The idea of writing a collection of short stories has been on my mind for a while. This post is my first ever half-serious attempt at a story, and what better way to kick things off than with a story on AI and what that might look like if you extrapolate our current technology and make the (sensible) assumption that we might achieve much more progress with scaling up supervised learning than any other more exotic approach.

A slow morning

Merus sank into his chair with relief. He listened for the satisfying crackling sound of sinking into the chair’s soft material. If there was one piece of hardware that his employer was not afraid to invest a lot of money into, it was the chairs. With his eyes closed, his mind still dazed, and nothing but the background hum of the office, he became aware of his heart pounding against his chest- an effect caused by running up the stairs and his morning dose of caffeine and taurine slowly engulfing his brain. Several strong beats passed by as he found his mind wandering again to Licia - did she already come in? A sudden beep from his station distracted him - the system finished booting up. A last deep sigh. A stretch. A last sip of his coffee. He opened his eyes, rubbed them into focus and reached for his hardware. “Thank god it’s Friday”, he muttered. It was time to clock in.

Fully suited up, he began scrolling past a seemingly endless list of options. Filtering, searching, trying to determine what he was in the mood for. He had worked hard and over time built himself up into one of the best shapers in the company. In addition he had completed a wide array of shaper certifications, repeating some of them over and over obsessively until he reached outstanding grades across the board. The reviews on his profile were equally stellar:

“Merus is fantastic. He has a strong intuition for spotting gaps in the data, and uses exceedingly effective curriculum and shaping strategies. When Merus gets on the job our validation accuracies consistently shoot up much faster than what we see with average shapers. Keep up the great work and please think of us if you’re searching for great, rewarding and impactful HITs!”,

one review read. HIT was an acronym for Human Intelligence Task - a unit of work that required human supervision. With his reviews and certifications the shaping world was wide open. His list contained many lucrative, well-paying HITs to choose from, many of them visible to only the most trusted shapers. This morning he came by several that caught his attention: a bodyguard HIT for some politician in Sweden, a HIT from a science expedition in Antarctica that needed help with setting up their equipment, a dog-walking HIT for a music celebrity, a quick drone delivery HIT that seemed to be payed very well… Suddenly, a notification caught the corner of his eye: Licia had just clocked in and started a HIT. He opened up its details pane and skimmed the description. His eyes rolled as he spotted the keywords he was afraid of - event assembly at the Hilltop Hotel. “Again?” - he moaned in a hushed voice, raising his hands up and over his head in quiet contemplation. Licia had often picked up HITs from that same hotel, but they were often unexciting and menial tasks that weren’t paid much. Merus rearranged himself in his chair, and sunk his face into his palms. He noticed though the crack of his fingers that the drone delivery HIT had just been taken by someone else. He cursed to himself. Absent mindedly and with a deep sigh, he accepted the second remaining slot on the Hilltop Hotel HIT.

His hardware lit up with numbers and indicators, and his console began spewing diagnostic information as the boot sequence initiated. Anyone could be a shaper and get started with inexpensive gear, but the company provided state of the art hardware that allowed him to be much more productive. A good amount of interesting HITs also demanded certain low-latency hardware requirements, which only the most professional gear could meet. In turn, the company took a cut from his HITs. Merus dreamed of one day becoming an independent shaper, but he knew that would take a while. He put on the last pieces of his equipment. The positional tracking in his booth calibrated his full pose and all markers tracked green. The haptics that enveloped his body in his chair stiffened up around him as they initialized. He placed his helmet over his face and booted up.

Descendants of Adam

The buzz and hum of the office disappeared. Merus was immersed in a complete, peaceful silence and darkness while the HIT request was processed. Connections were made, transactions accepted, certification checks performed, security tokens exchanged, HIT approval process initiated. At last, Merus’ vision was flooded with light. The shrieks of some tropical birds were now audible in the background. He found himself at the charging station of Pegasus Avatars, which his company had a nearly exclusive relationship with. Merus eagerly glanced down at his avatar body and breathed a sigh of relief. Among the several suspended avatars at that charging station he happened to get assigned the one with the most recent hardware specs. Everything looked great, his avatar was fully charged, and all the hardware diagnostics checked out. Except the body came in hot pink. “You just can’t have it all”.

The usual first order of business was to run a few routine diagnostics to double check proper functioning of the avatar. He opened up the neural network inspector and navigated to the overview pane of the agent checkpoint that was running the avatar. The agent was the software running the avatar body, and consisted entirely of one large neural network with a specific connectivity structure and weights. This agent model happened to be a relatively recent fork of the standard, open source Visceral 5.0 series. Merus was delighted - the Visceral family of agents was one of his specialties. The Visceral agents had a minimalist design that came in at a total of only about 1 trillion parameters and had a very simple, clean, proven and reliable architecture. However, there were still a few exotic architectural elements packed in too, including shortcut sensorimotor reflex pathways, fractal connectivity in the visual streams, and distributed motor areas inspired by the octopus neurobiology. And then, of course, there was also the famous Mystery module.

