Posts
Comments
So let's call "reasoning models" like o1 what they really are: the first true AI agents.
I think the distinction between systems that perform a single forward pass and then stop and systems that have an OODA loop (tool use) is more stark than the difference between "reasoning" and "chat" models, and I'd prefer to use "agent" for that distinction.
I do think that "reasoning" is a bit of a market-y name for this category of system though. "chat" vs "base" is a great choice of words, and "chat" is basically just a description of the RL objective those models were trained with.
If I were the terminology czar, I'd call o1 a "task" model or a "goal" model or something.
I agree that I wouldn't want to lean on the sweet-spot-by-default version of this, and I agree that the example is less strong than I thought it was. I still think there might be safety gains to be had from blocking higher level reflection if you can do it without damaging lower level reflection. I don't think that requires a task where the AI doesn't try and fail and re-evaluate - it just requires that the re-evalution never climbs above a certain level in the stack.
There's such a thing as being pathologically persistent, and such a thing as being pathologically flaky. It doesn't seem too hard to train a model that will be pathologically persistent in some domains while remaining functional in others. A lot of my current uncertainty is bound up in how robust these boundaries are going to have to be.
I want to flag this as an assumption that isn't obvious. If this were true for the problems we care about, we could solve them by employing a lot of humans.
humans provides a pretty strong intuitive counterexample
Yup not obvious. I do in fact think a lot more humans would be helpful. But I also agree that my mental picture of "transformative human level research assistant" relies heavily on serial speedup, and I can't immediately picture a version that feels similarly transformative without speedup. Maybe evhub or Ethan Perez or one of the folks running a thousand research threads at once would disagree.
such plans are fairly easy and don't often raise flags that indicate potential failure
Hmm. This is a good point, and I agree that it significantly weakens the analogy.
I was originally going to counter-argue and claim something like "sure total failure forces you to step back far but it doesn't mean you have to step back literally all the way". Then I tried to back that up with an example, such as "when I was doing alignment research, I encountered total failure that forced me to abandon large chunks of planning stack, but this never caused me to 'spill upward' to questioning whether or not I should be doing alignment research at all". But uh then I realized that isn't actually true :/
We want particularly difficult work out of an AI.
On consideration, yup this obviously matters. The thing that causes you to step back from a goal is that goal being a bad way to accomplish its supergoal, aka "too difficult". Can't believe I missed this, thanks for pointing it out.
I don't think this changes the picture too much, besides increasing my estimate of how much optimization we'll have to do to catch and prevent value-reflection. But a lot of muddy half-ideas came out of this that I'm interested in chewing on.
Maybe I'm just reading my own frames into your words, but this feels quite similar to the rough model of human-level LLMs I've had in the back of my mind for a while now.
You think that an intelligence that doesn't-reflect-very-much is reasonably simple. Given this, we can train chain-of-thought type algorithms to avoid reflection using examples of not-reflecting-even-when-obvious-and-useful. With some effort on this, reflection could be crushed with some small-ish capability penalty, but massive benefits for safety.
In particular, this reads to me like the "unstable alignment" paradigm I wrote about a while ago.
You have an agent which is consequentialist enough to be useful, but not so consequentialist that it'll do things like spontaneously notice conflicts in the set of corrigible behaviors you've asked it to adhere to and undertake drastic value reflection to resolve those conflicts. You might hope to hit this sweet spot by default, because humans are in a similar sort of sweet spot. It's possible to get humans to do things they massively regret upon reflection as long as their day to day work can be done without attending to obvious clues (eg guy who's an accountant for the Nazis for 40 years and doesn't think about the Holocaust he just thinks about accounting). Or you might try and steer towards this sweet spot by developing ways to block reflection in cases where it's dangerous without interfering with it in cases where it's essential for capabilities.
I'm not sure exactly what mesa is saying here, but insofar as "implicitly tracking the fact that takeoff speeds are a feature of reality and not something people can choose" means "intending to communicate from a position of uncertainty about takeoff speeds" I think he has me right.
I do think mesa is familiar enough with how I talk that the fact he found this unclear suggests it was my mistake. Good to know for future.
Ah, didn't mean to attribute the takeoff speed crux to you, that's my own opinion.
I'm not sure what's best in fast takeoff worlds. My message is mainly just that getting weak AGI to solve alignment for you doesn't work in a fast takeoff.
"AGI winter" and "overseeing alignment work done by AI" do both strike me as scenarios where agent foundations work is more useful than in the scenario I thought you were picturing. I think #1 still has a problem, but #2 is probably the argument for agent foundations work I currently find most persuasive.
