Alignment is mostly about making cognition aimable at all
(Epistemic status: attempting to clear up a misunderstanding about points I have attempted to make in the past. This post is not intended as an argument for those points.)
I have long said that the lion’s share of the AI alignment problem seems to me to be about pointing powerful cognition at anything at all, rather than figuring out what to point it at.
It’s recently come to my attention that some people have misunderstood this point, so I’ll attempt to clarify here.
In saying the above, I do not mean the following:
(1) Any practical AI that you’re dealing with will necessarily be cleanly internally organized around pursuing a single objective. Managing to put your own objective into this “goal slot” (as opposed to having the goal slot set by random happenstance) is a central difficult challenge. [Reminder: I am not asserting this]
Instead, I mean something more like the following:
(2) By default, the first minds humanity makes will be a terrible spaghetti-code mess, with no clearly-factored-out “goal” that the surrounding cognition pursues in a unified way. The mind will be more like a pile of complex, messily interconnected kludges, whose ultimate behavior is sensitive to the particulars of how it reflects and irons out the tensions within itself over time.
Making the AI even have something vaguely nearing a ‘goal slot’ that is stable under various operating pressures (such as reflection) during the course of operation, is an undertaking that requires mastery of cognition in its own right—mastery of a sort that we’re exceedingly unlikely to achieve if we just try to figure out how to build a mind, without filtering for approaches that are more legible and aimable.
Separately and independently, I believe that by the time an AI has fully completed the transition to hard superintelligence, it will have ironed out a bunch of the wrinkles and will be oriented around a particular goal (at least behaviorally, cf. efficiency—though I would also guess that the mental architecture ultimately ends up cleanly-factored (albeit not in a way that creates a single point of failure, goalwise)).
(But this doesn’t help solve the problem, because by the time the strongly superintelligent AI has ironed itself out into something with a “goal slot”, it’s not letting you touch it.)
Furthermore, insofar as the AI is capable of finding actions that force the future into some narrow band, I expect that it will tend to be reasonable to talk about the AI as if it is (more-or-less, most of the time) “pursuing some objective”, even in the stage where it’s in fact a giant kludgey mess that’s sorting itself out over time in ways that are unpredictable to you.
I can see how my attempts to express these other beliefs could confuse people into thinking that I meant something more like (1) above (“Any practical AI that you’re dealing with will necessarily be cleanly internally organized around pursuing a single objective…”), when in fact I mean something more like (2) (“By default, the first minds humanity makes will be a terrible spaghetti-code mess…”).
In case it helps those who were previously confused: the “diamond maximizer” problem is one example of an attempt to direct researchers’ attention to the challenge of cleanly factoring cognition around something a bit like a ‘goal slot’.
As evidence of a misunderstanding here: people sometimes hear me describe the diamond maximizer problem, and respond to me by proposing training regimes that (for all they know) might make the AI care a little about diamonds in some contexts.
This misunderstanding of what the diamond maximizer problem was originally meant to be pointing at seems plausibly related to the misunderstanding that this post intends to clear up. Perhaps in light of the above it’s easier to understand why I see such attempts as shedding little light on the question of how to get cognition that cleanly pursues a particular objective, as opposed to a pile of kludges that careens around at the whims of reflection and happenstance.
I’d be curious to understand why you believe this happens. Humans (the only general intelligence we have so far) seems to preserve some uncertainty over goal distributions. So it is unclear to me that generality will necessarily clarify goals.
To be a bit more concrete: I find it plausible that the AGI will encounter possible fine grained (concrete) goals that map into the same high level representation of its goal, whatever it may be. Then you have to refine what the goal representation was meant to mean. After all, a representation of the goal is not the goal itself necessarily. I believe this is what humans face, and why human goals are often a small mess.
If I understand you right, you’re thinking of scenarios like “the AI initially tries to create lots of watery looking stuff, but then it later realizes that watery looking stuff can be made of different substances (e.g., oxygen paired with protium vs. deuterium)”. We can imagine different outcomes here, like:
Some part of the AI feels like protium is important for “real water”, while another part feels that deuterium is important for “real water”. So the AI spends a lot of its resources going back and forth between the two goals, undoing its own work regularly.
The AI thinks about its values, and realizes that (for some complicated reason related to how it does reflection and how its goals work) it’s really deuterium-containing water that it likes, not protium-containing water. So it switches to making heavy water exclusively.
The AI thinks about its values, and realizes that (for some complicated reason related to how it does reflection and how its goals work) it wants to put 90% of its resources into producing heavy water, and 10% into producing light water.
Whether 1 counts as “one agent that’s internally conflicted” versus “multiple agents in a tug-of-war for control” might turn out to be a matter of semantics, depending on whether there turns out to be a crisp and natural interpretation of the word “agent”.
Whether 2 counts as “the agent self-modifying to change its goals” versus “the agent keeping the same goals but changing its probability distribution about which physical things those goals are pointing at”, may also turn out to be an unimportant or arbitrary distinction. It at least doesn’t seem very important from a human perspective: the first kind of agent may have a different internal design than the second kind of agent, but the behaviors are likely to look the same from the outside, since sufficiently coherent agents optimize expected utility (probability times utility) in practice, and it may be hard to say from the outside which parts of the expected utility are probability vs. utility, especially if the agent’s doing a bunch of complicated reflection and self-modification.
Similarly, whether 3 counts as “normative uncertainty about what’s best” versus “complete certainty in a meta-level goal that assigns some utility to heavy water and some utility to light water” may turn out to be a somewhat arbitrary distinction.
I understood Nate’s post to be saying that sufficiently capable agents tend to stop looking like 1, not that they necessarily tend to stop looking like 2 or like 3.
In principle it’s possible for an agent to stably consist of some sub-agent that optimizes heavy water on Mondays and Tuesdays and light water on the other days of the week. But because the first sub-agent will tend to want to disempower the second sub-agent (so it can produce more heavy water on other days of the week), and the second sub-agent will tend to want to disempower the first sub-agent, there are many scenarios where one sub-agent or the other ends up “winning”.
(Or, failing that, the two sub-agents will tend to eventually agree to effectively merge into a new agent that values a compromise between the goals of the original two sub-agents, since both agents can get more of what they want if they spend fewer resources on undoing the other agent’s hard work.)
You understood me correctly. To be specific I was considering the third case in which the agent has uncertainty about is preferred state of the world. It may thus refrain from taking irreversible actions that may have a small upside in one scenario (protonium water) but large negative value in the other (deuterium) due to eg decreasing returns, or if it thinks there’s a chance to get more information on what the objectives are supposed to mean.
I understand your point that this distinction may look arbitrary, but goals are not necessarily defined at the physical level, but rather over abstractions. For example, is a human with high level of dopamine happier? What is exactly a human? Can a larger human brain be happier? My belief is that since these objectives are built over (possibly changing) abstractions, it is unclear whether a single agent might iron out its goal. In fact, if “what the representation of the goal was meant to mean” makes reference to what some human wanted to represent, you’ll probably never have a clear cut unchanging goal.
Though I believe an important problem in this case is how to train an agent able to distinguish between the goal and its representation, and seek to optimise the former. I find it a bit confusing when I think about it.