This approach ignores the fact that if we use advanced LLMs to make new paradigm advancements that are extremely effective RL sociopaths, we’ll at that point have the help of the relatively harmless but still very powerful LLMs to do safety work on the RL agents — this is a major help with mitigating autonomy risks! Of course, there’s always the risk that new RL architecture discoveries create economic incentives to scale the scary RL agents without sufficient safety work, but the prospect of using HHH AI to align scary AI is weirdly under-explored when talking about that exact advanced LLM + advanced RL learner world.
Phaedrus
Karma: 104
Strange that no one is mentioning that Ada Palmer also wrote one of the most innovative and well-crafted sci-fi series of the 2000s, the Terra Ignota tetralogy. Very clearly an unusual women with an extremely big brain!
I think even if they hit some insane targets in the near term, the act of claiming explosive growth in a legible (and legally serious) growth estimate might be shocking to a lot of third parties, and have some wider memetic ripple effects. While it feels like the public has become “situationally aware” at a rapid pace in the last year, most people have not grappled deeply with the implications of possible transformative AI within the next few years.
I think this is true for sufficiently well-narrated nonfiction as well — I think a great deal of my psychology was shaped by reading about the classical world as a youth. Biography is probably the paradigmatic example of this genre — Ron Chernow’s book Titan, about the life of John D. Rockefeller, made the America of the late nineteenth century far more “real” to me than a more boradly informative textbook could have.
Historical fiction also capitalizes on this same effect, as it’s able both to bootstrap off the narrative richness and detail of real history and offer the reader an general education in the lived experience of that time.
I think essays like this are not very helpful for the AI safety agenda. In fact, they seem quite likely to do more harm than good?
I see dozens of arguments for and against specific AI risk models every day. A large fraction of these arguments (especially in the Twitter and podcast circles where people think deeply about AI and often work in AI risk or frontier labs) are against the Yudkowsky positions in IABIED. These are often arguments made by very smart people, well versed in the literature (including Yudkowsky’s writing), who have significant meta-cognitive awareness when it comes to their own intellectual weaknesses.
All this is to say: of all these “live” debates that I see play out every day, none involve the type of IQ/intelligence skepticism that is criticized in this argument. People debate the effects on power-seeking of differing RL task lengths, the speed of super-exponential AI-assisted research takeoff, and the emergence of AI situational awareness with regard to testing environments in frontier models, but no one that I can see, in any significant position of power, is falling for the farcical reasoning employed by Mr. Humman. Even Simon Lermen’s example of “misuse” of comparative advantage (https://www.lesswrong.com/posts/tBr4AtpPmwhgfG4Mw/comparative-advantage-and-ai) is a fine educational example for the specific labor dynamics of the period we’re entering now, where AI use and human labor coexist. In any event other than the (unlikely) event of a ASI takeoff in hours, there will be a period of years if not decades where comparative advantage holds for human and AI labor. Is there anyone who’s arguing from Ricardo about the likelihood of AI risk? Maybe Marc Andresseen types? IQ is even more lopsided — who in silicon valley in the year of our lord 2025 is an IQ denier? Is the Trump admin somehow unfamiliar with the concept that people or machines can be very smart?
Anyways, all this is to say: who is this post for? Why write and promote arguments against irrelevant incorrect positions instead of addressing the massive substantive issues where intelligent, informed people disagree? Why take a continual “outside view” of the AI risk debate instead of grappling with the substantive issues that sway the opinions of informed insiders at labs and in politics? Will smart critics of AI safety arguments resent that their side is represented in Yudkowsky posts by a cretin with the reflective acumen of a kindergartener?
I agree with your sentiment — I suppose I was implicitly presenting the bull case (or paradigmatic case) of cultural drift, wherein the future values are supported by future people but despised by their ancestors.
I think your example is closer to the familiar “Moloch” dynamic, where social and material technology leads to collective outcomes that are obviously undesirable to all involved. Moloch is certain to be an possible issue in any future world!
Although you don’t explicitly mention it, I feel like this whole post is about value drift. The doomers are generally right on the facts (and often on the causal pathways), and we do nonetheless consider the post-doom world better, but the 1-nth order effects of these new technologies reciprocally change our preferences and worldviews to favor the (doomed?) world created by the aforementioned new technologies.
The question of value drift is especially strange given that we have a “meta-intuition” that moral/social values evolving and changing is good in human history. BUT, at the same time, we know from historical precedent that we ourselves will not approve of the value changes. One might attempt to square the circle here by arguing that perhaps if we were, hypothetically, able to see and evaluate future changed values, that we would in reflective equilibrium accept these new values. Sadly, from what I can gather this is just not borne out by the social science: when it comes to questions of value drift, society advances by the deaths of the old-value-havers and the maturation of a next generation with “new” values.
