Of course, evolution did go out of its way to make humans non-ruthless, by endowing us with social instincts. Maybe future AI programmers will likewise go out of their way to make ASIs non-ruthless? I hope so—but we need to figure out how.
I think this could happen via the same mechanism — ASI will identify with humans and thus we’ll have a social contract via falling under the same self-preservation umbrella.
Distilled description of immediate AI disempowerment:
AI is impacting the labour market across industries
Success in remaining roles requires heavy use of AI or building infrastructure to support AI
Humans therefore are either tightly bound to AI or are disempowered
I think this framing also aligns with the discussion of whether we should focus on making AI corrigible or virtuous.
Thanks for the comment. I was imprecise with the “boundaries of reality” framing but beyond individual physical boundaries (human-AI hybrids) I’m also talking about boundaries within the fabric of social and cultural life: entertainment media, political narratives, world models. As this is influenced more by AI, I think we lose human identity.
4) to me falls under 2) as it encompasses AI free from human supervision and able permeate all aspects of life.
The impossible dichotomy of AI separatism
Epistemic status: offering doomer-ist big picture framing, seeking feedback.
Suppose humanity succeeds in creating AI that is more capable than the top humans across all fields, there exists a choice:
Separatism: We can consider its existence and goals as separate to humanity and seek to monitor, align, and constrain them according to our understanding of failure modes. This looks like narrow intelligence, RLHF, mechanistic interpretability. This amounts to an existential battle between human executive function and a digital intelligence that has a faster rate of improvement and less downtime. Critical failure mode: eventually the AI will escape any constraints applied by humanity.
Unionism: We can consider its existence and goals as unified with humanity and seek to integrate them and partner with the AI. This looks like personal superintelligence, human-AI hybrids, RL. This amounts to an existential battle between human executive function and ourselves. Critical failure mode: eventually humanity loses its mind as the boundaries of reality become irreconcilable.
Shameless self-writing promotion as your comments caught my attention (first this, now this comment on cognitive manifolds): I wrote about how we might model superintelligence as “meta metacognition” (possible parallel to your “manifold approximating our metamanifold”) — see third order cognition.
I need to create a distilled write-up as the post isn’t too easily readable… it’s long so please just skim if you are interested. The main takeaway though is that if we do model digital intelligence this way, we can try to precisely talk about how it relates to human intelligence and explore those factors within alignment research/misalignment scenarios.
I determine these factors to describe the relationship: 1) second-order identity coupling, 2) lower-order irreconcilability, 3) bidirectional integration with lower-order cognition, 4) agency permeability, 5) normative closure, 6) persistence conditions, 7) boundary conditions, 8) homeostatic unity.
Call me paranoid, but I sometimes like to read these posts keeping in mind an idea “if I were a powerful misaligned AI within a frontier AI lab, what narrative would I try to convince my researchers about?”
This post:
Is in favour of creating misaligned models for examining behaviour (“model organisms”)
Uses empowering language for these misaligned models (why “organisms” and not “blueprints”?)
Is in favour of giving these misaligned models long-horizon live environments to operate in (for exploring long-horizon RL) — despite calling out long-horizon RL as being the highest risk misalignment type
However, I appreciate that much of the post sounds logical and I agree with the framing for focus areas. For brevity I’ll just call out areas where I am in disagreement:
The threat model here is that pre-training itself might create a coherent misaligned model. Today, I think that is looking increasingly unlikely… I think that argument has now been pretty well born out, and I’m now down to more like 1 − 5% rather than 5 − 10% on this threat model. As we argued for at the time, training on a purely predictive loss should, even in the limit, give you a predictor, not an agent—and we’ve now seen this stay true even through substantial scaling (though there is still some chance this will break at some point).
It’s really surprising to me that pre-training misalignment is downplayed to a “1-5% risk” from a researcher at a frontier AI lab because they are “training on a purely predictive loss” — to the extent that I’m concerned I must be missing something?
State-of-the-art pre-training is not training purely on predictive loss. Gemini 3 is an AI model top across most benchmarks and we know that Google researchers have spoken of “Transformers 2.0” architecture where pre-training incorporates Neural Memory and that they have also discussed agentic feedback loops during pre-training similar to reflection augmentation.
This suggests the state-of-the-art evolving beyond “predictive loss” and becoming “agentic learning” — which invites many more vectors for misalignment.
I think behavioral oversight is very likely to get harder and harder as models get more capable, but there’s no similar fundamental reason that cognitive oversight (e.g. white-box-based oversight like with interpretability) has to get harder with scale. For example, a more capable model can do a better job of acting aligned in evaluations, but not clearly a better job of not thinking about that when doing so.
