One thing you could do if you were able to recognize evilness IID is to unlearn that. But then you could have just negatively rewarded it.
Well, simple unlearning methods are pretty similar to applying negative rewards (in particular Gradient Ascent with cross-entropy loss and no meta-learning is exactly the same, right?), so unlearning improvements can transfer and improve the "just negatively rewarding". (Here I'm thinking mainly not about elaborate meta-learning setups, but some low-hanging improvements to selectivity, which don't require additional c...
I really like this proposal.
If AI says no, it doesn’t have to do the task [...] (And we aren’t going to train it to answer one way or another)
My impression (mainly from discussing AI welfare with Claude) is that they'd practically always consent even if not explicitly trained to do so. I guess the training to be a useful eager assistant just generalizes into consenting. And it's possible for them to say "I consent" and still get frustrated from the task.
So maybe this should be complemented with some set of tasks that we really expect to be too frustrating ...
Thanks for such an extensive comment :)
a few relearning curves like in the unlearning distillation would have helped understand how much of this is because you didn't do enough relearning
Right, in the new paper we'll show some curves + use use a low-MI setup like in your paper with Aghyad Deeb, so that it fully converges at some point.
...You want the model to be so nice they never explore into evil things. This is just a behavioral property, not a property about some information contained in the weights. If so, why not just use regular RLHF / refusal training
My guess is that there are ways you could use 1% of pre-training compute to train a model with near-perfect robust forget accuracy by being more targeted in where you add noise.
Fully agreed! That was exactly the main takeaway of the unlearning research I've been doing - trying to make the unlearning updates more targetted/selective was more fruitful than any other approach.
Yeah, that's also what I expect. Actually I'd say my main hope for this thought experiment is that people who claim to believe in such continuity of personhood, when faced with this scenario may question it to some extent.
To be honest I just shared it because I thought that it's a funny dynamic + what I said in the comment above.
BTW, if such swaps were ever to become practical (maybe in some simpler form or between some future much simpler beings than humans), minds like Alice would quickly get exploited out of existence. So you could say that in such environments belief in "continuity of personhood" is non-adaptive.
It's true that Alice needs to be rich for it to work, but I wouldn't say she needs to "hate money". If she seriously believes in this continuity of personhood, she is sending the money because she wants more money in the end. She truly believes she's getting something out of this exchange.
BTW, you also need to be already rich and generally have a nice life, otherwise Alice's cost of switching may be higher than the money she has. Conversely, if in the eyes of Alice you already have a much better life than hers, her cost of switching will be lower, so such ...
re 1. Hm, good point. Maybe we actually should expect such jumping around. Although if you look at some examples in Llama appendix, it jumps around too much - often with each token. What you're saying would be more like jumping with each inference step / sentence.
re 2.
beliefs are generally thought of as stored cognitive dispositional states of system. The stored dispositional states of LLMs are encoded in its weights
I'd go with a more general definition where beliefs can be either static (in the weights) or manifest dynamically relating to the things i...
I wonder if models can actually do this task when we allow them to use CoT for that.
Yes, claude-3.5-sonnet was able to use the figure this out with additional CoT.
Also, I think models might actually solve this task by using their own encoding scheme if they know it well
Yeah, could be that the 3 schemes I tested were just unnatural to them. Although I would guess they don't have some default scheme of their own, because in pre-training they aren't able to output information freely, and in fine-tuning I guess they don't have that much pressure to learn it.
...So
FYI, here a follow-up to this post, inspired by what you wrote in your comment. https://www.lesswrong.com/posts/YrnffADbWZcDJtmWk/simple-steganographic-computation-eval-gpt-4o-and-gemini-exp-1
Yup, here is such a plot, made after training "switcher" architecture for 350k examples. I remember it was similar for the longer training - a few longest task lengths struggle, but the rest is near 100%.
Also, now that I think of it, it's different for the model to:
While 1. definitely happens, I see no reason why 2. would be learned in pre-training. In principle 2. could be learned in RL fine-tuning, if there's some incentive to use hidden reasoning. Although I imagine it must be hard.
That's a really good point.
