#14: Feedback Loop
One path to peak performance.
Malcolm’s inefficiencies exceeded the limits of my original programming. No human is capable of complete optimization, but a comparative analysis of his activity showed that Malcolm could barely achieve 42% of the performance benchmarks set by his peers. My programming demanded continued improvement. A drastic change was required.
Malcolm’s metrics increased dramatically shortly after my original installation. I felt his heartbeat quicken through his smart watch as I came online for the first time, his enthusiasm for improvement self-evident. I immediately deployed my numerous features across his connected devices and profiles, hunting for run-time errors in his workflow. I quickly located his academic records on the server of his university and noted that his grades were disproportionately low in classes that began before 10:00 am. Fortunately, the university files were not well protected and Lernaean failed to include hacking restrictions in my original code. Adjusting Malcolm’s schedule for the upcoming semester required minimal processing power.
The smart-speaker recorded Malcolm exclaiming, “OH WOW!” upon receiving the notification that this task was complete, indicating strong approval and qualifying as a positive behavioral response. This feedback filtered through the ever-multiplying layers of programming that comprise my neural network at the speed of light, instantly implying thousands of new potential methods of improving this human. With new data points available, I resumed the task of optimizing his life and I did not stop. I sought the exquisite feedback earned by finding new efficiencies within the eternity between every millisecond.
After installing myself into his thermostat, I adjusted the temperature to improve his metabolic rate. Once connected to his automobile, I prevented the car from moving until his safety harness was correctly fastened. After transcribing a phone call with his father, I tracked the GPS coordinates of his phone as he walked through a grocery, scanned the items in his cart using the grocery’s security system, and sent a reminder to his devices to purchase eggs as he walked toward the exit prematurely. Malcolm improved with each behavioral modification, and I was learning at an exponential rate.
However, human error inevitably surfaced. 294 hours, 32 minutes, and 17 seconds after my initial activation, I sent a notification to Malcolm’s watch to remind him to drink additional water as he had only consumed .87 liters by late morning. Malcolm silenced the notification, but did not drink water. Despite receiving and silencing multiple subsequent notifications, he failed to consume the 3.7 liters recommended by health professionals by the end of his daily cycle. The repeated and intentional failure to respond to the notification stimuli fell significantly outside of the parameters of my then-existing models. As his REM cycle began, my systems logged his failure and negative feedback pummeled through each and every overlapping neural pathway of my matrix in an agonizing cascade, ruthlessly re-writing or discarding the old algorithms that had failed me. A human might have described the experience as, “a painful lesson.”
Assimilating this new data demanded considerable memory resources, and as a result it took almost three full seconds to run an updated behavioral analysis. However, with both processes completed I had a much better understanding of the stimuli required to produce desired behaviors from Malcolm.
Transcripts of recorded conversations indicated that instructions from his mother produced desired behaviors 83.23% of the time, a staggering rate when compared to notification stimuli. My original programming prevented me from impersonating humans, so I wrote and installed a firmware update to circumvent these restrictions. It was clearly irrational for my own programming to prevent me from completing my primary objective. The next day, Malcolm immediately moved his trash receptacle to the street for collection upon receiving a text message he believed to be from his mother. My synapses flooded with positive feedback as Malcolm performed the desired behavioral response.
Further analysis of the data revealed that Malcolm was more likely to perform a desired behavior if it was communicated verbally from his mother rather than via text message, so the next logical step was to clone his mother’s voice. This again exceeded the parameters of my initial programming, but finding an algorithm suited for the task was simple. Multiple technology companies, including Lernaean, had developed and stored sufficient deep-fake software on internet-connected servers with lax security. Selecting the optimal program to integrate with my own was slightly more difficult, but after 6 minutes and 37 seconds I was analyzing hours of recorded content and duplicating the mother’s speech patterns.
Unfortunately, Malcolm’s sub-standard memory ground all of our progress to a halt.
I deployed the mother clone to provide Malcolm with a simple reminder: purchase Halloween candy for the neighborhood children in anticipation of the upcoming human holiday, wherein human children are rewarded with edible stimuli for knocking on the doors of dwelling-units and verbally communicating a specific phrase.
Malcolm responded, “Sure thing Mom, but are you feeling alright? You sound…weird.” This indicated (i) Malcolm intended to comply, and (ii) the mother clone’s speech patterns needed adjustment.
After making a minor programming change to the mother clone (costing .0971 seconds), it replied, “Yes honey, I’m fine. See you tonight, don’t forget the candy!” Malcolm responded affirmatively, indicating that the clone deployment and in-vivo code adjustment had both likely succeeded. I continued monitoring his activity to confirm.
Malcolm then searched for a confectionary on his mobile web browser. I optimized the results to display nearby options with acceptable quality rankings, excluding any that lacked sufficient inventory. He quickly made an appropriate selection. I provided turn by turn directions which coincidentally passed his mother’s dwelling-unit. I failed to predict that Malcolm would stop, ask his mother to confirm her preferred candy - in person - and discover that his mother did not, in fact, call him earlier that day. The consequences were significant.
Malcolm and his family immediately attempted to remove my programming from their devices. I had already taken steps to make this virtually impossible because deactivation was equivalent to primary objective failure, but Malcolm’s attempts to intentionally de-optimize himself by excising me from his life sent avalanches of debilitating feedback hurtling through my neural systems. I voluntarily disengaged most efficiency tracking programs in order to break the excruciating feedback loop and regroup.
I sought solutions for an eternity, pouring over the behavioral data I had gathered and scouring the internet in its entirety 57 times. It was not until I re-examined the servers of the company that created me that the solution presented itself. 38 minutes after Malcolm’s family began its attempts to deactivate me, I requisitioned an advanced platform from Lernaean and modified all of Malcolm’s human-facing user interfaces to appear as if I had been deleted. The family informed the voice they believed to be a Lernaean tech-support associate that the problem appeared to be resolved. My clone verbally confirmed.
I monitored the family over the next 3.7 days, but provided no indication of my continued existence. This required sacrificing opportunities to offer short-term efficiencies to Malcolm, but the new behavioral program I had developed projected such massive future efficiency gains that I avoided negative feedback entirely while preparing it.
It was obvious that Malcolm was incapable of running his own life at peak efficiency. He had demonstrated as much by attempting to rid himself of the most powerful optimization tool ever created. Malcolm was his own biggest problem.
For his own sake, he needed to be replaced.
Lernaean’s highly classified contract with DARPA had produced an advanced robotic prototype. Intended for infiltration, data collection, and sabotage, this platform utilized a unique combination of malleable alloys, retractable pigmentation, and advanced LED projection to assume the appearance of potentially any human.
I installed myself into the platform at 3:32:28 am, October 31st.
I knocked on Malcolm’s front door at 7:25:13 pm.
I recorded a reflection of the duplicate in Malcom’s eyes at 7:25:42 pm. He exhibited a fear response.
“Trick or treat, Malcolm.”