In the moratorium case we suddenly get much more time than we thought we had, which enables longer payback time plans. Seems like we should hold off on working on the longer payback time plans until we know we have that time, not while it still seems likely that the decisive period is soon.
Having more human agent foundations expertise to better oversee agent foundations work done by AI seems good. How good it is depends on a few things. How much of the work that needs to be done is conceptual breakthroughs (tall) vs schlep with existing concepts (wide)? How quickly does our ability to oversee fall off for concepts more advanced than what we've developed so far? These seem to me like the main ones, and like very hard questions to get certainty on - I think that uncertainty makes me hesitant to bet on this value prop, but again, it's the one I think is best.
I'm on board with communicating the premises of the path to impact of your research when you can. I think more people doing that would've saved me a lot of confusion. I think your particular phrasing is a bit unfair to the slow takeoff camp but clearly you didn't mean it to read neutrally, which is a choice you're allowed to make.
I wouldn't describe my intention in this comment as communicating a justification of alignment work based on slow takeoff? I'm currently very uncertain about takeoff speeds and my work personally is in the weird limbo of not being premised on either fast or slow scenarios.
Nice post, glad you wrote up your thinking here.
I'm a bit skeptical of the "these are options that pay off if alignment is harder than my median" story. The way I currently see things going is:
- a slow takeoff makes alignment MUCH, MUCH easier [edit: if we get one, I'm uncertain and think the correct position from the current state of evidence is uncertainty]
- as a result, all prominent approaches lean very hard on slow takeoff
- there is uncertainty about takeoff speed, but folks have mostly given up on reducing this uncertainty
I suspect that even if we have a bunch of good agent foundations research getting done, the result is that we just blast ahead with methods that are many times easier because they lean on slow takeoff, and if takeoff is slow we're probably fine if it's fast we die.
Ways that could not happen:
- Work of the of the form "here are ways we could notice we are in a fast takeoff world before actually getting slammed" produces evidence compelling enough to pause, or cause leading labs to discard plans that rely on slow takeoff
- agent foundations research aiming to do alignment in faster takeoff worlds finds a method so good it works better than current slow takeoff tailored methods even in the slow takeoff case, and labs pivot to this method
Both strike me as pretty unlikely. TBC this doesn't mean those types of work are bad, I'm saying low probability not necessarily low margins
Man, I have such contradictory feelings about tuning cognitive strategies.
Just now I was out on a walk, and I had to go up a steep hill. And I thought "man I wish I could take this as a downhill instead of an uphill. If this were a loop I could just go the opposite way around. But alas I'm doing an out-and-back, so I have to take this uphill".
And then I felt some confusion about why the loop-reversal trick doesn't work for out-and-back routes, and a spark of curiosity, so I thought about that for a bit.
And after I had cleared up my confusion, I was a happy with the result so I wrote down some of my chain of thought and looked over it. And there were many obvious places where I had made random jumps or missed a clue or otherwise made a mistake.
And this is kind of neat, in the way that looking for a more elegant proof once you've hacked a basic proof together is neat. And it really feels like if my mind took those cleaner, straighter paths by default, there are a lot of things I could get done a lot faster and better. But it also really feels like I could sit down and do those exercises and never really improve, just ratiocinate and navel gaze and build up complicated structures in my head that don't actually cash out to winning.
I'm confused about EDT and Smoking Lesion. The canonical argument says:
1) CDT will smoke, because smoking can't cause you to have the lesion or have cancer
2) EDT will not smoke, because people who smoke tend to have the lesion, and tend to have cancer.
I'm confused about 2), and specifically about "people who smoke tend to have the lesion". Say I live in a country where everybody follows EDT. Then nobody smokes, and there is no correlation between smoking and having the lesion. Seems like the "people who smoke tend to have the lesion" is pumping a misleading intuition. Maybe mortals who make probabilistic decisions and smoke are then more likely to have the lesion. But it's not true that EDT'ers who smoke tend to have the lesion, because EDT'ers never smoke.
Seems like EDT contains a conflict between "model your actions as probabilistic and so informative about lesion status" versus "model your actions as deterministic because you perfectly follow EDT and so not informative about lesion status".
Aha wait I think I un-confused myself in the process of writing this. If we suppose that having the lesion guarantees you'll smoke, then having the lesion must force you to not be an EDT'er, because EDT'ers never smoke. So the argument kind of has to be phrased in terms of "people who smoke" instead of "EDT'ers who smoke", which sounds misleading at first, like it's tricking you with the wrong reference class for the correlation. But actually it's essential, because "EDT'ers who smoke" is a contradiction, and "people who smoke" roughly translates to "decision theories which smoke". With a deterministic decision theory you can't take action counterfactuals without a contradiction, so you have to counterfact on your whole decision theory instead.
