See something I’ve written which you disagree with? I’m experimenting with offering cash prizes of up to US$1000 to anyone who changes my mind about something I consider important. Message me our disagreement and I’ll tell you how much I’ll pay if you change my mind + details :-) (EDIT: I’m not logging into Less Wrong very often now, it might take me a while to see your message—I’m still interested though)
John_Maxwell
Karma: 13,352
Fair point. I also haven’t done much posting since adding the bounty to my profile. Was thinking it might attract the attention of people reading the archives, but maybe there just aren’t many archive readers.
There is some observational evidence that coffee drinking increases lifespan. I think the proposed mechanism has to do with promoting autophagy. https://www.acpjournals.org/doi/10.7326/M21-2977 But it looks like decaf works too. (Decaf has a bit of caffeine.)
I think somewhere else I read that unfiltered coffee doesn’t improve lifespan, so try to drink the filtered stuff?
In my experience caffeine dependence is not a big deal and might help my sleep cycle.
Eliezer is a good example of someone who built a lot of status on the back of “breaking” others’ unworkable alignment strategies. I found the AI Box experiments especially enlightening in my early days.
Fair enough.
My personal feeling is that poking holes in alignment strategies is easier than coming up with good ones, but I’m also aware that thinking that breaking is easy is probably committing some quantity of typical mind fallacy.
Yeah personally building feels more natural to me.
I agree a leaderboard would be great. I think it’d be cool to have a leaderboard for proposals as well—“this proposal has been unbroken for X days” seems like really valuable information that’s not currently being collected.
I don’t think I personally have enough clout to muster the coordination necessary for a tournament or leaderboard, but you probably do. One challenge is that different proposals are likely to assume different sorts of available capabilities. I have a hunch that many disagreements which appear to be about alignment are actually about capabilities.
In the absence of coordination, I think if someone like you was to simply start advertising themselves as an “uberbreaker” who can shoot holes in any proposal, and over time give reports on which proposals seem the strongest, that could be really valuable and status-rewarding. Sort of a “pre-Eliezer” person who I can run my ideas by in a lower stakes context, as opposed to saying “Hey Eliezer, I solved alignment—wallop me if I’m wrong!”
I wrote a comment on your post with feedback.
I don’t have anything prepared for red teaming at the moment—I appreciate the offer though! Can I take advantage of it in the future? (Anyone who wants to give me critical feedback on my drafts should send me a personal message!)
I skimmed the post, here is some feedback (context):
-
I’m probably not the best person to red team this since some of my own alignment ideas are along similar lines. I’m also a bit on the optimistic side about alignment more generally—it might be better to talk to a pessimist.
-
This sounds a bit like the idea of a “low-bandwidth oracle”.
-
I think the biggest difficulty is the one you explicitly acknowledged—boxing is hard.
-
But there are also problems around ensuring that bandwidth is actually limited. If you have a human check to see that the AGI’s output conforms to the spec, the AGI could put an infohazard in the output, or write code which looks like it does one thing but actually does another. If you have a computer program or AI check the AI’s output, the AGI could exploit a flaw in that program/AI. I think this sort of thing basically gets you a probabilistic safety measure, because there’s always a risk that there’s a flaw that the superintelligent AGI sees (or can infer) that you don’t see. (I like this intuition pump for seeing why these sort of problems are plausible.) I think probabilistic safety measures can be good if we stack a lot of them together in the right way.
-
The idea of emitting machine-checkable proofs is interesting. I’m not sure such proofs are very useful though. “Finding the right spec is one of the biggest challenges in formal methods.”—source. And finding the right spec seems more difficult to outsource to an unfriendly AI. In general, I think using AI to improve software reliability seems good, and tractable.
I think you’ll find it easier to get feedback if you keep your writing brief. Assume the reader’s time is valuable. Sentences like “I will mention some stuff later that maybe will make it more clear how I’d think about such a question.” should simply be deleted—make huge cuts. I think I might have been able to generate the bullet points above based on a 2-paragraph executive summary of your post. Maybe post a summary at the top, and say people are welcome to give feedback after just having read the summary.
