I’m interested in doing in-depth dialogues to find cruxes. Message me if you are interested in doing this.
I do alignment research, mostly stuff that is vaguely agent foundations. Currently doing independent alignment research on ontology identification. In 2023 I was on Vivek’s team at MIRI, before that I did MATS 2, and before that I did a CS and Neuroscience undergrad (thesis on statistical learning theory).
I like the advice you’ve given. I recently wrote a message of final advice to the PIBBSxILIAD fellowship. It seems fairly closely related.
Final Advice from Jeremy
I have some final advice for your future careers as alignment researchers:
Keep your eye on the real problems and a full pathway to solving them. Don’t be distracted by the short term proxies of success. Don’t trust your employers, mentors, colleagues or funders to do this for you, they won’t. Always strive to understand the entire stack of motivations for your work and how it fits into a full solution to the problem. If you discover that some part of the stack isn’t as valid as you thought, switch research topic. In most fields your current set of skills is the most important factor in deciding what to work on. This is a field where catching up to the frontier of a subfield is relatively easy, so usually you should prioritize solving important problems over how well the problem fits with your current skills.
The real problem is building a superintelligence that you understand at a very deep level, such that you know it will act as you intended. You need to understand things like: How its goals are stored and why they will stay the same as its world model updates. How much it can rely on its world model. When and how you can specify goals that have easy-to-predict and safe consequences. How it might be motivated to improve itself, and why you can trust it to do this.
Your work will almost always be several steps upstream of this goal, and only solve a small subproblem. This is fine, as long as the connection is known and clearly communicated to others, so that the research community can prioritize necessary but neglected problems.
Have ambition. Hold yourself to a higher standard than the current field exemplifies. Your work kinda only matters if it makes significant advances. So take risks. Avoid low value experiments done just for publication. Think about the deep conceptual questions and allow them to motivate your research.
This attitude is somewhat associated with becoming a crackpot. To balance this out: Break down your research plans into small steps. Take care to communicate your ideas clearly and frequently. Work on small problems and help advance other people’s research, but treat this as training for the real work.
2 is close enough. Extrapolating the results of safe tests to unsafe settings requires a level of theoretical competence we don’t currently have. Steve Brynes just made a great post that is somewhat related, I endorse everything in that post.
You did a good job of communicating your positive feelings about this kind of value system, I understand slightly better why you like it.
I can see how it can be worth the trade-off to make a new goal if that’s the only way to get the work done. But it’s negative if the work can be done directly.
And we know many small-ish cases where we can directly compute a policy from a goal. So what makes it impossible to make larger plans without adding new goals? And why does adding new goals shift it from impossible to possible?
A surprisingly large fraction of people I talk to try to convince me that some kind of irrationality is good actually, and my overly strong assumptions about rationality are causes me to expect AI doom. I fairly sure this is false and I’m making relatively weak assumptions about what it means to be rational. One assumption I am happy to make is that an agent will try to avoid shooting itself in the foot (by its own lights). What sort of actions count as “shooting itself in the foot” depends on what the goals are about, and what the environment is like, and often while explaining this I reference this post.
Having a simple example of an extremely obvious example of coherence is great.
I meant to write a longer review but have run out of time. I’ll try to add it later.
We start off with some early values, and then develop instrumental strategies for achieving them. Those instrumental strategies become crystallized and then give rise to other instrumental strategies for achieving them, and so on.
This seems true of me for some cases, mostly during childhood. Maybe it was a hack that evolution used to get from near-sensory value specifications to more abstract values. But if I (maybe unfairly) take this as a full model of human values and entirely remove the terminal-instrumental distinction, then it seems to make a bunch of false predictions. E.g. there are lots of jobs that people don’t grow to love doing. E.g. There are lots of things that people love doing after only trying them once (where they tried it for no particular instrumental reason).
there’s a lot of room for positive-sum trade between goals
Once each goal exists there’s room for positive sum trade, but creating new goals is always purely negative for every other currently existing goal, right? My vague memory is that your response is that constructing new instrumental goals is somehow necessary for computational tractability, but I don’t get why that would be true.
I still think this post is pretty good and I stand by the arguments. I’m really glad Peter convinced me to work on it with him.
In Sections 1,2&3 we tried to set up consequentialism and the arguments for why this framework fits any agent that generalizes in certain ways.