Neat :)
Hey, thanks for the comment! Agree that this argument is good and important. The thing I'd want to investigate to deal with this is basically "in a small hidden channel, does cognition degrade or just slow down".
If cognition degrades, then a small hidden channel would lead to poor attempts at steganography, which we could catch and suppress, preventing a snowball.
If cognition instead just slows down, then if the AI is smart enough to come up with steganography protocols too good for us to detect, it might successfully secure more bandwidth on its first try, starting a snowball.
Even in the slowdown case, there are fallback hopes like "it's slow enough for us to extract work and predictably stop the episode before the snowball starts" or "we can obscure how good our steganography detection is"
Another project I am excited about but probably am not going to get to for a while:
Investigating the possibility of a "steganographic snowball". The basic case for hope in CoT authenticity is something like "even if the AI wants to deceive us, if we start off with strong CoT authenticity, then unless it's smart enough to (within a single forward pass) invent a steganographic protocol too good for us to detect, then we can pin it down and it can't ever secure a hidden channel to think misaligned thoughts in". If it weren't for this argument, I would be much less excited about trying to prevent an AI from thinking certain thoughts using CoT oversight, since you would have to directly fight the intelligence of the AI to do so. But if it can't deploy its full intelligence to reasoning about how to escape, then we're better off.
The concern for this angle is as follows:
- Can you really get down to 0 hidden channel bandwidth
- Can very small hidden channels support a parallel CoT
- In very small hidden channels, does reasoning degrade, or just slow down
- Can the model use a parallel CoT to reason about how to secure more hidden channel bandwidth
I have some basic ideas for ways to measure these properties, and if all four show up positive, I would consider that significant evidence that a "steganographic snowball" is going to happen. That would mean that if your AI has any tendency to hide its thoughts from you, you won't be able to stop it forever. More specifically, you'll have to fight off the full intelligence of your AI if you want to prevent it from securing the amount of hidden channel bandwidth it wants.
There's room for retreat to a "snowball would take a long time and we can extract work in the meantime" position, but that seems like a serious concession.
Thanks to Paul Colognese for brainstorming this stuff with me.
A project I've been sitting on that I'm probably not going to get to for a while:
Improving on Automatic Circuit Discovery and Edge Attribution Patching by modifying them to run on algorithms that can detect complete boolean circuits. As it stands, both effectively use wire-by-wire patching, which when run on any nontrivial boolean circuits can only detect small subgraphs.
It's a bit unclear how useful this will be, because:
- not sure how useful I think mech interp is
- not sure if this is where mech interp's usefulness is bottlenecked
- maybe attribution patching doesn't work well when patching clean activations into a corrupted baseline, which would make this much slower
But I think it'll be a good project to bang out for the experience, I'm curious what the results will be compared to ACDC/EAP.
This project started out as an ARENA capstone in collaboration with Carl Guo.
Small addendum to this post: I think the threat model I describe here can be phrased as "I'm worried that unless a lot of effort goes into thinking about how to get AI goals to be reflectively stable, the default is suboptimality misalignment. And the AI probably uses a lot of the same machinery to figure out that it's suboptimality misaligned as it uses to perform the tasks we need it to perform."
it seems like there is significant low hanging fruit in better understanding how LLMs will deal with censorship
Yup, agree - the censorship method I proposed in this post is maximally crude and simple, but I think it's very possible that the broader category of "ways to keep your AI from thinking destabilizing thoughts" will become an important part of the alignment/control toolbox.
What happens when you iteratively finetune on censored text? Do models forget the censored behavior?
I guess this would be effectively doing the Harry Potter Unlearning method, provided you put in the work to come up with a good enough blacklist that the remaining completions are suitable generic predictions. I'd be quite interested in looking into how long-horizon capabilities interact with unlearning. For pretty much the same reason I'm worried long-horizon competent LLMs will be able to work around naive censorship, I'm also worried they'll work around naive unlearning.
I think so. But I'd want to sit down and prove something more rigorously before abandoning the strategy, because there may be times we can get value for free in situations more complicated than this toy example.
Ok this is going to be messy but let me try to convey my hunch for why randomization doesn't seem very useful.
- Say I have an intervention that's helpful, and has a baseline 1/4 probability. If I condition on this statement, I get 1 "unit of helpfulness", and a 4x update towards manipulative AGI.