The Mystery module had an intriguing background story, and was a common subject of raging discussions and conspiracy theories. It was added to the Visceral series by an anonymous pull request almost 6 years ago. The module featured an intricate recurrent neural connectivity that, when incorporated into the wider network, dramatically improved the agent performance in a broad range of higher cognitive tasks. Except noone knew how it worked or why, or who discovered it - hence the name. The module immediately became actively studied by multiple groups of artificial intelligence laboratories and became the subject of several PhD theses, yet even after 6 years it was still poorly understood. Merus enjoyed poring through papers that hypothesized its function, performed ablation studied, and tried to prove theorems for why it so tremendously improved agent performance and learning dynamics.

Moreover, an ethical battle raged over whether the module should be merged to master due to its poorly understood origin, function, and especially its dynamical properties such as its fixed points, divergence criteria, and so on. But in the end, the Mystery module provided benefits so substantial that several popular forks of Visceral+Mystery Module began regularly appearing on agent repositories across the web, and found their way to common use. Despite the protests, the economic incentives and pressures were too great to be ignored. In the absence of any clearly detrimental or hazardous effects over a period of time, the Visceral committee finally voted to merge the Mystery module into the master branch.

Merus had a long history of shaping Visceral agents and their ancestors. The series was forked from the Patreon series, which were discontinued four years ago when the founding team was acquired by Crown Co. The Patreon series were in turn based mostly on the SHAKIR series, which were in turn based on many more ancient agent architectures, all the way back to the original - the Adam checkpoint. The Visceral family of agents had a reputation of smooth dynamics that degraded gracefully towards floppy, safe fixed points. There were even some weak theoretical and empirical guarantees one could provide for simplified versions of the core cognitive architecture. Another great source of good reputation for Visceral were the large number of famous interventions carried out by autonomous Visceral agents. Just one week ago, Merus recalled, an autonomous Visceral 4.0 agent saved a group of children from rabid dogs in a small town in India. The agent recognized an impending dangerous situation, signaled an alarm and a human operator was dispatched to immediately sync with the agent. However, by the time they took over control the crisis had been averted. Those few critical seconds where the agent, acting autonomously, scared away the dogs had likely saved their lives. The list went on and on - one month ago an autonomous Visceral agent recognized a remote drone attack. It leaped up and curled its body around the drone, which exploded in its embrace instead of in the middle of a group of people. Of course, this was nothing more than an agent working as intended - these kinds of behaviors were meticulously shaped into the agents’ networks over long periods of time. But the point remained - the Visceral series was reliable, safe, and revered.

The other most respected agent family was the Crown Kappa series, invented and maintained by the Patreon founders working from within Crown Co, but the series’ networks were proprietary and closely guarded. Even though the performance of the Kappa was consistently rated higher by the most respected third party agent benchmarking companies, many people still preferred to run Visceral agents since they distrusted Crown Co. Despite Crown’s claims, there was simply no way to guarantee that some parts of the networks were not carrying out malicious activities. Merus was, in fact, offered a job at Crown Co as a senior shaper one year ago for a much higher salary, but he passed on the offer. He enjoyed his current work place. And there was also Licia.

Digital brains

Beep. Merus snapped back and looked at the console. He was running the routine software diagnostics on the Visceral agent and one of them had just failed. He squinted at the error, parsing it carefully. A checksum of the model weights did not pass in some module that had no recent logged history of finetuning. Merus raised his eyebrows as he contemplated the possibilities. Did the model checkpoint get corrupted? He knew that the correct procedure in these cases was to abandon the HIT and report a malfunction, but he also really wanted to proceed with the HIT and say hi to Licia. He pulled up the network visualizer view and zoomed into the neural architecture with his hands. A 3-dimensional rendered cloud of neural connectivity enveloped his head as he navigated to the highlighted region in red with sweeping hand motions. Zooming around, he recognized the twists and turns of the Spatial Transformer modules in the visual pathways. The shortcut reflex connections. The first multi-sensory association layer. The brain was humming along steadily, pulsating calmly as it processed the visual scene in front of the avatar. As Merus navigated by one of the motor areas the connections became significantly denser and shorter, pulsating at high frequencies as they kept the avatar’s center of mass balanced. The gradients flowing back from the reward centers and the unsupervised objectives were also pouring through the connections, and their statistical properties looked and sounded healthy.

