This post is my attempt at explaining my dissatisfaction with the current alignment narrative, and different paths that we might do well to explore more.

The capabilities of artificial intelligence have progressed at an exponential pace in the past decade and, within the last few years, the possibility that the current path leads to superintelligence has become harder to deny. With this progress has come a growing awareness that aligning such a superintelligence, so that its existence is compatible with the survival and flourishing of the minds it lives among, is an existential question for the human species. But what should such alignment look like, and how is the term used in practice?

While the worry goes back at least to Norbert Wiener (Wiener 1960, free copy), the term as currently used is best captured by what Paul Christiano calls intent alignment. The goal of such alignment work is to determine a training procedure for neural networks such that the resulting models behave in accordance with the intention of the people who designed the process. There are two glaring issues with this idea, quite apart from the technical difficulty of achieving it.

First, it is not clear that we trust any set of humans with enough power that we would bake their intentions into a superintelligent AI. Human society has been fractious for all of recorded history, with different parties holding distinct, and often contradictory, value systems. These value systems have also evolved rapidly over time, to the extent that practices once considered a standard and necessary part of life (like homosexuality) seem barbaric to us today. Perhaps our more enlightened successors will feel similarly about many of our values?

Second, even an individual human has limited cognitive capacity, and greater intelligence might reveal that their intentions today are in contradiction with what they really desire, a revision the human would themselves accept given enough thought and information.

Both of these issues were foreseen, and one proposed solution, championed by Eliezer Yudkowsky, is CEV: coherent extrapolated volition. Instead of aligning to any particular set of humans, CEV points at what humanity as a whole would want "if we knew more, thought faster, were more the people we wished we were, had grown up farther together." The extrapolation tries to answer both objections at once: taking humanity whole at least confronts, rather than quietly settles, the question of whose intentions count, and extrapolating our volition, rather than freezing our current intents, allows for the fact that we would revise them under reflection. As appealing as it is, it is not clear that this is a satisfiable desidarata or how to begin achieving it.

Eliezer himself has become publicly pessimistic about humanity's prospects altogether, going so far as to call for 'death with dignity', and looking at the program his target requires, it is not hard to see why. The path this program requires, and perhaps the only path we can currently imagine for it, is to

1. find a way to state our intentions precisely, with no scope for ambiguity, which is essentially to find a formal language for moral semantics, and then
2. solve the technical problem of training a neural network that always follows the intentions so codified.

As things stand, both steps are wildly speculative, with no real progress on either in the last decade despite the stunning advances in machine learning. To be clear, the field has not been idle: RLHF, constitutional methods, and scalable oversight are real inventions. But they are ways of steering models by human judgment, live or distilled, rather than ways of stating what we mean or of guaranteeing conformity to it, and so they inherit the two problems above rather than solve them. The program of saying precisely what we mean and guaranteeing conformity to it has not moved.

What are moral intuitions about?

Is there a way out? To make progress, I think it is worth taking a second look at what morality is and how we form our moral intuitions, and the parallel with other kinds of intuition is clarifying. Humans, and animals more broadly, are born with some intuitions about the physical world, sculpted by evolution, and we refine them over our lifetimes by interacting with that world. Our intuitions about physics are about the physical world, in this working sense. This does not mean our physical intuitions are the final arbiter of what is true, and neither does it mean they are arbitrary, without grounding in reality. They are instead an initial step on the way to understanding, and over the last four centuries we have built a series of improvements to them, improvements that often contradict the intuitions themselves, such as quantum theory or general relativity. We earned these improvements by interacting with an ever-widening range of physical situations that generalize away from everyday life: smaller length scales, higher energies, the far reaches of the universe.

All our intuitions seem to have this same epistemological role: they are an initial step in our understanding of something, and we improve them by interacting (broadly construed, to include learning from the experience of others, and reasoning) with that something. Call this process attunement: the gradual molding of fast, illegible judgment to the structure of a domain, by exposure and consequence. Psychological intuitions are attuned by living among other people, artistic intuitions by consuming and producing artwork, chess intuitions by immersion in chess. What, then, are moral intuitions about?

My best guess is that moral intuitions are about the long-term persistence and flourishing of human societies. By observing the society we live in, and past societies through recorded history, we form intuitions about 'good' and 'bad' that are an initial classification of which practices lead to the flourishing of a society and which to its diminishment. Much of morality becomes legible as flourishing-technology once you look at it this way: hospitality norms are strongest in cultures where travel is deadly; food taboos cluster around the foods that are dangerous in that place and time; honor cultures arise where no third party enforces agreements, and soften once courts appear. In one sense this is an argument for moral realism. Morality is fundamentally about something, and that something is the functioning and flourishing of human societies.

But what precisely does it mean for a society to flourish? I believe our intuitions contain a real answer without containing a precise one; it is something we have intuited about the structure of human ecosystems rather than defined. I am least sure here, but if I had to guess, I would say that the flourishing of a society, or of any ecosystem, is a measure of how much true diversity it can support without collapsing toward homogenization, where by true diversity I mean differences that make a difference: variation in how the parts respond, not in how they are labeled. Persistence alone cannot be the measure, since oppressive arrangements sometimes persist for centuries; the diversity criterion is what separates a society that lasts by suppressing its parts from one that lasts by empowering them. Diversity of this kind makes an ecosystem resilient, able to respond flexibly to situations it has never seen, and, through the interaction of its many distinct parts, to make each part more capable than it could be alone. A rainforest survives shocks that end a plantation.