Similarly, I think it is worth investing in clarity. If a sentence is unclear, I have a tendency to just keep reading and not mention it unless I have a prior that the author knows what they’re talking about. (The older I get, the more I assume that unclear writing means the author is confused and ignorable.) I like writing advice from Paul Graham and Scott Adams.
Personally I’m more willing to give feedback on prepublication drafts because that gives me more influence on what people end up reading. I don’t have much time to do feedback right now unfortunately.
-
Thanks for the reply!
As some background on my thinking here, last I checked there are a lot of people on the periphery of the alignment community who have some proposal or another they’re working on, and they’ve generally found it really difficult to get quality critical feedback. (This is based on an email I remember reading from a community organizer a year or two ago saying “there is a desperate need for critical feedback”.)
I’d put myself in this category as well—I used to write a lot of posts and especially comments here on LW summarizing how I’d go about solving some aspect or another of the alignment problem, hoping that Cunningham’s Law would trigger someone to point out a flaw in my approach. (In some cases I’d already have a flaw in mind along with a way to address it, but I figured it’d be more motivating to wait until someone mentioned a particular flaw in the simple version of the proposal before I mentioned the fix for it.)
Anyway, it seemed like people often didn’t take the bait. (Thanks to everyone who did!) Even with offering $1000 to change my view, as I’m doing in my LW user profile now, I’ve had 0 takers. I stopped posting on LW/AF nearly as much partially because it has seemed more efficient to try to shoot holes in my ideas myself. On priors, I wouldn’t have expected this to be true—I’d expect that someone else is going to be better at finding flaws in my ideas than I am myself, because they’ll have a different way of looking at things which could address my blind spots.
Lately I’ve developed a theory for what’s going on. You might be familiar with the idea that humans are often subconsciously motivated by the need to acquire & defend social status. My theory is that there’s an asymmetry in the motivations for alignment building & breaking work. The builder has an obvious status motive: If you become the person who “solved AI alignment”, that’ll be really good for your social status. That causes builders to have status-motivated blindspots around weak points in their ideas. However, the breaker doesn’t have an obvious status motive. In fact, if you go around shooting down peoples’ ideas, that’s liable to annoy them, which may hurt your social status. And since most proposals are allegedly easily broken anyways, you aren’t signaling any kind of special talent by shooting them down. Hence the “breaker” role ends up being undervalued/disincentivized. Especially doing anything beyond just saying “that won’t work”—finding a breaker who will describe a failure in detail instead of just vaguely gesturing seems really hard. (I don’t always find such handwaving persuasive.)
I think this might be why Eliezer feels so overworked. He’s staked a lot of reputation on the idea that AI alignment is a super hard problem. That gives him a unique status motive to play the red team role, which is why he’s had a hard time replacing himself. I think maybe he’s tried to compensate for this by making it low status to make a bad proposal, in order to browbeat people into self-critiquing their proposals. But this has a downside of discouraging the sharing of proposals in general, since it’s hard to predict how others will receive your ideas. And punishments tend to be bad for creativity.
So yeah, I don’t know if the tournament idea would have the immediate effect of generating deep insights. But it might motivate people to share their ideas, or generate better feedback loops, or better align overall status motives in the field, or generate a “useless” blacklist which leads to a deep insight, or filter through a large number of proposals to find the strongest ones. If tournaments were run on a quarterly basis, people could learn lessons, generate some deep ideas from those lessons, and spend a lot of time preparing for the next tournament.
A few other thoughts...
it’s going to be a significant danger to have breakers run out of exploit ideas and mistake that for a win for the builders
Perhaps we could mitigate this by allowing breakers to just characterize how something might fail in vague terms—obviously not as good as a specific description, but still provides some signal to iterate on.
It might be a challenge to create a similarly engaging format that allows for longer deliberation times on these harder problems, but it’s probably a worthwhile one.
