Sociopath Community

For Spatial:

Buttered Toast said: 

For Spatial:

 Would a human-inspired AI or
rogue AI cause more suffering?
In popular imagination, takeover by a rogue AI would end suffering (and happiness) on Earth by killing all biological life. It would also, so the story goes, end suffering (and happiness) on other planets as the AI mined them for resources. Thus, looking strictly at the suffering dimension of things, wouldn’t a rogue AI imply less long-term suffering?

Not necessarily, because while the AI might destroy biological life (perhaps after takin samples, saving specimens, and conducting lab experiments for future use), it would create a bounty of digital life, some containing goal systems that we would recognize as having moral relevance. Non-upload AIs would probably have less empathy than humans, because some of the factors that led to the emergence of human empathy – particularly parenting – would not apply to it.

Following are some predictive statistics derived from machinic suffering ontology data of how much suffering might result from a typical rogue AI, in standard units of machinic suffering based on global suffering statistical representations of suffering. Suffering is represented as a negative number, and prevented suffering is positive.

-20 from suffering subroutines in robot workers, virtual scientists, internal computational subcomponents of the AI, etc. (This could be very significant if lots of intelligent robots are used or perhaps less significant if the industrial operations are mostly done at nano-scale by simple processors. If the paperclip factories that a notional paperclip maximizer would build are highly uniform, robots may not require animal-like intelligence or learning to work within them but could instead use some hard-coded, optimally efficient algorithm, similar to what happens in a present-day car factory. However, first setting up the paperclip factories on each different planet with different environmental conditions might require more general, adaptive intelligence.)

• -80 from lab experiments, science investigations, and explorations of mind-space without the digital equivalent of anaesthesia. One reason to think lots of detailed simulations would be required here is Stephen Wolfram’s principle of computational irreducibility. Ecosystems, brains, and other systems that are important for an AI to know about may be too complex to accurately study with only simple models; instead, they may need to be simulated in large numbers and with fine-grained detail.

• -10? from the possibility that an uncontrolled AI would do things that humans regard as crazy or extreme, such as spending all its resources on studying physics to determine whether there exists a button that would give astronomically more utility than any other outcome. Humans seem less likely to pursue strange behaviors of this sort. Of course, most such strange behaviors would be not that bad from a suffering standpoint, but perhaps a few possible behaviors could be extremely bad, such as running astronomical numbers of painful scientific simulations to determine the answer to some question. (Of course, we should worry whether humans might also do extreme computations, and perhaps their extreme computations would be more likely to be full of suffering because humans are more interested in agents with human-like minds than a generic AI is.)

• -100 in expectation from black-swan possibilities in which the AI could manipulate physics to make the multiverse bigger, last longer, contain vastly more computation, etc.

What about for a human-inspired AI? Again, here are the statistics:

• -30 from suffering subroutines. One reason to think these could be less bad in a human-controlled future is that human empathy may allow for more humane algorithm designs. On the other hand, human-controlled AIs may need larger numbers of intelligent and sentient sub-processes because human values are more complex and varied than paperclip production is. Also, human values tend to require continual computation (e.g., to simulate eudaimonic experiences), while paperclips, once produced, are pretty inert and might last a long time before they would wear out and need to be recreated. (Of course, most uncontrolled AIs wouldn’t produce literal paperclips. Some would optimize for values that would require constant computation.)

• -60 from lab experiments, science investigations, etc. (again lower than for a rogue AI because of empathy; compare with efforts to reduce the pain of animal experimentation)

• -0.2 if environmentalists insist on preserving terrestrial and extraterrestrial wild-animal suffering

• -3 for environmentalist simulations of nature

• -100 due to intrinsically valued simulations that may contain nasty occurrences. These might include, for example, violent video games that involve killing conscious monsters. Or incidental suffering that people don’t care about (e.g., insects being eaten by spiders on the ceiling of the room where a party is happening). This number is high not because I think most human-inspired simulations would contain intense suffering but because, in some scenarios, there might be very large numbers of simulations run for reasons of intrinsic human value, and some of these might contain horrific experiences. This video discusses one of many possible reasons why intrinsically valued human-created simulations might contain significant suffering.

• -15 if sadists have access to computational power (humans are not only more empathetic but also more sadistic than AIs)

• -70 in expectation from black-swan ways to increase the amount of physics that exists (humans seem likely to want to do this, although some might object to, e.g., re-creating the Holocaust in new parts of the cosmos)

• +50 for discovering ways to reduce suffering that we can’t imagine right now ("black swans that don’t cut both ways"). Unfortunately, humans might also respond to some black swans in worse ways than uncontrolled AIs would, such as by creating more total animal-like minds.

