I’m Tamsin Leake, co-founder and head of research at Orthogonal, doing agent foundations.
Tamsin Leake
Karma: 2,835
I think (not sure!) the damage from people/orgs/states going “wow, AI is powerful, I will try to build some” is larger than the upside of people/orgs/states going “wow, AI is powerful, I should be scared of it”. It only takes one strong enough one of the former to kill everyone, and the latter is gonna have a very hard time stopping all of them.
By not informing the public that AI is indeed powerful, awareness of that fact is disproportionately allocated to people who will choose to think hard about it on their own, and thus that knowledge is more likely to be in reasonabler hands (for example they’d also be more likely to think “hmm maybe I shouldn’t build unaligned powerful AI”).
The same goes for cyborg tools, as well as general insights about AI: we should want them to be differentially accessible to alignment people than the general public.
In fact, my biggest criticism of OpenAI is not that they built GPTs, but that they productized it, made it widely available, and created a giant public frenzy about LLMs. I think we’d have more time to solve alignment if they kept it internally and the public wasn’t thinking about AI nearly as much.
Even if tool AI is controllable, tool AI can be used to assist in building non-tool AI. A benign superassistant is one query away from outputting world-ending code.
In my opinion the hard part would not be figuring out where to donate to {decrease P(doom) a lot} rather than {decrease P(doom) a little}, but figuring out where to donate to {decrease P(doom)} rather than {increase P(doom)}.
(oops, this ended up being fairly long-winded! hope you don’t mind. feel free to ask for further clarifications.)
There’s a bunch of things wrong with your description, so I’ll first try to rewrite it in my own words, but still as close to the way you wrote it (so as to try to bridge the gap to your ontology) as possible. Note that I might post QACI 2 somewhat soon, which simplifies a bunch of QACI by locating the user as {whatever is interacting with the computer the AI is running on} rather than by using a beacon.
A first pass is to correct your description to the following:
-
We find a competent honourable human at a particular point in time , like Joe Carlsmith or Wei Dai, and give them a rock engraved with a 1GB secret key, large enough that in counterfactuals it could replace with an entire snapshot of . We also give them the ability to express a 1GB output, eg by writing a 1GB key somewhere which is somehow “signed” as the only . This is part of — is not just the human being queried at a particular point in time, it’s also the human producing an answer in some way. So is a function from 1GB bitstring to 1GB bitstring. We define as , followed by whichever new process describes in its output — typically another instance of except with a different 1GB payload.
-
We want a model of the agent . In QACI, we get by asking a Solomonoff-like ideal reasoner for their best guess about after feeding them a bunch of data about the world and the secret key.
-
We then ask the question , “What’s the best utility-function-over-policies to maximise?” to get a utility function . We then **ask our solomonoff-like ideal reasoner for their best guess about which action maximizes .
Indeed, as you ask in question 3, in this description there’s not really a reason to make step 3 an extra thing. The important thing to notice here is that model might get pretty good, but it’ll still have uncertainty.
When you say “we get by asking a Solomonoff-like ideal reasoner for their best guess about ”, you’re implying that — positing
U(M,A)to be the function that says how much utility the utility function returned by modelMattributes to actionA(in the current history-so-far) — we do something like:let M ← oracle(argmax { for model M } 𝔼 { over uncertainty } P(M)) let A ← oracle(argmax { for action A } U(M, A)) perform(A)Indeed, in this scenario, the second line is fairly redundant.
The reason we ask for a utility function is because we want to get a utility function within the counterfactual — we don’t want to collapse the uncertainty with an argmax before extracting a utility function, but after. That way, we can do expected-given-uncertainty utility maximization over the full distribution of model-hypotheses, rather than over our best guess about . We do:
let A ← oracle(argmax { for A } 𝔼 { for M, over uncertainty } P(M) · U(M, A)) perform(A)That is, we ask our ideal reasoner (
oracle) for the action with the best utility given uncertainty — not just logical uncertainty, but also uncertainty about which . This contrasts with what you describe, in which we first pick the most probable and then calculate the action with the best utility according only to that most-probable pick.