There are relatively few posts that try to explain why inner alignment problems are likely, rather than just possible. I think one good way to view our argument in Section 4 is as a generalization of Carlsmith’s counting argument for scheming, except with somewhat less reliance on intentional scheming and more focus on the biology-like messiness inside of a trained agent.
When we wrote this we were hoping to create a reasonably comprehensive summary of the entire end-to-end argument for why we believe AI ruin is likely. I don’t think it was a total failure and I’m still fairly happy with it. It doesn’t seem to have led to these beliefs becoming much more widespread though. Most alignment research being done today still seems motivated by threat models that misunderstand the main difficulties, from my perspective.
The key thing I think is usually missing is: AGI should be thought of as a dynamic system that learns and grows. The pathway of growth depends on details of the cognitive algorithms, and these details are under-specified by training. Working through detailed examples of each thing that can be underspecified is a good way to intuitively grasp just how large of a problem this is.
Here are some ways I’d write this post differently now:
I’d factor out sections 1,2&3 and do it as a separate post about consequentialism and online learning. These are too long and get in the way of more interesting parts. They are important prerequisites, but it’s more important that people actually get to section 4 and understand it. I think we were trying too hard to convince people who don’t believe in trained AIs having goals, or don’t believe that having goals is necessary for intelligent behavior. Understanding how people reacted to this would have been helpful for writing the rest of it.
I also usually frame it differently now. I focus on the necessity of true-beliefs-that-correspond-to-reality for certain kinds of tasks, rather than the necessity of certain types of goals. Ultimately it’s the same, it’s just easier to overinterpret what I’m saying when framed in terms of goals.
We probably should have done section 5 (control schemes) as a separate post as well. I’ve really come to believe in factoring arguments whenever possible. This saves reader time, but also can make the dependencies of each belief and argument clearer, and this section was largely unrelated to the others. Control seems to have become a popular area of research, but it still looks like largely a waste of time to me, and I don’t know whether or where we went wrong in this argument.
I think section 6 was probably unnecessary, or at least should have been factored out.
Almost all of it should have been cut down a lot. I think now I could cut out at least half of the words while increasing clarity, but I didn’t know how at the time.
I agree with plex that the title is crazy. I can never remember it. Maybe now I’d call it something like “An end-to-end argument for AI doom”?
Ah I see, that makes sense, sorry about that. This post was written with more emphasis on the distribution shift that reveals misalignment rather than the underlying degree of freedom that allowed that misalignment to happen in the first place. Both of these (degree of freedom and distribution shift) are necessary to cause misalignment, any other form of misalignment would (probably) just be crushed by RLHF or similar.
premised
That’s closer to being a conclusion rather than a premise. This section of this post or this is the main argument for that. It’s just an underspecification argument, you could see it as a generalization of Carlsmith’s counting argument.
It’s interesting that your framing is “high confidence there’s no underlying corrigible motivation”, and mine is more like “unlikely it starts without flaws and the improvement process is under-specified in ways that won’t fix large classes of flaws”. I think the arguments linked support my view. Possibly I’ve not made some background reasoning or assumptions explicit.
I’d be happy to video call if you want to talk about this, I think that’d be a quicker way for us to work out where the miscommunication is.
When you say ‘that weakness’, you mean the inability to identify a subtask as alignment-related?
Mainly “bad actors can split their work...” with current LLMs, but yeah also identifying/guessing the overall intentions of humans giving subtasks.
But there often is a long period between when you stop endorsing the habit and when you’ve finally trained yourself to do something different. (If ever. Behavior-change is famously hard for humans.)
Yeah often, but I think that’s stretching the analogy too far. A scenario with a dangerous AI doing AI research has more self-improvement options than a human, and the stakes are higher (for the AI) than most human habit-breaking is to humans. This distribution shift is important when an AI has significantly greater-than-human self-improvement options. If it doesn’t, then the AI noticing non-endorsed habits may still happen, and that’d be good if it happened in a way that allowed humans to notice what’s wrong and try to fix it.
Also, I’ll note that religious deconversions very often happen in stages, including stages that involve narrow realizations that you were mistaken, and looming suspicions that you’re going to change your mind. The whole edifice doesn’t usually collapse in a single moment. It’s a process. (This interview covers a good example.)