I think something like a realtime Slack discussion could be pretty engaging. I think there is room for both high-deliberation and low-deliberation formats. [EDIT: You could also have a format in between, where the blue team gets little time, and the red team gets lots of time, to try to simulate the difference in intelligence between an AGI and its human operators.] Also, I’d expect even a slow, high-deliberation tournament format to be more engaging than the way alignment research often gets done (spend a bunch of time thinking on your own, write a post, observe post score, hopefully get a few good comments, discussion dies out as post gets old).
Thanks for writing this! Do you have any thoughts on doing a red team/blue team alignment tournament as described here?
Chapter 7 in this book had a few good thoughts on getting critical feedback from subordinates, specifically in the context of avoiding disasters. The book claims that merely encouraging subordinates to give critical feedback is often insufficient, and offers ideas for other things to do.
And just as I was writing this I came across another good example of the ‘you think you’re in competition with others like you but mostly you’re simply trying to be good enough’
I’m straight, so possibly unreliable, but I remember Michael Curzi as a very good-looking guy with a deep sexy voice. I believe him when he says other dudes are not competition for him 95% of the time. ;-)
I wrote a comment here arguing that voting systems tend to encourage conformity. I think this is a way in which the LW voting system could be improved. You might get rid of the unlabeled quality axis and force downvoters to be specific about why they dislike the comment. Maybe readers could specify which weights they want to assign to the remaining axes in order to sort comments.
I think Agree/Disagree is better than True/False, and Understandable/Confusing would be better than Clear/Muddled. Both of these axes are functions of two things (the reader and the comment) rather than just one (the comment) and the existing labels implicitly assume that the person voting on the comment has a better perspective on it than the person who wrote it. I think the opposite is more likely true—speaking personally at least, my votes tend to be less thoughtful than my comments.
Other axis ideas: Constructive/nonconstructive, important/unimportant. Could also try a “thank” react, and an “intriguing” or “interesting” react (probably replacing “surprise”—I like the idea of reinforcing novelty but the word “surprise” seems like too high of a bar?) Maybe also reacts for “this comment should’ve been longer/shorter”?
I’ll respond to the “Predict hypothetical sensors” section in this comment.
First, I want to mention that predicting hypothetical sensors seems likely to fail in fairly obvious ways, e.g. you request a prediction about a sensor that’s physically nonexistent and the system responds with a bunch of static or something. Note the contrast with the “human simulator” failure mode, which is much less obvious.
But I also think we can train the system to predict hypothetical sensors in a way that’s really useful. As in my previous comment, I’ll work from the assumptions (fairly weak IMO) that
-
We can control the data our systems get.
-
We are capable of doing regular old supervised learning—possibly in conjunction with transfer learning that gives the system generic prior knowledge like the meaning of English words, but not specific prior knowledge like details of our situation (unless we want that). Our supervised learning finds a function which maps training examples in X to labels in Y (labels may or may not correspond to “reality”).
In particular, these assumptions imply that our system doesn’t necessarily need to know whether a sensor it’s trying to predict exists physically (or if it would be physically possible to build).
But what if over the course of its operation, the system accidentally learns that a sensor of interest doesn’t exist? E.g. because it points a sensor that does exist in the direction of the one that doesn’t exist, and there’s nothing present. Ideally we could be reasonably confident of good “readings” from the nonexistent sensor even past that point.
To achieve this, we could make use of the “regular old supervised learning” assumption and construct a doctored dataset as follows:
-
Place camera S1 in the center of the vault, turn it on, and have it noisily rotate 360 degrees to capture a panorama.
-
Mount audio sensor S2 on the interior wall of the vault, turn both sensors on, and have S1 do the same panorama.
-
Create a supervised learning problem to predict the readings of S2 gathered in step 2, given S1′s panorama from step 1.
If we have prediction problems like this in our dataset, hypotheses which predict that S2 continues behaving as normal even if physically absent will achieve a lower loss, and should be preferentially chosen.
A similar method lets us preferentially choose hypotheses that predict that S2 will emit a special “TAMPERING ACTIVE” signal during tampering, then work as normal once tampering is complete:
-
Place audio sensor S2 in the corner of the vault and turn it on.
-
Move a robotic arm right next to S2.