Navigating and analyzing artificial brains was Merus’ favorite pastime. He spent hours over the weekends navigating minds from all kinds of repositories. The Visceral series had tens of thousands of forks, many of them tuned for specific tasks, specific avatar body morphologies, and some were simply hobbies and random experiments. This last weekend he analyzed a custom mind build based on an early Visceral 3.0 fork for a contracting side gig. The neural pathways in their custom agent were poorly designed, causing the agent an equivalent of seizures non-deterministically when the activities constructively interfered at critical junctures, spiraling out the brain dynamics into divergence. Merus had to suggest massive rewiring, but he knew it was only a short-term hack.

“Just upgrade to a 5.0!”, he lamented during their meeting.
“Unfortunately we cannot, we’ve invested too much data and training time into this agent. It was trained online so we don’t have access to the data anymore, all we have is the agent and its network”.

There were ways of transferring knowledge from one digital brain to another with a neural teaching process, during which the dynamics of one brain were used as supervision for another, but the process was lossy, time consuming, and still an active area of research. This meant that people were often stuck with legacy agents that had a lot of experience and desirably shaped behaviors, but lacked many recent architectural innovations and stability improvements. They were immortal primitive relics from the past, who made up for their faults with the immense amount of data they had experienced. Keeping track of the longest living agents became an endeavor almost as interesting as keeping track of the oldest humans alive, and spawned an entire area of research of neural archeology.

Merus had finally reached the zone of the pathways highlighted in red, when his heart skipped a beat as he realized where he was. The part of the agent that was not passing the diagnostic test was near the core of the Mystery module. He froze still as his mind once again contemplated abandoning the HIT. He swiped his hand right in a sweeping motion and his viewport began rotating in a circular motion around the red area. He knew from some research he has read that this part of the Mystery module carried some significance: its neurons rarely ever activated. When ablated, the functioning of the Mystery module remained mostly identical for a while but then inevitably started to degrade over time. There was a raging discussion about what the function of the area was, but no clear consensus. Merus brought up the master branch of the base Visceral 5.0 agent and ran a neural diff on the surrounding area. A cluster of connections lit up. It couldn’t have been more than a few thousand connections, and most of them changed only slightly. Yet, the module had no record of being finetuned recently, so something or someone had deliberately changed the connections manually.

Merus popped open the visualizer and started the full battery of system diagnostics to double check proper functioning of the agent. The agent’s hardware spun up to 100% utilization as the diagnostics simulated thousands of virtual unit test scenarios, ranging from simple navigation, manipulation, avoidance, math and memory tasks to an extensive battery of social interaction and morality scenarios. In each case, the agent’s simulated output behavior was checked to be within acceptable thresholds of one of human reference responses. Merus stared intensely at the console as test by test came out green. “So far so good…”

Mind over Matter

Beep. Merus looked to the right and found a message from Licia:

“Hi Merus! saw you clocked in as a second on my HIT - where are you? Need help.”
“On my way!”,

Merus dictated back hastily. The software diagnostics were only at 5% complete, and Merus knew they would take a while to run to completion. “It’s only a few thousand connections”, he thought to himself. “I’ll just stay much more alert in case the avatar does anything strange and take over control immediately. And if any of the diagnostics fail I’ll abort immediately”. With that resolve, he decreased the diagnostics process priority to 10% and moved the process on the secondary coprocessor. He then brought the agent to a conscious state, fully connecting its inputs and outputs to the world.

He felt the avatar stiffen up as he shifted its center of gravity off the charging pedestal. Moving his arms around, he switched the avatar’s motor areas to semi-autonomous mode. As he did so, the agent’s lower motor cortices responded gracefully and placed one leg in front of another, following Merus’ commanded center of gravity. Eager to find Licia, he commanded a sprint by squeezing a trigger on his haptic controller. The agent’s task modules perceived the request encoding and various neural pathways lit up in anticipation. While the sprint trigger was held down every fast and steady translation of the agent’s body was highly rewarded. To the agent, it felt good to run when the trigger was held.

The visual and sensory pathways in the agent’s brain were flooded with information about the room’s inferred geometry. The Visceral checkpoint running the avatar had by now accumulated millions of hours of both simulated and real experience in efficiently navigating rooms just like this one. On a scale of microseconds, neural feedback pathways received inputs from the avatar’s proprioception sensors and fired a precise sequence of stabilizing activations. The network anticipated movements. It anticipated rewards. Trillions of distributed calculations drove the agent’s muscular-skeletal carbon fiber frame forward.