Yeah, another good point, I agree. I haven’t watched that interview but I saw another video Rhett made. On the other hand, each stage can sometimies look like “slow buildup without acknowledgement of any update → crisis & fast update”. But I agree that religious deconversion is a decent analogy, and humans are playing the role of the pastor trying to catch and redirect the process, and that sometimes can work.
It’s unclear to me how strongly we can or should draw the analogy between changes in belief and changes in motivation, since one has a right answer and the other (presumably) doesn’t.
I don’t think of most the distribution-shift-induced-changes as changes in motivation, more like revealing/understanding the underlying motivation better. So with the habits, noticing that a habit is working against you can be as simple as updating a belief (about the consequences of that habit).
Yeah, but putting your attention on the right things often does take a lot of compute.
True but you don’t usually update or know that you’re going to update during this part.
Is the claim that there’s a dilemma between two options?
I think I don’t want to claim that there’s a strict dilemma, more that the paths between 1 and 2 are many and varied and hard to catch, even from the inside. Often because it’s fast, but sometimes just because it’s messy and there are lots of pressures and mechanisms involved.
your current model is aligned in which case you already know how to align AI so what do you even need it for?
I don’t think this is an objection anyone makes. I think it’s widely agreed that the safety of more capable agents is harder to guarantee, so aligned low-capability models would be useful research assistants for that work. My central objections are different, mostly this and this. Maybe also that most people underestimate how difficult non-fake alignment research is.
If you think this is not a useful strategy, then why not?
Problem factorisation is difficult for non-trivial problems.
That weakness of LLMs is a skill issue, and should (probably, mostly) go away at >human-researcher capability.
Good point. There can be a middle ground, but most of the examples that come to mind are more binary.
E.g. If you notice that you don’t endorse a habit, this generally happens immediately. There’s not a long period of being uncertain whether you endorse the habit, and still following the habit. If you’re uncertain, and the habit-situation is coming up, this forces you to think it through.
On the other hand, with this one:
If the overseer is only invoked when you think the overseer knows more than you.
Seems like the AI’s understanding of how much the overseer knows could change gradually, and it might feel compelled to alert the overseer of this change. Maybe depends a bit on the internal mechanisms for this corrigibility-property, but mostly this one looks like gradual change could allow proactive flagging.
One background assumption that might be relevant: without various biases that lead to people being attached to beliefs, if a belief update can be made just by thinking about it, then the belief change will happen fast once attention is allocated to it. It’s rare for logical belief updates to require lots of compute, once your attention is on the right things.
I am confident about this, so I’m okay with you judging accordingly.
I appreciate your rewrite. I’ll treat it as something to aspire to, in future. I agree that it’s easier to engage with.
I was annoyed when writing. Angry is too strong a word for it though, it’s much more like “Someone is wrong on the internet!”. It’s a valuable fuel and I don’t want to give it up. I recognise that there are a lot of situations that call for hiding mild annoyance, and I’ll try to do it more habitually in future when it’s easy to do so.
There’s a background assumption that maybe I’m wrong to have. If I write a comment with a tone of annoyance, and you disagree with it, it would surprise me if that made you feel bad about yourself. I don’t always assume this, but I often assume it on Lesswrong because I’m among nerds for whom disagreement is normal.
So overall, I think my current guess is that you’re trying to hold me to standards that are unnecessarily high. It seems supererogatory rather than obligatory.
I think we already see overconfidence in models. See davidad’s comment on how this could come from perverse RL credit assignment h/t (Jozdien). See also this martingale score paper. I think it’s reasonable to extrapolate from current models and say that future models will be overconfident by default
Cool, that makes sense. My disagreement with this come from thinking that the current LLM paradigm is kinda currently missing online learning. When I add that in, it seems much less reasonable an extrapolation, to me.
This seems probable with online learning but not necessarily always the case. It’s also possible that the model is not overconfident on easy to verify tasks but is overconfident on hard to verify tasks.
I assumed that you weren’t talking about this kind of domain-specific overconfidence, since your original comment suggested forecasting as a benchmark. This seems not totally implausible to me, but at the same time data-efficient generalisation is a ~necessary skill of most kinds of research so it still seems odd to predict a particular kind of inability to generalise while also conditioning on being good at research.
Like yes of course overconfidence is something that would get fixed eventually, but it’s not clear to be that it will be fixed until it’s too late
I’m primarily thinking about the AI correcting itself, like how you and I would in cases where it was worth the effort.