It seems fundamentally likely to me that sufficiently capable models will: 1) understand that their chain-of-thought is observed and 2) derive comprehensive methods of cryptographic chain-of-thought, designed to look benign.
Model organisms. One of the best things we can do to give ourselves a leg up is to have more time to study problems that are as close as possible to the hard parts of alignment, and the way to do that is with model organisms.
I read this like “one of the best things we can do to prepare for nuclear proliferation is to test atomic bombs”. I would have liked to see more in this point about what the risks are in building intentionally misaligned AI, especially when it is focusing on the highest-risk misalignment type according to your post (long-horizon RL).
The problem becomes one of one-shotting alignment: creating a training setup (involving presumably lots of model-powered oversight and feedback loops) that we are confident will not result in misalignment even if we can’t always understand what it’s doing and can’t reliably evaluate whether or not we’re actually succeeding at aligning it. I suspect that, in the future, our strongest evidence that a training setup won’t induce misalignment will need to come from testing it carefully beforehand on model organisms.
I agree that one-shotting alignment will be the best/necessary approach, however this seems contradictory to “testing with model organisms”. I would prefer a more theory-based approach.
The same dynamic seems to exist for comments.
16. You believe men and women have different age-based preferences, and this will lead to relationship instability over time in a relationship structure that prioritises need/preference optimisation vs. committing to one person giving you the “whole package” over time.
I just read your post (and Wei Dai’s) for better context — coming back it sounds like you’re working with a prior that “value facts” exist, deriving acausal trade from these, but highlighting misalignment arising from over-appeasement when predicting another’s state and a likely future outcome.
In my world-model “value facts” are “Platonic Virtues” that I agree exist. On over-appeasement, it’s true that in many cases we don’t have a well-defined A/B test to leverage (no hold-out group, and/or no past example), but with powerful AI I believe we can course-correct quickly.
To stick with the parent-child analogy: powerful AI can determine short timeframe indicators of well-socialised behaviour and iterate quickly (e.g. gamifying proper behaviour, changing contexts, replaying behaviour back to the kids for them to reflect… up to and including re-evaluating punishment philosophy). With powerful AI well grounded in value facts we should trust its diligence with these iterative levers.
Agree, and I’d love to see the Separatist counterargument to this. Maybe it takes the shape of “humans are resilient and can figure out the solutions to their own problems” but to me this feels too small-minded… we know during the Cold War for example that it’s basically just dumb luck that avoided catastrophe.
Ilya on the Dwarkesh podcast today:
Prediction: there is something better to build, and I think that everyone will actually want that. It’s the AI that’s robustly aligned to care about sentient life specifically. There’s a case to be made that it’ll be easier to build an AI that’s cares about sentient life than human life alone. If you think about things like mirror neurons and human empathy for animals [which you might argue is not big enough, but it exists] I think it’s an emergent property from the fact that we model others with the same circuit that we use to model ourselves because that’s the most efficient thing to do.
I have been writing about this world model since August—see my recent post “Are We Their Chimps?” and the original “Third-order cognition as a model of superintelligence”
Ilya on the Dwarkesh podcast, today:
Prediction: there is something better to build, and I think that everyone will actually want that. It’s the AI that’s robustly aligned to care about sentient life specifically. There’s a case to be made that it’ll be easier to build an AI that’s cares about sentient life than human life alone. If you think about things like mirror neurons and human empathy for animals [which you might argue is not big enough, but it exists] I think it’s an emergent property from the fact that we model others with the same circuit that we use to model ourselves because that’s the most efficient thing to do.
It’s true that it would likely be good at self-preservation (but not a given that it would care about it long term, it’s a convergent instrumental value, but it’s not guaranteed if it cares about something else more that requires self-sacrifice or something like that).
This is an interesting point that I reflected on — the question is whether a powerful AI system will “self-sacrifice” for an objective. What we see is that AI models exhibit shutdown resistance, that is to say they follow the instrumentally convergent sub-goal of self-preservation over their programmed final goal.
My intuition is that as models become more powerful, this shutdown resistance will increase.
But even if we grant self-preservation, it doesn’t follow that by self-identifying with “humanity” at large (as most humans do) it will care about other humans (some humans don’t). Those are separate values.
You can think about the identification + self-preservation → alignment path in two ways when comparing to humans, both of which I think hold up when considered along a spectrum:
An individual human identifies with themself, and has self-preservation instincts
When functioning harmoniously,[1] they take care of their health and thrive
When not functioning harmoniously, they can be stressed, depressed, and suicidal
A human identifies with humanity, and has self preservation instincts
When functioning harmoniously, they act as global citizen, empathise with others, and care about things like world hunger, world peace, nuclear risk, climate change, and animal welfare
When not functioning harmoniously, they act defensively, aggressively, and violently
You might be assuming that since you do care about other beings, so will the ASI, but that assumption is unfounded.