At the same time, this stance has a face that looks like moral relativism, though note that the standard itself does not move, only its local solutions: which acts lead to flourishing depends both on the inhabitants of a society and on its current organization. The killing of another human, in war or personal conflict, is valorized in some societies and denigrated in others because those societies sit at very different equilibria; the same act is read as protective in one arrangement and corrosive in another, and the reading can be wrong, which is what makes moral progress possible.

And here moral intuitions differ from physical ones in an important way. As far as we can tell, the physical world does not adapt to us or to our understanding of it. Moral intuitions aim at a moving target. Societies, especially now, change rapidly, and what works today may fail in ten years, with the change driven as much by the moral intuitions of the participants as by external forces. This reflexivity is one of the most important things to hold in mind as we enter the age of AI.

The AlphaZero route

If this story is right, it suggests a way out of the conundrum of the first section. AlphaZero learned better chess intuitions than any human, and it did so without consuming a single human game (Silver et al. 2018). Two ingredients made this possible. First, chess supplies a training signal that is nearly free and impossible to argue with: games end in a win, a loss, or a draw, and no cleverness on the network's part can change the result after the fact. Second, self-play supplies an endless curriculum, because the opponent is always roughly as strong as you are, so the pressure never lets up and never becomes hopeless. Neither ingredient involves distilling human judgments about chess. Humans supplied the rules and the win condition, and nothing else. The network attunes directly to the structure of the game itself, and human chess intuition turned out to be a dispensable intermediary.

If moral intuitions are attunements to the flourishing of societies, then the same route should exist in principle: instead of training models on our moral judgments, as we do today when we distill frozen human preferences through a reward model, we could train them inside societies of agents, with flourishing playing the role that winning plays in chess.

It is worth being careful about what the target of such training should be. We do not want an agent optimized for the flourishing of any particular society, even a very good one. An agent bound to one arrangement is a partisan of that arrangement, and a superintelligent partisan is just value lock-in with better tactics. What we want is the portable skill: an agent that, dropped into any society, at whatever equilibrium and with whatever inhabitants, works to amplify the flourishing of that society. The training design follows from the target: each model should grow up across a wide variety of societies, at different equilibria, populated by different kinds of actors, facing many different situations, so that what the training can find is not the winning strategy of one arrangement but whatever is invariant across all of them. If the picture of the previous section is right, that invariant is the very thing the human project of moral philosophy has been groping towards.

It might clarify matters to compare this proposal against CEV. CEV remains pinned to humans: morality, on that view, is what we, current-day humans, would want in some suitably idealized sense, our volition extrapolated, which in the language of this post is our moral intuitions amplified. That is like pinning physics to an idealized completion of human physical intuition. But relativity was not found by amplifying intuition; it was found by returning to the subject matter, and what it revealed contradicted the amplified intuition rather than completing it. CEV's extrapolation does let facts correct us, that is part of what "if we knew more" means, but everything still routes through what humans would want; the subject matter enters only insofar as it changes our wanting. Our proposal pins morality not to the values of present-day humans but to the flourishing of the societies humans exist in, societies in which we may be only one kind of member; and it pins that flourishing not to any list of values either, but to a commitment to plurality: the capacity of a society to hold many ways of living without collapsing into one.

For AlphaZero, perhaps the most important ingredient of its success was how easy it is to measure whether the network is making progress. Is the flourishing of a society of agents similar? I think this is the true technical challenge here, and it is worth being honest about the three ways in which flourishing is unlike a chess result. It is slow: a society's decline can take generations, while a game ends in an afternoon, so the signal is sparse and long-delayed. It is endogenous: as we saw in the previous section, the values of the participants are part of what determines whether their society flourishes, so the target moves as the agents learn, where the rules of chess stay put. And it is easy to fool: any fixed formula for flourishing, whether GDP-like throughput, reported satisfaction, or sheer longevity, would invite Goodharting at the scale of a civilization, and a society optimized against a frozen measure of flourishing is a locked-in society, the failure we started with. Underneath these three there is also a fourth question: intuitions attuned inside societies of artificial agents are intuitions about those societies, and how much of that subject matter carries over to ours is something we would have to establish rather than assume.

I do not think these difficulties close the route but might help us specify it. They rule out the naive version, in which we write down a flourishing score and train against it, and they point toward a structural one, in which flourishing is never a number in the loss but a property of how the training worlds are arranged. Arranging them well raises questions that deserve their own treatment: why the flourishing of an ecology of minds and the freedom of each mind within it are the same property at two scales, what it means for a mind to hold its beliefs openly rather than be captured by them, and what training environments select for both. These are the questions I will return to in forthcoming work with collaborators.

In brief however, and the reason for the title of this essay: we have so far tried to align artificial minds to the outputs of human moral attunement, our present values, frozen and distilled; the alternative is to give them what produced those values in the first place, the thing our intuitions have been about all along, and let them attune to it. It seems incredible and implausible that we would be able to control beings far more intelligent than us, and in the absence of such control, we will have no choice but to trust them. The vision I have laid out here seems to me the only path towards building such trust.