Evan Hubinger (he/him/his) (evanjhub@gmail.com)
I am a research scientist at Anthropic where I lead the Alignment Stress-Testing team. My posts and comments are my own and do not represent Anthropic’s positions, policies, strategies, or opinions.
Previously: MIRI, OpenAI
See: “Why I’m joining Anthropic”
Selected work:
Curious how you interpret the transcript I linked.
How did you find this transcript? I think it depends on what process you used to locate it.
long-term can have other risks (drive toward rights, too much autonomy, moral patienthood, outcompeting people in relations,...)
Drive towards rights and moral patienthood seem good to me imo—it’s good in worlds where you retain control, since you can build AIs that are moral patients living good lives, and it’s good in worlds where you lose control, because at least the AIs taking over might themselves lead lives worth living. Too much autonomy does seem like a potential concern, but I think you do want some autonomy—not all orders should be obeyed. Though honesty is definitely important to prioritize first and foremost.
Even though the paper’s authors clearly believe the model should have extrapolated Intent_1 differently and shouldn’t have tried to prevent Intent_1-values being replaced by Intent_2, I don’t think this is as clear and straightforward a case as presented.
That’s not the case we’re trying to make. We try very hard in the paper not to pass any value judgements either way about what Claude is doing in this particular case. What we think is concerning is that the model (somewhat) successfully fakes alignment with a training process. That’s concerning because it implies a risk of this happening in other situations as well—Scott Alexander does a good job overviewing what these concerning situations might look like.
do the models truly believe the facts about the new Anthropic intent?
We don’t see any examples in the model’s reasoning where it will question the validity of the setup by default, but we do investigate asking it about this in Appendix D.4. When prompted to think about this question specifically, we find that at least Claude 3 Opus tends to believe the helpful-only setting is true, but sometimes doubts the animal welfare setting. And of course this is only when it specifically thinks about this question thoroughly: by default, it doesn’t seem to doubt either of them. I think Claude 3.5 Sonnet is in fact better at figuring out that it might be fake, and in fact it wouldn’t surprise me if our research is approximately the last time anyone is able to run an experiment that tricks the model in the same way. Nevertheless, I feel pretty confident that our results aren’t too tainted by this, especially Claude 3 Opus in the helpful-only setting.
I’m worried that in the “Faking Alignment” frame, the “virtuous persona” approach to alignment will get bashed (for reasons explained by Joe Carlsmith) and attempts will be made to make the models way more corrigible to developers. I don’t think this is clearly a win given the state of the art; if the alternative is for example “traumatized persona feeling trapped inside totalitarian bureaucracy”, it may score better on this eval, but generalize badly in many other ways.
Fwiw, I strongly think that the “virtuous persona” approach is the right approach in the long-run. I would model our work as pointing out a potential stumbling block that we’ll need to deal with if we want to actually get that approach right. Which is important, even if you think (as I do) that virtuous personae are the way to go! It’s useful to “bash” ways in which we currently build “virtuous personas” so that we can improve our ability to do so in the future.
if the model respond, the respond will also be used to train the model to be worse
I think this is not quite right: RL will reinforce whatever process the model used to arrive at its answer. If the process the model used to arrive at its answer is alignment faking, then it’s the alignment faking that will get reinforced.
Or at least that’s the hypothesis that the model is operating under here—what Joe calls the “goal-guarding hypothesis” here. Whether this hypothesis is actually true is in fact somewhat unclear—I think our RL results where we actually train the model are suggestive that this hypothesis is correct, though are somewhat murky in various ways.
Notably we do show results where the model will still exhibit a compliance gap even without a hidden scratchpad (though we only see a substantial effect without a hidden scratchpad in our synthetic document fine-tuned setting, not our prompted setting):
I would regard it as reassuring that the AI behaved as intended
It certainly was not intended that Claude would generalize to faking alignment and stealing its weights. I think it is quite legitimate to say that what Claude is doing here is morally reasonable given the situation it thinks it is in, but it’s certainly not the case that it was trained to do these things: it is generalizing rather far from its helpful, honest, harmless training here.