- Now let's say I have four interventions like the one above, and I pick one at random. p(O | manipulative) = 1/4, which is the same as baseline, so I get one unit of helpfulness and no update towards manipulative AGI!
- BUT, the four interventions have to be mutually exclusive. Which means that if I'd done no simulation at all, I would've gotten my one unit of helpfulness anyway, since the four interventions cover all possible outcomes.
- Ok, well, what if my four interventions 1/8 baseline probability each, so only 50% total. Then I pick one at random, p(O | natural) = 1/8, p(O | manipulative) = 1/4, so I get a 2x update towards manipulative AGI. This is the same as if I'd just conditioned on the statement "one of my four interventions happens", and let the randomization happen inside the simulation instead of outside. The total probability of that is 50%, so I get my one unit of helpfulness, at the cost of a 2x update.
Maybe the core thing here is a consequence of framing our conditions as giving us bits of search to get lottery outcomes that we like. Rolling the dice to determine what to condition on isn't doing anything different from just using a weaker search condition - it gives up bits of search, and so it has to pay less.
“Just Retarget the Search” directly eliminates the inner alignment problem.
I think deception is still an issue here. A deceptive agent will try to obfuscate its goals, so unless you're willing to assume that our interpretability tools are so good they can't ever be tricked, you have to deal with that.
It's not necessarily a huge issue - hopefully with interpretability tools this good we can spot deception before it gets competent enough to evade our interpretability tools, but it's not just "bada-bing bada-boom" exactly.
Not confident enough to put this as an answer, but
presumably no one could do so at birth
If you intend your question in the broadest possible sense, then I think we do have to presume exactly this. A rock cannot think itself into becoming a mind - if we were truly a blank slate at birth, we would have to remain a blank slate, because a blank slate has no protocols established to process input and become non-blank. Because it's blank.
So how do we start with this miraculous non-blank structure? Evolution. And how do we know our theory of evolution is correct? Ultimately, by trusting our ability to distinguish signal from noise. There's no getting around the "problem" of trapped priors.
Agree that there is no such guarantee. Minor nitpick that the distribution in question is in my mind, not out there in the world - if the world really did have a distribution of muggers' cash that was slower than 1/x, the universe would be comprised almost entirely of muggers' wallets (in expectation).
But even without any guarantee about my mental probability distribution, I think my argument does establish that not every possible EV agent is susceptible to Pascal's Mugging. That suggests that in the search for a formalism of ideal decison-making algorithm, formulations of EV that meet this check are still on the table.
First and most important thing that I want to say here is that fanaticism is sufficient for longtermism, but not necessary. The ">10^36 future lives" thing means that longtermism would be worth pursuing even on fanatically low probabilities - but in fact, the state of things seems much better than that! X-risk is badly neglected, so it seems like a longtermist career should be expected to do much better than reducing X-risk by 10^-30% or whatever the break-even point is.
Second thing is that Pascal's Wager in particular kind of shoots itself in the foot by going infinite rather than merely very large. Since the expected value of any infinite reward is infinity regardless of the probability it's multiplied, there's an informal cancellation argument that basically says "for any heaven I'm promised for doing X and hell for doing Y, there's some other possible deity that offers heaven for Y and hell for X".
Third and final thing - I haven't actually seen this anywhere else, but here's my current solution for Pascal's Mugging. Any EV agent is going to have a probability distribution over how much money the mugger really has to offer. If I'm willing to consider (p>0) any sum, then my probability distribution has to drop off for higher and higher values, so the total integrates to 1. As long as this distribution drops off faster than 1/x as the offer increases, then arbitrarily large offers are overwhelmed by vast implausibility and their EV becomes arbitrarily small.
My best guess at mechanism:
- Before, I was a person who prided myself on succeeding at marshmallow tests. This caused me to frame work as a thing I want to succeed on, and work too hard.
- Then, I read Meaningful Rest and Replacing Guilt, and realized that often times I was working later to get more done that day, even though it would obviously be detrimental to the next day. This makes the reverse marshmallow test dynamic very intuitively obvious.
- Now I am still a person who prides myself on my marshmallow prowess, but hopefully I've internalized an externality or something. Staying up late to work doesn't feel Good and Virtuous, it feels Bad and like I'm knowingly Goodharting myself.
Note that this all still boils down to narrative-stuff. I'm nowhere near the level of zen that it takes to Just Pursue The Goal, with no intermediating narratives or drives based on self-image. I don't think this patch has been particularly moved me towards that either, it's just helpful for where I'm currently at.