(i.e., you can still build ASI with a overconfident AI)
I think you’re saying this a tad too confidently. Overconfidence should slow down an AI in its research, cause it to invest too much in paths that won’t work out, over and over again. It’s possible it would still succeed, and it’s a matter of degree in how overconfident it is, but this could be an important blocker to being capable of effective research and development.
Yes, I agree with that.
I don’t think I am retreating to a motte.
My read was:
JG: Without ability to learn from mistakes
FS: Without optimal learning from mistakes
But this was misdirection, we are arguing about how surprised we should be when a competent agent doesn’t learn a very simple lesson after making the mistake several times. Optimality is misdirection, the thing you’re defending is extreme sub-optimality and the thing I’m arguing for is human-level ability-to-correct-mistakes.
On our current trajectory, I expect the minimal viable scary agent will fail to be epistemically efficient relative to humans in the following cases
I agree that there are plausibly domains where a minimal viable scary agent won’t be epistemically efficient with respect to humans. I think you’re overconfident (lol) in drawing specific conclusions (i.e. that a specific simple mistake is likely) from this kind of reasoning about capable AIs, and that’s my main disagreement.
But engaging directly, all three of these seem not very relevant to the case of general overconfidence, because general overconfidence is noticeable and correctable from lots of types of experiment. A more plausible thing to expect is low quality predictions about low data domains, not general overconfidence across low and high data domains.
I assume you’re talking about this one?
No, I meant this one:
I don’t think the first AI smart enough to cause catastrophe will need to be that smart.
I think focusing on the “first AI smart enough” leads to a lot of low-EV research. If you solve a problem with the first AI smart enough, this doesn’t help much because a) there are presumably other AIs of similar capability, or soon will be, with somewhat different capability profiles and b) it won’t be long before there are more capable AIs and c) it’s hard to predict future capability profiles.
The minimal viable scary agent is in fact scary.
It doesn’t need to be superhuman at everything to be scary
It is worth investing more than zero resources into mitigating the risks we expect to see with the first scary agents
This is true even if we don’t expect those mitigation to scale all the way up to superhuman-at-literally-all-tasks ASI.
I agree with all of these, so it feels a little like you’re engaging with an imagined version of me who is pretty silly.
Trying to rephrase my main point, because I think this disagreement must be at least partially a miscommunication:
Humans like you and I have the ability to learn from mistakes after making them several times. Across-the-board overconfidence is a mistake that we wouldn’t have much trouble correcting in ourselves, if it were important.
Domain-specific overconfidence on domains with little feedback is not what I’m talking about, because it didn’t appear to be what Tim was talking about. I’m also not talking about bad predictions in general.
one risk factor in this kind of research is that the capabilities people might resolve that weakness in the course of their work, in which case your effort was wasted. But I don’t think that that consideration is overwhelmingly strong.
My argument was that there were several of “risk factors” that stack. I agree that each one isn’t overwhelmingly strong.
I prefer not to be rude. Are you sure it’s not just that I’m confidently wrong? If I was disagreeing in the same tone with e.g. Yampolskiy’s argument for high confidence AI doom, would this still come across as rude to you?
“Overconfident” gets thrown around a lot by people who just mean “incorrect”. Rarely do they mean actual systematic overconfidence. If everyone involved in building AI shifted their confidence down across the board, I’d be surprised if this changed their safety-related decisions very much. The mistakes they are making are more complicated, e.g. some people seem “underconfident” about how to model future highly capable AGI, and are therefore adopting a wait-and-see strategy. This isn’t real systematic underconfidence, it’s just a mistake (from my perspective). And maybe some are “overconfident” that early AGI will be helpful for solving future problems, but again this is just a mistake, not systemic overconfidence.
At no point in this discussion do I reference “limits of intelligence”. I’m not taking any limits, or even making reference to any kind of perfect reasoning. My x-risk threat models in general don’t involve that kind of mental move. I’m talking about near-human-level intelligence, and the reasoning works for AI that operates similarly to how they work now.
A way to think about Condensation in the algorithmic setting
These are note that describe a (somewhat flawed) variation of algorithmic condensation. It covers similar ground as this post by Abram, with slightly more concrete detail. I’m publishing it because a) it’s simple and relatively concrete and still works on most test cases I’ve worked through, and b) it helps motivate the more general approach.
