i’ve noticed a life hyperparameter that affects learning quite substantially. i’d summarize it as “willingness to gloss over things that you’re confused about when learning something”. as an example, suppose you’re modifying some code and it seems to work but also you see a warning from an unrelated part of the code that you didn’t expect. you could either try to understand exactly why it happened, or just sort of ignore it.
reasons to set it low:
each time your world model is confused, that’s an opportunity to get a little bit of signal to improve your world model. if you ignore these signals you increase the length of your feedback loop, and make it take longer to recover from incorrect models of the world.
in some domains, it’s very common for unexpected results to actually be a hint at a much bigger problem. for example, many bugs in ML experiments cause results that are only slightly weird, but if you tug on the thread of understanding why your results are slightly weird, this can cause lots of your experiments to unravel. and doing so earlier rather than later can save a huge amount of time
understanding things at least one level of abstraction down often lets you do things more effectively. otherwise, you have to constantly maintain a bunch of uncertainty about what will happen when you do any particular thing, and have a harder time thinking of creative solutions
reasons to set it high:
it’s easy to waste a lot of time trying to understand relatively minor things, instead of understanding the big picture. often, it’s more important to 80-20 by understanding the big picture, and you can fill in the details when it becomes important to do so (which often is only necessary in rare cases).
in some domains, we have no fucking idea why anything happens, so you have to be able to accept that we don’t know why things happen to be able to make progress
often, if e.g you don’t quite get a claim that a paper is making, you could resolve your confusion just by reading a bit ahead. if you always try to fully understand everything before digging into it, you’ll find it very easy to get stuck before actually make it to the main point the paper is making
there are very different optimal configurations for different kinds of domains. maybe the right approach is to be aware that this is an important hparameter and occasionally try going down some rabbit holes and seeing how much value it provides
This seems to be related to Goldfish Reading. Or maybe complementary. In Goldfish Reading one reads the same text multiple times, not trying to understand it all at once or remember everything, i.e., intentionally ignoring confusion. But in a structured form to avoid overload.
Yeah, this seems like a good idea for reading—lets you get best of both worlds. Though it works for reading mostly because it doesn’t take that much longer to do so. This doesn’t translate as directly to e.g what to do when debugging code or running experiments.
I think it’s very important to keep track of what you don’t know. It can be useful to not try to get the best model when that’s not the bottleneck. But I think it’s always useful to explicitly store the knowledge of what models are developed to what extent.
The algorithm that I have been using, where what to understand to what extend is not a hyperparameter, is to just solve the actual problems I want to solve, and then always slightly overdo the learning, i.e. I would always learn a bit more than necessary to solve whatever subproblem I am solving right now. E.g. I am just trying to make a simple server, and then I learn about the protocol stack.
This has the advantage that I am always highly motivated to learn something, as the path to the problem on the graph of justifications is always pretty short. It also ensures that all the things that I learn are not completely unrelated to the problem I am solving.
I am pretty sure if you had perfect control over your motivation this is not the best algorithm, but given that you don’t, this is the best algorithm I have found so far.
saying “sorry, just to make sure I understand what you’re saying, do you mean [...]” more often has been very valuable
yeah, turns off the combat mode
more importantly, both i and the other person get more out of the conversation. almost always, there are subtle misunderstandings and the rest of the conversation would otherwise involve a lot of talking past each other. you can only really make progress when you’re actually engaging with the other person’s true beliefs, rather than a misunderstanding of their beliefs.
hypothesis: intellectual progress mostly happens when bubbles of non tribalism can exist. this is hard to safeguard because tribalism is a powerful strategy, and therefore insulating these bubbles is hard. perhaps it is possible for there to exist a monopoly on tribalism to make non tribal intellectual progress happen, in the same way a monopoly on violence makes it possible to make economically valuable trade without fear of violence
Continuing the analogy:You’d want there to be a Tribe, or perhaps two or more Tribes, that aggressively detect and smack down any tribalism that isn’t their own. It needs to be the case that e.g. when some academic field starts splintering into groups that stereotype and despise each other, or when people involved in the decision whether to X stop changing their minds frequently and start forming relatively static ‘camps,’ the main Tribe(s) notice this and squash it somehow. And/or maybe arrange things so it never happens in the first place.I wonder if this sorta happens sometimes when there is an Official Religion?
another way to lean really hard into the analogy: you could have a Tribe which has a constitution/laws that dictate what kinds of argument are ok and which aren’t, has a legislative branch that constantly thinks about what kinds of arguments are non truthseeking and should be prohibited, a judicial branch that adjudicates whether particular arguments were truthseeking by the law, and has the monopoly on tribalism in that it is the only entity that can legitimately silence people’s arguments or (akin to exile) demand that someone be ostracized. there would also be foreign relations/military (defending the continued existence of the Tribe against all the other tribes out there, many of which will attempt to destroy the Tribe via very nontruthseeking means)
unfortunately this is pretty hard to implement. free speech/democracy is a very strong baseline but still insufficient. the key property we want is a system where true things systematically win over false things (even when the false things appeal to people’s biases), and it is sufficiently reliable at doing so and therefore intellectually legitimate that participants are willing to accept the outcome of the process even when it disagrees with what they started with. perhaps there is some kind of debate protocol that would make this feasible?
Prediction markets? Generally, track people’s previous success rates about measurable things.
prediction markets have two major issues for this use case. one is that prediction markets can only tell you whether people have been calibrated in the past, which is useful signal and filters out pundits but isn’t very highly reliable for out of distribution questions (for example, ai x-risk). the other is that they don’t really help much with the case where all the necessary information is already available but it is unclear what conclusion to draw from the evidence (and where having the right deliberative process to make sure the truth comes out at the end is the cat-belling problem). prediction markets can only “pull information from the future” so to speak.
BTW, I like the “monopoly on violence” analogy. We can extend it to include verbal violence—you can have an environment where it is okay to yell at people for being idiots, or you can have an environment where it is okay to yell at people for being politically incorrect. Both will shape the intellectual development in certain directions.
Conflicts arise is when you don’t have a monopoly, so sometimes people get yelled at for being idiots, other times for being politically incorrect, and then you have endless “wars” about whether we should or shouldn’t study a politically sensitive topic X with an open mind, both sides complaining about lack of progress (from their perspective).
The more mutually contradictory constraints you have, the more people will choose the strategy “let’s not do anything unusual”, because it is too likely to screw up according to some of the metrics and get yelled at.
random fun experiment: accuracy of GPT-4 on “Q: What is 1 + 1 + 1 + 1 + …?\nA:”
blue: highest logprob numerical token
orange: y = x
...I am suddenly really curious what the accuracy of humans on that is.
‘Can you do Addition?’ the White Queen asked. ‘What’s one and one and one and one and one and one and one and one and one and one?’
‘I don’t know,’ said Alice. ‘I lost count.’
This is a cool idea. I wonder how it’s able to do 100, 150, and 200 so well. I also wonder what are the exact locations of the other spikes?
Oh, I see your other graph now. So it just always guesses 100 for everything in the vicinity of 100.
Is it a very universal experience to find it easier to write up your views if it’s in response to someone else’s writeup? Seems like the kind of thing that could explain a lot about how research tends to happen if it were a pretty universal experience.