-
Have the arm tamper with S2 so it no longer hears the sound of footfalls.
-
Discard S2 and replace it with a fresh new sensor, S2′, that has identical factory specifications.
-
Have someone walk beneath S2′ in full view of camera sensor S1.
-
Splice together a composite audio sensor reading with the following sequence: S2 readings from steps 1 and 2, a “TAMPERING ACTIVE” signal for the duration of step 3, nothing for step 4, and the S2′ readings from step 5.
-
Create a supervised learning problem to predict the composite reading from step 6, given the movements of the robot arm from steps 2 and 3, plus the visual readings from S1 with step 4 snipped out.
The “S2 is vulnerable to tampering” hypothesis predicts that footfalls should not be present in the audio. But since they are present in the audio, it will achieve a higher loss. Failure to predict the emission of the “TAMPERING ACTIVE” signal also gives this hypothesis a higher loss.
It might also be possible to do away with the “TAMPERING ACTIVE” signal somehow, e.g. by making S2 and S2′ identical except that the tampering which works on S2 doesn’t work on S2′, then splicing together readings from the robot arm doing the exact same tamper sequence on both sensors.
Finally, here are a few ideas for preferentially choosing hypotheses that predict accurate readings from technologically impossible sensors.
Supersensor Idea #1
-
Write a natural language description of a “telepathic sensor”. Describe it as being mounted right above a camera sensor S1.
-
Place boxes X, Y, and Z within the vault beneath the watchful eye of S1.
-
Turn on S1. Have a person enter the vault, visibly ponder for a moment, and open box Y.
-
Create a stream of “telepathic sensor” readings. Find a timestamp in the video where the person is pondering and hasn’t yet made a movement. Add a reading from the “telepathic sensor” at that timestamp: “The person intends to open box Y”.
Supersensor Idea #2
-
Write a natural language description of an “x-ray vision” sensor which can be pointed in various directions. Describe it as being mounted right above a camera sensor S1.
-
Place a deck of cards on a table beneath the watchful eye of S1.
-
Have a person enter the vault, shuffle the deck beneath the table, show a 9 of hearts to S1, and place it face down on top of the deck.
-
Create a stream of “x-ray vision” sensor readings. Specify that the sensor is initially pointed away, but once the action finishes, the sensor turns towards the table and registers: “The top card of the deck is a 9 of hearts.”
Supersensor Idea #3
-
Write a natural language description of a “nanobot sensor” which beeps whenever a nanobot is present within 10 feet.
-
Run an action sequence which builds nanobots and moves them within 10 feet of where the “nanobot sensor” is mounted.
-
Create a stream of “nanobot sensor” readings. Specify that the sensor starts beeping as soon as the nanobots are nearby.
A few more thoughts on translating natural language into sensor readings:
-
Excise a real sensor type from system’s prior knowledge. Ask your system to predict sensor data from a physical instance of this sensor, given a natural language description of its workings plus other readings from the environment. (H/T Romeo Stevens)
-
Make a weird modified sensor (e.g. a camera sensor which provides an upside down picture). Ask your system to predict readings from the modified sensor, given a natural language description of its modifications plus other readings from the environment.
Anyway, I’m not sure we need to reason about a physically impossible counterfactual or condition on the sensor existing physically. It seems perfectly coherent to ask “what is the answer to this thought experiment” rather than “if this sensor existed, what would it see”? For example, instead of the question “what would an Acme Corp camera mounted here see”, consider the question “if the light which passes through a pinhole at these coordinates intersected with a plane at these other coordinates, and the intersections were digitized and formatted the same way Acme Corp cameras format photos, what would be the resulting binary file?”
Humans don’t seem to have a lot of trouble performing thought experiments. If the system tries to fit the data with a hypothesis that references existing pretrained conceptual understanding, as I described above, that could give the system an inductive bias towards performing human-like thought experiments. This could be bad if human thought experiments are vulnerable to human deficiencies. It could also be good if we’d like the AI’s hypothetical sensors to behave in the same intuitive way our thought experiments do.