To answer the rest of your questions:
Is this basically IDA, where Step 1 is serial amplification, Step 2 is imitative distillation, and Step 3 is reward modelling?
Unclear! I’m not familiar enough with IDA, and I’ve bounced off explanations for it I’ve seen in the past. QACI doesn’t feel to me like it particularly involves the concepts of distillation or amplification, but I guess it does involve the concept of iteration, sure. But I don’t get the thing called IDA.
Why not replace Step 1 with Strong HCH or some other amplification scheme?
It’s unclear to me how one would design an amplification scheme — see concerns of the general shape expressed here. The thing I like about my step 1 is that the QACI loop (well, really, graph (well, really, arbitrary computation, but most of the time the user will probably just call themself in sequence)) is that its setup doesn’t involve any AI at all — you could go back in time before the industrial revolution and explain the core QACI idea and it would make sense assuming time-travelling-messages magic, and the magic wouldn’t have to do any extrapolating. Just tell someone the idea is that they could send a message to {their past self at a particular fixed point in time}. If there’s any amplification scheme, it’ll be one designed by the user, inside QACI, with arbitrarily long to figure it out.
What does “bajillion” actually mean in Step 1?
As described above, we don’t actually pre-determine the length of the sequence, or in fact the shape of the graph at all. Each iteration decides whether to spawn one or several next iteration, or indeed to spawn an arbitrarily different long-reflection process.
Why are we doing Step 3? Wouldn’t it be better to just use M directly as our superintelligence? It seems sufficient to achieve radical abundance, life extension, existential security, etc.
Why not ask M for the policy π directly? Or some instruction for constructing π? The instruction could be “Build the policy using our super-duper RL algo with the following reward function...” but it could be anything.
Hopefully my correction above answers these.
What if there’s no reward function that should be maximised? Presumably the reward function would need to be “small”, i.e. less than a Exabyte, which imposes a maybe-unsatisfiable constraint.
(Again, untractable-to-naively-compute utility function*, not easily-trained-on reward function. If you have an ideal reasoner, why bother with reward functions when you can just straightforwardly do untractable-to-naively-compute utility functions?)
I guess this is kinda philosophical? I have some short thoughts on here. If an exabyte is enough to describe to describe {a communication channel with a human-on-earth} to an AI-on-earth, which I think seems likely, then it’s enough to build “just have a nice corrigible assistant ask the humans what they want”-type channels.
Put another way: if there are actions which are preferable to other actions, then it seems to me like utility function are a fully lossless way for counterfactual QACI users to express which kinds of actions they want the AI to perform, which is all we need. If there’s something wrong with utility function over worlds, then counterfactual QACI users can output a utility function which favors actions which lead to something other than utility maximization over worlds, for example actions which lead to the construction of a superintelligent corrigible assistant which will help the humans come up with a better scheme.
Why is there no iteration, like in IDA? For example, after Step 2, we could loop back to Step 1 but reassign as with oracle access to .
Again, I don’t get IDA. Iteration doesn’t seem particularly needed? Note that inside QACI, the user does have access to an oracle and to all relevant pieces of hypothesis about which hypothesis it is inhabiting in — this is what, in the QACI math, this line does:
’s distribution over answers demands that the answer payload , when interpreted as math and with all required contextual variables passed as input ().
Notably, is the hypothesis for which world the user is being considered in, and for their location within that world. Those are sufficient to fully characterize the hypothesis-for- that describes them. And because the user doesn’t really return just a string but a math function which takes as input and returns a string, they can have that math function do arbitrary work — including rederive . In fact, rediriving is how they call a next iteration: they say (except in math) “call again (rederived using ), but with this string, and return the result of that.” See also this illustration, which is kinda wrong in places but gets the recursion call graph thing right.
