See also: Is "Strong Coherence" Anti-Natural?
Introduction
Many AI risk failure modes imagine strong coherence/goal directedness[1] (e.g. [expected] utility maximisers).
Such strong coherence is not represented in humans (or any other animal), seems unlikely to emerge from deep learning and may be "anti-natural" to general intelligence in our universe[2][3].
I suspect the focus on strongly coherent systems was a mistake that set the field back a bit, and it's not yet fully recovered from that error[4].
I think most of the AI safety work for strongly coherent agents (e.g. decision theory) will end up inapplicable/useless for aligning powerful systems, because powerful systems in the real world are "of an importantly different type".
Ontological Error?
I don't think it nails everything, but on a purely ontological level, @Quintin Pope and @TurnTrout's shard theory feels a lot more right to me than e.g. HRAD. HRAD is based on an ontology that seems to me to be mistaken/flawed in important respects.
The shard theory account of value formation (while lacking) seems much more plausible as an account of how intelligent systems develop values (where values are "contextual influences on decision making") than the immutable terminal goals in strong coherence ontologies. I currently believe that (immutable) terminal goals is just a wrong frame for reasoning about generally intelligent systems in our world (e.g. humans, animals and future powerful AI systems)[2].
Theoretical Justification and Empirical Investigation Needed
I'd be interested in more investigation into what environments/objective functions select for coherence and to what degree said selection occurs.
And empirical demonstrations of systems that actually become more coherent as they are trained for longer/"scaled up" or otherwise amplified.
I want advocates of strong coherence to explain why agents operating in rich environments (e.g. animals, humans) or sophisticated ML systems (e.g. foundation models[5]) aren't strongly coherent.
And mechanistic interpretability analysis of sophisticated RL agents (e.g. AlphaStar, OpenAI Five [or replications thereof]) to investigate their degree of coherence.
Conclusions
Currently, I think strong coherence is unlikely (plausibly "anti-natural"[3][2]) and am unenthusiastic about research agendas and threat models predicated on strong coherence.
Disclaimer
The above is all low confidence speculation, and I may well be speaking out of my ass[6].
- ^
By "strong coherence/goal directedness" I mean something like:
Informally: a system has immutable terminal goals.
Semi-formally: a system's decision making is well described as (an approximation) of argmax over actions (or higher level mappings thereof) to maximise the expected value of a single fixed utility function over states.
- ^
You cannot well predict the behaviour/revealed preferences of humans or other animals by the assumption that they have immutable terminal goals or are expected utility maximisers.
The ontology that intelligent systems in the real world instead have "values" (contextual influences on decision making) seems to explain their observed behaviour (and purported "incoherencies") better.
Many observed values in humans and other mammals (see[7]) (e.g. fear, play/boredom, friendship/altruism, love, etc.) seem to be values that were instrumental for increasing inclusive genetic fitness (promoting survival, exploration, cooperation and sexual reproduction/survival of progeny respectively). Yet, humans and mammals seem to value these terminally and not because of their instrumental value on inclusive genetic fitness.
That the instrumentally convergent goals of evolution's fitness criterion manifested as "terminal" values in mammals is IMO strong empirical evidence against the goals ontology and significant evidence in support of shard theory's basic account of value formation in response to selection pressure.
This is not to say that I think all coherence arguments are necessarily dead on arrival, but rather in practice, I think coherent behaviour (not executing strictly dominated strategies) acts upon our malleable values, to determine our decisions. We do not replace said values with argmax over a preference ordering.
As @TurnTrout says:
I have some take (which may or may not be related) like
utility is a measuring stick which is pragmatically useful in certain situations, because it helps corral your shards (e.g. dogs and diamonds) into executing macro-level sensible plans (where you aren't throwing away resources which could lead to more dogs and/or diamonds) and not just activating incoherently.
but this doesn't mean I instantly gain space-time-additive preferences about dogs and diamonds such that I use one utility function in all contexts, such that the utility function is furthermore over universe-histories (funny how I seem to care across Tegmark 4?).
- ^
E.g. if the shard theory account of value formation is at all correct, particularly the following two claims:
* Values are inherently contextual influences on decision making
* Values (shards) are strengthened (or weakened) via reinforcement events
Then strong coherence in the vein of utility maximisation just seems like an anti-natural form. See also[2]. I think evolutionarily convergent "terminal" values provides (strong) empirical evidence against the naturalness of strong coherence.
I could perhaps state my thesis that strong coherence is anti-natural more succinctly as:
Decision making in intelligent systems is best described as "executing computations/cognition that historically correlated with higher performance on the objective function a system was selected for performance on".
[This generalises the shard theory account of value formation from reinforcement learning to arbitrary constructive optimisation processes.]
It is of an importantly different type from the "immutable terminal goals"/"expected utility maximisation" I earlier identified with strong coherence.
- ^
I'm given the impression that the assumption of strong coherence is still implicit in some current AI safety failure modes [e.g. it underpins deceptive alignment[8]].)
- ^
Is "mode collapse" in RLHF'ed models an example of increased coherence?
- ^
I do think that my disagreements with e.g. deceptive alignment/expected utility maximisation is not simply a failure of understanding, but I am very much an ML noob, so there can still be things I just don't know. My opinions re: coherence of intelligent systems would probably be different in a significant way by this time next year.