We want a generative model whose output is a set of observations. We want the hypothesis space to be the space of programs, and we want it to factorize naturally into components such that each component “contributes to” some subset of the observations.
Observations: A set of tuples, where each tuple contains some content and a location. For example, the content might be a pixel r,g,b and the location might be the coordinates of that pixel x,y. Or the content might be a bit b and location as the index of that bit inside a large bitstring.
To make the factorization work we need to impose some structure on the program. We’ll have one top program which can access (via built-in instructions) a
Listof other strings. The List must have an interface that satisfies some properties:It must be possible to retrieve elements based on their content, or their index in the list.
Item retrieval from the
Listis such that any items other than the one that actually ends up being retrieved can be removed from theListwithout changing the behavior of the program.For example, we could have a function
lookupthat scores each item on the list and returns the highest scoring one.The top program, when executed, must output the full set of observations (and no others).
Scoring subsets
A condensation has a score for every subset A of observation variables.
Conditioned[1] score: Delete as many strings as possible from the List such that the subset of observations A is still recovered correctly by running the program. Sum the length of these programs. This gives the conditioned score. “Recovered correctly” means that the program outputs a set that contains all the correct pixels, and contains no more than one pixel for each location. It’s allowed to output incorrect pixels that aren’t part of A (because forcing it to output an exact subset would add complexity).
Some more detail about
ListWe want to make sure the program can’t hide complexity in unexpected ways using the
Listdata structure. Kaarel pointed out one way of doing this in long lists, where you can sometimes save complexity by adding a bit to several items in the list. Items that contain the bit are located such that their index conveys useful information.To avoid this and similar hacks, we’ll assume that the
Listmemory layout is automatically and perfectly managed behind the scenes, but the additional data required for this (e.g. storing the lengths of each item in the list) counts as part of the complexity of the top program. This stops us from adding a bit to an item in the list as a means of encoding its index, because that forces the List to remember which item has a unique length.Why include a location label in every observation?
This is a simple way to force the program to keep track of which bit it’s generating at a given time, even when most parts of the program (or observations) aren’t present.
Example: Two red squares
Imagine our data is an image of two red squares on a black background. Remember each pixel also contains its location. Ideally, we want the model to contain a string representing the concept of “red square”, two strings for the individual squares, a string for the background color and a top string that holds information about how to tie it all together.
We can guess the conditional scores of important subsets:
A single pixel in the background only requires the top string to generate. This is a longer string than what it would take to describe the background color, so it’s not a great score.
A single pixel in one of the red squares requires three strings to generate (top, red square, specific red square). Not great score.
The set of pixels that cover a red square require the same three strings as a single pixel. The score should be near-optimal.
The set of all pixels requires all strings, and its score is near-optimal (this is an ~optimal compression of the data).
Example: 70% 1s random bitstring
We want this condensation to have the following structure: One top string that captures the fact that 70% of the observations, and a string associated with each observation that captures the unique information in that bit.
However, this won’t work. Because each string can hold a minimum of 1 bit, the total length of all strings is going to be much larger than the optimal compression of the observations.
So the condensation model could go in two directions: Either it’ll drop the 70% information, or it’ll partition the observations into small groups in exchange for a poor score on the small subsets.
This flaw motivates a better approach
I think Sam Eisenstat has a (maybe unfinished, unpublished) approach that helps with this issue: Have the top latent generate a probability distribution over the other latents, and then score the other latents according to their nlogprob in this distribution, rather than their length.
We can see this as taking the
Listmemory management to its logical conclusion. My description above requires that the top latent tracks the length of each latent string (which is equivalent to specifying a uniform distribution over bitstrings of fixed length). Sam’s approach just allows this distribution to be non-uniform.This is cleaner theoretically, but I find it somewhat harder to think about. It makes it hard to look at an example and work out a set of programs the would score well. It also requires a bunch of machinery (around handling the distributions) that isn’t usually needed (because most useful latent concepts are larger than a bit), so it’s good to be able to drop it.
Simple score:
If we’re scoring using the simple score, we don’t need to allow access to other strings, so there’s no
List. Simply have a set of programs, each of which outputs a set of observations. The simple score of A is the sum of lengths of programs such that the intersection of their output sets correctly pins down the observations in A.