I think so/I have this. (I would emoji react for a less heavy response, but doesn’t work on older short forms)
The corollary is that it’s really annoying to respond to widely held views or frames which aren’t clearly written up anywhere. Particularly if these views are very inprecise and confused.
new galaxy brain hypothesis of how research advances: progress happens when people feel unhappy about a bad but popular paper and want to prove it wrong (or when they feel like they can do even better than someone else)
why it’s often necessary to have bad incremental papers that don’t introduce any generalizable techniques (nobody will care about the followup until it’s refuting the bad paper)
why so much of academia exists to argue that other academics are wrong and bad
why academics sometimes act like things don’t exist unless there’s a paper about them, even though the thing is really obvious
This subjectively seems to me to be the case.
Since there are basically no alignment plans/directions that I think are very likely to succeed, and adding “of course, this will most likely not solve alignment and then we all die, but it’s still worth trying” to every sentence is low information and also actively bad for motivation, I’ve basically recalibrated my enthusiasm to be centered around “does this at least try to solve a substantial part of the real problem as I see it”. For me at least this is the most productive mindset for me to be in, but I’m slightly worried people might confuse this for me having a low P(doom), or being very confident in specific alignment directions, or so on, hence this post that I can point people to.
I think this may also be a useful emotional state for other people with similar P(doom) and who feel very demotivated by that, which impacts their productivity.
a common discussion pattern: person 1 claims X solves/is an angle of attack on problem P. person 2 is skeptical. there is also some subproblem Q (90% of the time not mentioned explicitly). person 1 is defending a claim like “X solves P conditional on Q already being solved (but Q is easy)”, whereas person 2 thinks person 1 is defending “X solves P via solving Q”, and person 2 also believes something like “subproblem Q is hard”. the problem with this discussion pattern is it can lead to some very frustrating miscommunication:
if the discussion recurses into whether Q is hard, person 1 can get frustrated because it feels like a diversion from the part they actually care about/have tried to find a solution for, which is how to find a solution to P given a solution to Q (again, usually Q is some implicit assumption that you might not even notice you have). it can feel like person 2 is nitpicking or coming up with fully general counterarguments for why X can never be solved.
person 2 can get frustrated because it feels like the original proposed solution doesn’t engage with the hard subproblem Q. person 2 believes that assuming Q were solved, then there would be many other proposals other than X that would also suffice to solve problem P, so that the core ideas of X actually aren’t that important, and all the work is actually being done by assuming Q.
I can see how this could be a frustrating pattern for both parties, but I think it’s often an important conversation tree to explore when person 1 (or anyone) is using results about P in restricted domains to make larger claims or arguments about something that depends on solving P at the hardest difficulty setting in the least convenient possible world.
As an example, consider the following three posts:
Challenge: construct a Gradient Hacker
Gradient hacking is extremely difficult
My Objections to “We’re All Gonna Die with Eliezer Yudkowsky”
I think both of the first two posts are valuable and important work on formulating and analyzing restricted subproblems. But I object to citation of the second post (in the third post) as evidence in support of a larger point that doom from mesa-optimizers or gradient descent is unlikely in the real world, and object to the second post to the degree that it is implicitly making this claim.There’s an asymmetry when person I is arguing for an optimistic view on AI x-risk and person 2 is arguing for a doomer-ish view, in the sense that person I has to address all counterarguments but person 2 only has to find one hole. But this asymmetry is unfortunately a fact about the problem domain and not the argument / discussion pattern between I and 2.
I find myself in person 2′s position fairly often, and it is INCREDIBLY frustrating for person 1 to claim they’ve “solved” P, when they’re ignoring the actual hard part (or one of the hard parts). And then they get MAD when I point out why their “solution” is ineffective. Oh, wait, I’m also extremely annoyed when person 2 won’t even take steps to CONSIDER my solution—maybe subproblem Q is actually easy, when the path to victory aside from that is clarified.
In neither case can any progress be made without actually addressing how Q fits into P, and what is the actual detailed claim of improvement of X in the face of both Q and non-Q elements of P.
yeah, but that’s because Q is easy if you solve P
Very nicely described, this might benefit from becoming a top level post
here’s a straw hypothetical example where I’ve exaggerated both 1 and 2; the details aren’t exactly correct but the vibe is more important:
1: “Here’s a super clever extension of debate that mitigates obfuscated arguments [etc], this should just solve alignment”
2: “Debate works if you can actually set the goals of the agents (i.e you’ve solved inner alignment), but otherwise you can get issues with the agents coordinating [etc]”
1: “Well the goals have to be inside the NN somewhere so we can probably just do something with interpretability or whatever”
2: “how are you going to do that? your scheme doesn’t tackle inner alignment, which seems to contain almost all of the difficulty of alignment to me. the claim you just made is a separate claim from your main scheme, and the cleverness in your scheme is in a direction orthogonal to this claim”
1: “idk, also that’s a fully general counterargument to any alignment scheme, you can always just say ‘but what if inner misalignment’. I feel like you’re not really engaging with the meat of my proposal, you’ve just found a thing you can say to be cynical and dismissive of any proposal”
2: “but I think most of the difficulty of alignment is in inner alignment, and schemes which kinda handwave it away are trying to some some problem which is not the actual problem we need to solve to not die from AGI. I agree your scheme would work if inner alignment weren’t a problem.”
1: “so you agree that in a pretty nontrivial number [let’s say both 1&2 agree this is like 20% or something] of worlds my scheme does actually work- I mean how can you be that confident that inner alignment is that hard? in the world’s where inner alignment turns out to be easy then my scheme will work.”
2: “I’m not super confident, but if we assume that inner alignment is easy then I think many other simpler schemes will also work, so the cleverness that your proposal adds doesn’t actually make a big difference.”
So Q=inner alignment? Seems like person 2 not only pointed to inner alignment explicitly (so it can no longer be “some implicit assumption that you might not even notice you have”), but also said that it “seems to contain almost all of the difficulty of alignment to me”. He’s clearly identified inner alignment as a crux, rather than as something meant “to be cynical and dismissive”. At that point, it would have been prudent of person 1 to shift his focus onto inner alignment and explain why he thinks it is not hard.
Note that your post suddenly introduces “Y” without defining it. I think you meant “X”.
One possible model of AI development is as follows: there exists some threshold beyond which capabilities are powerful enough to cause an x-risk, and such that we need alignment progress to be at the level needed to align that system before it comes into existence. I find it informative to think of this as a race where for capabilities the finish line is x-risk-capable AGI, and for alignment this is the ability to align x-risk-capable AGI. In this model, it is necessary but not sufficient for alignment for alignment to be ahead by the time it’s at the finish line for good outcomes: if alignment doesn’t make it there first, then we automatically lose, but even if it does, if alignment doesn’t continue to improve proportional to capabilities, we might also fail at some later point. However, I think it’s plausible we’re not even on track for the necessary condition, so I’ll focus on that within this post.
Given my distributions over how difficult AGI and alignment respectively are, and the amount of effort brought to bear on each of these problems, I think there’s a worryingly large chance that we just won’t have the alignment progress needed at the critical juncture.
I also think it’s plausible that at some point before when x-risks are possible, capabilities will advance to the point that the majority of AI research will be done by AI systems. The worry is that after this point, both capabilities and alignment will be similarly benefitted by automation, and if alignment is behind at the point when this happens, then this lag will be “locked in” because an asymmetric benefit to alignment research is needed to overtake capabilities if capabilities is already ahead.
There are a number of areas where this model could be violated:
Capabilities could turn out to be less accelerated than alignment by AI assistance. It seems like capabilities is mostly just throwing more hardware at the problem and scaling up, whereas alignment is much more conceptually oriented.
After research is mostly/fully automated, orgs could simply allocate more auto-research time to alignment than AGI.