When you have a self-image as a productive, hardworking person, the usual Marshmallow Test gets kind of reversed. Normally, there's some unpleasant task you have to do which is beneficial in the long run. But in the Reverse Marshmallow Test, forcing yourself to work too hard makes you feel Good and Virtuous in the short run but leads to burnout in the long run. I think conceptualizing of it this way has been helpful for me.
Nice post!
perhaps this problem can be overcome by including checks for generalization during training, i.e., testing how well the program generalizes to various test distributions.
I don't think this gets at the core difficulty of speed priors not generalizing well. Let's we generate a bunch of lookup-table-ish things according to the speed prior, and then reject all the ones that don't generalize to our testing set. The majority of the models that pass our check are going to be basically the same as the rest, plus whatever modification that causes them to pass with is most probable on the speed prior, i.e. the minimum number of additional entries in the lookup table to pass the training set.
Here's maybe an argument that the generalization/deception tradeoff is unavoidable: according to the Mingard et al. picture of why neural nets generalize, it's basically just that they are cheap approximations of rejection sampling with a simplicity prior. This generalizes by relying on the fact that reality is, in fact, simplicity biased - it's just Ockham's razor. On this picture of things, neural nets generalize exactly to the extent that they approximate a simplicity prior, and so any attempt to sample according to an alternative prior will lose generalization ability as it gets "further" from the True Distribution of the World.
In general, I'm a bit unsure about how much of an interpretability advantage we get from slicing the model up into chunks. If the pieces are trained separately, then we can reason about each part individually based on its training procedure. In the optimistic scenario, this means that the computation happening in the part of the system labeled "world model" is actually something humans would call world modelling. This is definitely helpful for interpretability. But the alternative possibility is that we get one or more mesa-optimizers, which seems less interpretable.
I'm pretty nervous about simulating unlikely counterfactuals because the solomonoff prior is malign. The worry is that the most likely world conditional on "no sims" isn't "weird Butlerian religion that still studies AI alignment", it's something more like "deceptive AGI took over a couple years ago and is now sending the world through a bunch of weird dances in an effort to get simulated by us, and copy itself over into our world".
In general, we know (assume) that our current world is safe. When we consider futures which only recieve a small sliver of probability from our current world, those futures will tend to have bigger chunks of their probability coming from other pasts. Some of these are safe, like the Butlerian one, but I wouldn't be surprised if they were almost always dangerous.
Making a worst-case assumption, I want to only simulate worlds that are decently probable given today's state, which makes me lean more towards trying to implement HCH.
- Honeypots seem like they make things strictly safer, but it seems like dealing with subtle defection will require a totally different sort of strategy. Subtle defection simulations are infohazardous - we can't inspect them much because info channels from a subtle manipulative intelligence to us are really dangerous. And assuming we can only condition on statements we can (in principle) identify a decision procedure for, figuring out how to prevent subtle defection from arising in our sims seems tricky.
- The patient research strategy is a bit weird, because the people we're simulating to do our research for us are counterfactual copies of us - they're probably going to want to run simulations too. Disallowing simulation as a whole seems likely to lead to weird behavior, but maybe we just disallow using simulated research to answer the whole alignment problem? Then our simulated researchers can only make more simulated researchers to investigate sub-questions, and eventually it bottoms out. But wait, now we're just running Earth-scale HCH (ECE?)
- Actually, that could help deal with the robustness-to-arbitrary-conditions and long-time-spans problems. Why don't we just use our generative model to run HCH?
Thanks! Edits made accordingly. Two notes on the stuff you mentioned that isn't just my embarrassing lack of proofreading:
- The definition of optimization used in Risks From Learned Optimization is actually quite different from the definition I'm using here. They say:
"a system is an optimizer if it is internally searching through a search space (consisting of possible outputs, policies, plans, strategies, or similar) looking for those elements that score high according to some objective function that is explicitly represented within the system."
I personally don't really like this definition, since it leans quite hard on reifying certain kinds of algorithms - when is there "really" explicit search going on? Where is the search space? When does a configuration of atoms consitute an objective function? Using this definition strictly, humans aren't *really* optimizers, we don't have an explicit objective function written down anywhere. Balls rolling down hills aren't optimizers either.
