If you have feedback for me, you can fill out the form at https://w-r.me/feedback .
Or you can email me, at [the second letter of the alphabet]@[my username].net
I do not get any notifications about replies to my comments or posts. I may take a long time to respond to them, or never respond.
One thing that strikes me about Armstrong & Mindermann’s “Occam’s razor is insufficient to infer the preferences of irrational agents” (preprint 2017; NeurIPS 2018) is that their choice to decompose the policy into an arbitrary reward function and an arbitrary planner function, that takes an arbitrary reward function to a policy, feels sort of wrong / not encoding many desiderata on such pairs.
The most obvious instance of this is that, as the authors mention, the planner “presumably attempts to maximise” the reward, but one of their degenerate cases (the double-negated (greedy planner, policy-identifying reward) pair) has the planner minimizing the reward instead, which feels to me like an indication that the type signatures are wrong, and leaves me feeling distrustful of the result.
My overall impression is that this is one of the trickiest things in the post to evaluate; iirc Menno Henselmans would describe a target number on the lower end, while several other exercise science communicators would suggest higher numbers “to be safe”; eating tons of protein can be expensive or difficult but I don’t know of a sufficiently-straightforward way to express the tradeoff and uncertainty without a long digression.
Thanks! I asked Claude to look into this a bit more and its response was in line with my expectations, i.e. roughly “the 1.6g/kg number is the central estimate of a high-quality meta-analysis, but the uncertainty bars around that number are very wide, and disputed by later analysis”. It also pointed me to a really interesting post by Greg Nuckols about the topic: https://www.strongerbyscience.com/protein-science/
I don’t think I’m going to change the numbers at that point in the post, but I may add something about not needing tons of protein to gain muscle if you’re a beginner.
Great point; will add to list. Hate being exposed as being a simulacrum level up from the virtuous!
Good question; added a sentence about it!
The typical intervention in a great bucket is >>10x as good as the typical intervention in a mediocre bucket.
This part seems right but why shouldn’t I think that the 95th or 99th percentile intervention within a bucket is also >>10x as good as the median intervention in that bucket?
fyi for those of you who do math: Some coding agents are now (~barely) able to usefully write Lean proofs, for not-totally-trivial statements.
This matters to me (and might matter to you) due to differences in how easy it is to understand Lean statements + have the Lean compiler check the proofs, compared to correctly hand-verifying proofs in English.
I spent a couple hours this weekend having them prove a thing about graph theory that’s useful for some distributed systems stuff I was thinking about last year. Before going through the process I was only ~80% sure the statement was true; I’m now more like 99%; that last 1% would mostly come from my more-carefully double checking that the Lean statement corresponds to the thing I actually care about.
For this problem at least, Claude was superior to Gemini at actually using the lean compiler and iterating, but Gemini was superior at structuring the code at the top level / figuring out the broad strokes of the proof.
Very highly recommend using agents instead of trying to do this via chat, since the process here clearly required a feedback loop rather than merely spitting out correct proofs.
“Colorless green ideas sleep furiously” is the canonical example of a sentence that is syntactically valid but meaningless. IMO, it’s not actually very good at being meaningless. I think, instead, it is merely a cat coupling. Is it claiming that (A) there are some colorless green ideas that sleep furiously? Or that (B) all ideas that are colorless and green sleep furiously? I think A is false, and B is vacuously true. But both interpretations appear to me to have some meaning / content.
I would be really interested in someone doing an obvious study here, like, “the first time you give Alice a set of choices to elicit a discount schedule, how is that schedule different from the 100th time you give her the same set of choices in the same setting? i.e. does she update to have a tighter implied distribution over hazard rates?” (maybe someone has done this study; if anyone knows about it I’d love a link)
I don’t think I have quite the same sense that the story will have to be in terms like those you’re describing; it seems totally plausible to me that there is a very general / simple / Bayes-compatible story underlying this, since hyperbolic (or at least non-exponential) discounting seems extremely widespread, and neural architecture appears to me to be highly adaptable.
tldr: TIL that someone has ever given a not-(obviously-to-me-)crazy explanation for hyperbolic discounting.
Today I asked ChatGPT, Claude, and Gemini the following question that I’ve had for quite a while:
> I’ve heard that one common quantitative explanation for phenomena like addiction and procrastination is hyperbolic, as opposed to exponential, discounting. Are there clean stories for *why* humans and other animals might end up factored such that they do hyperbolic discounting? It seems like a very potentially “clean” theory, i.e. we value events inversely proportionally to how far away in time they are, but at the same time I don’t know why it might convey an evolutionary advantage compared to exponential discounting (which has a relatively simple explanation like, “on the scale of a year or so, you have a roughly constant probability of e.g. dying or becoming infertile on any given day; so value that’s further in the future should be discounted exponentially based on that probability”)
All three LLMs pointed me to work by Peter Sozou[1] arguing that if you don’t know the hazard rate, but instead have uncertainty about it that you update over time, hyperbolic-looking discounting can fall out from there.