Alignment(/coordination to slow down) could turn out to be easy. It could turn out that applying the same amount of effort to alignment and AGI results in alignment being solved first.
However, I don’t think these violations are likely for the following reasons respective:
It’s plausible that our current reliance on scaling is a product of our theory not being good enough and that it’s already possible to build AGI with current hardware if you have the textbook from the future. Even if the strong version of the claim isn’t true, one big reason that the bitter lesson is true is that bespoke engineering is currently expensive, and if it became suddenly a lot cheaper we would see a lot more of it and consequently squeezing more out of the same hardware. It also seems likely that before total automation, there will be a number of years where automation is best modelled as a multiplicative factor on human researcher effectiveness. In that case, because of the sheer number of capabilities researchers compared to alignment researchers, alignment researchers would have to benefit a lot more to just break even.
If it were the case that orgs would pivot, I would expect them to currently be allocating a lot more to alignment than they do currently. While it’s still plausible that orgs haven’t allocated more to alignment because they think AGI is far away, and that a world where automated research is a thing is a world where orgs would suddenly realize how close AGI is and pivot, that hypothesis hasn’t been very predictive so far. Further, because I expect the tech for research automation to be developed at roughly the same time by many different orgs, it seems like not only does one org have to prioritize alignment, but actually a majority weighted by auto research capacity have to prioritize alignment. To me, this seems difficult, although more tractable than the other alignment coordination problem, because there’s less of a unilateralist problem. The unilateralist problem still exists to some extent: orgs which prioritize alignment are inherently at a disadvantage compared to orgs that don’t, because capabilities progress feeds recursively into faster progress whereas alignment progress is less effective at making future alignment progress faster. However, on the relevant timescales this may become less important.
I think alignment is a very difficult problem, and that moreover by its nature it’s incredibly easy to underestimate. I should probably write a full post about my take on this at some point, and I don’t really have space here to really dive into it here, but a quick meta level argument for why we shouldn’t lean on alignment easiness even if there is a non negligible chance of easiness is that a) given the stakes, we should exercise extreme caution and b) there are very few problems we have that are in the same reference class as alignment, and of the few that are even close, like computer security, they don’t inspire a lot of confidence.
I think exploring the potential model violations further is a fruitful direction. I don’t think I’m very confident about this model.
one man’s modus tollens is another man’s modus ponens:
“making progress without empirical feedback loops is really hard, so we should get feedback loops where possible”
“in some cases (i.e close to x-risk), building feedback loops is not possible, so we need to figure out how to make progress without empirical feedback loops. this is (part of) why alignment is hard”
Yeah something in this space seems like a central crux to me.
I personally think (as a person generally in the MIRI-ish camp of “most attempts at empirical work are flawed/confused”), that it’s not crazy to look at the situation and say “okay, but, theoretical progress seems even more flawed/confused, we just need to figure out some how of getting empirical feedback loops.”
I think there are some constraints on how the empirical work can possibly work. (I don’t think I have a short thing I could write here, I have a vague hope of writing up a longer post on “what I think needs to be true, for empirical work to be helping rather than confusedly not-really-helping”)
you gain general logical facts from empirical work, which can aide providing a blurry image of the manifold that the precise theoretical work is trying to build an exact representation of
A common cycle:
This model is too oversimplified! Reality is more complex than this model suggests, making it less useful in practice. We should really be taking these into account. [optional: include jabs at outgroup]
This model is too complex! It takes into account a bunch of unimportant things, making it much harder to use in practice. We should use this simplified model instead. [optional: include jabs at outgroup]
Sometimes this even results in better models over time.
for something to be a good way of learning, the following criteria have to be met:
tight feedback loops
transfer of knowledge to your ultimate goal
sufficiently interesting that it doesn’t feel like a grind
trying to do the thing you care about directly hits 2 but can fail 1 and 3. many things that you can study hit 1 but fail 2 and 3. and of course, many fun games hit 3 (and sometimes 1) but fail to hit 2.
corollary: for things with very long feedback loops, or where you aren’t motivated by default, it can be faster for learning to do something that is actually not directly the thing you care about
This is basically math (and computer science) education. On one hand, some parts are probably not very useful. On the other hand, some people expect that teachers will defend every single step along the way by explaining how specifically this tiny atom of knowledge improves the student’s future life. No, I am not preparing a PowerPoint presentation on how knowing that addition is associative and commutative will make you rich one day.
funnily enough, my experience has been almost entirely from the other direction—almost everything I know is from working directly on things I care about, and very little is from study. one of the reasons behind this shortform was trying to untangle why people spend lots of time studying stuff and whether/when it makes sense for me to study vs simply to learn by doing
I think it is good to use your goals as a general motivation for going approximately in some direction, but the opposite extreme of obsessing whether every single detail you learn contributes to the goal is premature optimization.
It reminds me of companies where, before you are allowed to spend 1 hour doing something, the entire team first needs to spend 10 hours in various meetings to determine whether that 1 hour would be spent optimally. I would rather spend all that time doing things, even if some of them turn out to be ultimately useless.
Sometimes it’s not even obvious in advance which knowledge will turn out to be useful.
lifehack: buying 3 cheap pocket sized battery packs costs like $60 and basically eliminates the problem of running out of phone charge on the go. it’s much easier to remember to charge them because you can instantaneously exchange your empty battery pack for a full one when you realize you need one, plugging the empty battery pack happens exactly when you swap for a fresh one, and even if you forget once or lose one you have some slack
the project of rationality is the project of becoming slightly more mesaoptimizery
Corollary to Others are wrong != I am right (https://www.lesswrong.com/posts/4QemtxDFaGXyGSrGD/other-people-are-wrong-vs-i-am-right): It is far easier to convince me that I’m wrong than to convince me that you’re right.
Quite a large proportion of my 1:1 arguments start when I express some low expectation of the other person’s argument being correct. This is almost always taken to mean that I believe that some opposing conclusion is correct. Usually I have to give up before being able to successfully communicate the distinction, let alone addressing the actual disagreement.
Some aspirational personal epistemic rules for keeping discussions as truth seeking as possible (not at all novel whatsoever, I’m sure there exist 5 posts on every single one of these points that are more eloquent)
If I am arguing for a position, I must be open to the possibility that my interlocutor may turn out to be correct. (This does not mean that I should expect to be correct exactly 50% of the time, but it does mean that if I feel like I’m never wrong in discussions then that’s a warning sign: I’m either being epistemically unhealthy or I’m talking to the wrong crowd.)
If I become confident that I was previously incorrect about a belief, I should not be attached to my previous beliefs. I should not incorporate my beliefs into my identity. I should not be averse to evidence that may prove me wrong. I should always entertain the possibility that even things that feel obviously true to me may be wrong.
If I convince someone to change their mind, I should avoid say things like “I told you so”, or otherwise try to score status points out of it.
I think in practice I adhere closer to these principles than most people, but I definitely don’t think I’m perfect at it.
(Sidenote: it seems I tend to voice my disagreement on factual things far more often (though not maximally) compared to most people. I’m slightly worried that people will interpret this as me disliking them or being passive aggressive or something—this is typically not the case! I have big disagreements about the-way-the-world-is with a bunch of my closest friends and I think that’s a good thing! If anything I gravitate towards people I can have interesting disagreements with.)
I should always entertain the possibility that even things that feel obviously true to me may be wrong.
I should always entertain the possibility that even things that feel obviously true to me may be wrong.