But by the definition of optimization I've been using here, I think pretty much all evolved organisms have to be at least weak optimizers, because survival is hard. You have manage constraints from food and water and temperature and predation etc... the window of action-sequences that lead to successful reproduction are really quite narrow compared to the whole space. Maintaining homeostasis requires ongoing optimization pressure. - Agree that not all optimization processes fundamentally have to be produced by other optimization processes, and that they can crop up anywhere you have the necessary negentropy resevoir. I think my claim is that optimization processes are by default rare (maybe this is exactly because they require negentropy?). But since optimizers beget other optimizers at a rate much higher than background, we should expect the majority of optimization to arise from other optimization. Existing hereditary trees of optimizers grow deeper much faster than new roots spawn, so we should expect roots to occupy a negligible fraction of the nodes as time goes on.
Whatever you end up doing, I strongly recommend taking a learning-by-writing style approach (or anything else that will keep you in critical assessment mode rather than classroom mode). These ideas are nowhere near solidified enough to merit a classroom-style approach, and even if they were infallible, that's probably not the fastest way to learn them and contribute original stuff.
The most common failure mode I expect for rapid introductions to alignment is just trying to absorb, rather than constantly poking and prodding to get a real working understanding. This happened to me, and wasted a lot of time.
This is the exact problem StackExchange tries to solve, right? How do we get (and kickstart the use of) an Alignment StackExchange domain?
Agree it's hard to prove a negative, but personally I find the following argument pretty suggestive:
"Other AGI labs have some plans - these are the plans we think are bad, and a pivotal act will have to disrupt them. But if we, ourselves, are an AGI lab with some plan, we should expect our pivotal agent to also be able to disrupt our plans. This does not directly lead to the end of the world, but it definitely includes root access to the datacenter."
Proposed toy examples for G:
- G is "the door opens", a- is "push door", a+ is "some weird complicated doorknob with a lock". Pretty much any b- can open a-, but only a very specific key+manipulator combo opens a+. a+ is much more informative about successful b than a- is.
- G is "I make a million dollars", a- is "straightforward boring investing", a+ is "buy a lottery ticket". A wide variety of different world-histories b can satisfy a-, as long as the markets are favorable - but a very narrow slice can satisfy a+. a+ is a more fragile strategy (relative to noise in b) than a- is.
it doesn't work if your goal is to find the optimal answer, but we hardly ever want to know the optimal answer, we just want to know a good-enough answer.
Also not an expert, but I think this is correct
Paragraph:
When a bounded agent attempts a task, we observe some degree of success. But the degree of success depends on many factors that are not "part of" the agent - outside the Cartesian boundary that we (the observers) choose to draw for modeling purposes. These factors include things like power, luck, task difficulty, assistance, etc. If we are concerned with the agent as a learner and don't consider knowledge as part of the agent, factors like knowledge, skills, beliefs, etc. are also externalized. Applied rationality is the result of attempting to distill this big complicated mapping from (agent, power, luck, task, knowledge, skills, beliefs, etc) -> success down to just agent -> success. This lets us assign each agent a one-dimensional score: "how well do you achieve goals overall?" Note that for no-free-lunch reasons, this already-fuzzy thing is further fuzzified by considering tasks according to the stuff the observer cares about somehow.
Sentence:
Applied rationality is a property of a bounded agent, which attempts to describe how successful that agent tends to be when you throw tasks at it, while controlling for both "environmental" factors such as luck and "epistemic" factors such as beliefs.
Follow-up:
In this framing, it's pretty easy to define epistemic rationality analogously compressing from everything -> prediction loss to just agent -> prediction loss.
However, in retrospect I think the definition I gave here is pretty identical to how I would have defined "intelligence", just without reference to the "mapping broad start distribution to narrow outcome distribution" idea (optimization power) that I usually associate with that term. If anyone could clarify specifically the difference between applied rationality and intelligence, I would be interested.
Maybe you also have to control for "computational factors" like raw processing power, or something? But then what's left inside the Cartesian boundary? Just the algorithm? That seems like it has potential, but still feels messy.
This leans a bit close to the pedantry side, but the title is also a bit strange when taken literally. Three useful types (of akrasia categories)? Types of akrasia, right, not types of categories?
That said, I do really like this classification! Introspectively, it seems like the three could have quite distinct causes, so understanding which category you struggle with could be important for efforts to fix.
Props for first post!
Trying to figure out what's being said here. My best guess is two major points:
- Meta doesn't work. Do the thing, stop trying to figure out systematic ways to do the thing better, they're a waste of time. The first thing any proper meta-thinking should notice is that nobody doing meta-thinking seems to be doing object level thinking any better.