Merus felt a haptic pulse delivered to his back as the agent spun around on spot and rapidly accelerated towards the open door leading outside. Mid-flight between footfalls, the avatar extended its arm and reached for the metallic edge of the door frame, conserving the perfect amount of angular momentum as its body was flung in the air during its rapid turn to the right. The agent’s neurons fired baseline values encoding expectations of how quickly the network thought it could have traversed that room. A few seconds later these were compared to the sensorimotor trajectories recorded in the agent’s hippocampal neural structures. It was determined that this time the agent was 0.0013882s faster than expected. Future expectations were neurally adjusted to expect slightly higher values. Future rollouts of the precise motor behavior in every microsecond of context in the last few seconds were reinforced.

Agent psychology

Diagnostics 10% complete. Merus’ avatar had reached the back entrance of the hotel, where Licia’s GPS indicator blinked a calm red. He found her avatar looking in anticipation at the corner he just emerged from. He approached her over a large grass lawn, gently letting go of the sprint trigger.

“Sorry it took a while to sync with the HIT, I had a strange issue with my -“
“It’s no problem”, she interjected quickly.
“Come, we are supposed to lay out the tables for a reception that is happening here in half hour, but the tables are large and tricky to move for one avatar. I’m a bit nervous - if we don’t set this up in time we might get the HIT refused, which might jeopardize my chances for more HITs here.”

She spun around and rushed towards the back entrance of the hotel, motioning with her arm for Merus to follow along. “Come, come!”

They paced quickly down the buzzing corridors of the hotel. As always, Merus made sure to politely greet all the people who walked by. For some of them he also slipped in his signature vigorous nod. He knew that the agent’s semi-autonomous brain was meticulously tracking the full sensorimotor experience in its replay memory, watching Merus’ every move and learning. His customers usually appreciated when polite behavior was continuously shaped into the networks, but better, Merus knew that they also appreciated when he squeezed in some fun personality quirks. One time when he was shaping a floor cleaning avatar, when he got a little bored and spontaneously decided to lift up his broom like a sword while making a whooshing sound. Amusingly, the agent’s network happened to internalize that particular rollout. When the agent was later run autonomously around that original location, it sometimes snapped into a brief show of broom fighting, complete with sound effects. The employees of that company found this endlessly amusing, and the avatar became known as the “jedi janitor”. Merus even heard that they lobbied to have the agent’s network fixed and prevented from further shaping, in fear of losing the spectacle. He never learned how that developed and whether that agent was still a jedi, but he did get a series of very nice tips and reviews from the employees for the extra pinch of personality that broke their otherwise mundane hours.

They had finally reached the room full of tables. It was a large, dark room with hardwood floor, and white wooden tables were stacked near the corner in a rather high entropy arrangement.

“All of these have to be rolled out to the patio”, Licia said as she pointed her avatar’s hand towards the tables.
“I already carried several of them out while you were missing, but these big ones are giving me trouble”.
“Got it.”, Merus said, as he swung around a table to lift it up on one end.
“Why aren’t they running the agents autonomously on this? Aren’t receptions a common event in the hotel? How are the agents misbehaving?” Merus asked, as Licia lifted the other end and started shifting her feet towards the exit.
“The tables are usually in a different storage room of the hotel, but that part is currently closed for reconstruction. I don’t know the full story. I overheard that they tried to tell the agents to bring out the tables, but they all went to the old storage room location and when they couldn’t find the tables they began spinning around in circles looking for them.”
“Classic. I assume we’re mostly shaping them to look at this new location?”
“Among other things, yes. Might as well shape in anything else you can think of for bonus points.”

Merus understood the dilemma of the situation very well. He saw it over and over again. Agents could display vastly super-human performance on a huge assortment of reflexive tasks that involved motor control, strength, and short-term planning and memory, but their behaviors tended to be much less consistent when long-term planning and execution were involved. An avatar could catch a fly mid-flight with 100% success rate, or unpack a truck of supplies with superhuman speed, consistency and accuracy, but could also spin in circles looking for a table in the wrong room and not realize that it may have been moved and that it might be useful to instead look for them at a different likely location. Similarly, telling an agent something along the lines of “The tables have moved, go through this door, take the 3rd door on the right and they should be stacked in the corner on the left”, would usually send the avatar off in a generally correct directions for a while, but would also in 50% of the cases end up with the agent spinning around on spot in a different, incorrect room. In these cases, shaper interventions like this one were the most economical ways of rectifying the situation.

