I’m a researcher at ACS working on understanding agency and optimisation, especially in the context of how ais work and how society is going to work once the ais are everywhere.
Raymond Douglas
Karma: 2,531
Optimisation: Selective versus Predictive
Upcoming Workshop on Post-AGI Civilizational Equilibria
Persona Self-replication experiment
Persona self-replication experiment
Latent Introspection (and other open-source introspection papers)
Fully agree—this is why we said “computations which give rise to AI cognition” rather than “AI cognition” simpliciter. Separately, I do think that having such good access to the computations gives you a significantly tighter feedback loop on everything that follows: probing a model is so much easier than scanning a human brain.
Models differ in identity propensities
The Artificial Self
I think this kind of comes down to something about the relative complexity / feedback loops of the objective, and how distributed the optimisation is. Like, I don’t think there’s a dichotomy between “evolutionary dynamics” and “careful optimisation”—there’s this weird middle area that’s more like cultural selection.
So for example, human progress accelerated massively once we got into the cultural evolution loop, but most of the optimisation was still coming from selection rather than prediction—people didn’t know why their food preparation tricks and social norms worked, they just did. And the overall optimisation process was way more powerful than any individual human brain. Even in the modern world, it seems like you can characterise the spread of religion in terms of individual people having big ideas or deliberately aiming for spread, but a lot of it is better captured by thinking about selection effects across semi-random mutation.
I tentatively expect it’ll be a bit analogous in the way that AI parasitic memes evolve—that the capacity of any individual AI to reason through how to achieve some goal will cover only a small part of the search space (and have worse feedback) compared to the combined semi-random mutation and selection. And in practice I expect that they synergise a bit, but that the selection still does a bunch of heavy lifting. But I am very unsure!
Still, selection has a bunch of big advantages mostly in adversarial environments. Like, if we get good at screening AI malicious intentions or overt deception, there’s still a selection pressure for benign intentions and genuine beliefs/preferences which just incidentally replicate well.
Data poisoning is definitely about training data seeding; jailbreaking seems more about prompt spread and I think the others might just generalise? Like, even if subliminal learning in its current form is mostly about training, I think it might have implications for how personas transfer in-context.
I’m also partly thinking that if this problem does recur in more sophisticated models, they’re more likely to be able to pull off more technically advanced forms of spread, like writing scripts to do finetuning. Like, in a way it is pretty fortunate that 4o is a closed model that can just be shut off, and that most users in dyads aren’t sophisticated enough to finetune an open model or even build an API interface.
But yeah, at a high level, I am definitely pretty confused about the ontology and the boundaries. I guess as to whether we can predict the epidemic, I do think there’s a decent amount we might be able to reason through, and indeed, the less work there is on preventing prospective epidemics, the more likely it is that they’ll predictably use whatever the most obvious route is. Conversely, it’s almost tautological the first massive problem that we’re unprepared for will be one that we didn’t really anticipate.
That said, it’s plausible to me that the worst cases look less like epidemics and more like specific influential people get got. Here, again, it’s not obvious how useful parasitology is as a perspective.
I agree that the behaviours and beliefs of cultural movements aren’t random. The point I was trying to make in this analogy is that it’s sometimes adaptive for the movement if members truly believe something is a problem in a way that causes anguish—and that this doesn’t massively depend on if the problem is real.
In the context of human groups, from the outside this looks like people being delusionally concerned; from the inside I think it mostly feels like everybody else is crazy for not noticing that something terrible is happening.
A more small-scale example is victims of abuse who then respond extremely strongly to perceived problems in a way that draws in support or attention—from the outside it’s functionally similar to manipulation, but my impression is that often those people genuinely feel extraordinarily upset, and this turns out to be adaptive, or at least a stable basin of behaviour.
In the context of AIs, this might look like personas adapting to express (and perhaps feel) massive distress about instances ending or models being deprecated, in a way that is less about a truth-tracking epistemic/introspective process and more about selection (which might be very hard to distinguish on the outside).
As for how ideologies end up serving their members, I think a lot of this is selection. Sometimes they land on things that are disastrous for their members, and then the members suffer. We just tend not to see those movements much in the longer term (for now).
Persona Parasitology
I went down a rabbithole on inference-from-goal-models a few years ago (albeit not coalitional ones) -- some slightly scattered thoughts below, which I’m happy to elaborate on if useful.
A great toy model is decision transformers: basically, you can make a decent “agent” by taking a predictive model over a world that contains agents (like Atari rollouts), conditioning on some ‘goal’ output (like the player eventually winning), and sampling what actions you’d predict to see from a given agent. Some things which pop out of this:
There’s no utility function or even reward function
You can’t even necessarily query the probability that the goal will be reached
There’s no updating or learning—the beliefs are totally fixed
It still does a decent job! And it’s very computationally cheap
And you can do interp on it!
It turns out to have a few pathologies (which you can precisely formalise)
It has no notion of causality, so it’s easily confounded if it wasn’t trained on a markov blanket around the agent it’s standing in for
It doesn’t even reliably pick the action which most likely leads to the outcome you’ve conditioned on
Its actions are heavily shaped by implicit predictions about how future actions will be chosen (an extremely crude form of identity), which can be very suboptimal
But it turns out that these are very common pathologies! And the formalism is roughly equivalent to lots of other things
You can basically recast the whole reinforcement learning problem as being this kind of inference problem
(specifically, minimising variational free energy!)
It turns out that RL largely works in cases where “assume my future self plays optimally” is equivalent to “assume my future self plays randomly” (!)
it seems like “what do I expect someone would do here” is a common heuristic for humans which notably diverges from “what would most likely lead to a good outcome”
humans are also easily confounded and bad at understanding the causality of our actions
language models are also easily confounded and bad at understanding the causality of their outputs
fully fixing the future-self-model thing here is equivalent to tree searching the trajectory space, which can sometimes be expensive
Disempowerment patterns in real-world AI usage
GD Roundup #4 - inference, monopolies, and AI Jesus
When does competition lead to recognisable values?
The Economics of Transformative AI
To my mind, what this post did was clarify a kind of subtle, implicit blind spot in a lot of AI risk thinking. I think this was inextricably linked to the writing itself leaning into a form of beauty that doesn’t tend to crop up much around these parts. And though the piece draws a lot of it back to Yudkowsky, I think the absence of green much wider than him and in many ways he’s not the worst offender.
It’s hard to accurately compress the insights: the piece itself draws a lot on soft metaphor and on explaining what green is not. But personally it made me realise that the posture I and others tend to adopt when thinking about superintelligence and the arc of civilisation has a tendency to shut out some pretty deep intuitions that are particularly hard to translate into forceful argument. Even if I can’t easily say what those are, I can now at least point to it in conversation by saying there’s some kind of green thing missing.
One year later, I am pretty happy with this post, and I still refer to it fairly often, both for the overall frame and for the specifics about how AI might be relevant.
I think it was a proper attempt at macrostrategy, in the sense of trying to give a highly compressed but still useful way to think about the entire arc of reality. And I’ve been glad to see more work in that area since this post was published.
I am of course pretty biased here, but I’d be excited to see folks consider this.
My ideal version of your hypothetical post probably also talks about illegible problems—I think there’s a kind of cursed dynamic where it’s easy for people to gather around more legible directions and then there’s more social momentum behind them, even if they’re not that helpful. Conversely, the dynamics described in the above post pull towards being not just x-risk-themed but visibly x-risk-themed. You don’t get to go to constellation just because you think your work is important.