Echoing interstice's sentiment here, but I feel like the core insight of this post was already understood by/implicit in what a bunch of AI safety people are doing. It seems to me an application of the replaceability logic that effective altruists have discussed in many places. Even I (who has been far away from AI safety discussions for a long time now) had essentially a "duh" reaction to this post (even though for a lot of your posts I have a "wow" reaction).

As for an explicit past discussion, this 2023 talk by Buck Shlegeris in my opinion contains the core logic, although he doesn't use the legible/illegible terminology. In particular, one of the central points of the talk is how he chooses what to work on:

So here's the question that I ask myself. Assume that a lab ends up in the situation described earlier [roughly: a lab is very close to creating and deploying a transformative AI, other labs are somewhat but not very far behind and even less thoughtful about risks] and they're being as smart as possible about handling the alignment problems. How can I produce helpful changes to their alignment plan by doing technical research now?

Translated into the legible/illegible terminology, I interpret this question as something like "What problems are legible to me but illegible to AI labs currently (evidenced by them not already working on them), but will probably become legible to AI labs by the time they are about to deploy transformative AI?" (I realize there are a bunch of unstated assumptions in Buck's talk, and also I am not Buck, so I am kind of doing quite a lot of my own interpretation here, so you might reasonably disagree that the talk contains your core logic. :)

If I'm right that the core insight of the post is not novel, then the disagreement between prosaic safety researchers and people like you might not be about whether to work on legible problems vs illegible problems vs make-problems-more-legible (although there's probably some of that, like in your footnote about Paul), but instead about:

• Which problems are currently legible to key decision-makers. You think prosaic safety work in general is legible, so the thing to do is to work on philosophical questions which are illegible to almost everyone, while perhaps many prosaic safety people think that there are many prosaic safety problems that are illegible to purely capabilities researchers and policymakers and lab executives, and that the thing to do is to work on those prosaic safety problems.
• Which problems will become legible to them by the time they are about to deploy transformative AI. You are pessimistic about people's ability to realize the importance of certain philosophical problems, so you don't expect currently-illegible problems to become legible by the time TAI is deployed, whereas perhaps many prosaic safety people think that as AI becomes more and more capable, the alignment and other problems with AIs will automatically become more and more apparent to everyone (and in fact, that this has already been happening, with things like the release of ChatGPT, reporting of LLM psychosis, etc), so it's less important to spend effort persuading people about illegible problems, than to just work on the currently-illegible-to-decision-makers problems, so that when the time comes for the research to be useful, it has already been done.

My own views are much closer to yours than to the prosaic-safety view I laid out above. In fact, after watching Buck's talk in 2023, I wrote the following in a private conversation:

i feel like (even non-capabilities-advancing) people working on prosaic alignment and mechanistic interpretability have this attitude of like "well, i guess this is how AGI is gonna get built, so we might as well prepare for it by trying to align these kinds of models", but this kind of attitude further entrenches this narrative and makes it more likely that prosaic AGI gets built. so there's like a self-fulfilling prophecy type dynamic going on that worries me, and i think it's a good chunk of why i've always been unexcited about working on prosaic alignment

i.e. not only does working on legible safety problems burn the remaining timeline, it is the very thing which hyperstitions "AI timelines" into existence in the first place.

I didn't log the time I spent on the original blog post, and it's kinda hard to assign hours to this since most of the reading and thinking for the post happened while working on the modeling aspects of the MTAIR project. If I count just the time I sat down to write the blog post, I would guess maybe less than 20 hours.

As for the "convert the post to paper" part, I did log that time and it came out to 89 hours, so David's estimate of "perhaps another 100 hours" is fairly accurate.

Did you end up running it through your internal infohazard review and if so what was the result?

With help from David Manheim, this post has now been turned into a paper. Thanks to everyone who commented on the post!

Rob, are you able to disclose why people at Open Phil are interested in learning more decision theory? It seems a little far away from the AI strategy reports they've been publishing in recent years, and it also seemed like they were happy to keep funding MIRI (via their Committee for Effective Altruism Support) despite disagreements about the value of HRAD research, so the sudden interest in decision theory is intriguing.

I was reading parts of Superintelligence recently for something unrelated and noticed that Bostrom makes many of the same points as this post:

If the frontrunner is an AI system, it could have attributes that make it easier for it to expand its capabilities while reducing the rate of diffusion. In human-run organizations, economies of scale are counteracted by bureaucratic inefficiencies and agency problems, including difficulties in keeping trade secrets. These problems would presumably limit the growth of a machine intelligence project so long as it is operated by humans. An AI system, however, might avoid some of these scale diseconomies, since the AI’s modules (in contrast to human workers) need not have individual preferences that diverge from those of the system as a whole. Thus, the AI system could avoid a sizeable chunk of the inefficiencies arising from agency problems in human enterprises. The same advantage—having perfectly loyal parts—would also make it easier for an AI system to pursue long-range clandestine goals. An AI would have no disgruntled employees ready to be poached by competitors or bribed into becoming informants.

So the existence of this interface implies that A is “weaker” in a sense than A’.

Should that say B instead of A', or have I misunderstood? (I haven't read most of the sequence.)

Does anyone know how Brian Christian came to be interested in AI alignment and why he decided to write this book instead of a book about a different topic? (I haven't read the book and looked at the Amazon preview but couldn't find the answer there.)

HCH is the result of a potentially infinite exponential process (see figure 1) and thereby, computationally intractable. In reality, we can not break down any task into its smallest parts and solve these subtasks one after another because that would take too much computation. This is why we need to iterate distillation and amplification and cannot just amplify.

In general your post talks about amplification (and HCH) as increasing the capability of the system and distillation as saving on computation/making things more efficient. But my understanding, based on this conversation with Rohin Shah, is that amplification is also intended to save on computation (otherwise we could just try to imitate humans). In other words, the distillation procedure is able to learn more quickly by training on data provided by the amplified system compared to just training on the unamplified system. So I don't like the phrasing that distillation is the part that's there to save on computation, because both parts seem to be aimed at that.

(I am making this comment because I want to check my understand with you or make sure you understand this point because it doesn't seem to be stated in your post. It was one of the most confusing things about IDA to me and I'm still not sure I fully understand it.)

I still don't understand how corrigibility and intent alignment are different. If neither implies the other (as Paul says in his comment starting with "I don't really think this is true"), then there must be examples of AI systems that have one property but not the other. What would a corrigible but not-intent-aligned AI system look like?

I also had the thought that the implicative structure (between corrigibility and intent alignment) seems to depend on how the AI is used, i.e. on the particulars of the user/overseer. For example if you have an intent-aligned AI and the user is careful about not deploying the AI in scenarios that would leave them disempowered, then that seems like a corrigible AI. So for this particular user, it seems like intent alignment implies corrigibility. Is that right?

The implicative structure might also be different depending on the capability of the AI, e.g. a dumb AI might have corrigibility and intent alignment equivalent, but the two concepts might come apart for more capable AI.