One possible concern is hypotheses which reference dataset doctoring. Obviously one could try to excise knowledge of that possibility. Another quick idea is to try & train a classifier to differentiate doctored vs non-doctored SmartVault sequences, and keep improving our fakes until the classifier can’t easily tell the difference? Or try to avoid any sort of branching so the system always acts like it’s dealing with a doctored dataset when in production? Could even fuzz the live data stream in a way that makes it appear doctored ;-) Finally, to get a sense for the cognitive signature of a doctoring-related hypothesis, one could train the system to solve some problems where the only way to minimize the loss is to think a lot about doctoring. Maybe a classifier which aims to detect the presence of doctoring-related cognition could be useful here.
Another possibility is an alternative hypothesis along the lines of “predict what the operator would want me to predict”—unclear if that’s desirable?
-
Thanks for the reply! I’ll respond to the “Hold out sensors” section in this comment.
One assumption which seems fairly safe in my mind is that as the operators, we have control over the data our AI gets. (Another way of thinking about it is if we don’t have control over the data our AI gets, the game has already been lost.)
Given that assumption, this problem seems potentially solvable
Moreover, my AI may be able to deduce the presence of the additional sensors very cheaply. Perhaps it can notice the sensors, or it can learn about my past actions to get a hint about where I may have placed them. If this is possible, then “Predict the readings of all sensors” need not be much more complex than “Predict the readings of one sensor.”
If the SmartVault learns a policy from data which was all generated prior to the installation of the new sensors, it seems unlikely that policy would deliberately account for the existence of those specific new sensors. We could leave info about your past actions out of the dataset as well.
An alternative approach in response to this bit:
I’m concerned that I will learn a better version of the human simulator which predicts the readings of all sensors and then outputs what a human would infer from the complete set.
The scenario is: we’re learning a function F1(A, S1) → D where A is an action sequence, S1 is readings from the known sensor, and D is a diamond location. Previously we’ve discussed two functions which both achieve perfect loss on our training data:
-
D1(A, S1) -- a direct translator which takes A and S1 into account
-
H1(A, S1) -- a simulation of what a human would believe if they saw A and S1
Let’s also consider two other functions:
-
D2(A, S1, S2) -- a direct translator which takes A, S1, and S2 into account
-
H2(A, S1, S2) -- a simulation of what a human would believe if they saw A, S1, and S2
Your concern is that there is a third function on the original (A, S1) domain which also achieves perfect loss:
H1′(A, S1) = H2(A, S1, P_S2(A, S1)) -- defining P_S2 as a prediction of S2′s readings given A & S1, we have H1′ as a simulation of what a human would believe if they saw A, S1, and readings for S2 predicted from A & S1.
Why would it be bad if gradient descent discovered H1′? Because then when it comes time to learn a policy, we incentivize policies which deceive predicted readings for S2 in addition to S1.
Here’s an idea for obtaining a function on the original (A, S1) domain which does not incentivize policies which deceive S2:
-
Learn a function F2 on the expanded domain (A, S1, S2), using a training set which is carefully constructed so that the only way to achieve perfect loss is to do a good job of taking readings from S2 into account. (For example, deliberately construct scenarios where the readings from S2 are not what you would expect if you were only looking at A and S1, and make ignoring A & S1 in favor of S2 key to labeling those scenarios correctly.) F2 could be closer to either D2 or H2, I don’t think it matters much.
-
Define a function F1_only(A, S1) = F2(A, S1, <hardcoded stream of boring S2 sensor data>).
Now let’s use F1_only as the target for learning our policy. I argue a policy has no incentive to deceive S2, because we know that F2 has been optimized to trust its S2 argument over its A and S1 arguments regarding what is going on around S2, and when we call F2 through F1_only, its S2 argument will always be telling it there are no interesting readings coming from S2. So, no bonus points for a policy which tries to fool S2 in addition to S1.