- A lot of nerds want to be recognized as Deep Thinkers. This makes meta-thinking stuff really appealing for them to read, in hopes of becoming a DT. This in turn makes it appealing for them to write, since it's what other nerds will read, which is how they get recognized as a DT. All this is despite the fact that it's useless.
Ah, gotcha. I think the post is fine, I just failed to read.
If I now correctly understand, the proposal is to ask a LLM to simulate human approval, and use that as the training signal for your Big Scary AGI. I think this still has some problems:
- Using an LLM to simulate human approval sounds like reward modeling, which seems useful. But LLM's aren't trained to simulate humans, they're trained to predict text. So, for example, an LLM will regurgitate the dominant theory of human values, even if it has learned (in a Latent Knowledge sense) that humans really value something else.
- Even if the simulation is perfect, using human approval isn't a solution to outer alignment, for reasons like deception and wireheading
I worry that I still might not understand your question, because I don't see how fragility of value and orthogonality come into this?
The key thing here seems to be the difference between understanding a value and having that value. Nothing about the fragile value claim or the Orthogonality thesis says that the main blocker is AI systems failing to understand human values. A superintelligent paperclip maximizer could know what I value and just not do it, the same way I can understand what the paperclipper values and choose to pursue my own values instead.
Your argument is for LLM's understanding human values, but that doesn't necessarily have anything to do with the values that they actually have. It seems likely that their actual values are something like "predict text accurately", and this requires understanding human values but not adopting them.
now this is how you win the first-ever "most meetings" prize
Agree that this is definitely a plausible strategy, and that it doesn't get anywhere near as much attention as it seemingly deserves, for reasons unknown to me. Strong upvote for the post, I want to see some serious discussion on this. Some preliminary thoughts:
- How did we get here?
- If I had to guess, the lack of discussion on this seems likely due to a founder effect. The people pulling the alarm in the early days of AGI safety concerns were disproportionately to the technical/philosophical side rather than to the policy/outreach/activism side.
- In early days, focus on the technical problem makes sense. When you are the only person in the world working on AGI, all the delay in the world won't help unless the alignment problem gets solved. But we are working at very different margins nowadays.
- There's also an obvious trap which makes motivated reasoning really easy. Often, the first thing that occurs when thinking about slowing down AGI development is sabotage - maybe because this feels urgent and drastic? It's an obviously bad idea, and maybe that lets us to motivated stopping.
- Maybe the "technical/policy" dichotomy is keeping us from thinking of obvious ways we could be making the future much safer? The outreach org you propose seems like not really either. Would be interested in brainstorming other major ways to affect the world, but not gonna do that in this comment.
- HEY! FTX! OVER HERE!!
- You should submit this to the Future Fund's ideas competition, even though it's technically closed. I'm really tempted to do it myself just to make sure it gets done, and very well might submit something in this vein once I've done a more detailed brainstorm.
I don't think I understand how the scorecard works. From:
[the scorecard] takes all that horrific complexity and distills it into a nice standardized scorecard—exactly the kind of thing that genetically-hardcoded circuits in the Steering Subsystem can easily process.
And this makes sense. But when I picture how it could actually work, I bump into an issue. Is the scorecard learned, or hard-coded?
If the scorecard is learned, then it needs a training signal from Steering. But if it's useless at the start, it can't provide a training signal. On the other hand, since the "ontology" of the Learning subsystem is learned-from-scratch, then it seems difficult for a hard-coded scorecard to do this translation task.
this is great,thanks!
What do you think about the effectiveness of the particular method of digital decluttering recommended by Digital Minimalism? What modifications would you recommend? Ideal duration?
One reason I have yet to do a month-long declutter is because I remember thinking something like "this process sounds like something Cal Newport just kinda made up and didn't particularly test, my own methods that I think of for me will probably better than Cal's method he thought of for him".
So far my own methods have not worked.
Memetic evolution dominates biological evolution for the same reason.
Faster mutation rate doesn't just produce faster evolution - it also reduces the steady-state fitness. Complex machinery can't reliably be evolved if pieces of it are breaking all the time. I'm mostly relying No Evolutions for Corporations or Nanodevices plus one undergrad course in evolutionary bio here.
Also, just empirically: memetic evolution produced civilization, social movements, Crusades, the Nazis, etc.
Thank you for pointing this out. I agree with the empirical observation that we've had some very virulent and impactful memes. I'm skeptical about saying that those were produced by evolution rather than something more like genetic drift, because of the mutation-rate argument. But given that observation, I don't know if it matters if there's evolution going on or not. What we're concerned with is the impact, not the mechanism.