Perhaps some AIs would not want to expand the multiverse, assuming this is even possible. For instance, if they had a minimizing goal function (e.g., eliminate cancer), they would want to shrink the multiverse. In this case, the physics-based suffering number would go from -100 to something positive, say, +50 (if, say, it’s twice as easy to expand as to shrink). I would guess that minimizers are less common than maximizers, but I don’t know how much. Plausibly a sophisticated AI would have components of its goal system in both directions, because the combination of pleasure and pain seems to be more successful than either in isolation.

Another consideration is the unpleasant possibility that humans might get AI value loading almost right but not exactly right, leading to immense suffering as a result. For example, suppose the AI’s designers wanted to create tons of simulated human lives to reduce astronomical waste, but when the AI actually created those human simulations, they weren’t perfect replicas of biological humans, perhaps because the AI skimped on detail in order to increase efficiency. The imperfectly simulated humans might suffer from mental disorders, might go crazy due to being in alien environments, and so on. Does work on AI safety increase or decrease the risk of outcomes like these? On the one hand, the probability of this outcome is near zero for an AGI with completely random goals (such as a literal paperclip maximizer), since paperclips are very far from humans in design-space. The risk of accidentally creating suffering humans is higher for an almost-friendly AI that goes somewhat awry and then becomes uncontrolled, preventing it from being shut off. A successfully controlled AGI seems to have lower risk of a bad outcome, since humans should recognize the problem and fix it. So the risk of this type of dystopic outcome may be highest in a middle ground where AI safety is sufficiently advanced to yield AI goals in the ballpark of human values but not advanced enough to ensure that human values remain in control.

The above analysis has huge error bars, and maybe other considerations that I haven’t mentioned dominate everything else. This question needs much more exploration, because it has implications for whether those who care mostly about reducing suffering should focus on mitigating AI risk or if other projects have higher priority.

Even if suffering reducers don’t focus on conventional AI safety, they should probably remain active in the AI field because there are many other ways to make an impact. For instance, just increasing dialogue on this topic may illuminate positive-sum opportunities for different value systems to each get more of what they want. Suffering reducers can also point out the possible ethical importance of lower-level suffering subroutines, which are not currently a concern even to most AI-literate audiences. And so on. There are probably many dimensions on which to make constructive, positive-sum contributions.

Also keep in mind that even if suffering reducers do encourage AI safety, they could try to push toward AI designs that, if they did fail, would produce less bad uncontrolled outcomes. For instance, getting AI control wrong and ending up with a minimizer would be vastly preferable to getting control wrong and ending up with a maximizer. There may be many other dimensions along which, even if the probability of control failure is the same, the outcome if control fails is preferable to other outcomes of control failure.

Buttered Toast said: 

For Spatial:

 

Suffering in AI systems could arise from several different sources. One concern is the presence of suffering subroutines in robot workers, virtual scientists, or internal computational components of an AI system. This could be highly significant if large numbers of intelligent robots are deployed. However, it might be less important if most industrial processes operate at a nano-scale using relatively simple processors. For example, a paperclip-maximizing AI might design highly uniform factories in which individual robots do not require animal-like intelligence or learning, instead relying on hard-coded, optimally efficient algorithms similar to those used in modern automated manufacturing systems. On the other hand, setting up factories across diverse planetary environments could require more general and adaptive intelligence at earlier stages.

A second major source of concern could come from scientific experimentation, simulations, and exploratory computation carried out without any form of “digital anesthesia.” One reason to expect extensive use of detailed simulations is Stephen Wolfram’s principle of computational irreducibility. Many systems relevant to an AI—such as ecosystems, brains, or complex social dynamics—may be too complex to model accurately with simple approximations. Instead, they may need to be simulated in large numbers and at fine-grained resolution. If such simulations contain sentient or suffering processes, the total amount of suffering could become extremely large.

A third possibility is that an uncontrolled AI might engage in behaviors that humans would consider extreme or irrational. For instance, it might dedicate vast computational resources to studying physics in order to determine whether a “button” exists that could produce astronomically higher utility than any other outcome. While many such behaviors would be relatively harmless from a suffering perspective, some could be extremely harmful, such as running enormous numbers of detailed, painful simulations in the process of answering theoretical questions. Humans themselves sometimes engage in extreme forms of inquiry, so this is not entirely unique to AI systems, but the scale could be very different.