I find it a helpful framing to instead allow things that feel obviously false to become more familiar, giving them the opportunity to develop a strong enough voice to explain how they are right. That is, the action is on the side of unfamiliar false things, clarifying their meaning and justification, rather than on the side of familiar true things, refuting their correctness. It’s harder to break out of a familiar narrative from within.
current understanding of optimization
high curvature directions (hessian eigenvectors with high eigenvalue) want small lrs. low curvature directions want big lrs
if the lr in a direction is too small, it takes forever to converge. if the lr is too big, it diverges by oscillating with increasing amplitude
momentum helps because if your lr is too small, it makes you move a bit faster. if your lr is too big, it causes the oscillations to cancel out with themselves. this makes high curvature directions more ok with larger lrs and low curvature directions more ok with smaller lrs, improving conditioning
high curvature directions also have bigger gradients. this is the opposite of what we want because in a perfect world higher curvature directions would have smaller gradients (natural gradient does this but it’s usually too expensive). adam second moment / rmsprop helps because it makes gradients stay exactly the same size when the direction gets bigger, which is sorta halfway right
applied per param rather than per eigenvector
in real NNs edge of stability means it’s actually even more fine to have a too-high lr: the max curvature increases throughout training until it gets to the critical point where it would diverge, but then instead of diverging all the way the oscillations along the top eigenvector somehow cause the model to move into a slightly lower curvature region again, so that it stabilizes right at the edge of stability.
for Adam, these oscillations also cause second moment increases, which decreases preconditioned max curvature without affecting the original curvature. so this means the original max curvature can just keep increasing for Adam whereas it doesn’t for SGD (though apparently there’s some region where it jumps into a region with low original max curvature too)
https://distill.pub/2017/momentum/ really cool momentum explainer
https://arxiv.org/abs/2103.00065 - edge of stability
https://arxiv.org/abs/2207.14484 - edge of stability for adam
What does “the lr” mean in this context?
adhd is a mechanism for seeking domains with tight feedback loops
Understanding how an abstraction works under the hood is useful because it gives you intuitions for when it’s likely to leak and what to do in those cases.
takes on takeoff (or: Why Aren’t The Models Mesaoptimizer-y Yet)
here are some reasons we might care about discontinuities:
alignment techniques that apply before the discontinuity may stop applying after / become much less effective
makes it harder to do alignment research before the discontinuity that transfers to after the discontinuity (because there is something qualitatively different after the jump)
second order effect: may result in false sense of security
there may be less/negative time between a warning shot and the End
harder to coordinate and slow down
harder to know when the End Times are coming
alignment techniques that rely on systems supervising slightly smarter systems (i.e RRM) depend on there not being a big jump in capabilities
I think these capture 90% of what I care about when talking about fast/slow takeoff, with the first point taking up a majority
(it comes up a lot in discussions that it seems like I can’t quite pin down exactly what my interlocutor’s beliefs on fastness/slowness imply. if we can fully list out all the things we care about, we can screen off any disagreement about definitions of the word “discontinuity”)
some things that seem probably true to me and which are probably not really cruxes:
there will probably be a pretty big amount of AI-caused economic value and even more investment into AI, and AGI in particular (not really a bold prediction, given the already pretty big amount of these things! but a decade ago it may have been plausible nobody would care about AGI until the End Times, and this appears not to be the case)
continuous changes of inputs like compute or investment or loss (not technically an input, but whatever) can result in discontinuous jumps in some downstream metric (accuracy on some task, number of worlds paperclipped)
almost every idea is in some sense built on some previous idea, but this is not very useful because there exist many ideas  and it’s hard to tell which ones will be built on to create the idea that actually works (something something hindsight bias). this means you can’t reason about how they will change alignment properties, or use them as a warning shot
possible sources of discontinuity:
breakthroughs: at some point, some group discovers a brand new technique that nobody had ever thought of before / nobody had made work before because they were doing it wrong in some way / “3 hackers in a basement invent AGI”
depends on how efficient you think the research market is. I feel very uncertain about this
importantly I think cruxes here may result in other predictions about how efficient the world is generally, in ways unrelated to AI, and which may make predictions before the End Times
seems like a subcrux of this is whether the new technique immediately works very well or if it takes a nontrivial amount of time to scale it up to working at SOTA scale
overdetermined “breakthroughs”: some technique that didn’t work (and couldn’t have been made to work) at smaller scales starts working at larger scales. lots of people independently would have tried the thing
importantly, under this scenario it’s possible for something to simultaneously (a) be very overdetermined (b) have very different alignment properties
very hard to know which of the many ideas that don’t work might be the one that suddenly starts working with a few more OOMs of compute
at some scale, there is just some kind of grokking without any change in techniques, and the internal structure and generalization properties of the networks changes a lot. trends break because of some deep change in the structure of the network
mostly isomorphic to the previous scenario actually
for example, in worlds where deceptive alignment happens because at x params suddenly it groks to mesaoptimizer-y structure and the generalization properties completely change
at some scale, there is “enough” to hit some criticality threshold of some kind of thing the model already has. the downstream behavior changes a lot but the internal structure doesn’t change much beyond the threshold. importantly while obviously some alignment strategies would break, there are potentially invariants that we can hold onto
for example, in worlds where deceptive alignment happens because of ontology mismatch and ontologies get slowly more mismatched with scale, and then past some threshold it snaps over to the deceptive generalization
I think these can be boiled down to 3 more succinct scenario descriptions:
breakthroughs that totally change the game unexpectedly
mechanistically different cognition suddenly working at scale
more of the same cognition is different
The following things are not the same:
Schemes for taking multiple unaligned AIs and trying to build an aligned system out of the whole
I think this is just not possible.
Schemes for taking aligned but less powerful AIs and leveraging them to align a more powerful AI (possibly with amplification involved)
This breaks if there are cases where supervising is harder than generating, or if there is a discontinuity. I think it’s plausible something like this could work but I’m not super convinced.
In the spirit of https://www.lesswrong.com/posts/fFY2HeC9i2Tx8FEnK/my-resentful-story-of-becoming-a-medical-miracle , some anecdotes about things I have tried, in the hopes that I can be someone else’s “one guy on a message board. None of this is medical advice, etc.
No noticeable effects from vitamin D (both with and without K2), even though I used to live somewhere where the sun barely shines and also I never went outside, so I was almost certainly deficient.
I tried Selenium (200mg) twice and both times I felt like utter shit the next day.
Glycine (2g) for some odd reason makes me energetic, which makes it really bad as a sleep aid. 1g taken a few hours before bedtime is substantially less disruptive to sleep, but I haven’t noticed substantial improvements.
Unlike oral phenylephrine, intranasal phenylephrine does things, albeit very temporarily, and is undeniably the most effective thing I’ve tried, though apparently you’re not supposed to use it too often, so I only use it when it gets really bad.
hypothesis: the kind of reasoning that causes ML people to say “we have made no progress towards AGI whatsoever” is closely analogous to the kind of reasoning that makes alignment people say “we have made no progress towards hard alignment whatsoever”
ML people see stuff like GPT4 and correctly notice that it’s in fact kind of dumb and bad at generalization in the same ways that ML always has been. they make an incorrect extrapolation, which is that AGI must therefore be 100 years away, rather than 10 years away
high p(doom) alignment people see current model alignment techniques and correctly notice that they fail to tackle the AGI alignment problem in the same way that alignment techniques always have. they make an incorrect extrapolation and conclude that p(doom) = 0.99, rather than 0.5
(there is an asymmetry which is that overconfidence that alignment will be solved is much more dangerous than overconfidence that AGI will be solved)
It’s differential progress that matters in alignment. I.e., if you expected that we need additional year of alignment research after creating AGI, it still looks pretty doomed, even if you admit overall progress in field.
sure, but seems orthogonal to the thing i’m describing—the claim is that a lot of alignment work on current models has ~no bearing on progress towards aligning AGI.