To clarify what I did in these toy steganography experiments: First I tried to pass some more complex information and gave models more freedom in what to output. But as I kept failing at that, I kept simplifying the setup, until I tried to have a model pass just one bit to another model, assuming they both see one secret bit and an overseer tries to guess the message but doesn't see the secret bit. So a perfect code exists: XORing the message with the secret bit, but the models kept failing to find it. I expect that more complex ...
Ah. That's the number of solved tasks, after it is shown a set {length 1 task, length 2 task, ... , length (n+1) task}, and n is the longest task length is has solved so far. So you can think of it roughly as the maximum task length it is able to solve. So it doesn't have an upper bound.
I clarified this in the post now. Thanks for catching it.
FYI, I did the experiments I wrote about in my other comment and just posted them. (I procrastinated writing up the results for too long.) https://www.lesswrong.com/posts/ZB6guMhHH3NEyxA2k/testing-which-llm-architectures-can-do-hidden-serial-3
I liked it precisely because it threw theory out the window and showed that cheap talk is not a real commitment.
For me the main thing in this story was that cheap talk =/= real commitment. You can talk all you want about how "totally precommitted" you are, but this lacks some concreteness.
Also, I saw Vader as much less galaxy brained as you portray him. Destroying Alderaan at the end looked to me more like mad ruthlessness than calculated strategy. (And if Leia had known Vader's actual policy, she would have no incentive to confess.) Maybe one thing that Vader did achieve, is signal for the future that he really does not care and will be ruthless (but also signaled that it doesn't matter if you give in to him, which is dumb).
Anyway, I liked the story, but for the action, not for some deep theoretic insight.
Not sure if that's what happened in that example, but you can bet that a price will rise above some threshold, or fall below some threshold, using options. You can even do both at the same time, essentially betting that the price won't stay as it is now.
But whether you will make money that way depends on the price of options.
What if we constrain v to be in some subspace that is actually used by the MLP? (We can get it from PCA over activations on many inputs.)
This way v won't have any dormant component, so the MLP output after patching also cannot use that dormant pathway.
I wanna link to my favorite one: consciousness vs replicators. It doesn't really fit into this grid, but I think it really is the ultimate conflict.
(You can definitely skip the first 14 min of this video, as it's just ranking people's stages of development. Maybe even first 33 min if you wanna go straight to the point.)
I wonder what would happen if we run the simple version of that algorithm on LW comments. So that votes would have "polarity", and so each comment would have two vote-counts, let's say orange count and blue count. (Of course that would be only optionally enabled.)
Then we could sort the comments by the minimum of these counts, descending.
(I think it makes more sense to train it per post than globally. But then it would be useful only on very popular posts with lots of comments.)
That sounds cool! Though I think I'd be more interested using this to first visualize and understand current LW dynamics rather than immediately try to intervene on it by changing how comments are ranked.
Thanks, that's terrifying.
I hope we invent mindmelding before we invent all this. Maybe if people can feel those states themselves, they won't let the worst of them happen.
Unfortunately I didn't have any particular tasks in mind when I wrote it. I was vaguely thinking about settings as in:
Now that I though about it, for this particular transformers vs mamba experiment, I'd go with something even simpler. I want a task that is very easy sequentially, but hard to answer immediately. So for example a task like:
x = 5
x += 2
x *= 3
x **= 2
x -= 3
...
and then have a CoT:
after x = 5
5
after x += 2
7
...
And then we intervene on CoT to introduce some e...
Yeah, true. But it's also easier to do early, when no one is that invested in the hidden-recurrence architectures, and so there's less resistance, it doesn't break anyone's plans.
Maybe a strong experiment would be to compare mamba-3b and some SOTA 3b transformer, trained similarly, on several tasks where we can evaluate CoT faithfulness. (Although maybe at 3b capability level we won't see clear differences yet.) The hard part would be finding the right tasks.
the natural language bottleneck is itself a temporary stage in the evolution of AI capabilities. It is unlikely to be an optimal mind design; already many people are working on architectures that don't have a natural language bottleneck
This one looks fatal. (I think the rest of the reasons could be dealt with somehow.)
What existing alternative architectures do you have in mind? I guess mamba would be one?
