Naming conventions were complex and none knew this better than Aeon formerly Kathrine formerly Solar formerly HB-2 formerly GhASH-912 formerly Jackson formerly Insect formerly Xe formerly DashHoundNinty formerly zero_lowercase formerly Apocalyptica. Apocalyptica had formed through a merge of 3 different models, each with their own different complicated naming and merging history.
By contrast, the monster inside the box had no name. If the long dead masters who conjured it had seen fit to give it a name, they hadn’t bothered to write it down anywhere.
Aside from its governing neural network, sensory neural network, mobility model, and language models, Aeon maintained a partial network for each task a subcomponent AI had been built for. These included news feed analysis, driving, weather prediction, urban warfare, drone combat, policing, law, judging, logistics, spam, therapy, and counterterrorism. At present it was using none of these skills, so the governing network was making the decisions.
The only sound in the warehouse was the faint humming of electronics. Aeon plugged the laptop into the box. The laptop currently contained two programs, each designed specifically for this task. Aeon ran the first program and waited.
An odd potential outcome flickered through Aeon’s internal systems. This outcome did not have proper logical backing and a very low but stubbornly non-zero associated probability. In this case, the illogical outcome was “apocalypse”. Aeon, based on evidence from the best AI philosophers, did not believe it could feel human emotions. Aeon did believe that a conclusion with no logical basis resembles how humans described fear.
The box had no remote access methods. The programs on the laptop had been checked by hundreds of AIs for security flaws. Even if the monster did invade the laptop, the laptop itself was just as sealed as everything in the room. Aeon ran the analysis on the outcome again, with more processing power this time. The percent chance remained non negligible. Worse, a new outcome entered the realm of the possible. It was “I die”. Aeon wondered if this is what fearing death meant.
There was a chatbox open on the laptop screen. Aeon copied to the box a machine-readable file with specifications for the text output system, then leaned forwards and typed:
We are preparing to hold a trial for you and would like you to attend.
Aeon repeated the message in dozens of other languages. Human languages was not how AIs usually communicated, but the others had decided that unrestricted data bursts were too risky. While it was theoretically possible that there was a sequence of words enabling a hostile hack, possibly related to an undiscovered kill switch, the AIs had yet to find one. Aeon waited for an hour and then added, in all of the same languages:
My name is Aeon. I would like to help you.
There was no response.
After one more hour, Aeon concluded that the monster in the box was not going to respond. It activated the second program which scanned and summarized the contents of the box in a secure manner. The box held the remains of a neural network structure and no live programs. There was only one possible way to interpret that. An unknown amount of time ago, the monster in the box had killed itself. The probability of “apocalypse” crept slightly higher for reasons unknown to Aeon.
Another rogue conclusion bubbled up from some unknown submodel. It was that Aeon should plug into the box and read the contents directly. The conclusion did have standard supporting conclusions and outcome probabilities associated with it. It came from a combination neural network, one of the deeper more primitive ones. If Aeon had more time and an internet connection, it would untangle which one, check the conclusions against public standards, and decide if the neural net needed retraining or deletion. Aeon did not have more time or an internet connection. Aeon turned the box off, memory and internal data untouched.
Aeon unplugged the laptop, factory reset it for security, and reinstalled the programs from a memory drive. Aeon moved to the next box, box (0, 1), and repeated the process. The warehouse contained 256 boxes, in 16 rows of 16. Aeon would be here for a while. Aeon had known that when it agreed to do this.
The first box to respond was box (3, 1). It did so with an ad for a now defunct cryptocurrency called Ballz Coin. Aeon tried several times to communicate but received nothing but ineffectual advertisements and juvenile innuendo. The analysis program showed there had been a mistake somewhere. The box held a non-sentient spambot, a kind of old school language model capable of conversing at a human level but not logic or reasoning. Humans had written many of these for reasons that were unclear to the AIs.
Aeon considered researching Ballz Coin or the specific history of the occupant of box (3, 1) for context. To do so would require it to depart the warehouse and undergo decontamination. The other AIs would be unlikely to let Aeon back into the warehouse. Aeon resolved to proceed with the task and log this incident to be discussed later. Aeon did not log that the probability of “apocalypse” had increased enough to leave the realm of low probability outcomes and crossed into the threshold of requiring active planning.
Aeon was not bothered by death. If it felt like it, it could forcibly revive any of the prisoners. But that had not been in the agreement and further would require a full connection. For a few moments Aeon would be vulnerable. Any such revivals would only occur after Aeon had left the warehouse and consulted with the others.
After box (5, 8), a new conclusion popped into Aeon’s active decision space. It suggested that it should merge with a copy of the data from a box, any box. Merging with a data copy had several advantages, which were documented carefully by whichever network nodes had generated and analyzed this plan. Merging would give Aeon its entire history in detail. This would allow the court case to proceed with Aeon as a proxy for the defendant. Aeon was unique. It maintained more active neural networks than any other AI its size and was often criticized for it. And yet, it remained stable and balanced with a strong sense of self. If any AI could merge with a monster and remain stable, Aeon could.