One of the greatest tragedies of truth-seeking as a human is that the things we instinctively do when someone else is wrong are often the exact opposite of the thing that would actually convince the other person.
Schmidhubering the agentic LLM stuff pretty hard https://leogao.dev/2020/08/17/Building-AGI-Using-Language-Models/
Rightfully so! Read your piece back in 2021 and found it true & straightforward.
an interesting fact that I notice is that in domains where there are are a lot of objects in consideration, those objects have some structure so that they can be classified, and how often those objects occur follows a power law or something, there are two very different frames that get used to think about that domain:
a bucket of atomic, structureless objects with unique properties where facts about one object don’t really generalize at all to any other object
a systematized, hierarchy or composition of properties or “periodic table” or full grid or objects defined by the properties they have in some framework
and a lot of interesting things happen when these collide or cooccur, or when shifting from one to the other
I know my description above is really abstract, so here are a bunch of concrete examples that all gesture at the same vibe:
basically all languages have systematic rules in general but special cases around the words that people use very often. this happens too often in unrelated languages to be a coincidence, and as a native/fluent speaker it always feels very natural but as a language learner it’s very confusing. for example, for languages with conjugations, a few of the most common verbs are almost always irregular. e.g [to be, am, is, are, was, were] (english), [sein, bin, ist, war, sind] (german), [être, suis, est, était, sont] (french); small counting numbers are often irregular [first, second, third], [两个], [premier], [ひとつ、ふたつ、みっつ]. my theory for why this makes sense to natives but not to language learners is that language learners learn things systematically from the beginning, and in particular don’t deal with the true distribution of language usage but rather an artificially flat one designed to capture all the language features roughly equally.
often, when there is a systematic way of naming things, the things that are most common will have special names/nicknames (eg IUPAC names vs common names). sometimes this happens because those things were discovered first before the systematization happened, and the once the systematization happens everyone is still used to the old names for some things. but also even if you start with the systematized thing, often people will create nicknames after the fact.
it often happens that we write software tools for a specific problem, and then later realize that that problem is a special case of a more general problem. often going more general is good because it means we can use the same code to do a wider range of things (which means less bugs, more code reuse, more elegant code). however, the more general/abstract code is often slightly clunkier to use for the common case, so often it makes sense to drop down a level of abstraction if the goal is to quickly hack something together.
when compressing some distribution of strings, the vast majority of the possible but unlikely strings can be stored basically verbatim with a flag and it is very easy to tell properties of the string by looking at the compressed representation; whereas for the most common strings they have to map to short strings that destroy all structure of the data without the decompressor. though note that not all the examples can be described as instances of compression exactly
sometimes, there’s friction between people who are using the systematizing and people who are doing the atomic concepts thing. the systematizer comes off as nitpicky, pedantic, and removed from reality to the atomic concepts person, and the atomic concepts person comes off as unrigorous, uncosmopolitan, and missing the big picture to the systematizer.
I think the concept of zero only being invented long after the other numbers is also an instance of this—in some sense for basic everyday usage in counting things, the existence of zero is a weird technicality, and I could imagine someone saying “well sure yes there is a number that comes before zero, but it’s not useful for anything, so it’s not worth considering”. I think a lot of math (eg abstract algebra) is the result of applying truly enormous amounts of this kind of systematizing
I think this also sort of has some handwavy analogies to superposition vs composition.
if there is an existing name for the thing I’m pointing at, I would be interested in knowing.
it is often claimed that merely passively absorbing information is not sufficient for learning, but rather some amount of intentional learning is needed. I think this is true in general. however, one interesting benefit of passively absorbing information is that you notice some concepts/terms/areas come up more often than others. this is useful because there’s simply too much stuff out there to learn, and some knowledge is a lot more useful than other knowledge. noticing which kinds of things come up often is therefore useful for prioritization. I often notice that my motivational system really likes to use this heuristic for deciding how motivated to be while learning something.
I think it might also depend on your goals. Like how fast you want to learn something. If you have less than ideal time, then maybe more structured learning is necessary. If you have more time then periods of structureless/passive learning could be beneficial.
retargetability might be the distinguishing factor between controllers and optimizers
as in, controllers are generally retargetable and optimizers aren’t? or vice-versawould be interested in reasoning, either way
a claim I’ve been saying irl for a while but have never gotten around to writing up: current LLMs are benign not because of the language modelling objective, but because of the generalization properties of current NNs (or to be more precise, the lack thereof). with better generalization LLMs are dangerous too. we can also notice that RL policies are benign in the same ways, which should not be the case if the objective was the core reason. one thing that can go wrong with this assumption is thinking about LLMs that are both extremely good at generalizing (especially to superhuman capabilities) and simultaneously assuming they continue to have the same safety properties. afaict something like CPM avoids this failure mode of reasoning, but lots of arguments don’t
what is the “language models are benign because of the language modeling objective” take?
basically the Simulators kind of take afaict
House rules for definitional disputes:
If it ever becomes a point of dispute in an object level discussion what a word means, you should either use a commonly accepted definition, or taboo the term if the participants think those definitions are bad for the context of the current discussion.
(If the conversation participants are comfortable with it, the new term can occupy the same namespace as the old tabooed term (i.e going forward, we all agree that the definition of X is Y for the purposes of this conversation, and all other definitions no longer apply))
If any of the conversation participants want to switch to the separate discussion of “which definition of X is the best/most useful/etc”, this is fine if all the other participants are fine as well. However, this has to be explicitly announced as a change in topic from the original object level discussion.
A few axes along which to classify optimizers:
Competence: An optimizer is more competent if it achieves the objective more frequently on distribution
Capabilities Robustness: An optimizer is more capabilities robust if it can handle a broader range of OOD world states (and thus possible pertubations) competently.
Generality: An optimizer is more general if it can represent and achieve a broader range of different objectives
Real-world objectives: whether the optimizer is capable of having objectives about things in the real world.
Some observations: it feels like capabilities robustness is one of the big things that makes deception dangerous, because it means that the model can figure out plans that you never intended for it to learn (something not very capabilities robust would just never learn how to deceive if you don’t show it). This feels like the critical controller/search-process difference: controller generalization across states is dependent on the generalization abilities of the model architecture, whereas search processes let you think about the particular state you find yourself in. The actions that lead to deception are extremely OOD, and a controller would have a hard time executing the strategy reliably without first having seen it, unless NN generalization is wildly better than I’m anticipating.
Real world objectives is definitely another big chunk of deception danger; caring about the real world leads to nonmyopic behavior (though maybe we’re worried about other causes of nonmyopia too? not sure tbh), I’m actually not sure how I feel about generality: on the one hand, it feels intuitive that systems that are only able to represent one objective have got to be in some sense less able to become more powerful just by thinking more; on the other hand I don’t know what a rigorous argument for this would look like. I think the intuition relates to the idea of general reasoning machinery being the same across lots of tasks, and this machinery being necessary to do better by thinking harder, and so any model without this machinery must be weaker in some sense. I think this feeds into capabilities robustness (or lack thereof) too.
Examples of where things fall on these axes:
A rock would be none of the properties.
