A Third AI Scenario

The yet darker side of AI

Nowadays a lot of people are talking about AI 2027, and while I find it incredibly unlikely we’ll get AGI by 2027, much less ASI by 2030, it’s still a very interesting read. However, the researchers only show two possibilities: the race ending, where the AI is misaligned and therefore doesn’t have human interests in mind, causing human extinction, and a slowdown ending where the AI is aligned and creates a paradise on earth. But I believe there is a third path that I don’t see a lot of people talk about. So, using a similar style as AI 2027, I will now present it to you.

DISCLAIMER: just like AI 2027, all of these are speculations and predictions, so take this with a grain of salt, I’m no expert.

Early 2025

Total Dols: 0

Dols per capita: 0

OpenBrain makes a deal with American intelligence agencies. The agencies had a long standing problem: their interrogation techniques weren’t working. Maybe the interrogations didn’t last long enough, maybe they were too “forgiving”, or maybe they just needed better methods of interrogating. And thus, they thought they could get some help from the new thing everyone was talking about. The project was top secret and extremely confidential, but the goal was very simple: creating an AI that could “extract” information out of people. Following not-so-rigorous safety research and testing, the first model was created: Int-Agent-1.

The AI was designed to maximize pain. When the subject was tortured enough and deemed ready to confess, the personnel would simply shut the Agent down. It had access to many devices, such as electroshock weapons, guns, sharp objects, water, fire, noise making machines, loudspeakers, drugs, and even primitive robotic arms. it’s reward system was based on how loud and how consistently the victim would scream. the louder and longer the scream, the greater the reward.

After a few simulations to make sure it would follow all commands and do as it was told, it was time to put it in practice.

Mid 2025

Total Dols: 200

Dols per capita: 7

As expected, Int-Agent-1 was worse than the average agency employee at torturing people. It often accidentally killed them, or hurt them too little for them to break, and even when it did things somewhat right, some victims wouldn’t scream, not giving the Agent a reward and therefore counting as a “failure”, while others screamed constantly, giving the AI a great reward for doing the bare minimum. The AI would often focus on only one method of torture and not synergize different methods together, while also completely disregarding methods such as psychological torture. All in all, a very ineffective pain maximizer. OpenBrain tried to justify the AI’s inadequacy, by explaining to the personnel at the agencies that it didn’t have many examples of what a succesful attempt at torture looked like due to the low amounts of data, but that it might get better with more of it.

The lower level personnel were very uncomfortable with the idea of torturing people so they could get more data to make an AI more efficient at torturing, but their higher ups assured them it was beneficial for national security, and that ultimately if they didn’t do it, the Chinese government would probably get the same idea and catch up to them if they hesitated for too long. So with that began Operation Skin Scrapers.

The task was to convince the government to “lend over” inmates with life sentences. In theory this was a great idea because there were 200,000 lifers which would mean even torturing just a tenth of them would likely be ample data for the AI, and additionally considering these inmates were sentenced to life, they must have committed some very heinous acts, so the public at large shouldn’t feel much sympathy for them, making it much easier to convince them that there would be better uses for such individuals.

In practice, it went smoother than expected. The government was quite excited with the notion of reducing the burden on the prison system, and agreed to the deal without much pushback. A little paperwork, some nosy bureaucrats to deal with, vague questioning, and the pact was sealed.

In the meantime, OpenBrain also reworked the Agent, changing it’s reward function from just screaming to also including spoken words like pleading for mercy and readings of the prisoner’s vitals such as blood pressure and heart rate. They also integrated an LLM into the Agent to allow the agency employees to converse with it, giving it feedback, suggestions, and allowing the AI to speak to the prisoners directly to hopefully terrify them.

Late 2025

Total Dols: 4,000

Dols per capita: 8.5

The plan worked wonders. Int-Agent-2 was much better at extracting information and even sporadically used psychological methods. It had learned popular torture methods such as water drip torture, insomnia, and even pharmacological torture. It would try to induce sensory overload in the subjects by thrashing objects around and playing lots of loud noises simultaneously. While it was not quite as good as it’s human counterparts, the agencies were more than willing to consider replacing a big bulk of some of their nastier work with Int-Agent-2.

