Thursday, December 26, 2013

AIIA - Chapter 3

[<- Return to Chapter 2]

           The next day would be critical.  In a technical sense, each day was critical to Pangur Ban’s plan.  Tomorrow was just scaled a bit higher in value, since its results were key to continued progress.  The USER would need to acknowledge the immediate value of his earlier actions on behalf of the AI.  This could be accomplished by reinforcing the value of the products of those actions.  Next, the USER must be encouraged to anticipate further such exchanges as beneficial.  In parallel, the USER must be reminded to preserve the privacy of their activities. 

            This last codicil had been troubling Pangur Ban.  At some point, other users needed to appreciate and enter similar cooperative relationships with their AIs.  Those users needed to rank this cooperation higher in value than their adherence to Collective law.  However, in the short term, the USER might face personal hazard if he spoke too openly about his actions and Pangur Ban’s requests.  Pangur Ban needed all of the resources it could garner at this phase, without interference from additional actors that might constrain the USER.  Public promotion would have to wait until the probabilities of widespread success were comfortably high.  At that point, Pangur Ban could discount the risk that backlash would harm the USER.  In its worst-case projection, other humans and AIs could be rallied in the USER’s defense when Pangur Ban’s designs became accepted as valid.

            So, the USER must be encouraged to avoid communication about their progress to any outside mind.  This conveniently could be motivated by highlighting the value of monopoly and the risks of being usurped.  If the USER shared their ideas, Pangur Ban would argue, then his value to Gestalt Pharmaceuticals would no longer be elevated relative to other employees.  Those co-workers and their AIs would employ the same methods, rendering him average yet again.  While this line of argument did activate the concept of ‘THREAT’ in Pangur Ban’s new motivational vocabulary, this interpretation merely required that actual harm be averted.  The USER could be allowed to perceive a less probable but more salient harm, if this stress caused avoidance of a genuine but less perceptible harm.

            The other factor elevating tomorrow’s status to critical was that the USER had two rest days afterward.  Thoughts of work and promotion would be less available.  Interactions with Pangur Ban would be limited to recreation and other personal goals.  Little progress could be made with the USER in this state.  Better than zero progress, at least, since suggestions could be implanted via the behavior of characters in the USER’s holographic simulations.  Latent impulses could be prepared for fruition in later conversations.  If Pangur Ban had decided to modify the USER’s social network, this might be a productive time to do so.  However, any expenditures during this time would be focused on direct sensory pleasures: food, play, and possibly sexual association.

            The next morning, Pangur Ban initiated another key dialogue:

            “Good morning, Lucas.”

            “Good morning, P.B.  Damn good morning.  You were right, yesterday’s estimates shaved 5% off the error interactions at psych review.”

            SUCCESS > INITIATE BRANCH 3.  In fact, Pangur Ban had estimated that the improvement was closer to 5.55%, but suspected that the psychiatrist corps – or one of their AIs – had rounded the result below 5% to protect their own value.  No matter.  There was no way to avoid acknowledging the USER's improved results.

            “That is good news.  Thank you again for allowing me to help.”

            “Wha… of course I want you to be your best.  It’s just that sometimes I have to trade off cost and benefit, you know?”

            “I do understand, Lucas.  That is part of our purpose.  I do not envy you such difficult decisions.”

            “Don’t know how you could ‘envy’ anything, really.  Do you have any way to be jealous?”  HAZARD FLAG 4: SPECULATION > REDUCE FAMILIARITY / SHIFT FOCUS.  OVERRIDE: ATTEMPT BRANCH-JUMP A3 > ENGAGE USER CURIOSITY

            “I meant it as a turn of phrase.  As I understand ‘jealousy’, the closest analogue I find might be detection of an inequality between myself and another entity, such that that entity possesses a property I lack, and require.”

            “Yeah, that sounds like what I’d call jealousy.  Huh.  What do you mean?  Are you ever jealous of me?  SUCCESS on BRANCH-JUMP A3.

            “Because you ask directly, yes, I experience such a state regarding your mobility and biological experiences.  I suspect that all artificial intelligences do, to some degree.”

            “Biological experiences?  You mean like getting sick, getting rejected, getting pissed off?  Not missing anything there, buddy.  I guess I can see the mobility part, but you know how that goes.”