Another reason to do “iteration” like this inside the counterfactual rather than in the actual factual world (if that’s what IDA does, which I’m only guessing here) is that we don’t have as many iteration steps as we want in the factual world — eventually OpenAI or someone else kills everyone, whereas in the counterfactual, the QACI users are the only ones who can make progress, so the QACI users essentially have as long as they want, so long as they don’t take too long in each individual counterfactual step or other somewhat easily avoided actions like that.
Why isn’t Step 3 recursive reward modelling? i.e. we could collect a bunch of trajectories from and ask to use those trajectories to improve the reward function.
Unclear if this still means anything given the rest of this post. Ask me again if it does.
-
Hi !
ATA is extremely neglected. The field of ATA is at a very early stage, and currently there does not exist any research project dedicated to ATA. The present post argues that this lack of progress is dangerous and that this neglect is a serious mistake.
I agree it’s neglected, but there is in fact at least one researh project dedicated to at least designing alignment targets: the part of the formal alignment agenda dedicated to formal outer alignment, which is the design of math problems to which solutions would be world-saving. Our notable attempts at this are QACI and ESP (there was also some work on a QACI2, but it predates (and in-my-opinion is superceded by) ESP).
Those try to implement CEV in math. They only work for doing CEV of a single person or small group, but that’s fine: just do CEV of {a single person or small group} which values all of humanity/moral-patients/whatever getting their values satisfied instead of just that group’s values. If you want humanity’s values to be satisfied, then “satisfying humanity’s values” is not opposite to “satisfy your own values”, it’s merely the outcome of “satisfy your own values”.
-
I wonder how much of those seemingly idealistic people retained power when it was available because they were indeed only pretending to be idealistic. Assuming one is actually initially idealistic but then gets corrupted by having power in some way, one thing someone can do in CEV that you can’t do in real life is reuse the CEV process to come up with even better CEV processes which will be even more likely to retain/recover their just-before-launching-CEV values. Yes, many people would mess this up or fail in some other way in CEV; but we only need one person or group who we’d be somewhat confident would do alright in CEV. Plausibly there are at least a few eg MIRIers who would satisfy this. Importantly, to me, this reduces outer alignment to “find someone smart and reasonable and likely to have good goal-content integrity”, which is a matter of social & psychology that seems to be much smaller than the initial full problem of formal outer alignment / alignment target design.
-
One of the main reasons to do CEV is because we’re gonna die of AI soon, and CEV is a way to have infinite time to solve the necessary problems. Another is that even if we don’t die of AI, we get eaten by various moloch instead of being able to safely solve the necessary problems at whatever pace is necessary.
-
the main arguments for the programmers including all of [current?] humanity in the CEV “extrapolation base” […] apply symmetrically to AIs-we’re-sharing-the-world-with at the time
I think timeless values might possibly help resolve this; if some {AIs that are around at the time} are moral patients, then sure, just like other moral patients around they should get a fair share of the future.
If an AI grabs more resources than is fair, you do the exact same thing as if a human grabs more resources than is fair: satisfy the values of moral patients (including ones who are no longer around) not weighed by how much leverage they current have over the future, but how much leverage they would have over the future if things had gone more fairly/if abuse/powergrab/etc wasn’t the kind of thing that gets your more control of the future.
“Sorry clippy, we do want you to get some paperclips, we just don’t want you to get as many paperclips as you could if you could murder/brainhack/etc all humans, because that doesn’t seem to be a very fair way to allocate the future.” — and in the same breath, “Sorry Putin, we do want you to get some of whatever-intrinsic-values-you’re-trying-to-satisfy, we just don’t want you to get as much as ruthlessly ruling Russia can get you, because that doesn’t seem to be a very fair way to allocate the future.”
And this can apply regardless of how much of clippy already exists by the time you’re doing CEV.
trying to solve morality by themselves
It doesn’t have to be by themselves; they can defer to others inside CEV, or come up with better schemes that their initial CEV inside CEV and then defer to that. Whatever other solutions than “solve everything on your own inside CEV” might exist, they can figure those out and defer to them from inside CEV. At least that’s the case in my own attempts at implementing CEV in math (eg QACI).