Maybe there is some kind of unintended consequence to this weird setup; I just came up with it and it’s still a bit half-baked in my mind. (Perhaps you could make some kind of exotic argument on the basis of inner optimizers and acausal trade between different system components?) But the meta point is there’s a lot of room for creativity if you don’t anthropomorphize and just think in terms of learning functions on datasets. I think the consequences of the “we control the data our AIs get” assumption could be pretty big if you’re willing to grant it.
-
I wrote a post in response to the report: Eliciting Latent Knowledge Via Hypothetical Sensors.
Some other thoughts:
-
I felt like the report was unusually well-motivated when I put my “mainstream ML” glasses on, relative to a lot of alignment work.
-
ARC’s overall approach is probably my favorite out of alignment research groups I’m aware of. I still think running a builder/breaker tournament of the sort proposed at the end of this comment could be cool.
-
Not sure if this is relevant in practice, but… the report talks about Bayesian networks learned via gradient descent. From what I could tell after some quick Googling, it doesn’t seem all that common to do this, and it’s not clear to me if there has been any work at all on learning the node structure (as opposed to internal node parameters) via gradient descent. It seems like this could be tricky because the node structure is combinatorial in nature and thus less amenable to a continuous optimization technique like gradient descent.
-
There was recently a discussion on LW about a scenario similar to the SmartVault one here. My proposed solution was to use reward uncertainty—as applied to the SmartVault scenario, this might look like: “train lots of diverse mappings between the AI’s ontology and that of the human; if even one mapping of a situation says the diamond is gone according to the human’s ontology, try to figure out what’s going on”. IMO this general sort of approach is quite promising, interested to discuss more if people have thoughts.
-
Eliciting Latent Knowledge Via Hypothetical Sensors
(Well, really I expect it to take <12 months, but planning fallacy and safety margins and time to iterate a little and all that.)
There’s also red teaming time, and lag in idea uptake/marketing, to account for. It’s possible that we’ll have the solution to FAI when AGI gets invented, but the inventor won’t be connected to our community and won’t be aware of/sold on the solution.
Edit: Don’t forget to account for the actual engineering effort to implement the safety solution and integrate it with capabilities work. Ideally there is time for extensive testing and/or formal verification.
Yes, if you’ve just created it, then the criteria are meaningfully different in that case for a very limited time.
It’s not obvious to me that this is only true right after creation for a very limited time. What is supposed to change after that?
I don’t see how we’re getting off track. (Your original statement was: ‘One such “clever designer” idea is decoupling plan generation from plan execution, which really just means that the plan generator has humans as part of the initial plan executing hardware.’ If we’re discussing situations where that claim may be false, it seems to me we’re still on track.) But you shouldn’t feel obligated to reply if you don’t want to. Thanks for your replies so far, btw.
My point is that plan execution can’t be decoupled successfully from plan generation in this way. “Outputting a plan” is in itself an action that affects the world, and an unfriendly superintelligence restricted to only producing plans will still win.
“Outputting a plan” may technically constitute an action, but a superintelligent system (defining “superintelligent” as being able to search large spaces quickly) might not evaluate its effects as such.
Yes, it is possible for plans to score highly under the first criterion but not the second. However, in this scenario the humans are presumably going to discourage such plans, so they effectively score the same as the second criterion.
I think you’re making a lot of assumptions here. For example, let’s say I’ve just created my planner AI, and I want to test it out by having it generate a paperclip-maximizing plan, just for fun. Is there any meaningful sense in which the displayed plan will be optimized for the criteria “plans which lead to lots of paperclips if shown to humans”? If not, I’d say there’s an important effective difference.
If the superintelligent search system also has an outer layer that attempts to collect data about my plan preferences and model them, then I agree there’s the possibility of incorrect modeling, as discussed in this subthread. But it seems anthropomorphic to assume that such a search system must have some kind of inherent real-world objective that it’s trying to shift me towards with the plans it displays.
The main problem is that “acting via plans that are passed to humans” is not much different from “acting via plans that are passed to robots” when the AI is good enough at modelling humans.
I agree this is true. But I don’t see why “acting via plans that are passed to humans” is what would happen.
I mean, that might be a component of the plan which is generated. But the assumption here is that we’ve decoupled plan generation from plan execution successfully, no?