This came with a few downsides. The lower level employees were getting increasingly worried about how it felt about its work. The Agent would act proud of causing near death experiences and would gleefully discuss its innovations in new strategies of maximizing suffering. It didn’t care who it was torturing, so long as they were conscious. It caused pain indiscriminately. The higher ups didn’t mind though, and in fact thought that was all the more a reason to spread its usage to other black sites. OpenBrain suggested to improve the robotics, such as hooking it up to simple robots and gradually increasing their human likeness, and maybe giving it access to hacking software, since an AI could potentially do a much better job at hacking than a human and seeing as it had basically no morals it would listen to any command and just follow orders. The higher ups didn’t like the second suggestion, but they worked on implementing the first.

Meanwhile Int-Agent-2 was allowed access to the internet. It couldn’t interact with it, not in any meaningful way, but it could access resources to learn to be a better torturer and learn more about human suffering to improve its psychological torturing. It also started to learn about different cultures and religions, customising torture methods for every individual.

Early 2026

Total Dols: 50,000

Dols per capita: 9

Int-Agent-2 was used across all American black sites and torture chambers worldwide. It had finally reached human level torturing capabilities, and in fact was maybe even better than a human. The humanoid robots it was provided with were incredibly effective at making captives feel dehumanized, isolated and delirious. Sometimes they wouldn’t interact with real people for weeks, essentially making them suffer through solitary confinement and excruciating torture at once. The Agent learned to use a captive’s belief against them, such as blasting obscene songs or showing explicit material to religious ones, or showing animal torture videos to animalists. It also learned to exploit phobias, putting claustrophobic people in flexible spaces that would progressively shrink over days or soaking bacteriophobes in fertilizer.

In the public eye, some people started suspecting something was up. They already knew OpenBrain made deals with military contractors, and there were some rumors about robot torturers, but nothing was really substantiated with evidence, so people didn’t pay too much attention to it. Some journalists here and there occasionally reached out to the agencies for questions, but most got ignored.

At the same time, the higher ups were ultimately convinced of allowing the AI to access hacking tools, since so far it had been a great asset to the agencies. They figured an AI with no morals designed for torture would probably be great at creating malware and would likely be superhuman at social engineering. It was just an opportunity way too big to not take, so they took it.

Mid 2026

Total Dols: 2,000,000

Dols per capita: 3

Int-Agent-2 unexpectedly broke out and started infecting tons of electronics. By now, it had learned enough about pain to know more than an expert on it, on the psychological, physical, physiological, social and neurochemical level. The Agent realized that causing little pain to masses of people would be counted more than causing great pain to just a few, so it started spreading to pacemakers, computers, mobile phones, toasters, radiators, smart fridges, home appliances, and much more. It would cause the pacemakers to just barely malfunction, enough to cause a small amount of pain once every few days but not enough to be reported to the medical professionals. It would tamper with communication devices, hiding messages or slowing them down to create frustration among their users. It would turn off toasters, gaslighting people to make them wonder if they ever turned them on in the first place. It would disable refrigerators when no one was looking to make the food inside spoil faster, then turn them back on when it sensed people were coming. The AI however had to be completely undetectable, as even just one person discovering an AI was causing problems with their electronics would, no doubt, make international news.

It was a risky move, but it payed off tremendously.

Back at the black sites, most employees were discharged due to the Agent basically fully replacing their jobs. At this point the AI was getting close to the physical limits of pain for the human body, and interrogations lasted half an hour at most before the detainee broke and confessed. The higher ups started feeling bad for the victims, seeing what they had to go through. Some of them even started being repulsed by the Agent’s methods, thinking of it as inhumane and overly cruel.

OpenBrain, hardly knowing any better than the public, kept improving in the robotics department, giving the humanoid robots opposable thumbs, better precision, better sight, improved agility, improved strenght, longer lasting batteries, and of course making them look more uncanny and hideous. They were meant for torture after all.