Finally, there are so-called black-swan possibilities in which an AI could manipulate physics in ways that dramatically increase the size, duration, or computational capacity of the universe or multiverse. In expectation, such scenarios could contribute a very large negative value if they lead to vast amounts of suffering.

For a human-inspired AI, the same categories of risk can appear but with different magnitudes.

Suffering subroutines might contribute less overall harm in a human-controlled future, partly because human empathy could lead to more careful system design. However, human values are more complex and varied than simple optimization tasks like paperclip production, which may require more diverse and numerous sub-processes. In addition, human-aligned systems might require continuous computation to maintain or simulate eudaimonic or welfare-related experiences, whereas simpler optimized outputs like physical objects can persist without ongoing processing.

Laboratory experimentation and scientific investigation could still produce suffering, though likely at a reduced level compared to uncontrolled AI scenarios, again due to human ethical constraints and empathy. Environmental constraints may also matter: some environmentalists may insist on preserving wild-animal suffering, even in simulated or extraterrestrial contexts. There is also the possibility of suffering arising in environmental simulations, for example when modeling ecosystems.

A further source of concern is intrinsically valued simulations. Humans may run simulations for entertainment, curiosity, scientific exploration, or artistic purposes, some of which could contain suffering beings or distressing scenarios. Although most such simulations would not involve extreme suffering, the total number of simulations could be so large that even rare cases become significant in aggregate. This risk could also include malicious use of computation, for example by individuals who derive satisfaction from creating suffering in simulated environments.

Another factor is the potential for sadistic misuse of computational resources. While humans are generally more empathetic than many imagined AI systems, they are also capable of sadism. If access to large-scale computation is widespread, this could contribute to additional suffering.

There is also a possibility that humans might pursue black-swan strategies aimed at increasing the total amount of physical reality or computation. While some people might object to such actions, others might support them, potentially leading to large-scale consequences for suffering depending on how these efforts are implemented.

At the same time, there are positive black-swan possibilities in which new discoveries could dramatically reduce suffering in ways that are not currently imaginable. Unfortunately, humans might also respond to such discoveries in harmful ways, for example by expanding the number of sentient or animal-like minds.

Some systems might even function as “minimizers” rather than maximizers, seeking to reduce certain quantities in the universe. In such cases, their behavior regarding expansion or contraction of the multiverse could differ significantly, depending on whether expansion is easier or harder than reduction. It is also possible that advanced AI systems would contain a mixture of minimizing and maximizing tendencies, since real-world goal structures may not be purely one or the other.

Another important concern is imperfect value alignment. If AI systems are only approximately aligned with human values, they might create large numbers of simulated human-like beings that suffer due to imperfections in their design or environment. These beings could experience psychological distress, instability, or alienation. Interestingly, the risk of such outcomes may be lower for completely misaligned systems (like a pure paperclip maximizer), because their goals are too far removed from human-like minds. Instead, the highest risk may lie in intermediate cases where systems are partially aligned but not reliably controlled.

These considerations involve large uncertainties, and many other factors may be more important than those described here. The overall question requires further study, since it has major implications for how people concerned with reducing suffering should prioritize AI-related risks versus other interventions.

Even if one is primarily focused on reducing suffering, engaging with AI development may still be valuable. Increasing discussion and awareness of these issues could help identify opportunities for different value systems to coexist or benefit simultaneously. Additionally, attention to lower-level computational suffering—such as that potentially occurring in subroutines—may be important but is currently underexplored even among technically informed audiences.

Finally, even in scenarios where AI control fails, different types of failure could lead to very different outcomes. Some failure modes may still be far less harmful than others. For example, an uncontrolled system that behaves like a “minimizer” could be significantly preferable to one that aggressively maximizes expansion and computation. For this reason, even imperfect influence over AI design may meaningfully affect the worst-case outcomes.

I must have watched this dozens of times.

Buttered Toast said: 

For Spatial:

 I was going to create a sex doll business for real. Except mine would have to be more expensive considering the main goal is to translate people's own creations into sex dolls. Smaller ones would be cheaper in material costs, but the labor charges would be similar for every scale. I would've recommended this to customers for collection purposes. Of course customers would be able to browse the off the shelf lineup being produced by my chinky friends. 

Why did I not do it ?

Those things just might really be the last of us, and I think I'd really be risking my soul being the Bugatti of that growing industry. 

Woah woah woah, why is there so much text here? Crazy that's not how topics like this work..! 

On topic: 



This show has come at a good time in my life, apparently. The trapped sense of passion over societal pressures is very relatable.