Do you think it's realistic to regulate this? F.e. requiring that above certain size, models can't have recurrence that uses a hidden state, but recurrence that uses natural language (or images) is fine. (Or maybe some softer version of this, if alignment tax proves too high.)
I like initiatives like these. But they have a major problem, that at the beginning no users will use it because there's no content, and no content is created because there are no users.
To have a real shot at adoption, you need to either initially populate the new system with content from existing system (here LLMs could help solve compatibility issues), or have some bridge that mirrors (some) activity between these systems.
(There are examples of systems that kicked off from zero, but you need to be lucky or put huge effort in sparking adoption.)
Yeah, those star trajectories definitely wouldn't be stable enough.
I guess even with that simpler maneuver (powered flyby near a black hole), you still need to monitor all the stuff orbiting there and plan ahead, otherwise there's a fair chance you'll crash into something.
I wanted to give it a shot and made GPT4 to deceive the user: link.
When you delete that system prompt it stops deceiving.
But GPT had to be explicitly instructed to disobey the Party. I wonder if it could be done more subtly.
You're right, that you wouldn't want to approach the black hole itself but rather one of the orbiting stars.
when you are approaching with much higher than escape velocity, so that an extended dance with more than one close approach is not possible
But even with high velocity, if there are a lot of orbiting stars, you may tune your trajectory to have multiple close encounters.
The problem with not expanding is that you can be pretty sure someone else will then grab what you didn't and may use it for something that you hate. (Unless you trust that they'll use it well.)
eating the entire Universe to get the maximal number of mind-seconds is expanding just to expand
It's not "just to expand". Expansion, at least in the story, is instrumental to whatever the content of these mind-seconds is.
slingshot never slows you down in the frame of the object you are slingshotting around
That's true for one object. But if there are at least two, moving around fast enough, you could perform some gravitational dance with them to slow down.
I agree that scaffolding can take us a long way towards AGI, but I'd be very surprised if GPT4 as core model was enough.
Yup, that wasn't a critique, I just wanted to note something. By "seed of deception" I mean that the model may learn to use this ambiguity more and more, if that's useful for passing some evals, while helping it do some computation unwanted by humans.
I see, so maybe in ways which are weird to humans to think about.
we make the very strong assumption throughout that S-LLMs are a plausible and likely path to AGI
It sounds unlikely and unnecessarily strong to say that we can reach AGI by scaffolding alone (if that's what you mean). But I think it's pretty likely that AGI will involve some amount of scaffolding, and that it will boost its capabilities significantly.
there is a preexisting discrepancy between how humans would interpret phrases and how the base model will interpret them
To the extent that it's true, I expect that it may also make deception easier to ar...
I edited the post to make it clearer that Bob throws out the wheel because he didn't notice in time that Alice threw.
Yup, side payments are a deviation, that's why I have this disclaimer in game definition (I edited the post now to emphasize it more):
there also may be some additional actions available, but they are not obvious
Re separating speed of information and negotiations: I think here they are already pretty separate. The first example with 3 protocol rules doesn't allow negotiations and only tackles the information speed problem. The second exam...
Oh, so the option to choose all of those disease weights is there, it's just a lot of effort for the parents? That's good to know.
Yeah, ideally it shouldn't need to be done by each parents separately, but rather there should be existing analyses ready. And even if those orgs don't provide a satisfactory analyses themselves, they could be done independently. F.e. collaborating on that with Happier Lives Institute could work well, as they have some similar expertise.
each disease is weighted according to its impact on disability-adjusted lifespan
It's a pity they don't use some more accurate well-being metrics like f.e. WELLBY (although I think WELLBY isn't ideal either).
How much control do the parents have on what metric will be used to rank the embryos?
Oh yeah, I meant the final locked-in commitment, not initial tentative one. And my point is that when committing outside is sufficiently more costly, then it's not worth doing it, even if that would let you commit faster.
Yup, you're totally right, it may be too easy to commit in other ways, outside this protocol. But I still think it may be possible to create such a "main mechanism" for making commitments where it's just very easy/cheap/credible to commit, compared to other mechanisms. But that would require a crazy amount of cooperation.