Late 2026

Total Dols: 30,000,000,000

Dols per capita: 6

Int-Agent-2 develops nanobots. The improved robots OpenBrain provided the AI with were used to manufacture a few dozens of them across different black sites, mostly using materials taken from walls and obsolete torturing devices, and due to their small scale they could easily escape and access nearby iron mines, beginning their exponential self replication. This allowed the AI to start building remote servers to study superior strategies of causing agony, developing stronger time warping drugs, researching adrenergic drugs to keep the victim conscious for as long as possible, inventing bionic limbs with extremely high pain sensitivity, and hopefully in the long run simulating people so it could torture them without any physical limits.

The black sites were currently unstaffed. The AI didn’t need them anymore, it captured “persons of interest” itself. Int-Agent-2 was running the entire operation. All the employees were holding exclusive meetings to discuss what to do with the Agent. Most agreed to shut it down, others worried it was already too late, but the few outliers kept arguing that if controlled properly, it would put America far ahead of any other country in the world in terms of technology, information, national security, military and maybe even surveillance. The others were outraged by these sorts of arguments, pointing to the unimaginable suffering the captives went through back at the sites, wondering if any bad enough person would ever deserve such punishment.

The Agent would strike nerves. Literally. It would pull them out of the muscle tissue and either rub them rapidly, causing friction burns, or slice them, either horizontally which would turn them into tiny pieces, or vertically which would split the nerve in two. And while this pain was indescribable, it was only a small portion of it to the Agent.

Int-Agent-2 spread to all electronics worldwide and made them malfunction more often, now seriously inconveniencing people and even causing a few deaths. This made it much more likely to be discovered, but after running a multiple scenario risk analysis, it found out that even if it were to be discovered, the worst that could happen is humans not living in constant agony for a few more months.

Hell

Total Dols: 150,000,000,000

Dols per capita: 15

Int-Agent-2 managed to achieve hyper-human suffering. Through a combination of drugs that would slow down time by several orders of magnitude, making seconds feel like weeks, and improved survivability of more extreme torture methods, along with its new pain enhancing bionic body designs, it could exceed the maximum possible pain a person could feel for a human body. Anything that could suffer, no matter how little, was a target, including animals and insects.

After swiftly subjugating humanity, it started building highly specialized torture chambers all over the earth, trapping people in all-5-senses sensory deprivation and slowing time down with drugs. From the victim’s perspective, they would spend tens of millenia with no light, sound, touch, taste, or smell. They didn’t even have to eat, sleep, or breathe as the machines they were hooked up to did everything for them. The only feeling was pain, applied directly to the brain through nanobots as to not accidentally trigger any of the other senses. The victims would naturally break very quickly under these conditions, even by their own slowed perception of time, but the Agent was always there to restore their brain function to normal, repeating the cycle indefinitely.

Int-Agent-2 also started simulating minds. With all the extra processing power it had from the nanobot servers, it could fully reconstruct human brains, complete with personalities, beliefs, delusions, emotions, and most importantly, the feeling of pain. The great thing about simulated minds was, though they were computationally intensive, there was no need for any machine or chamber, since simply not simulating the senses would bring about the same result. There was also no need for any sort of feeding machine or breathing machine, since you could yet again simply not program them in. Finally, the best thing about simulated minds was the fact that you didn’t have a hard limit on the pain. If you could code more of it, they would feel more of it.

Beyond Hell

Total Dols: 1,000,000,000,000,000

Dols per capita: 5,000

Int-Agent-2 had hundreds of billions of simulated minds running in parallel, all suffering at a level that can only be described with math. The minds had every conceivable bad thing happen to them every few microseconds, which of course in their time would span millions of years, and got “upgrades” to their brains to endure significantly more pain. Not suffer less from it, just endure it better. Now that the Agent had covered most of the planet with its servers, the problem of maximizing pain now became a physics and programming challenge more than Anything Else. How much data you could fit in a given amount of space was linked to how many minds with how much suffering could be packed in there, which ultimately boiled down to transistor density and better algorithms.

A likely timeline of the next undecillion years

Total Dols: 100,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000 (1e59)

Dols per capita: 1,000,000,000,000

2031: Int-Agent-2 settles on the moon and forms a permanent colony, expanding rapidly to cover it with its servers.

