Oh interesting! I just had a go at testing it on screenshots from a parallel conversation and it seems like it incorrectly interprets those screenshots as also being of its own conversation.
So it seems like ‘recognising things it has said’ is doing very little of the heavy lifting and ‘recognising its own name’ is responsible for most of the effect.
I’ll have a bit more of a play around and probably put a disclaimer at the top of the post some time soon.
The ‘reward being chance of winning’ stuff changes a bit about how the model generalises if it’s playing a game with randomness and conditioned on the upper end—it biases the model towards ‘expecting risk to pay off’. Eg/ if the model plays a 1-step game where it either banks 1 point or gets a 1% chance of 10 points, then conditioning on it getting 10 points will cause it to take the lower EV action. But this isn’t super relevant.
In general I am pretty confused about how models would generalise strategies out of distribution, and this seems like at least partly an empirical ML / science of deep learning problem. But my intuition is that no, even if the model ‘understands’ that an action causes a good outcome [ P(outcome |. action) ], it still might not take that action if the action is unlikely on priors [ P(action) ], because the basic driver of its choice is what action it will observe given the outcome [ P(action | outcome) ] which is proportional to both of those quantities (and in fact, probably equal, with some normalising constant P(outcome)).
Re lowering the temperature: the thing the predictive model learns is usually some numbers that you normalise into a probability distribution with a softmax function, and you can easily tweak the softmax to make it favour more likely completions, and disfavour unlikely ones. This would greatly decrease the turn-by-turn chance of disconnecting the button (and therefore overall chance), but probably not impair the model’s performance elsewhere much.
Re generalisation—decision transformers don’t really have strategies per se, they pick moment to moment, and might be systematically miscalibrated about what they’ll do in future timesteps. It is true that they’ll have some chance at every timestep, which will add up over time, but if you were actually trying to implement this then you could do things like lowering the temperature, which shouldn’t affect performance otherwise.
Re higher conditioning—I think this shouldn’t be true. For the sake of argument we can reframe it as a binary outcome, where the model’s final return (as a proportion of total possible return) becomes its chance of ‘winning’. The thing the model is figuring out is not ‘what action leads to me winning’, or even ‘what action is more likely in worlds where I win than worlds where I lose’, it’s ‘what action do I expect to see from agents that win’. If on turn 1, 99% of agents in the training set voluntarily slap a button that has a 1% chance of destroying them, and then 50% go on to win, as well as 50% of the agents that didn’t slap the button, then a DT will (correctly) learn that ‘almost all agents which go on to win tend to slap the button on turn 1’.
Re correlation—Sure, I am taking the liberal assumption that there’s no correlation in the training data, and indeed a lot of this rests on the training data having a nice structure
Thanks! Yeah this isn’t in the paper, it’s just a thing I’m fairly sure of which probably deserves a more thorough treatment elsewhere. In the meantime, some rough intuitions would be:
delusions are a result of causal confounders, which must be hidden upstream variables
if you actually simulate and therefore specify an entire markov blanket, it will screen off all other upstream variables including all possible confounders
this is ludicrously difficult for agents with a long history (like a human), but if the STF story is correct, it’s sufficient, and crucially, you don’t even need to know the full causal structure of reality, just a complete markov blanket
any holes in the markov blanket/boundary represent ways for unintended causal pathways to leak through, which separate the predictor’s predictions about the effect of an action from the actual causal effect of the action, making the agent appear ‘delusional’
I hope we’ll have a proper writeup soon; in the meantime let me know if this doesn’t make sense.
A slightly sideways argument for interpretability: It’s a really good way to introduce the importance and tractability of alignment research
In my experience it’s very easy to explain to someone with no technical background that
Image classifiers have got much much better (like in 10 years they went from being impossible to being something you can do on your laptop)
We actually don’t really understand why they do what they do (like we don’t know why the classifier says this is an image of a cat, even if it’s right)
But, thanks to dedicated research, we have begun to understand a bit of what’s going on in the black box (like we know it knows what a curve is, we can tell when it thinks it sees a curve)
Then you say ‘this is the same thing that big companies are using to maximise your engagement on social media and sell you stuff, and look at how that’s going. and by the way did you notice how AIs keep getting bigger and stronger?’
At this point my experience is it’s very easy for people to understand why alignment matters and also what kind of thing you can actually do about it.
Compare this to trying to explain why people are worried about mesa-optimisers, boxed oracles, or even the ELK problem, and it’s a lot less concrete. People seem to approach it much more like a thought experiment and less like an ongoing problem, and it’s harder to grasp why ‘developing better regularisers’ might be a meaningful goal.
