take them quite a while to diagnose. Left to their own devices, that is."

"You didn't." Definitive, this time.

"I had you going there, didn't I?" Phil smiled. "But seriously, I am leaving once the fellowship runs out. You seem to get hooked up with some pretty curious stuff -- I just wanted to say, whatever comes next, keep me in mind."

"Thanks, but ... really, don't let the work here go to waste. I --"

"Look, I've thought plenty about it. I just wanted to let you know, you know? Hey, let me get out of your way -- you really look like you could use some down time ... sorry. Tomorrow."

{more ...}