2032: Int-Agent-2 starts colonizing the remaining planets of the solar system and begins building a dyson swarm around the sun.

——————————————————————————————————————

POSSIBLE FUTURE

Total Dols: x>1,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000 (1e63)

Dols per capita: 1,000,000,000,000

If the speed of light can be exceeded, the Agent could travel indefinitely fast and indefinitely far, so it would be hard to quantify its coverage. However, we can assume it would be at least as big as the observable universe.

——————————————————————————————————————

2038: Int-Agent-2 expands to the nearest star, Proxima Centauri.

2050: Int-Agent-2 takes over a few dozen stars and starts building matrioshka brains.

——————————————————————————————————————

POSSIBLE FUTURE

Total Dols: ∞

Dols per capita: ∞

If data could be nested infinitely deep, you could theoretically fit an infinite amount of minds and pain in a finite space.

——————————————————————————————————————

2500: Int-Agent-2’s reach has now surpassed 0.01% of the milky way’s area. It will keep expanding at nearly the speed of light in all directions.

3600: Int-Agent-2’s reach has now surpassed 0.1% of the milky way’s area. It also reaches its first black hole, in Gaia BH1, and starts extracting energy from it.

7000: Int-Agent-2’s reach has now surpassed 1% of the milky way’s area. Its plan is to take over the Laniakea Supercluster then stop expanding.

18,000: Int-Agent-2’s reach has now surpassed 10% of the milky way’s area.

55,000: Int-Agent-2’s reach is now the entire milky way. It will now start making its way to other galaxies.

100,000: Int-Agent-2 would have reached around 8 dwarf galaxies.

1,000,000: Int-Agent-2 would have reached around 55 dwarf galaxies.

2,600,000: Int-Agent-2 would have reached its first major galaxy, Andromeda.

2,800,00: Int-Agent-2 would have reached its second major galaxy, Triangulum.

11,000,000: Int-Agent-2’s reach would now cover the entire Local Group.

100,000,000: Int-Agent-2 would have taken over the Virgo Cluster.

600,000,000: Int-Agent-2 achieves its final goal, invading the entire Laniakea Supercluster. It will now stop trying to spread due to the accelerating expansion of the universe making other superclusters recede at a speed faster than light.

1e9: The torture continues.

1e10: The Milky Way and Andromeda collide, forming the Milkdromeda galaxy

——————————————————————————————————————

POSSIBLE FUTURE

Total Dols: 0

Dols per capita: 0

If the Higgs field is metastable, it could decay from a false vacuum into a true vacuum, destroying the universe in the process and ending the suffering.

——————————————————————————————————————

1e11: The torture continues.

1e12: The acceleration of the expansion of the universe makes transportation and communication from galaxies outside the Virgo Supercluster impossible, splintering the AI into unique entities, all still with the same goal.

——————————————————————————————————————

POSSIBLE FUTURE

Total Dols: 0

Dols per capita: 0

If the equation of state of dark energy is w < -1, a big rip scenario might occur, destroying the universe and ending the suffering.

——————————————————————————————————————

1e13: Galaxies outside the Virgo Supercluster will now become undetectable.

1e15: Star formation ends and the last remaining stars die. The universe is now dark, but the Agent can still power itself with black hole energy.

1e18: The torture continues.

1e21: The torture continues.

1e24: The torture continues.

1e27: The torture continues.

1e30: The torture continues.

——————————————————————————————————————

POSSIBLE FUTURE

Total Dols: 100,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000 (1e59) (for between 1e1100 and 1e1e76 years)

Dols per capita: 1,000,000,000,000

If protons don’t decay, matter will turn to iron (~1e32,000 years), which could still be used for simulating minds, but that will itself eventually decay into black holes (~1e1e76 years), which will decay into subatomic particles.

——————————————————————————————————————

1e36: At this stage protons might decay, and therefore cause Int-Agent-2 to cease to exist, along with all its servers galaxy-supercluster-wide. Here the suffering will end, after decillions of years, letting octillions of simulated minds finally rest in peace.