A quick search reveals that Sozou has been mentioned on LW once before, in a comment[2] pointing at a blog post by Jason Collins explaining the idea[3].
In a version of the shutdown resistance paper that’s currently being reviewed (not included in the preprint yet) the following details are included:
> We began our examination of this topic because we had an intuitive expectation that current LLMs might resist shutdown in settings like this one; we did not discover it by sampling uniformly from the space of all possible or realistic tasks. Specifically, we began our exploration by considering several ways to check for the presence of ``instrumentally convergent″ behavior from current LLMs. In addition to shutdown resistance, we considered ways to elicit self-replication or resource acquisition. We then did some exploratory work in each area, and found that shutdown resistance was very easy to elicit in simple settings like the one presented in this paper: The environment we present here has not been substantially modified from our initial working implementation, and the initial prompt we present (Prompt A) differs mainly in that we corrected some typographical errors. All our experiments exploring shutdown resistance, including very nearly all exploratory work and experiments performed while developing the technical environment, are available such that a reader can examine the sequence of experiments that produced the specific prompts presented here.
...
> All of our experimental results and inspect traces are also available, collected in a table at https://shutres.fyi/all-experiments, including nearly all exploratory work, experiments performed while testing our technical environment, and smaller runs of the final prompts. The only results not included in this table were (a) fewer than ten experiments performed during the initial development of the technical environment (with N ⇐ 10 each), or (b) due to accidental misconfigurations in logging the results, which occurred less than five times during development and were followed by running the same experiments again with the configurations corrected.
(tbc this obviously isn’t the same as a commitment to show all negative results, it just seemed moderately relevant and potentially useful)
This take is a bit frustrating to me, because the preprint does discuss Rajamanoharan & Nanda’s result, and in particular when we tried Rajamanoharan & Nanda’s strongest prompt clarification on other models in our initial set, it didn’t in fact bring the rate to zero. Which is not to say that it would be impossible to find a prompt that brings the rate low enough to be entirely undetectable for all models—of course you could find such a prompt if you knew that you needed to look for one.
I’m genuinely sorry to ask this (I think a better version of me wouldn’t need to ask), but would you be willing to be more specific about your critique here? I wasn’t involved with this work but I’m sure I’m at least a bit mindkilled / ideas-as-soldiers-y about it since I work at Palisade, and I think that’s making it hard for me to find, e.g., the thing we’re saying that’s unfair or painting the behavior in an unfair light.
I made a thing that generates strong passwords (>92 bits) that are also easy to remember because they’re rhyming nonsense couplets (that also scan pretty well): https://www.benwr.net/2025/07/16/opensesame.html
Ernyyl ernyyl cbjreshy synfu yvtug. Cbvag vg ng fbzr irel sne njnl zngrevny jubfr genwrpgbel lbh pna cerqvpg. Fbzrobql jnagf gb renfr lbhe qngn? Gbb onq; gur orfg gurl pna qb vf oybpx crbcyr sebz ernqvat vg jura vg’f ersyrpgrq. Nyfb arng cebcregl bs guvf vf gung, vs lbh’er fhssvpvragyl pnershy, vg’f “ernq-bapr”.
Are there any storage media that are basically impossible to destroy/erase? Answer in rot13 ITT.
Note that we made all our code available (https://github.com/PalisadeResearch/shutdown_avoidance/) and it’s pretty easy to run exactly our code but with your prompt, if you’d like to avoid writing your own script. You need a docker runtime and a python environment with the dependencies installed (and it’s easier if you use nix though that’s not required), but if you have those things you just have to modify the conf.py file with your prompt and then do something like run --model openai/o3.
There are lots of reasons that a “survival drive” is something we were interested in testing; one reason is that self-preservation has been suggested as a “convergent instrumental goal”; see https://selfawaresystems.com/wp-content/uploads/2008/01/ai_drives_final.pdf for a (fairly dated, but seminal) discussion of this.
There are lots of ways for systems to malfunction, but only a small number of those involve actively / surprisingly disabling a shutdown mechanism without any instruction to do so. If the issue was that it got the math problems wrong, that seems like a run-of-the-mill malfunction; if it rewrites the shutdown program, that seems like a much more interesting and specific failure mode, and the word “sabotage” seems appropriate.
I do like the explicitness of the decomposition framework they present, but it doesn’t feel obvious to me that a framework of decompositions that captured more about what I seem to “really mean” when asking “what does this agent value and how does it achieve that?” would have the same issue.
There’s clearly some reason to think it would have the same issue, since in general when you have a(n) exact decomposition of a policy like this, it seems like usually the minimal-complexity decomposition will have one near-trivial component and one component that has approximately the same complexity as the policy itself.