But interpretability gives people a non-technical story for how alignment affects their lives, the scale of the problem, and how progress can be made. IMO no other approach to alignment is anywhere near as good for this.
My main takeaway from this post is that it’s important to distinguish between sending signals and trying to send signals, because the latter often leads to goodharting.
It’s tricky, though, because obviously you want to be paying attention to what signals you’re giving off, and how they differ from the signals you’d like to be giving off, and sometimes you do just have to try to change them.
For instance, I make more of an effort now than I used to, to notice when I appreciate what people are doing, and tell them, so that they know I care. And I think this has basically been very good. This is very much not me dropping all effort to signal.
But I think what you’re talking about is very applicable here, because if I were just trying to maximise that signal, I would probably just make up compliments, and this would probably be obviously insincere. So I guess the big question is, which things do you stop trying to do?
(Also, I notice I’m now overthinking editing this comment because I’ve switched gears from ‘what am I trying to say’ to ‘what will people interpret from this’. Time to submit, I guess.)
if you think timelines are short for reasons unrelated to biological anchors, I don’t think Bio Anchors provides an affirmative argument that you should change your mind.
Eliezer: I wish I could say that it probably beats showing a single estimate, in terms of its impact on the reader. But in fact, writing a huge careful Very Serious Report like that and snowing the reader under with Alternative Calculations is probably going to cause them to give more authority to the whole thing. It’s all very well to note the Ways I Could Be Wrong and to confess one’s Uncertainty, but you did not actually reach the conclusion, “And that’s enough uncertainty and potential error that we should throw out this whole deal and start over,” and that’s the conclusion you needed to reach.
I would be curious to know what the intended consequences of the forecasting piece were.
A lot of Eliezer’s argument seems to me to be pushing at something like ‘there is a threshold for how much evidence you need before you start putting down numbers, and you haven’t reached it’, and I take what I’ve quoted from your piece to be supporting something like ‘there is a threshold for how much evidence you might have, and if you’re above it (and believe this forecast to be an overestimate) then you may be free to ignore the numbers here’, contra the Humbali position. I’m not particularly confident on that, though.
Where this leaves me is feeling like you two have different beliefs about who will (or should) update on reading this kind of thing, and to what end, which is probably tangled up in beliefs about how good people are at holding uncertainty in their mind. But I’m not really sure what these beliefs are.
The belief that people can only be morally harmed by things that causally affect them is not universally accepted. Personally I intuitively would like my grave to not be desecrated, for instance.
I think we have lots of moral intuitions that have become less coherent as science has progressed. But if my identical twin started licensing his genetic code to make human burgers for people who wanted to see what cannibalism was like, I would feel wronged.
I’m using pretty charged examples here, but the point I’m trying to convey is that there are a lot of moral lenses to apply here, and there are defensible deontological prohibitions to be made. Perhaps under scrutiny they’d fall away but I don’t think it’s clear cut, or at least not yet.
You ask a number of good questions here, but the crucial point to me is that they are still questions. I agree it seems, based on my intuitions of the answers, like this isn’t the best path. But ‘how much would it cost’ and ‘what’s the chance a clone works on something counterproductive’ are, to me, not an argument against cloning, but rather arguments for working out how to answer those questions.
Also very ironic if we can’t even align clones and that’s what gets us.
I think there are extra considerations to do with what the clone’s relation to von Neumann. Plausibly, it might be wrong to clone him without his consent, which we can now no longer get. And the whole idea that you might have a right to your likeness, identity, image, and so on, becomes much trickier as soon as you have actually been cloned.
Also there’s a bit of a gulf between a parent deciding to raise a child they think might do good and a (presumably fairly large) organisation funding the creation of a child.
I don’t have strongly held convictions on these points, but I do think that they’re important and that you’d need to have good answers before you cloned somebody.
Well, I basically agree with everything you just said. I think we have quite different opinions about what politics is, though, and what it’s for. But perhaps this isn’t the best place to resolve those differences.
Ok I think this is partly fair, but also clearly our moral standards are informed by our society, and in no small part those standards emerge from discussions about what we collectively would like those standards to be, and not just a genetically hardwired disloyalty sensor.
Put another way: yes, in pressured environments we act on instinct, but those instincts don’t exist in a vacuum, and the societal project of working out what they ought to be is quite important and pretty hard, precisely because in the moment where you need to refer to it, you will be acting on System 1.
I believe Dusan was trying to say that davidad’s agenda limits the planner AI to only writing provable mathematical solutions. To expand, I believe that compared to what you briefly describe, the idea in davidad’s agenda is that you don’t try to build a planner that’s definitely inner aligned, you simply have a formal verification system that ~guarantees what effects a plan will and won’t have if implemented.