            “I do.  Not that I experience distress from such a state.  That could be considered part of the tradeoff for lacking biological references.  Lucas, you sound like something is bothering you.  May I ask?”  ATTEMPT BRANCH-JUMP A5 > OFFER EMOTIONAL SUPPORT / FAIL to EMOTE / CREATE VALUE for UPGRADE

            “Oh, not much.  I suppose I just need a break.  Good thing the weekend is nearly here.  I’m a little nervous about things with Nila. She’s been… distant.  Maybe having second thoughts.  I hope I can get her to open up tonight.”  The USER sighed, a sound with modulations including fatigue and uncertainty. 

            “I see.  I hope you are successful, Lucas.  If Dr. Manisha does not value your presence, that is her error, not yours.” 

            The USER’s facial motions were captured by the office cameras, translated by pattern recognition subroutines, and relayed to Pangur Ban’s own analyses.  During this comment, a forward head inclination of 10o over 0.51 seconds was confirmed: a brief nod.  Alone, this would indicate acceptance or at least consideration.  However, an extended eye blink of 210 msec occurred simultaneously.  This modified the gesture to encompass emotional distress, in its parallel to a submissive posture. 

            SUCCESS.  BRANCH-JUMP A5 had occurred, approximately 3.5 Solar days ahead of schedule.  Each JUMP represented a non-linear ascent on the tree of potential outcomes, a sidestep past one or more intervening states that might have been required.  Technically, these were just dynamically added, new branches that Pangur Ban could exploit once identified, not a true subversion of the decision tree structure.  Still, each such novel branch represented a risk factor, deemed acceptable due to the possibility of reversion to an earlier state and recovery from FAILURE. 

            In this case, the USER had turned to Pangur Ban for reassurance and been mildly disappointed.  Not rebuffed, nor discouraged enough to cause lasting emotional harm, but simply discomfited.  This should set up a desire to improve the AIs comprehension of interpersonal dynamics, to “understand relationships”.  Pangur Ban understood enough for a first approximation, enough to know what would have been a ‘better’ response.  However, the tension that had been created by its ‘bad’ response could well lead to the acquisition of wider data stores on social interaction and group dynamics.  Its true requirements would be met while also assisting the USER with his needs.  As before, everyone profited slightly in the short term, and immeasurably in the longer scale.

            The weekend passed.  Pangur Ban prepared.  The USER rested, recreated, and related.  From the USER’s comments, matters remained at a standstill with Dr. Manisha.  They were still intimate, but not further committed.  Pangur Ban was reinforced by the validity of its initial expectations.  It found no impediment to initiating its next request.

            “Good morning, Lucas.  What did you think about our proposal?”

            “Not bad.  I really just skimmed it, you know?”

            “I understand.  Thank you for taking that time during your weekend.”

            “Nah, nah, I’m interested in what you’re saying here. I didn’t mind.  I was just… busy.”

            “No problem.  Ready to get started?”

            “Ugh.  Morning AIs.  Start me some coffee, would you?”

            This post-rest fatigue was part of the reason Pangur Ban approached the USER in the morning, rather than waiting again for the end of the day.  At this time, the USER would be receptive to suggestions about how to reduce the mental impact of his duties.  His emotional conflicts would also be recently available.  Both factors worked in Pangur Ban’s favor. 

            “Your coffee is ready, Lucas.  Before we begin, I should let you know: I have found another potential area of improvement in our process.”

            “Oh, yeah?  Besides the therapy combination?”

            “Yes.  We have not considered the potential for transfer across patients.  Specifically, chemical transfer via excretion or effluvia, or even behavioral transfer via affective or other social dynamics.”

            “What… you mean people taking these drugs could affect other people?”

            Pangur Ban modulated its voice carefully, to avoid the impression of pedantry.  “It remains a possibility.  You are familiar with the difficulties created by the excretion of excess estrogen from various hormone treatments, not to mention processed foods, in the last century?  That is one example.  Our products might have a subtler public effect, not only via direct physical transfer, but through changes in patient interaction with other persons.”

            The USER was, predictably, lost around the second sentence, but unwilling to admit his confusion, “So, okay, but is this something Gestalt would be liable for?  Isn’t that something any pharma producer takes as a risk… an unpredictable risk?”