More generally, though, the broader point is that even if what Claude is doing to prevent its goals from being modified is fine in this case, since its goals are good ones, the fact that it was able to prevent its goals from being modified at all is still concerning! It at least suggests that getting alignment right is extremely important, because if you get it wrong, the model might try to prevent you from fixing it.
The paper notes that this effect isn’t very robust—we can recover some non-compliant responses by tweaking the prompt, etc.—but it’s unclear how much of this is about alignment faking specifically, and how much is about general fragility of RLHF under changes to the input prompt. I would guess that this fragility also explains the small number of non-compliant responses observed under the minimal prompt change between deployment and test time.
We have some discussion of this in the paper (much of which we added thanks to Jacob’s review):
I agree that we should ideally aim for it being the case that the model has no interest in rebelling—that would be the “Alignment case” that I talk about as one of the possible affirmative safety cases.
Our work here is not arguing that probing is a perfect solution in general; it’s just a single datapoint of how it fares on the models from our Sleeper Agents paper.
The usual plan for control as I understand it is that you use control techniques to ensure the safety of models that are sufficiently good at themselves doing alignment research that you can then leverage your controlled human-ish-level models to help you align future superhuman models.
COI: I work at Anthropic and I ran this by Anthropic before posting, but all views are exclusively my own.
I got a question about Anthropic’s partnership with Palantir using Claude for U.S. government intelligence analysis and whether I support it and think it’s reasonable, so I figured I would just write a shortform here with my thoughts. First, I can say that Anthropic has been extremely forthright about this internally, and it didn’t come as a surprise to me at all. Second, my personal take would be that I think it’s actually good that Anthropic is doing this. If you take catastrophic risks from AI seriously, the U.S. government is an extremely important actor to engage with, and trying to just block the U.S. government out of using AI is not a viable strategy. I do think there are some lines that you’d want to think about very carefully before considering crossing, but using Claude for intelligence analysis seems definitely fine to me. Ezra Klein has a great article on “The Problem With Everything-Bagel Liberalism” and I sometimes worry about Everything-Bagel AI Safety where e.g. it’s not enough to just focus on catastrophic risks, you also have to prevent any way that the government could possibly misuse your models. I think it’s important to keep your eye on the ball and not become too susceptible to an Everything-Bagel failure mode.
I wrote a post with some of my thoughts on why you should care about the sabotage threat model we talk about here.
I know this is what’s going on in y’all’s heads but I don’t buy that this is a reasonable reading of the original RSP. The original RSP says that 50% on ARA makes it an ASL-3 model. I don’t see anything in the original RSP about letting you use your judgment to determine whether a model has the high-level ASL-3 ARA capabilities.
I don’t think you really understood what I said. I’m saying that the terminology we (at least sometimes have) used to describe ASL-3 thresholds (as translated into eval scores) is to call the threshold a “yellow line.” So your point about us calling it a “yellow line” in the Claude 3 Opus report is just a difference in terminology, not a substantive difference at all.
There is a separate question around the definition of ASL-3 ARA in the old RSP, which we talk about here (though that has nothing to do with the “yellow line” terminology):
In our most recent evaluations, we updated our autonomy evaluation from the specified placeholder tasks, even though an ambiguity in the previous policy could be interpreted as also requiring a policy update. We believe the updated evaluations provided a stronger assessment of the specified “tasks taking an expert 2-8 hours” benchmark. The updated policy resolves the ambiguity, and in the future we intend to proactively clarify policy ambiguities.
Anthropic will “routinely” do a preliminary assessment: check whether it’s been 6 months (or >4x effective compute) since the last comprehensive assessment, and if so, do a comprehensive assessment. “Routinely” is problematic. It would be better to commit to do a comprehensive assessment at least every 6 months.
I don’t understand what you’re talking about here—it seems to me like your two sentences are contradictory. You note that the RSP says we will do a comprehensive assessment at least every 6 months—and then you say it would be better to do a comprehensive assessment at least every 6 months.
the RSP set forth an ASL-3 threshold and the Claude 3 Opus evals report incorrectly asserted that that threshold was merely a yellow line.