For a concrete example, consider that most Americans have historically been Christians. In fact, the history of the early United States is deeply influenced by Christianity, sometimes swelling in certain periods to fanatical levels. If those Americans could see the secular American republic of 2025, with little religious belief and no respect for the moral authority of Christian scripture, they would most likely be morally appalled. Perhaps they might view the loss of “traditional God-fearing values” as a harm that in itself outweighs the cumulative benefits of industrial modernity. As a certain Nazarene said: “For what shall it profit a man, if he shall gain the whole world, and lose his own soul?” (Mark 8:36)
With this in mind, as a final exercise I’d like you, dear reader, to imagine a future where humanity has advanced enormously technologically, but has undergone such profound value shifts that every central moral and social principle that you hold dear has been abandoned, replaced with mores which you find alien and abhorrent. In this scenario, do you obey your moral intuitions that the future is one of Lovecraftian horror? Or do you obey your historical meta-intuitions that future people probably know better than you do?
I lean more towards the Camp A side, but I do understand and think there’s a lot of benefit to the Camp B side. Hopefully I can, as a more Camp A person, help explain to Camp B dwellers why we don’t reflexively sign onto these kinds of statements.
I think that Camp B has a bad habit of failing to model the Camp A rationale, based on the conversations I see in in Twitter discussions between pause AI advocates and more “Camp A” people. Yudkowsky is a paradigmatic example of the Camp B mindset, and I think it’s worth noting that a lot of people in the public readership of his book found the pragmatic recommendations therein to be extremely unhelpful. Basically, they (and I) see Yudkowsky’s plan as calling for a mass-mobilization popular effort against AI. But his plans, both in IABIED and his other writings, fail to grapple at all with the existing political situation in the United States, or with the geopolitical difficulties involved in political enforcement of an AI ban.
Remaining in this frame of “we make our case for [X course of action] so persuasively that the world just follows our advice” does not make for a compelling political theory on any level of analysis. Looking at the nuclear analogy used in IABIED, if Yudkowsky had advocated for a pragmatic containment protocol like the New START nuclear weapons deal or the Iran-US JCPOA deal, then we (the readers) could see that the Yudkowsky/Camp B side had thought deeply about the complexity of using political power to achieve actions in the full messiness of the real world. But Yudkowsky leaves the details of how a multinational treaty far more widely scoped than any existing multinational agreement would be worked out as an exercise for the reader! When Russia is actively involved in major war with a European country and China is preparing for an semi-imminent invasion of an American ally, the (intentionally?) vague calls for a multinational AI ban ring hollow. Why is there so little Rat brainpower devoted to the pragmatics of how AI safety could be advanced within the global and national political contexts?*
There are a few other gripes that I (speaking in my capacity as a Camp A denizen) have with the Camp B doctrine. Beyond inefficacy/unenforcability, the idea that the development of a superintelligence is a “one-shot” action without the ability to fruitfully learn from near-ASI non-ASI models seems deeply implausible. Also various staples of the Camp B platform — orthogonality and goal divergence out of distribution, notably — seem pretty questionable, or at least undersupported by existing empirical and theoretic work by the MIRI/PauseAI/Camp B faction.
*I was actually approached by an SBF representative in early 2022, who told me that SBF was planning on buying enough American congressional votes via candidate PAC donations that EA/AI safetyists could dictate US federal policy. This was by far the most serious AI safety effort I’ve personally witnessed come out of the EA community, and one of only a few that connected the AI safety agenda to the “facts on the ground” of the American political system.
As someone who’s done a fair amount of meditation and read a couple dozen books on the topic, I’d just like to flag the fact that this is pretty well examined in the community, and while meditation as a whole is quite pre-paradigmatic, there seems to be an emerging consensus on some of the ways that meditation harm can manifest.
First off, it’s obviously true that if you have a pre-existing tendency towards schizophrenia or any general mental instability, then in a very similar way to psychedelics, meditation can cause a psychotic break or similar episode of mental instability.
Secondly, and to me more interestingly, there’s an emerging consensus that one of the things meditation does is relieve subconscious mental tension accumulated by either large-scale or small-scale traumas in the course of one’s life. It’s very in vogue to use the term karma to refer to this accumulation of mental pathology that one can analogize to stuck priors or misfiring circuits in the synaptic map. This is also, of course, pre-paradigmatic, so it’s not great to take anything on this front super seriously, but it’s a very useful frame.
Now, when it comes to meditation harm, one of the things you see is that going very deep, very fast in meditation without processing this kind of mental tension or trauma can result in the trauma coming up in overwhelming or counterproductive ways. Often, people don’t talk about this in explicit terms, but I personally think it’s very obvious that the fastest and most straightforward ways of “making progress” on the meditative path are the Burmese Mahasi Sayadaw method, which also has by far the highest rate of negative side effects in meditation. To me, this indicates strongly that the Sayadaw method, because it is focused on blowing through to meditative insight without a lot of emotional hippie processing along the way, involves getting hit with all this kind of accumulated tension all at once in a very intensive fashion.
There is no “you” before coming into existence under closed/empty conditions, so it is categorically false that you are “leaving” anything or “finding out” anything. This makes your assertion that “from behind the veil of ignorance, we’d want the people in society to be making sure we don’t enter into bad experiences when we leave the veil of ignorance” obviously false, because your implication that there is a really existing veil of influence rests on birth functioning as a kind of “soul lottery” rather than the creation of a sentient entity.
Understood correctly, as Rawls writes in A Theory of Justice, the veil of ignorance is a hypothetical thought experiment. Your attempts to ground this in an analogy to the process of coming into existence muddle the ontological issues and provide a Pareto downgrade from Rawl’s original formulation.