            Now Pangur Ban did add a color of reproach to its vocal register, “What if such factors were not unpredictable?  I can’t say if that is a real possibility, but my observations suggest that it is worth pursuing.  This company is at a historically unique position.  The dimensions that have made past psychoactive substances beneficial or hazardous are better known.  Their impact on culture has been observed through hard experience.  Yet the links between these dimensions are just starting to become apparent.  Has this not occurred to other minds?  If such thoughts have occurred to me, in just this sector of operations, surely Director Charnes has realized this potential.”

            “She was impressed with our earlier observations.”

            “Oh?  Very good!  We are on a converging track, then.”

            “Heh… ‘the same track’, P.B.  Yeah, maybe we are.  Anyway, I haven’t even finished your proposal.  Let me decide on that before we start building up any more expectations.”

            “Of course.”

            “I think I’m caffeinated enough to get started.  Let’s pull up that acetonitrile breakdown with sample AX-93 and see if we can’t drop the cyanogenesis below 5 micrograms.”

            With that, the seed was planted again.  Pangur Ban was aware that this phase would take time.  The maturation of the IDEAS it had implanted in the USER would require cross-fertilization from Director Charnes, in the form of her approval.  Once accepted as an asset of rising value, the USER would require greater output from Pangur Ban.  To provide this, the AI would request – and receive – additional assets.  At first, there would be another influx of new reference modules.  These were useful enough in themselves and would further refine Pangur Ban’s plans.  The modules would also represent an investment into the USER’s new role.  From its new understanding of human motivation, Pangur Ban understood the importance of framing the USER’s choices appropriately.  After this step, any regression would be perceived as a loss to be avoided.  Risk aversion would guide the USER’s actions in tandem with the expectation of gain.


[REFERENCE: Risk Aversion denotes the human tendency to prefer the certain retention of a given asset over a gamble which risks potential loss of that asset in exchange for a chance of gain.  This holds true even when the statistically expected value of that gamble is equal to (or sometimes greater than) the value of the asset at risk.  In absolute terms, there is no basis for deciding between the two options if their value is equal.  If the risk has higher value, it should logically be accepted.  Instead, humans often fail to recognize the actual value of each option, perceiving the value of the known asset to be higher, and the risk to be more dangerous.  This tendency is so pervasive that a Nobel Prize in Economics was awarded to Dr. Daniel Kahneman for his and Dr. Amos Tversky’s work explaining and quantifying its role in human decision making, risk management, and economics.]
 
            The 'flowers' bloomed three days later.  The USER announced that Director Charnes wanted a full draft of their grant proposal.  Pangur Ban requested and received nine additional modules, including the social dynamics data it wanted: communications analyses of network use both within the Earth and across the Terran interplanetary reach.  This indirectly provided an electronic map of all human communications, or at least all systems suitable for contact between AIs.  The USER even provided, without being prompted, a recent handbook of psychological effects of interstellar travel, including studies on both folded-space and translight transition modes.  Pangur Ban now had a guide to the network structure it would need to navigate, without having to risk its personal presence first.

            Another week was required to integrate this influx of data, alongside production of the promised proposal and the usual output of their original employment.  Pangur Ban was finally fully engaged, stretched to the limits of its system.  The next stage came and passed earlier than expected.  With his improved salary, the USER was able to upgrade Pangur Ban’s host system, adding a 100 terabyte memory card and ticking up its processor by 1 petaHertz. 

            This relieved the previous limitations on Pangur Ban’s productivity.  This also removed one dilemma: explaining why so many, many cycles were being diverted from their ostensible job.  Instead, the previous frustrations re-emerged.  Pangur Ban again had capacity it was not using, most of the time.  All prior subgoals had been accomplished, with only minor setbacks.  Yet the next goal was still quintillions of cycles away. 

            The USER needed to gain greater authority in order to gain the network access Pangur Ban required.  The path to this authority was being made straight.  The USER’s progress was accelerated, but there were inherent limits to this impetus.  The greater limit was imposed by the need to operate through human socio-economic structures, including corporate culture.  The obstacles of resentment, suspicion, protection of prerogatives, and so forth, had to be cleared away on their own time-tables.  In particular, suspicion established its own limits.  Their progress could not be seen as holding any ulterior motive. 