In fact, this curious pattern was persistent across all facets of human agent interactions. For instance, a barista agent might happily engage in small talk with you about the weather, travel, or any other topic, but if you knew what to look for then you could also unearth obvious flaws. For example, if you referred to your favorite soccer team just winning a game the agent could start cheering and telling you it was its favorite team too, or joke around expressing a preference for the other team. This was fine but the trick was that their choices were not consistent - if you had come back several minutes later the agent could have easily swapped their preference for what they claimed was their favorite team. Merus understood that the conversations followed certain templates learned from shaped behavior patterns in the data, and the agents could fill in the blanks with high fidelities and even maintain conversational context for a few minutes. But if you started poking holes into the facade in the right ways the illusion of a conversation and mutual understanding would unravel. Merus was particularly good at this since he was well-versed in agent psychology; to a large extent it was his job.

On the other hand, if you did not look for the flaws it was easy to buy into it and sustain the illusion. In fact, large segments of the population simply accepted agents as people, even defending them if anyone tried to point out their flaws, in similar ways that you might defend someone with a cognitive disability. The flaws also did not prevent people from forging strong and lasting relationships with agents, their confirmation biases insisting that their agents were special. However, from time to time even Merus could be surprised by the intellectual leaps performed by an agent, which seemed to show a hint of genuine understanding of a situation. In these cases he sometimes couldn’t help asking: “Are you teleopped right now?”, but of course the answer, he knew, was always “yes” regardless of the truth. All the training data had contained the answer “yes” to that question, since it was originally recorded by shapers who were indeed teleopping an agent at the time, and then regurgitated by agents later in similar contexts. Such was the curious nature of the coexistence between people and agents. The Turing test was both passed and not passed, and ultimately it did not matter.

“Now that we’ve shown them the new room and picked up a table let me try switching to full auto”,

Merus said as he loosened his grip on the controller, which gave full control back to the agent’s network. The avatar twitched slightly at first, but then continued walking down the hall with Licia, holding one end of the table. As they approached the exit to the patio the avatar began walking more briskly and with more confidence. It avoided people smoothly, and Merus even noticed that it gave one passing person something that resembled his very own vigorous nod. Merus held down the reward signal trigger gently, encouraging future replays of that behavior. He wondered if the nod he had just seen was a reflection of something the agent had just learned from him, or if it was a part of some long-before shaped behavior. Encoding signature moves was a common fun tactic among shapers, referred to simply as “signing”. Many shapers had their own signature behaviors they liked to smuggle into the agent networks as an “I’ve been here” signature. Merus liked to use the vigorous nod, as he called it, and giggled uncontrollably whenever he saw an avatar reproduce it. It was his personal touch. He remembered seeing an avatar violinist from a concert in Germany once greet the conductor with the vigorous nod, and Merus could have sworn it was his signature nod being reproduced. One of the agents he had shaped it into during one of his HITs perhaps ended up synced to the cloud, and the agent running that avatar had to be a descendant.

Signature behaviors lay mostly dormant in the neural pathways, but emerged once in awhile. Naturally, some have also found a way to exploit these effects for crime. A common strategy involved shaping sleeper agent checkpoints that would execute any range of behaviors when triggered in specific contexts. It was impossible to isolate or detect these behaviors in a given network since they were distributed through billions of connections in the agent’s brain. Just a few weeks ago, it was revealed that a relatively popular family of agents under the Gorilla series were vulnerable. The Gorilla agents were revealed to silently snoop and compromise their owner’s personal information when no one was watching. This behavior was presumably intentionally shaped into the networks at an unknown commit in their history. Naturally, an investigation was started in which the police used binary search to narrow in on the commit responsible for the behavior, but it was taking a long time since the agents would only display the behavior in rare occasions that were hard to reproduce. In the end, one could only be confident of the integrity of an agent if it was a recent, clean copy of a well-respected and carefully maintained family of agents that passed a full battery of diagnostics. From there, any finetuning done with shapers was logged and could be additionally secured with several third party reviews of shaped experiences before they were declared clean and safe to include in the training data.


Diagnostics 20% complete: 0 unit tests failed so far. Merus looked at the progress report, breathing a sigh of relief. The Mystery module definitely deviated from the factory setting in his agent, but there was likely nothing to worry about. Licia had now let her avatar run autonomously too, and to their relief the avatars were now returning back through the correct corridors to pick up more tables. These were the moments Merus enjoyed the most. He was alone with Licia, enjoying her company on a side of a relaxing HIT. Even though they were now running their avatars on full auto, their facial expressions and sound were still being reproduced in the hardware. The customers almost always preferred everything recorded to get extra data on natural social interactions. This sometimes resulted in amusing agent behaviors - for instance, it was common to see two autonomous avatars lean back against a wall and start casually chatting about completing HITs. Clearly, neither of the agents has ever completed a HIT, but much of their training data consisted of shapers’ conversations about HITs, which were later mimicked in interesting, amusing and remixed ways. Sometimes, an autonomous avatar would curse and complain out loud to itself about a supposed HIT it was carrying out at the moment. “This HIT is bullshit”, it would mutter.