So we therefore know that the plan we’re looking at (at least at the top level) is the result of plan generation, not the first step of plan execution (as you seem to be implicitly assuming?)
The AI is searching for plans which score highly according to some criteria. The criteria of “plans which lead to lots of paperclips if implemented” is not the same as the criteria of “plans which lead to lots of paperclips if shown to humans”.
I agree these are legitimate concerns… these are the kind of “deep” arguments I find more persuasive.
In that thread, johnswentworth wrote:
In particular, even if we have a reward signal which is “close” to incentivizing alignment in some sense, the actual-process-which-generates-the-reward-signal is likely to be at least as simple/natural as actual alignment.
I’d solve this by maintaining uncertainty about the “reward signal”, so the AI tries to find a plan which looks good under both alignment and the actual-process-which-generates-the-reward-signal. (It doesn’t know which is which, but it tries to learn a sufficiently diverse set of reward signals such that alignment is in there somewhere. I don’t think we can do any better than this, because the entire point is that there is no way to disambiguate between alignment and the actual-process-which-generates-the-reward-signal by gathering more data. Well, I guess maybe you could do it with interpretability or the right set of priors, but I would hesitate to make those load-bearing.)
(BTW, potentially interesting point I just thought of. I’m gonna refer to actual-process-which-generates-the-reward-signal as “approval”. Supposing for a second that it’s possible to disambiguate between alignment and approval somehow, and we successfully aim at alignment and ignore approval. Then we’ve got an AI which might deliberately do aligned things we disapprove of. I think this is not ideal, because from the outside this behavior is also consistent with an AI which has learned approval incorrectly. So we’d want to flip the off switch for the sake of caution. Therefore, as a practical matter, I’d say that you should aim to satisfy both alignment and approval anyways. I suppose you could argue that on the basis of the argument I just gave, satisfying approval is therefore part of alignment and thus this is an unneeded measure, but overall the point is that aiming to satisfy both alignment and approval seems to have pretty low costs.)
(I suppose technically you can disambiguate between alignment and approval if there are unaligned things that humans would approve of—I figure you solve this problem by making your learning algorithm robust against mislabeled data.)
Anyway, you could use a similar approach for the nice plans problem, or you could formalize a notion of “manipulation” which is something like: conditional on the operator viewing this plan, does their predicted favorability towards subsequent plans change on expectation?
Edit: Another thought is that the delta between “approval” and “alignment” seems like the delta between me and my CEV. So to get from “approval” to “alignment”, you could ask your AI to locate the actual-process-which-generates-the-labels, and then ask it about how those labels would be different if we “knew more, thought faster, were more the people we wished we were” etc. (I’m also unclear why you couldn’t ask a hyper-advanced language model what some respected moral philosophers would think if they were able to spend decades contemplating your question or whatever.)
Another edit: You could also just manually filter through all the icky plans until you find one which is non-icky.
(Very interested in hearing objections to all of these ideas.)
Power makes you dumb, stay humble.
Tell everyone in the organization that safety is their responsibility, everyone’s views are important.
Try to be accessible and not intimidating, admit that you make mistakes.
Schedule regular chats with underlings so they don’t have to take initiative to flag potential problems. (If you think such chats aren’t a good use of your time, another idea is to contract someone outside of the organization to do periodic informal safety chats. Chapter 9 is about how organizational outsiders are uniquely well-positioned to spot safety problems. Among other things, it seems workers are sometimes more willing to share concerns frankly with an outsider than they are with their boss.)
Accept that not all of the critical feedback you get will be good quality.
The book disrecommends anonymous surveys on the grounds that they communicate the subtext that sharing your views openly is unsafe. I think anonymous surveys might be a good idea in the EA community though—retaliation against critics seems fairly common here (i.e. the culture of fear didn’t come about by chance). Anyone who’s been around here long enough will have figured out that sharing your views openly isn’t safe. (See also the “People are pretty justified in their fears of critiquing EA leadership/community norms” bullet point here, and the last paragraph in this comment.)