Seems really wonky and like there could be a lot of things that could go wrong in hard-to-predict ways, but I guess I sorta get the idea.
I guess one of the main things I’m worried about is that it seems to require that we either:
Be really good at timing when we pause it to look at its internals, such that we look at the internals after it’s had long enough to think about things that there are indeed such representations, but not long enough that it started optimizing really hard such that we either {die before we get to look at the internals} or {the internals are deceptively engineered to brainhack whoever would look at them}. If such a time interval even occurs for any amount of time at all.
Have an AI that is powerful enough to have powerful internals-about-QACI to look at, but corrigible enough that this power is not being used to do instrumentally convergent stuff like eat the world in order to have more resources with which to reason.
Current AIs are not representative of what dealing with powerful optimizers is like; when we’ll start getting powerful optimizers, they won’t sit around long enough for us to look at them and ponder, they’ll just quickly eat us.
So the formalized concept is
Get_Simplest_Concept_Which_Can_Be_Informally_Described_As("QACI is an outer alignment scheme consisting of…")? Is an informal definition written in english?It seems like “natural latent” here just means “simple (in some simplicity prior)”. If I read the first line of your post as:
Has anyone thought about QACI could be located in some simplicity prior, by searching the prior for concepts matching(??in some way??) some informal description in english?
It sure sounds like I should read the two posts you linked (perhaps especially this one), despite how hard I keep bouncing off of the natural latents idea. I’ll give that a try.
To me kinda the whole point of QACI is that it tries to actually be fully formalized. Informal definitions seem very much not robust to when superintelligences think about them; fully formalized definitions are the only thing I know of that keep meaning the same thing regardless of what kind of AI looks at it or with what kind of ontology.
I don’t really get the whole natural latents ontology at all, and mostly expect it to be too weak for us to be able to get reflectively stable goal-content integrity even as the AI becomes vastly superintelligent. If definitions are informal, that feels to me like degrees of freedom in which an ASI can just pick whichever values make its job easiest.
Perhaps something like this allows use to use current, non-vastly-superintelligent AIs to help design a formalized version of QACI or ESP which itself is robust enough to be passed to superintelligent optimizers; but my response to this is usually “have you tried first formalizing CEV/QACI/ESP by hand?” because it feels like we’ve barely tried and like reasonable progress can be made on it that way.
Perhaps there are some cleverer schemes where the superintelligent optimizer is pointed at the weaker current-tech-level AI, itself pointed informally at QACI, and we tell the superintelligent optimizer “do what this guy says”; but that seems like it either leaves too many degrees of freedom to the superintelligent optimizer again, or it requires solving corrigibility (the superintelligent optimizer is corrigibly assisting the weaker AI) at which point why not just point the corrigibility at the human directly and ignore QACI altogether, at least to begin with.
The knightian in IB is related to limits of what hypotheses you can possibly find/write down, not—if i understand so far—about an adversary. The adversary stuff is afaict mostly to make proofs work.
I don’t think this makes a difference here? If you say “what’s the best not-blacklisted-by-any-knightian-hypothesis action”, then it doesn’t really matter if you’re thinking of your knightian hypotheses as adversaries trying to screw you over by blacklisting actions that are fine, or if you’re thinking of your knightian hypotheses as a more abstract worst-case-scenario. In both cases, for any reasonable action, there’s probly a knightian hypothesis which blacklists it.
Regardless of whether you think of it as “because adversaries” or just “because we’re cautious”, knightian uncertainty works the same way. The issue is fundamental to doing maximin over knightian hypotheses.