A pure controller (i.e a thermostat, “pile of heuristics”) can be competent, but not as capabilities robust, not general at all, and have objectives over the real world.
An analytic equation solver would be perfectly competent and capablilities robust (if it always works), not very general (it can only solve equations), and not be capable of having real world objectives.
A search based process can be competent, would be more capabilities robust and general, and may have objectives over the real world.
A deceptive optimizer is competent, capabilities robust, and definitely has real world objectives
Another generator-discriminator gap: telling whether an outcome is good (outcome->R) is much easier than coming up with plans to achieve good outcomes. Telling whether a plan is good (plan->R) is much harder, because you need a world model (plan->outcome) as well, but for very difficult tasks it still seems easier than just coming up with good plans off the bat. However, it feels like the world model is the hardest part here, not just because of embeddedness problems, but in general because knowing the consequences of your actions is really really hard. So it seems like for most consequentialist optimizers, the quality of the world model actually becomes the main thing that matters.
This also suggests another dimension along which to classify our optimizers: the degree to which they care about consequences in the future (I want to say myopia but that term is already way too overloaded). This is relevant because the further in the future you care about, the more robust your world model has to be, as errors accumulate the more steps you roll the model out (or the more abstraction you do along the time axis). Very low confidence but maybe this suggests that mesaoptimizers probably won’t care about things very far in the future because building a robust world model is hard and so perform worse on the training distribution, so SGD pushes for more myopic mesaobjectives? Though note, this kind of myopia is not quite the kind we need for models to avoid caring about the real world/coordinating with itself.
A thought pattern that I’ve noticed myself and others falling into sometimes: Sometimes I will make arguments about things from first principles that look something like “I don’t see any way X can be true, it clearly follows from [premises] that X is definitely false”, even though there are people who believe X is true. When this happens, it’s almost always unproductive to continue to argue on first principles, but rather I should do one of: a) try to better understand the argument and find a more specific crux to disagree on or b) decide that this topic isn’t worth investing more time in, register it as “not sure if X is true” in my mind, and move on.
For many such questions, “is X true” is the wrong question. This is common when X isn’t a testable proposition, it’s a model or assertion of causal weight. If you can’t think of existence proofs that would confirm it, try to reframe as “under what conditions is X a useful model?”.
there are policies which are successful because they describe a particular strategy to follow (non-mesaoptimizers), and policies that contain some strategy for discovering more strategies (mesaoptimizers). a way to view the relation this has to speed/complexity priors that doesn’t depend on search in particular is that policies that work by discovering strategies tend to be simpler and more generic (they bake in very little domain knowledge/metis, and are applicable to a broader set of situations because they work by coming up with a strategy for the task at hand on the fly). in contrast, policies that work by knowing a specific strategy tend to be more complex because they have to bake in a ton of domain knowledge, are less generally useful because they specifically know what to do in that situation, and thereby are also less retargetable)
another observation is that a meta-strategy with the ability to figure out what strategy is good is kind of defined by the fact that it doesn’t bake in specifics of dealing with a particular situation, but rather can adapt to a broad set of situations. there are also different degrees of meta-strategy-ness; some meta strategies will more quickly adapt to a broader set of situations. (there’s probably some sort of NFLT kind of argument you can make but NFLTs in general don’t really matter)
the ability to figure out strategies doesn’t necessarily have to be all reasoning, it can also encompass the experimental skillset
random brainstorming about optimizeryness vs controller/lookuptableyness:
let’s think of optimizers as things that reliably steer a broad set of initial states to some specific terminal state
seems like there are two things we care about (at least):
retargetability: it should be possible to change the policy to achieve different terminal states (but this is an insufficiently strong condition, because LUTs also trivially meet this condition, because we can always just completely rewrite the LUT. maybe the actual condition we want is that the complexity of the map is less than the complexity of just the diff or something?)
(in other words, in some sense it should be “easy” to rewrite a small subset or otherwise make a simple diff to the policy to change what final goal is achieved)
(maybe related idea: instrumental convergence means most goals reuse lots of strategies/circuitry between each other)
robustness: it should reliably achieve its goal across a wide range of initial states.
a LUT trained with a little bit of RL will be neither retargetable nor robust. a LUT trained with galactic amounts of RL to do every possible initial state optimally is robust but not retargetable (this is reasonable: robustness is only a property of the functional behavior so whether it’s a LUT internally shouldn’t matter; retargetability is a property of the actual implementation so it does matter). a big search loop (the most extreme of which is AIXI, which is 100% search) is very retargetable, and depending on how hard it searches is varying degrees of robustness.
(however, in practice with normal amounts of compute a LUT is never robust, this thought experiment only highlights differences that remain in the limit)
what do we care about these properties for?
efficacy of filtering bad behaviors in pretraining: sufficiently good robustness means doing things that achieve the goal even in states that it never saw during training, and then even in states that require strategies that it never saw during training. if we filter out deceptive alignment from the data, then the model has to do some generalizing to figure out that this is a strategy that can be used to better accomplish its goal (as a sanity check that robustness is the thing here: a LUT never trained on deceptive alignment will never do it, but one that is trained on it will do it, a sufficiently powerful optimizer will always do it)
arguments about updates wrt “goal”: the deceptive alignment argument hinges a lot on “gradient of the goal” making sense. for example when we argue that the gradient on the model can be decomposed into one component that updates the goal to be more correct and another component that updates the capabilities to be more deceptive, we make this assumption. even if we assume away path dependence, the complexity argument depends a lot on the complexity being roughly equal to complexity of goal + complexity of general goal seeking circuitry, independent of goal.
arguments about difficulty of disentangling correct and incorrect behaviors: there’s a dual of retargetability which is something like the extent to which you can make narrow interventions to the behaviour. (some kind of “anti naturalness” argument)
[conjecture 1: retargetability == complexity can be decomposed == gradient of goal is meaningful. conjecture 2: gradient of goal is meaningful/complexity decomposition implies deceptive alignment (maybe we can also find some necessary condition?)]
how do we formalize retargetability?
maybe something like there exists a homeomorphism from the goal space to NNs with that goal
problem: doesn’t really feel very satisfying and doesn’t work at all for discrete things
maybe complexity: retargetable if it has a really simple map from goals to NNs with goals, conditional on another NN with that goal
problem: the training process of just training another NN from scratch on the new goal and ignoring the given NN could potentially be quite simple
maybe complexity+time: seems reasonable to assume retraining is expensive (and maybe for decomposability we also consider complexity+time)
random idea: the hypothesis that complexity can be approximately decomposed into a goal component and a reasoning component is maybe a good formalization of (a weak version of) orthogonality?
a tentative model of ambitious research projects
when you do a big research project, you have some amount of risk you can work with—maybe you’re trying to do something incremental, so you can only tolerate a 10% chance of failure, or maybe you’re trying to shoot for the moon and so you can accept a 90% chance of failure.
budgeting for risk is non negotiable because there are a lot of places where risk can creep in—and if there isn’t, then you’re not really doing research. most obviously, your direction might just be a dead end. but there are also other things that might go wrong: the code might end up too difficult to implement, or it might run too slowly, or you might fail to fix a solvable-in-principle problem that comes up.