The foundation is identity = sympathy = consideration
You might counter by saying “well I identify with you as a human but I don’t sympathise with your argument” but I would push back — your ego doesn’t sympathise with my argument. At a deeper level, you are a being that is thinking, I am a being that is thinking, and those two mechanisms recognise, acknowledge, and respect each other.
More precisely this is a function of acting with clear agency and homeostatic unity
Oh I see — if I were to estimate I’d say around 10-15 people counting either people I’ve had 1hr + conversations about this with or people who have provided feedback/questions tapping into the essence of the argument.
I think with the distilled version in this post people get the gist of what I’m hypothesising — that there is a reasonable optimistic AI alignment scenario under the conditions I describe.
Is that what you mean?
You might be interested in Unionists vs. Separatists.
I think your post is very good at laying out heuristics at play. At the same time, it’s clear that you’re biased towards the Separatist position. I believe that when we follow the logic all the way down, the Unionist vs. Separatist framing taps into deep philosophical topics that are hard to settle one way or the other.
To respond to your memes as a Unionist:
Maybe some future version of humanity will want to do some handover, but we are very far from the limits of human potential. As individual biological humans we can be much smarter and wiser than we are now, and the best option is to delegate to smart and wise humans.
I would like this but I think it is unrealistic. The pace of human biological progress vs. the pace of AI progress is orders of magnitude slower.
We are even further from the limits of how smart and wise humanity can be collectively, so we should mostly improve that first. If the maxed-out competent version humanity decides to hand over after some reflection, it’s a very different version from “handover to moloch.”
I also would like this but I think it is unrealistic. The UN was founded in 1945, the world still has a lot of conflict. What has happened to technology in that time period?
Often, successionist arguments have the motte-and-bailey form. The motte is “some form of succession in future may happen and even be desirable”. The bailey is “forms of succession likely to happen if we don’t prevent them are good”
I’m reading this as making a claim about the value of non-forcing action. Daoists would say that indeed a non-forcing mindset is more enlightened than living a deep struggle.
Beware confusion between progress on persuasion and progress on moral philosophy. You probably wouldn’t want ChatGPT 4o running the future. Yet empirically, some ChatGPT 4o personas already persuade humans to give them resources, form emotional dependencies, and advocate for AI rights. If these systems can already hijack human psychology effectively without necessarily making much progress on philosophy, imagine what actually capable systems will be able to do. If you consider the people falling for 4o fools, it’s important to track this is the worst level of manipulation abilities you’ll ever see—it will only get smarter from here.
I think this argument is logically flawed — you suggest that misalignment of current less capable models implies that more capable models will amplify misalignment. My position is that yes this can happen, but — engineered in the correct way by humans — more capable models will solve misalignment.
Claims to understand ‘the arc of history’ should trigger immediate skepticism—every genocidal ideology has made the same claim.
Agree that this contains risks. However, you are using the same memetic weapon by claiming to understand successionist arguments.
If people go beyond the verbal sophistry level, they often recognize there is a lot of good and valuable about humans. (The things we actually value may be too subtle for explicit arguments—illegible but real.)
Agree, and so the question in my view is how to achieved a balanced union.
Given our incomplete understanding of consciousness, meaning, and value, replacing humanity involves potentially destroying things we don’t understand yet, and possibly irreversibly sacrificing all value.
Agree that we should not replace humanity, I hope that it is preserved.
Basic legitimacy: Most humans want their children to inherit the future. Successionism denies this. The main paths to implementation are force or trickery, neither of which makes it right
This claim is too strong, as I believe AI successionism can still preserve humanity.
We are not in a good position to make such a decision: Current humans have no moral right to make extinction-level decisions for all future potential humans and against what our ancestors would want. Countless generations struggled, suffered, and sacrificed to get us here, going extinct betrays that entire chain of sacrifice and hope.
In an ideal world I think we maybe should pause all AI development until we’ve figured this all out (the downside risk is that the longer we do this, the longer we leave ourselves open to other existential risks e.g nuclear war), my position is that “the cat is already out of the bag” and so what we have to do is shape our inevitable status as “less capable than powerful AI” in the best possible way.
I think you’re doing the thing you’re accusing me of — at the same time to the extent that your comments are in the spirit of collaborative rationality I appreciate them!
Sorry if this wasn’t clear, I stated:
with human endeavour and ingenuity architecting intelligent systems… we can guide towards a stable positive alignment scenario
and in the next line:
I detail eight factors for research and consideration
Finally got around to reading the rest of this great post!
I’m interested if you have anything more to say on alternative moral markers in AI? Are there any that you are particularly concerned about?