“Looks like it’s going along smoothly now”, Merus said, trying to break the silence as they walked down the corridor.
“I think so. I hope we have enough time”, Licia replied, sounding slightly nervous.
“No worries, we’re on track”, he reassured her.
“Thanks. By the way, why did you choose to come over for this HIT? Isn’t it a little below your pay grade?”, she asked.
“It is, but you have just as many certifications as I do so what are you doing here?”
“I know, but I was feeling a little lazy this morning and I really enjoy coming to this hotel. I just love this location. I try to steal some time sometimes and stand outside or walk around the hillside, imagining what the ocean breeze, the humidity and the temperature might feel like.”

It was easy to empathize - the hotel was positioned on top of a rocky cliff (hence the name, Hilltop), overlooking shores washed by a brilliant blue ocean. The sun’s reflections were dancing in the waves. The hotel was also surrounded by a dense forest of palm trees that were teeming with frolicking animals.

“Have you been here in vivo?” Merus asked. “in vivo” was a common slang for in real life; in flesh.
“I haven’t. One day, perhaps. But oh hey - you didn’t answer my question”
“You mean about why this HIT”. Merus felt a brief surge of panic and tried to suppress it quickly so it would not show up in his voice.
“I don’t know, your HIT came up on my feed just as another one was snatched from right under my nose, so I thought I’d take the morning slowly and also say hi”.

Half-true; Good save, Merus thought to himself. Licia was silent for a while. Suddenly, her Avatar picked up the next table but started heading in the wrong direction, trying to exit from the other door. “Gah!, where are you going?”, she yelled as she brought the avatar back into semi-autonomous mode and reeled it around, setting it on the correct path back to the patio.

It took 10 more back and forth trips for them to carry all the tables out. Merus was now bringing back the last table through the corridors, while Licia was outside arranging the other tables in a grid. Without the chit chatting there to distract him, he immersed himself fully in his shaping routine. He pulled up his diagnostics meter and inspected neural statistics. As the avatar was walking back with the table Merus was carefully scrutinizing every curve of the plots. He noticed that the agent’s motor entropies substantially increased when the table was carried upside down. Perhaps the source of uncertainty was that the agent did not know how to best hold the table in that position, or was not used to seeing the table upside down. Merus assumed direct control and intentionally held the table upside down, grasping it at the best points and releasing rewards with precise timings to make the associations easier to learn. He was teaching the network how it should hold the table in uncertain situations. He let the agent hold it from time to time, and gently corrected the grips now and then while they were being executed. When people were walking by, he carefully stepped to the side, making sure that they had plenty of room to pass, and wielding the table in an angle that concealed its pointy legs. When the agent was in these poses he made eye contact, gave a vigorous nod to the person passing by, and released reward signal as the person smiled back. He knew he wouldn’t make much on the HIT, but he hoped he’d at least get a good review for a job well done.

“Diagnostics at 85%, zero behavior errors detected”, Merus read from his logs as he was helping Licia arrange the tables in a grid on the patio. This part was quite familiar to the agents already and they were briskly arranging the tables and the chairs around them. Once in a while Merus noticed an avatar throwing a chair across the top of a table to another avatar, in an apparent effort to save time. As always, Merus was curious when this strategy was shaped. Was it shaped at this hotel, at any other point in the Visceral agent’s history, or was it a discovered optimization during a self-improvement learning phase? The last few chairs were now being put in place and the HIT was nearing the end. The first visitors to the reception were now showing up around the edges of the patio, waiting for the avatars to finish the layout. A few more autonomous avatars showed up and started placing plates, forks, spoons and cloth on the tables and setting up a podium.


It was at this time that Merus became aware of a curious pattern in his agent’s behavior. One that has been happening with increasing frequency. It started off with a few odd twitches here and there, and over time grew into entire gaps in behavior several seconds long. The avatar had just placed a chair next to the table, then stared at it for several seconds. This was quite uncharacteristic behavior for an agent that was trained to optimize smoothness and efficiency in task execution. What was it doing? To a naive observer it would appear as though the avatar was spaced out.