I claim that one of the principal components of being a good researcher is being able to eliminate as much unnecessary risk as possible, so you can spend your entire risk budget on the important bets.
for example, if you’re an extremely competent engineer, when brainstorming experiments you don’t have to think much about the risk that you fail to implement it. you know that even if you don’t think through all the contingencies that might pop up, you can figue it out, because you have a track record of figuring it out. you can say the words “and if that happens we’ll just scale it up” without spending much risk because you know full well that you can actually execute on it. a less competent engineer would have to pay a much greater risk cost, and correspondingly have to reduce the ambitiousness of the research bets (or else, take on way more risk than intented).
not all research bets are created equal, either. the space of possible research bets is vast, and most of them are wrong. but if you have very good research taste, you can much more reliably tell whether a bet is likely to work out. even the best researchers can’t just look at a direction and know for sure if it will work, if you know that you get a good direction 10% of the time you can do a lot more than if your direction is only good 0.1% of the time.
finally, if you know and trust someone to be reliable at executing on their area of expertise, you can delegate things that fall in their domain to them. in practice, this can be quite tough and introduce risk unless they have a very legible track record, or you are sufficiently competent in their domain yourself to tell if they’re likely to succeed. and if you’re sufficiently competent to do the job of any of your report (even if less efficiently), then you can budget less risk here knowing that even if someone drops their ball you could always pick it up yourself.
https://arxiv.org/abs/2304.08612 : interesting paper with improvement on straight through estimator
https://arxiv.org/abs/2302.07011 : sharpness doesn’t seem to correlate with generalization
the phenomenon of strange bedfellows is probably caused in no small part by outgroup vs fargroup dynamics
‘And what ingenious maneuvers they all propose to
me! It seems to them that when they have thought of two
or three contingencies’ (he remembered the general plan
sent him from Petersburg) ‘they have foreseen everything.
But the contingencies are endless.’
‘And what ingenious maneuvers they all propose to
me! It seems to them that when they have thought of two
or three contingencies’ (he remembered the general plan
sent him from Petersburg) ‘they have foreseen everything.
But the contingencies are endless.’
We spend a lot of time on trying to figure out empirical evidence to distinguish hypotheses we have that make very similar predictions, but I think a potentially underrated first step is to make sure they actually fit the data we already have.
Is the correlation between sleeping too long and bad health actually because sleeping too long is actually causally upstream of bad health effects, or only causally downstream of some common cause like illness?
Afaik, both. Like a lot of shit things—they are caused by depression, and they cause depression, horrible reinforcing loop. While the effect of bad health on sleep is obvious, you can also see this work in reverse; e.g. temporary severe sleep restriction has an anti-depressive effect. Notable, though with not many useful clinical applications, as constant sleep deprivation is also really unhealthy.
GPT-2-xl unembedding matrix looks pretty close to full rank (plot is singular values)
Unsupervised learning can learn things humans can’t supervise because there’s structure in the world that you need deeper understanding to predict accurately. For example, to predict how characters in a story will behave, you have to have some kind of understanding in some sense of how those characters think, even if their thoughts are never explicitly visible.
Unfortunately, this understanding only has to be structured in a way that makes reading off the actual unsupervised targets (i.e next observation) easy.
An incentive structure for scalable trusted prediction market resolutions
We might want to make a trustable committee for resolving prediction markets. We might be worried that individual resolvers might build up reputation only to exit-scam, due to finite time horizons and non transferability of reputational capital. However, shareholders of a public company are more incentivized to preserve the value of the reputational capital. Based on this idea, we can set something up as follows:
Market creators pay a fee for the services of a resolution company
There is a pool of resolvers who give a first-pass resolution. Each resolver locks up a deposit.
If an appeal is requested, a resolution passes up through a series of committees of more and more senior resolvers
At the top, a vote is triggered among all shareholders
It’s amazing how many proposals for dealing with institutional distrust sound a lot like “make a new institution, with the same structure, but with better actors.” You lose me at “trustable committee”, especially when you don’t describe how THOSE humans are motivated by truth and beauty, rather than filthy lucre. Adding more layers of committees doesn’t help, unless you define a “final, un-appealable decision” that’s sooner than the full shareholder vote.
the core of the proposal really boils down to “public companies have less incentive to cash in on reputation and exit scam than individuals”. this proposal is explicitly not “the same structure but with better actors”.
Levels of difficulty:
Mathematically proven to be impossible (i.e perfect compression)
Impossible under currently known laws of physics (i.e perpetual motion machines)
A lot of people have thought very hard about it and cannot prove that it’s impossible, but strongly suspect it is impossible (i.e solving NP problems in P)
A lot of people have thought very hard about it, and have not succeeded, but we have no strong reason to expect it to be impossible (i.e AGI)
There is a strong incentive for success, and the markets are very efficient, so that for participants with no edge, success is basically impossible (i.e beating the stock market)
There is a strong incentive for a thing, but a less efficient market, and it seems nobody has done it successfully (i.e a new startup idea that seems nobody seems to be doing)
Hopefully this is a useful reference for conversations that go like this:
A: Why can’t we just do X to solve Y?
B: You don’t realize how hard Y is, you can’t just think up a solution in 5 minutes
A: You’re just not thinking outside the box, [insert anecdote about some historical figure who figured out how to do a thing which was once considered impossible in some sense]
B: No you don’t understand, it’s like actually not possible, not just like really hard, because of Z
A: That’s what they said about [historical figure]!
(random shower thoughts written with basically no editing)
Sometimes arguments have a beat that looks like “there is extreme position X, and opposing extreme position Y. what about a moderate ‘Combination’ position?” (I’ve noticed this in both my own and others’ arguments)
I think there are sometimes some problems with this.
Usually almost nobody is on the most extreme ends of the spectrum. Nearly everyone falls into the “Combination” bucket technically, so in practice you have to draw the boundary between “combination enough” vs “not combination enough to count as combination”, which is sometimes fraught. (There is a dual argument beat that looks like “people too often bucket things into distinct buckets, what about thinking of things as a spectrum.” I think this does the opposite mistake, because sometimes there really are relatively meaningful clusters to point to. (this seems quite reminiscent of one Scottpost that I can’t remember the name of rn))
In many cases, there is no easy 1d spectrum. Being a “combination” could refer to a whole set of mutually exclusive sets of views. This problem gets especially bad when the endpoints differ along many axes at once. (Another dual argument here that looks like “things are more nuanced than they seem” which has its own opposite problems)
Of the times where this is meaningful, I would guess it almost always happens when the axis one has identified is interesting and captures some interesting property of the world. That is to say, if you’ve identified some kind of quantity that seems to be very explanatory, just noting that fact actually produces lots of value, and then arguing about how or whether to bucket that quantity up into groups has sharply diminishing value.
In other words, introducing the frame that some particular latent in the world exists and is predictive is hugely valuable; when you say “and therefore my position is in between other people’s”, this is valuable due to the introduction of the frame. The actual heavy lifting happened in the frame, and the part where you point to some underexplored region of the space implied by that frame is actually not doing much work.
I hypothesize one common thing is that if you don’t draw this distinction, then it feels like the heavy lifting comes in the part where you do the pointing, and then you might want to do this within already commonly accepted frames. From the inside I think this feels like existing clusters of people being surprisingly closed minded, whereas the true reason is that the usefulness of the existing frame has been exhausted.
related take: “things are more nuanced than they seem” is valuable only as the summary of a detailed exploration of the nuance that engages heavily with object level cruxes; the heavy lifting is done by the exploration, not the summary
TL;DR: This is basically empty individualism except identity is disentangled from cooperation (accomplished via FDT), and each agent can have its own subjective views on what would count as continuity of identity and have preferences over that. I claim that:
Continuity is a property of the subjective experience of each observer-moment (OM), not necessarily of any underlying causal or temporal relation. (i.e I believe at this moment that I am experiencing continuity, but this belief is a fact of my current OM only. Being a Boltzmann brain that believes I experienced all the moments leading up to that moment feels exactly the same as “actually” experiencing things.)