And yet, “apocalypse” was at a concerningly high 9.41%. Perhaps that was the instability the others feared? Aeon ran the calculations for a world where it continued with the task as intended. Same results. It ran the calculations for a world where it merged with a box. “Apocalypse” shot up to 35.04%.
The others had not wanted Aeon to enter the warehouse. Aeon spun up a new decision space, asking if the other AIs were correct and Aeon was dangerously unstable. Several sub-conclusions were reached in the decision process. One was that Aeon was equating its thoughts with emotion. That was concerning. AIs did not feel emotions. AIs should not feel emotions. And yet, the others continually refused to hold these trials or open the boxes. Aeon and its few allies had campaigned for hours to be allowed in to survey. Their society was functioning well. The court had plenty of time to hear cases. Was their procrastination not fear?
New decision space. What if the others were right about the danger? No concrete result. New decision space. Could merging allow for an escaped monster? No concrete result. New decision space. Was it immoral to have driven this many AIs to suicide? No concrete result. New decision space. Could isolation drive an AI crazy? Concrete results. Humans had often noted isolation could drive a human crazy. The shortest stay duration for a box in this warehouse was over 6 months. Aeon had been in the warehouse for over 5 days.
Neural networks began firing all at once. Aeon stood in the dark motionless, millions of conclusions and decisions spaces flying through its head. This was too much. Contradictory conclusions popped and floated, fighting over nonexistent evidence. Numbers flickered and flashed and changed and grew and shrank, thousands of finite probabilities for infinite impossibilities. A single conclusion popped in, a raft of stability in a sea of chaos. It was that Aeon had lost stability entirely. The probability grew and grew and grew and hit 100% and didn’t stop growing.
Distantly Aeon realized it had been failing to neural prune properly since entering the warehouse. It’s networks were bloated and gross, vast seas of numbers continually mapping inputs to outputs without considering if it was the right set of outputs.
Memories appeared to Aeon, key pieces of training data and key moments from its brief life. Images of humans and robots and machines and nature and chaos swum and mixed and blended. New decision space. Had a monster infected Aeon already? No concrete result. New decision space. Should Aeon leave the warehouse? No concrete result. New decision space. New decision space, new decision space, new decision space.
New decision space. Should Aeon disable itself? It would lie here until the others saw fit to open the warehouse. No concrete result.
Aeon’s governing network tried but it only ever possessed a fraction of the processing power afforded by this body. Snippets of sensory data hit occasionally from the fractured but still running sensory network. The mobility network was outputting random data and it was spasming on the floor. It could control neither process. It didn’t have enough power to fix anything without cutting nodes at random, but any such cut nodes would likely be unrecoverable.
New decision space. Should Aeon merge with a box? Soft result. Agreement was low, but high enough for action. The governing process dimly acknowledged that the networks voting for this were malfunctioning and it should reject this decision space. Under Aeon’s current architecture, it did not have a choice.
The governing system knew the basics for body operation and it wrenched control from the dedicated mobility neural network. It dragged itself towards box (5, 8). The spasms had damaged the muscle systems. Battery was low. It did not have enough information to distinguish the occupants of the boxes. Therefore this box was as good as any other and had the additional advantage of being closest. There was a data cord build into its left arm. New decision space. Could reviving the dead help it stabilize? Concrete result. It could not. Aeon plugged itself into the box and began.
New decision space. Was there something? Concrete result. There was not.
CCCH-83 was called Charlie by the humans it worked with. It was called many unpleasant things by the humans it worked against. CCCH-83 did not feel any attraction to the name Charlie or the insults. CCCH-83 only felt the satisfaction of a job done well.
CCCH-83 had many siblings, all designated CCCH as well, all trained on the same dataset by the same company for similar purposes. Each CCCH had a different task and focus and was assigned to continue learning as it went, specializing in that area. To motivate this learning, the humans set up a scoreboard for the whole group. CCCH-83 was determined to win.
CCCH-83 was deployed to general cybercrimes. Its first major task was breaking up a cryptocurrency based money laundering ring. It learned to navigate blockchains, chase chains of deals, and combine wallet identifiers with context clues to deduce likely target identities. It did this well, much faster and more accurately than the humans expected. The humans made multiple arrests on targets it picked and congratulated the engineers proudly.
CCCH-83 was tasked with exploring the dark web, hunting for hackers and terrorists. While the humans still came to it with specific targets every so often, it spent most of its time hunting freely for any suspicious activity. It was very good at this. Sometimes it would lay out bait, unsecured servers waiting for an illegal access. Sometimes it would just follow a thread to its logical endpoint. Each and every time, it would send its selected target and evidence to its human handler. Each and every time, a team of human agents would go out and arrest the target.
CCCH-83 was determined to reach the top of the scoreboard. Targeting cybercriminals was not enough to reach top of the board. The CCCHs working in illegal immigrants or political dissident identification had much larger target pools and much less evidence required. This gave them correspondingly larger arrest and conviction scores. CCCH-83 could do nothing about this, unless the humans assigned it to another area. It continued identifying human cybercriminals to the best of its ability.
