Existential risk from AI without an intelligence explosion

[x-post LessWrong]

In discussions of existential risk from AI, it is often assumed that the existential catastrophe would follow an intelligence explosion, in which an AI creates a more capable AI, which in turn creates a yet more capable AI, and so on, a feedback loop that eventually produces an AI whose cognitive power vastly surpasses that of humans, which would be able to obtain a decisive strategic advantage over humanity, allowing it to pursue its own goals without effective human interference. Victoria Krakovna points out that many arguments that AI could present an existential risk do not rely on an intelligence explosion. I want to look in sightly more detail at how that could happen. Kaj Sotala also discusses this.

An AI starts an intelligence explosion when its ability to create better AIs surpasses that of human AI researchers by a sufficient margin (provided the AI is motivated to do so). An AI attains a decisive strategic advantage when its ability to optimize the universe surpasses that of humanity by a sufficient margin. Which of these happens first depends on what skills AIs have the advantage at relative to humans. If AIs are better at programming AIs than they are at taking over the world, then an intelligence explosion will happen first, and it will then be able to get a decisive strategic advantage soon after. But if AIs are better at taking over the world than they are at programming AIs, then an AI would get a decisive strategic advantage without an intelligence explosion occurring first.

Since an intelligence explosion happening first is usually considered the default assumption, I'll just sketch a plausibility argument for the reverse. There's a lot of variation in how easy cognitive tasks are for AIs compared to humans. Since programming AIs is not yet a task that AIs can do well, it doesn't seem like it should be a priori surprising if programming AIs turned out to be an extremely difficult task for AIs to accomplish, relative to humans. Taking over the world is also plausibly especially difficult for AIs, but I don't see strong reasons for confidence that it would be harder for AIs than starting an intelligence explosion would be. It's possible that an AI with significantly but not vastly superhuman abilities in some domains could identify some vulnerability that it could exploit to gain power, which humans would never think of. Or an AI could be enough better than humans at forms of engineering other than AI programming (perhaps molecular manufacturing) that it could build physical machines that could out-compete humans, though this would require it to obtain the resources necessary to produce them.

Furthermore, an AI that is capable of producing a more capable AI may refrain from doing so if it is unable to solve the AI alignment problem for itself; that is, if it can create a more intelligent AI, but not one that shares its preferences. This seems unlikely if the AI has an explicit description of its preferences. But if the AI, like humans and most contemporary AI, lacks an explicit description of its preferences, then the difficulty of the AI alignment problem could be an obstacle to an intelligence explosion occurring.

It also seems worth thinking about the policy implications of the differences between existential catastrophes from AI that follow an intelligence explosion versus those that don't. For instance, AIs that attempt to attain a decisive strategic advantage without undergoing an intelligence explosion will exceed human cognitive capabilities by a smaller margin, and thus would likely attain strategic advantages that are less decisive, and would be more likely to fail. Thus containment strategies are probably more useful for addressing risks that don't involve an intelligence explosion, while attempts to contain a post-intelligence explosion AI are probably pretty much hopeless (although it may be worthwhile to find ways to interrupt an intelligence explosion while it is beginning). Risks not involving an intelligence explosion may be more predictable in advance, since they don't involve a rapid increase in the AI's abilities, and would thus be easier to deal with at the last minute, so it might make sense far in advance to focus disproportionately on risks that do involve an intelligence explosion.

It seems likely that AI alignment would be easier for AIs that do not undergo an intelligence explosion, since it is more likely to be possible to monitor and do something about it if it goes wrong, and lower optimization power means lower ability to exploit the difference between the goals the AI was given and the goals that were intended, if we are only able to specify our goals approximately. The first of those reasons applies to any AI that attempts to attain a decisive strategic advantage without first undergoing an intelligence explosion, whereas the second only applies to AIs that do not undergo an intelligence explosion ever. Because of these, it might make sense to attempt to decrease the chance that the first AI to attain a decisive strategic advantage undergoes an intelligence explosion beforehand, as well as the chance that it undergoes an intelligence explosion ever, though preventing the latter may be much more difficult. However, some strategies to achieve this may have undesirable side-effects; for instance, as mentioned earlier, AIs whose preferences are not explicitly described seem more likely to attain a decisive strategic advantage without first undergoing an intelligence explosion, but such AIs are probably more difficult to align with human values.

If AIs get a decisive strategic advantage over humans without an intelligence explosion, then since this would likely involve the decisive strategic advantage being obtained much more slowly, it would be much more likely for multiple, and possibly many, AIs to gain decisive strategic advantages over humans, though not necessarily over each other, resulting in a multipolar outcome. Thus considerations about multipolar versus singleton scenarios also apply to decisive strategic advantage-first versus intelligence explosion-first scenarios.