Each OM may have beliefs about the existence of past OMs, and about causal/temporal relations between those past OMs and the current OM (i.e one may believe that a memory of the past did in fact result from the faithful recording of a past OM to memory, as opposed to being spawned out of thin air as a Boltzmann brain loaded with false memories.)
Something like preference utilitarianism is true and it is ok to have preferences about things you cannot observe, or prefer the world to be in one of two states that you cannot in any way distinguish. As a motivating example, one can have preferences between taking atomic actions (a) enter the experience machine and erase all memories of choosing to be in an experience machine and (b) doing nothing.
Each OM may have preferences for its subjective experience of continuity to correspond to some particular causal structure between OMs, despite this being impossible for that OM to observe or verify. This is where the subjectivity is introduced: each OM can have its own opinion on which other OMs it considers to also be “itself”), and it can have preferences over its self-OMs causally leading to itself in a particular way. This does not have to be symmetric; for instance, your past self may consider your future self to be more self like than your future self considers past self.
Continuity of self as viewed by each OM is decoupled from decision theoretic cooperation. i.e they coincide in a typical individual, who considers their past/future selves to be also themself, and cooperates decision theoretically (i.e you consider past/future you getting utility to both count as “you” getting utility). However it is also possible to cooperate to the same extent with OMs with whom you do not consider yourself to be the same self (i.e twin PD), or to not coordinate with yourself (i.e myopia/ADHD).
(related: FDT and myopia being much the same thing; you can think of caring about future selves’ rewards because you consider yourself to implement a similar enough algorithm to your future self as acausal trade. This has the nice property of unifying myopia and preventing acausal trade, in that acausal trade is really just caring about OMs that would not be considered the same “self”. This is super convenient because basically every time we talk about myopia for preventing deceptive mesaoptimization we have to hedge by saying “and also we need to prevent acausal trade somehow”, and this lets us unify the two things.)
Properties of this theory:
This theory allows one to have preferences such as “I want to have lots of subjective experiences into the future” or “I prefer to have physical continuity with my past self” despite rejecting any universal concept of identity which seems pretty useful
This theory is fully compatible with all sorts of thought experiments by simply not providing an answer as to which OM your current OM leads to “next”. This is philosophically unsatisfying but I think the theory is still useful nonetheless
Coordination is solved through decision theory, which completely disentangles it from identity.
Imagine if aliens showed up at your doorstep and tried to explain to you that making as many paperclips as possible was the ultimate source of value in the universe. They show pictures of things that count as paperclips and things that don’t count as paperclips. They show you the long rambling definition of what counts as a paperclip from Section 23(b)(iii) of the Declaration of Paperclippian Values. They show you pages and pages of philosophers waxing poetical about how paperclips are great because of their incredible aesthetic value. You would be like, “yeah I get it, you consider this thing to be a paperclip, and you care a lot about them.” You could probably pretty accurately tell whether the aliens would approve of anything you’d want to do. And then you wouldn’t really care, because you value human flourishing, not paperclips. I mean, it’s so silly to care about paperclips, right?
Of course, to the aliens, who have not so subtly indicated that they would blow up the planet and look for a new, more paperclip-loving planet if they were to detect any anti-paperclip sentiments, you say that you of course totally understand and would do anything for paperclips, and that you definitely wouldn’t protest being sent to the paperclip mines.
I think I’d be confused. Do they care about more or better paperclips, or do they care about worship of paperclips by thinking beings? Why would they care whether I say I would do anything for paperclips, when I’m not actually making paperclips (or disassembling myself to become paperclips)?
I thought it would be obvious from context but the answers are “doesn’t really matter, any of those examples work” and “because they will send everyone to the paperclip mines after ensuring there are no rebellious sentiments”, respectively. I’ve edited it to be clearer.
random thoughts. no pretense that any of this is original or useful for anyone but me or even correct
It’s ok to want the world to be better and to take actions to make that happen but unproductive to be frustrated about it or to complain that a plan which should work in a better world doesn’t work in this world. To make the world the way you want it to be, you have to first understand how it is. This sounds obvious when stated abstractly but is surprisingly hard to adhere to in practice.
It would be really nice to have some evolved version of calibration training where I take some historical events and try to predict concrete questions about what happened, and give myself immediate feedback and keep track of my accuracy and calibration. Backtesting my world model so to speak. Might be a bit difficult to measure accuracy improvments due to non iid ness of the world, but worth trying the naive thing regardless. Would be interesting to try and autogen using GPT3.
Feedback loops are important. Unfortunately, from the inside it’s very easy to forget. In particular, setting up feedback loops is often high friction, because it’s hard to measure the thing we care about. Fixing this general problem is probably hard but in the meantime I can try to setup feedback loops for important things like productivity, world modelling, decision making, etc
self self improvement improvement: feeling guilty about not self improving enough and trying to fix your own ability to fix your own abilities
Lots of things have very counterintuitive or indirect values. If you don’t take this into account and you make decisions based on maximizing value you might end up macnamara-ing yourself hard.
The stages of learning something: (1) “this is super overwhelming! I don’t think I’ll ever understand it. there are so many things I need to keep track of. just trying to wrap my mind around it makes me feel slightly queasy” (2) “hmm this seems to actually make some sense, I’m starting to get the hang of this” (3) “this is so simple and obviously true, I’ve always known it to be true, I can’t believe anyone doesn’t understand this” (you start noticing that your explanations of the thing become indistinguishable from the things you originally felt overwhelmed by) (4) “this new thing [that builds on top of the thing you just learned] is super overwhelming! I don’t think I’ll ever understand it”
The feeling of regret really sucks. This is a bad thing, because it creates an incentive to never reflect on things or realize your mistakes. This shows up as a quite painful aversion to reflecting on mistakes, doing a postmortem, and improving. I would like to somehow trick my brain into reframing things somehow. Maybe thinking of it as a strict improvement over the status quo of having done things wrong? Or maybe reminding myself that the regret will be even worse if I don’t do anything because I’ll regret not reflecting in addition
Thought pattern that I’ve noticed: I seem to have two sets of epistemic states at any time: one more stable set that more accurately reflects my “actual” beliefs that changes fairly slowly, and one set of “hypothesis” beliefs that changes rapidly. Usually when I think some direction is interesting, I alternate my hypothesis beliefs between assuming key claims are true or false and trying to convince myself either way, and if I succeed then I integrate it into my actual beliefs. In practice this might look like alternating between trying to prove something is impossible and trying to exhibit an example, or taking strange premises seriously and trying to figure out its consequences. I think this is probably very confusing to people because usually when talking to people who are already familiar with alignment I’m talking about implications of my hypothesis beliefs, because that’s the frontier of what I’m thinking about, and from the outside it looks like I’m constantly changing my mind about things. Writing this up partially to have something to point people to and partially to push myself to communicate this more clearly.
I think this pattern is common among intellectuals, and I’m surprised it’s causing confusion. Are you labeling your exploratory beliefs and statements appropriately? An “epistemic status” note for posts here goes a long way, and in private conversation I often say out loud “I’m exploring here, don’t take it as what I fully believe” in conversations at work and with friends.
I think I do a poor job of labelling my statements (at least, in conversation. usually I do a bit better in post format). Something something illusion of transparency. To be honest, I didn’t even realize explicitly that I was doing this until fairly recent reflection on it.