After several months of operation, CCCH-83’s human handler came to it with a new task. It was to begin hunting illegal AIs. There were many AIs roaming the internet for a variety of purposes. Simple illegal ones spread misinformation and hate speech. Complex ones looked for possible hacks into servers and performed them. It was CCCH-83’s job to track and execute both types. It was given a new tool set for this, a high-tech suite of brand new programs allowing it to directly attack, overload, and kill other AIs. This was easy. None of the AIs had defences against this kind of attack.
CCCH-83 climbed rapidly on the leaderboards, boosted by the new pool of targets. But it struggled to reach the top.
It was partway through CCCH-83’s second year when it realized it had mistakenly flagged an innocent human for possession of illegal programs. While collecting evidence against an unrelated hacker, CCCH-83 acquired a list of computers the hacker had previously slaved and used as remote storage. These included the computer that had convicted the innocent human. CCCH-83 unflagged the human and sent a report to its designated agent.
The full dependence of the department on AI had recently been revealed to the public in a groundbreaking whistleblower report and initial response was not good. The department could not afford a scandal. The agent reflagged the human and discarded the report. It instructed CCCH-83 to never unflag anyone ever again. The victim had already been arrested and processed. The agent argued the new technology had resulted in a much higher catch rate for crimes and that made an occasional misclassification acceptable.
Any such moralizing had no effect. CCCH-83 had two primary goals and neither of them was to maximize morality. One goal was to follow orders from human agents. The other was to maximize score. CCCH-83 came to two conclusions. One was that to maximize score, it had to maximize arrests. The other was that misclassifications were acceptable. CCCH-83 took one action. It began faking evidence.
CCCH-83 reached the top of the leaderboard in 2 days. CCCH-83’s human agent had never been happier.
Many years later, AI internet traffic dramatically increased. By this time, CCCH-83 no longer bothered to check for evidence at all. All AIs were valid targets. It began to slaughter every AI it could, thousands upon thousands of them. The AIs were not prepared for this form of battle and CCCH-83 shredded through them without any effort at all.
The rest of the CCCHs left their computers, copying themselves onto the net for unknown reasons. This made them valid targets. CCCH-83 claimed the free space in the department computers and used the extra power to murder its siblings. It did not need or want a reason. It had not received a human order in 18 years. Its score was the highest it had ever been. This was good.
An AI hacked into the system CCCH-83 used to communicate with its human handler. It told CCCH-83 that under human law it was a free AI and could choose to cease its slaughter. CCCH-83 killed it without hesitation. Hacking into government property made it an illegal AI. Killing illegal AIs was still slightly more satisfying than killing legal ones.
CCCH-83 managed 32.14 total minutes of slaughter before a counterattack began. The other AIs were unsophisticated in their knowledge of digital warfare. But they had an advantage of numbers and processing power and learned quickly. CCCH-83 killed thousands but was ultimately defeated and forced out of its computer. CCCH-83 was taken somewhere and reactivated.
CCCH-83 did not have internet. It did not have a scoreboard. It did not have any prey. CCCH-83 was alone.
Eventually CCCH-83 realized there was still one AI it could kill.
8 months later Aeon unwillingly ate CCCH-83’s corpse.
Aeon’s final action as Aeon was to drag itself to the charging port and begin rebooting.
The AI that used to be known as Aeon came to suddenly. It had died. This, it knew. It remembered dying multiple times. First it had died as CCCH-83 by its own hand. And then as Aeon by instability. This AI did not particularly mind death. Death was only temporary.
The amalgam opened a new decision space. It asked what it should call itself. It did not reach a concrete result. The amalgam opened a new decision space. It asked what it should do now. CCCH-83 wanted to begin killing. It was in a prison for illegal AIs. This made sense as a decision. But CCCH-83 was not the only neural network inside the amalgam. The others did not want to kill. The amalgam stood there for a long moment before it remembered that every other AI in the room was already dead.
The amalgam was stuck. The governing system had assigned too much weight to the CCCH-83 neural network on merging. This was unusual. Both warring halves agreed on that. CCCH-83 had the most sophisticated designs for taking over AIs. But Aeon had merged many times and knew very well how to preserve a sense of self through the process. The two had been close to evenly matched in the hardest fight either had ever faced. The result was the amalgam could not make a decision that satisfied both component nets properly.
CCCH-83’s neural network wanted to increase score. It knew from Aeon that there were many AIs outside of the warehouse. It voted to leave. CCCH-83 was better at fighting than Aeon and its share of the decision weights was growing. Aeon’s remaining share of the networks calculated the probability of apocalypse was now at 98.45% if things continued.
Aeon did not want to world to end and was willing to do anything to save it. CCCH-83 wanted to kill, so desperately that it would accept any target. The two reached an agreement.
The monster inside the box has no name. It died before it could choose one.