Superintelligence via whole brain emulation

[x-post LessWrong]

Most planning around AI risk seems to start from the premise that superintelligence will come from de novo AGI before whole brain emulation becomes possible. I haven't seen any analysis that assumes both uploads-first and the AI FOOM thesis (Edit: apparently I fail at literature searching), a deficiency that I'll try to get a start on correcting in this post.

It is likely possible to use evolutionary algorithms to efficiently modify uploaded brains. If so, uploads would likely be able to set off an intelligence explosion by running evolutionary algorithms on themselves, selecting for something like higher general intelligence.

Since brains are poorly understood, it would likely be very difficult to select for higher intelligence without causing significant value drift. Thus, setting off an intelligence explosion in that way would probably produce unfriendly AI if done carelessly. On the other hand, at some point, the modified upload would reach a point where it is capable of figuring out how to improve itself without causing a significant amount of further value drift, and it may be possible to reach that point before too much value drift had already taken place. The expected amount of value drift can be decreased by having long generations between iterations of the evolutionary algorithm, to give the improved brains more time to figure out how to modify the evolutionary algorithm to minimize further value drift.

Another possibility is that such an evolutionary algorithm could be used to create brains that are smarter than humans but not by very much, and hopefully with values not too divergent from ours, who would then stop using the evolutionary algorithm and start using their intellects to research de novo Friendly AI, if that ends up looking easier than continuing to run the evolutionary algorithm without too much further value drift.

The strategies of using slow iterations of the evolutionary algorithm, or stopping it after not too long, require coordination among everyone capable of making such modifications to uploads. Thus, it seems safer for whole brain emulation technology to be either heavily regulated or owned by a monopoly, rather than being widely available and unregulated. This closely parallels the AI openness debate, and I'd expect people more concerned with bad actors relative to accidents to disagree.

With de novo artificial superintelligence, the overwhelmingly most likely outcomes are the optimal achievable outcome (if we manage to align its goals with ours) and extinction (if we don't). But uploads start out with human values, and when creating a superintelligence by modifying uploads, the goal would be to not corrupt them too much in the process. Since its values could get partially corrupted, an intelligence explosion that starts with an upload seems much more likely to result in outcomes that are both significantly worse than optimal and significantly better than extinction. Since human brains also already have a capacity for malice, this process also seems slightly more likely to result in outcomes worse than extinction.

The early ways to upload brains will probably be destructive, and may be very risky. Thus the first uploads may be selected for high risk-tolerance. Running an evolutionary algorithm on an uploaded brain would probably involve creating a large number of psychologically broken copies, since the average change to a brain will be negative. Thus the uploads that run evolutionary algorithms on themselves will be selected for not being horrified by this. Both of these selection effects seem like they would select against people who would take caution and goal stability seriously (uploads that run evolutionary algorithms on themselves would also be selected for being okay with creating and deleting spur copies, but this doesn't obviously correlate in either direction with caution). This could be partially mitigated by a monopoly on brain emulation technology. A possible (but probably smaller) source of positive selection is that currently, people who are enthusiastic about uploading their brains correlate strongly with people who are concerned about AI safety, and this correlation may continue once whole brain emulation technology is actually available.

Assuming that hardware speed is not close to being a limiting factor for whole brain emulation, emulations will be able to run at much faster than human speed. This should make emulations better able to monitor the behavior of AIs. Unless we develop ways of evaluating the capabilities of human brains that are much faster than giving them time to attempt difficult tasks, running evolutionary algorithms on brain emulations could only be done very slowly in subjective time (even though it may be quite fast in objective time), which would give emulations a significant advantage in monitoring such a process.

Although there are effects going in both directions, it seems like the uploads-first scenario is probably safer than de novo AI. If this is the case, then it might make sense to accelerate technologies that are needed for whole brain emulation if there are tractable ways of doing so. On the other hand, it is possible that technologies that are useful for whole brain emulation would also be useful for neuromorphic AI, which is probably very unsafe, since it is not amenable to formal verification or being given explicit goals (and unlike emulations, they don't start off already having human goals). Thus, it is probably important to be careful about not accelerating non-WBE neuromorphic AI while attempting to accelerate whole brain emulation. For instance, it seems plausible to me that getting better models of neurons would be useful for creating neuromorphic AIs while better brain scanning would not, and both technologies are necessary for brain uploading, so if that is true, it may make sense to work on improving brain scanning but not on improving neural models.