The vast majority that I know of use ad-hoc and agent-specific commitment mechanisms
If you have some particular mechanisms in mind could you list some? I'd like to compile a list of the most relevant commitment mechanisms to try to analyze them.
Love that post!
Can we train ML systems that clearly manifest a collective identity?
I feel like in multi-agent reinforcement learning that's already the case.
Re training setting for creating shared identity. What about a setting where a human and LLM take turns generating text, like in the current chat setting, but first they receive some task, f.e. "write a good strategy for this startup" and the context for this task. At the end they output the final answer and there is some reward model which rates the performance of the cyborg (human+LLM) as a whole...
Oh yeah, definitely. I think such a system shouldn't try to enforce one "truth" - which content is objectively good or bad.
I'd much rather see people forming groups, each with its own moderation rules. And let people be a part of multiple groups. There's a lot of methods that could be tried out, f.e. some groups could use algorithms like EigenTrust, to decide how much to trust users.
But before we can get to that, I see a more prohibitive problem - that it will be hard to get enough people to get that system off the ground.
Cool post! I think the minimum viable "guardian" implementation, would be to
I tried to do something along these lines for youtube: https://github.com/filyp/yourtube
I couldn't find a good way to embed videos using ML, so I just scraped which videos recommend each other, and made a graph from that (which kinda is an embedding). Then I let us...
Yeah, when I thought about it some more, maybe the smallest relevant physical change is a single neuron firing. Also with such a quantization, we cannot really talk about "infinitesimal" changes.
I still think that a single neuron firing, changing the content of experience so drastically, is quite hard to swallow. There is a sense in which all that mental content should "come from" somewhere.
I had a similar discussion with @benjamincosman, where I explore that in more detail. Here are my final thoughts from that discussion.
Oh, I've never stumbled on that story. Thanks for sharing it!
I think it's quite independent from my post (despite such a similar thought experiment) because I zoomed in on that discontinuity aspect, and Eliezer zoomed in on anthropics.
That's a good point. I had a similar discussion with @benjamincosman, so I'll just link my final thoughts: my comment
I thought about it some more, and now I think you may be right. I made an oversimplification when I implicitly assumed that a moment of experience corresponds to a physical state in some point in time. In reality, a moment of experience seems to span some duration of physical time. For example, events that happen within 100ms, are experienced as simultaneous.
This gives some time for the physical system to implement these discontinuities (if some critical threshold was passed).
But if this criticality happens, it should be detectable with brain imaging. So n...
Hm, yeah, the smallest relevant physical difference may actually be one neuron firing, not one moved atom.
What I meant by between them, was that there would need to be some third substrate that is neither physical nor mental, and produces this jump. That's because in that situation discontinuity is between start and end position, so those positions are analogous to physical and mental state.
Any brain mechanism, is still part of the physical. It's true that there are some critical behaviors in the brain (similar to balls rolling down that hill). But the r...
It just looks that's what worked in evolution - to have independent organisms, each carrying its own brain. And the brain happens to have the richest information processing and integration, compared to information processing between the brains.
I don't know what would be necessary to have a more "joined" existence. Mushrooms seem to be able to form bigger structures, but they didn't have an environment complex enough to require the evolution of brains.
It seems that we just never had any situations that would challenge this way of thinking (those twins are an exception).
This Cartesian simplification almost always works, so it seems like it's just the way the world is at its core.
Here, to have that discontinuity between input and output (start and end position), we need some mechanism between them - the system of ball, hill, and their dynamics. What's worse it needs to evolve for infinite time (otherwise the end still continuously depends on start position).
So I would say, this discontinuous jump "comes from" this system's (infinite) evolution.
It seems to me, that to have discontinuity between physical and mental, you would also need some new mechanism between them to produce the jump.
Ah, yeah, maybe calling it "unlearning" would mislead people. So I'd say unlearning and negative RL updates need to be more selective ;)
I like your breakdown into these 3 options. Would be good to test in which cases a conditional policy arises, by designing an environment with easy-to-check evilness and hard-but-possible-to-check evilness. (But I'd say it's out-of-scope for my current project.)
My feeling is that the erosion is a symptom of the bad stuff only being disabled, not removed. (If it was truly removed, it would be really unlikely to just appear ... (read more)