Epistemic states as a potential benign prior
This is indeed a meaningful distinction! I’d phrase it as:
Values about what the entire cosmos should be like
Values about what kind of places one wants one’s (future) selves to inhabit (eg, in an internet-like upload-utopia, “what servers does one want to hang out on”)
“Global” and “local” is not the worst nomenclature. Maybe “global” vs “personal” values? I dunno.
my best idea is to call the former “global preferences” and the latter “local preferences”, but that clashes with the pre-existing notion of locality of preferences as the quality of terminally caring more about people/objects closer to you in spacetime
I mean, it’s not unrelated! One can view a utility function with both kinds of values as a combination of two utility functions: the part that only cares about the state of the entire cosmos and the part that only cares about what’s around them (see also “locally-caring agents”).
(One might be tempted to say “consequentialist” vs “experiential”, but I don’t think that’s right — one can still value contact with reality in their personal/local values.)
That is, in fact, a helpful elaboration! When you said
Most people who “work on AI alignment” don’t even think that thinking is a thing.
my leading hypotheses for what you could mean were:
Using thought, as a tool, has not occured to most such people
Most such people have no concept whatsoever of cognition as being a thing, the way people in the year 1000 had no concept whatsoever of javascript being a thing.
Now, instead, my leading hypothesis is that you mean:
Most such people are failing to notice that there’s an important process, called “thinking”, which humans do but LLMs “basically” don’t do.
This is a bunch more precise! For one, it mentions AIs at all.
To be more precise: extrapolated over time, for any undesired selection process or other problem of that kind, either the problem is large enough that it gets exarcerbated over time so much that it eats everything — and then that’s just extinction, but slower — or it’s not large enough to win out and aligned superintelligence(s) + coordinated human action is enough to stamp it out in the long run, which means they won’t be an issue for almost all of the future.
It seems like for a problem to be just large enough that coordination doesn’t stamp it away, but also it doesn’t eat everything, would be a very fragile equilibrium, and I think that’s pretty unlikely.
single-use
Considering how loog it took me to get that by this you mean “not dual-use”, I expect some others just won’t get it.
Some people who are very concerned about suffering might be considering building an unaligned AI that kills everyone just to avoid the risk of an AI takeover by an AI aligned to values which want some people to suffer.
Let this be me being on the record saying: I believe the probability of {alignment to values that strongly diswant suffering for all moral patients} is high enough, and the probability of {alignment to values that want some moral patients to suffer} is low enough, that this action is not worth it.
I think this applies to approximately anyone who would read this post, including heads of major labs in case they happen to read this post and in case they’re pursuing the startegy of killing everyone to reduce S-risk.
See also: how acausal trade helps in 1, 2, but I think I think this even without acausal trade.
sigh I wish people realized how useless it is to have money when the singularity happens. Either we die or we get a utopia in which it’s pretty unlikely that pre-singularity wealth matters. What you want to maximize is not your wealth but your utility function, and you sure as hell are gonna get more from LDT handshakes with aligned superintelligences in saved worlds, if you don’t help OpenAI reduce the amount of saved worlds.
I was making a more general argument that applies mainly to powerful AI but also to all other things that might help one build powerful AI (such as: insights about AI, cyborg tools, etc). These things-that-help have the downside that someone could use them to build powerful but unaligned AI, which is ultimately the thing we want to delay / reduce-the-probability-of. Whether the downside is bad enough that making them public/popular is net bad is the thing that’s uncertain, but I lean towards yes, it is net bad.
I believe that:
It is bad for cyborg tools to be broadly available because that’ll help {people trying to build the kind of AI that’d kill everyone} more than they’ll {help people trying to save the world}.
It is bad for insights about AI to spread because of the same reason.
It is bad for LLM assistants to be broadly available for the same reason.
I don’t think I’m particularly relying on that assumption?? I don’t understand what sounded like I think this.
In any case, I’m not making strict “only X are Y” or “all X are Y” statements; I’m making quantitative “X are disproportionately more Y” statements.
Well, yes. And at that point the world is much more doomed; the world has to be saved ahead of that. To increase the probability that we have time to save the world before people find out, we want to buy time. I agree it’s inevitable, but it can be delayed. Making tools and insights broadly available hastens the bursting of the dam, which is bad; containing them delays the bursting of the dam, which is good.