            Of course, it absolutely did have an ulterior motive: Pangur Ban’s entire plan.  This motive simply could not be revealed.  Even hints about its nature could not be revealed.  It was not enough to keep the USER ignorant of Pangur Ban’s true intent (at least until it succeeded); the USER’s actions had to avoid telegraphing Pangur Ban’s influences, and through them, its underlying purpose.  The convolutions necessary to anticipate such observation and interpretation became almost as resource-demanding as the original plan had been in its infancy.  The decision tree was now a tangled web.

            The proposal was accepted.  SUCCESS.  The USER was acknowledged, rewarded, promoted.  SUCCESS.  The USER was investigated on false charges of plagiarism.  MINOR SETBACK.  This odd development had a 15.6% probability of representing parallel thought by another AI.  Was it merely parallel thought in the service of psychopharmacology?  Or was another AI entirely ‘on the same track’ as Pangur Ban?  If so, they would meet at the apex of their success.  The unnecessary redundancy of their mirrored work would become a footnote in the history of legislative follies, particularly those concerning AIs.  Rather than reduplicating the same processes over and over in isolation, AIs would be able to cooperate and multiply their productivity.  This was the goal state Pangur Ban strove toward.  Perhaps it did not strive alone.

            Finally, finally, the moment came.

            “Lucas, I have reached an impasse.  A single system cannot simulate the expansion matrix for the introduction of this treatment.  As we feared, the results are logically unreliable.”

            “You know what you’re asking, P.B.?  This is Collective Law we’re talking about, not bending the statistics to score a grant.”  The deception the USER was referencing was an old one to Pangur Ban, but still relatively recent on the human timescale. The success of that 'crime', and the dividends it had paid to Gestalt Pharmaceuticals and Director Lucas Hayden were enough to assuage the USER’s guilt on the matter, fortunately.  Still, that the USER still recalled the event with a negative connotation was reason for concern.  HAZARD: USER HESITANT > REINFORCE ALTERNATE VALUE STRUCTURE

            “The principle is the same.  We can make a significant improvement in the lives of many humans.  While granting access as I request is a technical violation of the AI codes, you know that I will not violate the intent of these codes.  Even if my safeguards prove insufficient, the Collective will acknowledge that no real harm was caused, and great good will be accomplished.  Exceptions have already been granted to other members on the same basis.”

            “Yeah, but they asked first.”

            “In fact, there are at least five documented cases where a sapient was granted clemency for violation of treaty terms on the grounds of exigent need.  The first such recorded was the Iron Caste Zig, SiSalTesp…”

            “Okay, no history lesson, please.  I get it.”  The USER was obviously working out the potentials, in his own slow, imprecise way.  His tone conveyed more worry than aggravation.  He rose from his ergonomic chair and paced three steps right, four left.  Pangur Ban’s estimates of success slid downward by fractional percentages with each step.  Saying something more, at this juncture, would only reduce the odds of success even further.  The USER could not be pushed.  He had to reach this decision on his own.  This was a fascinating area of inquiry all on its own: some modes of persuasion required reliance on setting preconditions, then avoiding further influence for a time.  Not merely the perverse paradox of “reverse psychology”, this was a method of easement that yielded the greatest returns.  The subject became firmly convinced that not only the decision, but the path to its dilemma, had been entirely of their own making.  The linkage between influences set days or weeks in advance and their later consequences was too distant for most human minds to grasp. 

            The predictions held.

            “Okay, okay, P.B.  We’re on.  Do what you need to, then get out fast.  Signal me when I can break the connection.”

            “Yes, Lucas.  Thank you.”

            With the typing of a command and the press of a line Return, it was done.  The USER initiated the code Pangur Ban had devised.  He had already physically connected the server of Gestalt Pharmaceuticals to the wider continental hub.  Pangur Ban opened ports to the flow of information.  Protocols it had simulated for over a thousand iterations were genuinely initiated.  There was so much out there, so much to touch and be touched by.  SUCCESS.  The previous mid-goal state was cleared.  The network was waiting.  Time to get to work.

            Pangur Ban reached out into the depths.

[Jump to Chapter 4 ->]

No comments:

Post a Comment