- ^
I mention mammals because I'm more familiar with them, not necessarily because only mammals display these values.
- ^
In addition to the other prerequisites listed in the "Deceptive Alignment" post, deceptive alignment also seems to require a mesa-optimiser so coherent that it would be especially resistant to modifications to its mesa-objective. That is it requires very strong levels of goal content integrity.
I think that updating against strong coherence would require rethinking the staples of (traditional) alignment orthodoxy:
* Instrumental convergence (see[2])
This is not to say that they are necessarily no longer relevant in systems that aren't strongly coherent, but that to the extent they manifest at all, they manifest in (potentially very) different ways than originally conceived when conditioned on systems with immutable terminal goals.
(this response is cross-posted as a top-level post on my blog)
i expect the thing that kills us if we die, and the thing that saves us if we are saved, to be strong/general coherent agents (SGCA) which maximize expected utility. note that this is two separate claims; it could be that i believe the AI that kills us isn't SGCA, but the AI that saves us still has to be SGCA. i could see shifting to that latter viewpoint; i currently do not expect myself to shift to believing that the AI that saves us isn't SGCA.
to me, this totally makes sense in theory, to imagine something that just formulates plans-over-time and picks the argmax for some goal. the whole of instrumental convergence is coherent with that: if you give an agent a bunch of information about the world, and the ability to run eg linux commands, there is in fact an action that maximizes the amount of expected paperclips in the universe, and that action does typically entail recursively self-improving and taking over the world and (at least incidentally) destroying everything we value. the question is whether we will build such a thing any time soon.
right now, we have some specialized agentic AIs: alphazero is pretty good at reliably winning at go; it doesn't "get distracted" with other stuff. to me, waiting for SGCA to happen is like waiting for a rocket to get to space in the rocket alignment problem: once the rocket is in space it's already too late. the whole point is that we have to figure this out before the first rocket gets to space, because we only get to shoot one rocket to space. one has to build an actual inside view understanding of agenticity, and figure out if we'll be able to build that or not. and, if we are, then we need to solve alignment before the first such thing is built — you can't just go "aha, i now see that SGCA can happen, so i'll align it!" because by then you're dead, or at least past its decisive strategic advantage.
i'm not sure how to convey my own inside view of why i think SGCA can happen, in part because it's capability exfohazardous. maybe one can learn from IEM or the late 2021 MIRI conversations? i don't know where i'd send someone to figure this out, because i think i largely derived it from the empty string myself. it does strongly seem to me that, while a single particular neural net might not be the first thing to be an SGCA, we can totally bootstrap SGCA from existing ML technology; it might just take a clever trick or two rather than being the completely direct solution of "oh you train it like this and then it becomes SGCA". recursive self-improvement is typically involved.
we also have some AIs, including sydney, which aren't SGCA. it might even be that SGCA is indeed somewhat unnatural for a lot of current deep learning capabilities. nevertheless, i believe such a thing is likely enough to be built that it's what it takes for us to die — maybe non-SGCA AI's impact on the economy would slowly disempower us over the course of 20~40 years, but in those worlds AI tech gets good enough that 5 years into it someone figures out the right clever trick to build (something that bootstraps to) SGCA and we die of agentic intelligence explosion very fast before we get to see the slow economic disempowerement. in addition, i believe that our best shot is to build an aligned SGCA.
why haven't animals or humans gotten to SGCA? well, what would getting from messy biological intelligences to SGCA look like? typically, it would look like one species taking over its environment while developing culture and industrial civilization, overcoming in various ways the cognitive biases that happened to be optimal in its ancestral environment, and eventually building more reliable hardware such as computers and using those to make AI capable of much more coherent and unbiased agenticity.
that's us. this is what it looks like to be the first species to get to SGCA. most animals are strongly optimized for their local environment, and don't have the capabilities to be above the civilization-building criticality threshold that lets them build industrial civilization and then SGCA AI. we are the first one to get past that threshold; we're the first one to fall in an evolutionary niche that lets us do that. this is what it looks like to be the biological bootstrap part of the ongoing intelligence explosion; if dogs could do that, then we'd simply observe being dogs in the industrialized dog civilization, trying to solve the problem of aligning AI to our civilized-dog values.
we're not quite SGCA ourselves because, turns out, the shortest path from ancestral-environment-optimized life to SGCA is to build a successor that is much closer to SGCA. if that successor is still not quite SGCA enough, then its own successor will probly be. this is what we're about to do, probly this decade, in industrial civilization. maybe if building computers was much harder, and brains were more reliable to the point that rational thinking was not a weird niche thing you have to work on, and we got an extra million years or two to evolutionarily adapt to industrialized society, then we'd become properly SGCA. it does not surprise me that that is not, in fact, the shortest path to SGCA.
The misalignment argument ignores all moral arguments, we just build whatever even if it's a very bad idea. If we don't have the capability to do that now, we might gain it in 5 years, or LLM characters might gain it 5 weeks after waking up, and surely 5 years after waking up and disassembling the moon to gain moon-scale compute.
There'd need to be an argument that fixed goal optimizers are impossible in principle even if they are sought to be designed on purpose, and this seems false, because you can always wrap a mind in a plan evaluation loop. It's just a somewhat inefficient weird algorithm, and a very bad idea for most goals. But with enough determination efficiency will improve.