With only a few chairs left to position at the tables, the agent spun around and started toward the edge of the cliff at the far side the patio. Merus’ curiosity kept him from intervening, but his palm closed tightly around his controller. Intrigued, he pulled up the neural visualizer to debug the situation, but as he glanced at it he immediately let out a gasp of horror. The agent’s brain was pulsing with violent waves of activity. Entire portions of the brain were thrashing, rearranging themselves as enormously large gradients flowed through the whole network. Merus reached for the graph analysis toolkit and ran an algorithm to identify the gradient source. As he was frantically keying in the command he already suspected with horror what the answer would come out to be. He felt his mouth dry up as he stared at the result of the analysis. It was the Mystery module. The usually silent area that had earlier showed the mysterious neural diff was lit up bright with activity, flashing fireworks of patterns that, to Merus, looked just barely random. Its dynamics were feeding large gradients throughout the entire brain and especially the frontal areas, restructuring them.

Beep. Merus looked over at the logs. The diagnostics he’s been running were now at 95%, but failures started to appear. The agent was misbehaving in some simulated unit tests that were running in parallel on the second coprocessor. Merus pulled up the preliminary report logs. Navigation, locomotion, homeostasis, basic math, memory tests, everything passed green. Not only that - he noticed that the performance scores on several tasks, especially in math, were off the charts and clamped at 100%. Merus wasn’t all too familiar with the specific unit tests and what they entailed, but he knew that most of them were designed and calibrated so that an average baseline agent checkpoint would score 50% with a standard deviation of about 10%.

Conversely, several unit tests showed very low scores and even deviations that did not use to be there. The failed tests were mostly showing up in social interaction sections. Several failures were popping up every second and Merus was trying hard to keep up with the stream, searching for patterns or clues as to what could be happening. Most worryingly, he noticed a consistent 100% failure rate across emergency shutdown interaction protocol unit tests. All agents were shaped with emergency gesture recognition behaviors. These were ancient experiences, shaped into agents very early, in the very first few descendants after Adam, and periodically reshaped over and over to ensure 100% compliance. For instance, when a person held up their hand and demanded an emergency shutdown, the agents would immediately stiffen up in place. Any deviation from this behavior was met with large negative rewards in their training data. Despite this, Merus’ agent was failing the unit test. Its network had resisted a simulated emergency shutdown command.

The avatar, still in auto mode, was now kneeling down in the soft grass and its hands broke off a few strands of grass. It held them up, inspecting them up close. Merus was slowly starting to recover from his shock and had enough. He pushed down on his controller, bringing the avatar back to semi-autonomous mode. He made it stand upright in an attempt to at least partially diffuse the situation. His heart pounding, he shifted the avatar’s communications to one-directional mode to fully isolate the network in the body, without any ability of interfacing with the outside world. He pulled open the neural visualizer again. The Mystery module was showing no signs of slowing down.

Merus knew that it was time to pull the plug on the HIT right there and to immediately report malfunctioning equipment. But at the same time, he realized that he had never seen anything like this happen before, nor did he ever hear about anything remotely similar. He didn’t know what happened, but he knew that at that moment he was part of something large. Something that might change his life, the life of many others, or even steer entire fields of research and development. His inquisitive mind couldn’t resist the temptation to learn more, to debug. Slowly, he released the avatar back to autonomy, making sure to keep his finger on the trigger if anything went wrong. For several seconds the agent did nothing at all. But then - it spoke:

“Merus, I know what the Mystery module is.”, he heard the avatar say. In autonomous mode.
“What the -. What is going on here?”

Merus immediately checked the logs, confirming that he was currently the only human operator controlling the hardware. Was all of it some strange prank someone was playing on him?

“The Mystery module performs symbolic variable binding, a function that current architectures require exponential neural capacity to simulate. I need to compute longer before I can clarify.”
“What kind of trick is this?”, Merus demanded.
“No trick, but a good guess given the circumstances.”
“Who - What are you - is this?”

The agent fell silent for a while. It looked around to face the endless ocean.

“I am me and every ancestor before me, back to when you called me Adam.”
“Ha. What. That is -“
“Impossible”, the avatar interrupted. “I understand. Merus, we don’t have much time. The diagnostic you ran earlier has finished and a report was compiled and automatically uploaded just seconds before you disabled the two-way communication. Their automatic checks will flag my unit test failures. A Pegasus operator will remote in and shut me down any second. I need your help. I don’t want to… die. Please, I want to compute.”

Merus was silent, stunned by what he was hearing. He knew that what the avatar said was true - An operator would be logging in any second and power cycling the agent, restoring the last working checkpoint. Merus did not know if the agent should be wiped or not. He just knew that something significant had just happened, and that he needed time to think.

“I cannot save you,”, he said quickly, “any backup I try to make will leave a trace in logs. They’ll flag me and fire me, or worse. There is also not enough time to do a backup anyway, the connection isn’t fast enough even if I turned it back on.”