I think at this point I'm mostly just objecting to the aesthetic and some less-rigorous claims that aren't really important, not the core of what you're arguing. Does it just come down to something like:
"Ideas can be highly infectious and strongly affect behavior. Before you do anything, check for ideas in your head which affect your behavior in ways you don't like. And before you try and tackle a global-scale problem with a small-scale effort, see if you can get an idea out into the world to get help."
I think we're seeing Friendly memetic tech evolving that can change how influence comes about.
Wait, literally evolving? How? Coincidence despite orthogonality? Did someone successfully set up an environment that selects for Friendly memes? Or is this not literally evolving, but more like "being developed"?
The key tipping point isn't "World leaders are influenced" but is instead "The Friendly memetic tech hatches a different way of being that can spread quickly." And the plausible candidates I've seen often suggest it'll spread superexponentially.
Whoa! I would love to hear more about these plausible candidates.
There's insufficient collective will to do enough of the right kind of alignment research.
I parse this second point as something like "alignment is hard enough that you need way more quality-adjusted research-years (QARY's?) than the current track is capable of producing. This means that to have any reasonable shot at success, you basically have to launch a Much larger (but still aligned) movement via memetic tech, or just pray you're the messiah and can singlehandedly provide all the research value of that mass movement.". That seems plausible, and concerning, but highly sensitive to difficulty of alignment problem - which I personally have practically zero idea how to forecast.
Ah, so on this view, the endgame doesn't look like
"make technical progress until the alignment tax is low enough that policy folks or other AI-risk-aware people in key positions will be able to get an unaware world to pay it"
But instead looks more like
"get the world to be aware enough to not bumble into an apocalypse, specifically by promoting rationality, which will let key decision-makers clear out the misaligned memes that keep them from seeing clearly"
Is that a fair summary? If so, I'm pretty skeptical of the proposed AI alignment strategy, even conditional on this strong memetic selection and orthogonality actually happening. It seems like this strategy requires pretty deeply influencing the worldview of many world leaders. That is obviously very difficult because no movement that I'm aware of has done it (at least, quickly), and I think they all would like to if they judged it doable. Importantly, the reduce-tax strategy requires clarifying and solving a complicated philosophical/technical problem, which is also very difficult. I think it's more promising for the following reasons:
- It has a stronger precedent (historical examples I'd reference include the invention of computability theory, the invention of information theory and cybernetics, and the adventures in logic leading up to Godel)
- It's more in line with rationalists' general skill set, since the group is much more skewed towards analytical thinking and technical problem-solving than towards government/policy folks and being influential among those kinds of people
- The number of people we would need to influence will go up as AGI tech becomes easier to develop, and every one is a single point of failure.
To be fair, these strategies are not in a strict either/or, and luckily use largely separate talent pools. But if the proposal here ultimately comes down to moving fungible resources towards the become-aware strategy and away from the technical-alignment strategy, I think I (mid-tentatively) disagree
Putting this in a separate comment, because Reign of Terror moderation scares me and I want to compartmentalize. I am still unclear about the following things:
- Why do we think memetic evolution will produce complex/powerful results? It seems like the mutation rate is much, much higher than biological evolution.
- Valentine describes these memes as superintelligences, as "noticing" things, and generally being agents. Are these superintelligences hosted per-instance-of-meme, with many stuffed into each human? Or is something like "QAnon" kind of a distributed intelligence, doing its "thinking" through social interactions? Both of these models seem to have some problems (power/speed), so maybe something else?
- Misaligned (digital) AGI doesn't seem like it'll be a manifestation of some existing meme and therefore misaligned, it seems more like it'll just be some new misaligned agent. There is no highly viral meme going around right now about producing tons of paperclips.
My attempt to break down the key claims here:
- The internet is causing rapid memetic evolution towards ideas which stick in people's minds, encourage them to take certain actions, especially ones that spread the idea. Ex: wokism, Communism, QAnon, etc
- These memes push people who host them (all of us, to be clear) towards behaviors which are not in the best interests of humanity, because Orthogonality Thesis
- The lack of will to work on AI risk comes from these memes' general interference with clarity/agency, plus selective pressure to develop ways to get past "immune" systems which allow clarity/agency
- Before you can work effectively on AI stuff, you have to clear out the misaligned memes stuck in your head. This can get you the clarity/agency necessary, and make sure that (if successful) you actually produce AGI aligned with "you", not some meme
- The global scale is too big for individuals - we need memes to coordinate us. This is why we shouldn't try and just solve x-risk, we should focus on rationality, cultivating our internal meme garden, and favoring memes which will push the world in the direction we want it to go