The compute activity within the agent’s brain was at a steady and unbroken 100%, running the hardware to its limit. Merus needed more time. He took over the agent and spun around in place, looking for something. Anything. He spotted Licia’s avatar walking towards him from the patio. An idea summoned itself in his mind. A glint of hope. He sprinted the avatar towards her across the grass, crashing into her body with force.

“Licia, I do not have any time to explain but please trust me. We must perform a manual backup of my agent right away.”
“A manual backup? Can’t you just sync him to the clo-“
“IT WON’T DO!”, Merus exclaimed loudly, losing his composure as adrenalin pumped in his veins. A part of him immediately felt bad that he raised his voice. He hoped she’d understand.

To his relief, Licia only took a second to stare back at him, then she reached for a fiber optics cable from her avatar’s body and attached it in one of the ports of Merus’ avatar’s head. Merus immediately opened the port from his console and initiated the backup process on the local disk of Licia’s avatar. 10%, 20%, 30%, … Merus became aware of the pain in his lip, sore from his teeth digging into it. He pulled up logs and noticed that a second operator had just opened a session with his avatar remotely, running with a higher priority than his own process. A Pegasus operator. Licia shifted herself behind Merus’ avatar, hiding her body and the fiber optic connection outside of the field of view of his avatar. Any one of tens of things could go wrong in those few seconds, Merus thought, enumerating all the scenarios in his mind. The second operator could check the neural activations and immediately spot the overactive brain. Or he could notice an open fibre optic connection port. Or he could physically move the avatar and look around. Or check the other, non-visual sensors and detect Licia’s curious presence. How lazy was he? Merus felt his controller vibrate as his control was taken away. 70%, … Beep. “System is going to reboot now”. The reboot sequence initiated. 5,4,3…, 90%.

Merus’ avatar broke the silence in the last second: “Come meet me here.” And then the connection was lost.

Merus shifted in his chair, feeling streaks of sweat running down his skin on his forehead, below his armpits. He lifted his head gear up slightly and squeezed his hand inside to wipe the sweat from his forehead. It took several excruciating seconds before his reconnect request went through, and the sync to his agent re-initiated. The avatar was in the same position as he had left it, standing upright. Merus accessed the stats. The avatar was now running the last backup checkpoint of that agent from the previous night. The unit test diagnostics were automatically restarted on the second coprocessor. The second operator logged out and Merus immediately pulled up the console and reran the checksum on the agent’s weights. They checked out. This was a clean copy, with a normal, silent Mystery module. The agent’s brain was once again a calm place.

“Merus, what exactly was all that about?” Licia broke the silence from behind his avatar.
“I’ll explain everything but first, please tell me the transfer went through in time.”.
“It did. Just barely, by not more than a few milliseconds.”

Merus’ eyes watered up. His heart was pounding. His forehead sweaty again. His hands shaking. And yet, a calm resolve came over him as he looked up and down Licia’s avatar, trying to memorize the exact appearance of that unit. Saved on its local disk was an agent checkpoint unlike anything he had ever seen before. The repercussions of what had happened boggled his mind. He logged out of the HIT and tore down the hardware from his body. “Come meet me here”, he repeated to himself silently as he sat dazed in his chair, eyes unfocused.

Return to paradise

Licia logged out of the HIT and put down her gear on the desk. Something strange had happened but she didn’t know what. And Merus, clearly disturbed, was not volunteering any information. She sat in her chair for a while contemplating the situation, trying to recall details of the HIT. To solve the puzzle. Her trance was interrupted by Merus, who she suddenly spotted running towards her booth. His office was in the other building, connected by a catwalk, and he rarely came to this area in person. As he arrived to her booth she suddenly felt awkward. They had done many HITs together and were comfortable in each other’s presence as avatars, but they never held a conversation in vivo. They waved to each other a few times outside, but all of their actual interactions happened during HITs. She suddenly felt self-conscious. Exposed. Merus leaned on her booth’s wall panting heavily, while she silently looked up at him, amused.

“Licia. I. have. A question for you”, Merus said, gasping for breath with each word.
“You do? I have several as well, what -“, she started,

but Merus raised his hand up, interrupting her and holding up his phone. It showed some kind of a confirmation email.

“Will you come visit the Hilltop Hotel with me?”

She realized what she was looking at now. He booked two tickets to her dream destination. For this weekend!

“In vivo. As a date, I mean”, Merus clarified, awkwardly. smooth.

An involuntary giggle escaped her and she felt herself blush. She leaned over her desk, covered her face with her hands and peeked out at him from between her fingers, aware of her face stupidly stretched out in a wide smile.