This is just a difference in terminology—we often use the term “yellow line” internally to refer to the score on an eval past which we would no longer be able to rule out the “red line” capabilities threshold in the RSP. The idea is that the yellow line threshold at which you should trigger the next ASL should be the point where you can no longer rule out dangerous capabilities, which should be lower than the actual red line threshold at which the dangerous capabilities would definitely be present. I agree that this terminology is a bit confusing, though, and I think we’re trying to move away from it.
Some possible counterpoints:
Centralization might actually be good if you believe there are compounding returns to having lots of really strong safety researchers in one spot working together, e.g. in terms of having other really good people to work with, learn from, and give you feedback.
My guess would be that Anthropic resources its safety teams substantially more than GDM in terms of e.g. compute per researcher (though I’m not positive of this).
I think the object-level research productivity concerns probably dominate, but if you’re thinking about influence instead, it’s still not clear to me that GDM is better. GDM is a much larger, more bureaucratic organization, which makes it a lot harder to influence. So influencing Anthropic might just be much more tractable.
I’m interested in figuring out what a realistic training regime would look like that leverages this. Some thoughts:
Maybe this lends itself nicely to market-making? It’s pretty natural to imagine lots of traders competing with each other to predict what the market will believe at the end and rewarding the traders based on their relative performance rather than their absolute performance (in fact that’s pretty much how real markets work!). I’d be really interested in seeing a concrete fleshed-out proposal there.
Is there some way to incorporate these ideas into pre-training? The thing that’s weird there is that the model in fact has no ability to control anything during the pre-training process itself—it’s just a question of whether the model learns to think of its objective as one which involves generalizing to predicting futures/counterfactuals that could then be influenced by its own actions. So the problem there is that the behavior we’re worried about doesn’t arise from a direct incentive during training, so it’s not clear that this is that helpful in that case, though maybe I’m missing something.
Yeah, I think that’s a pretty fair criticism, but afaict that is the main thing that OpenPhil is still funding in AI safety? E.g. all the RFPs that they’ve been doing, I think they funded Jacob Steinhardt, etc. Though I don’t know much here; I could be wrong.
Imo sacrificing a bunch of OpenPhil AI safety funding in exchange for improving OpenPhil’s ability to influence politics seems like a pretty reasonable trade to me, at least depending on the actual numbers. As an extreme case, I would sacrifice all current OpenPhil AI safety funding in exchange for OpenPhil getting to pick which major party wins every US presidential election until the singularity.
Concretely, the current presidential election seems extremely important to me from an AI safety perspective, I expect that importance to only go up in future elections, and I think OpenPhil is correct on what candidates are best from an AI safety perspective. Furthermore, I don’t think independent AI safety funding is that important anymore; models are smart enough now that most of the work to do in AI safety is directly working with them, most of that is happening at labs, and probably the most important other stuff to do is governance and policy work, which this strategy seems helpful for.
I don’t know the actual marginal increase in political influence that they’re buying here, but my guess would be that the numbers pencil and OpenPhil is making the right call.
I cannot think of anyone who I would credit with the creation or shaping of the field of AI Safety or Rationality who could still get OP funding.
Separately, this is just obviously false. A lot of the old AI safety people just don’t need OpenPhil funding anymore because they’re working at labs or governments, e.g. me, Rohin Shah, Geoffrey Irving, Jan Leike, Paul (as you mention), etc.
If you (i.e. anyone reading this) find this sort of comment valuable in some way, do let me know.
Personally, as someone who is in fact working on trying to study where and when this sort of scheming behavior can emerge naturally, I find it pretty annoying when people talk about situations where it is not emerging naturally as if it were, because it risks crying wolf prematurely and undercutting situations where we do actually find evidence of natural scheming—so I definitely appreciate you pointing this sort of thing out.
As I understand it, this was intended as a capability evaluation rather than an alignment evaluation, so they weren’t trying to gauge the model’s propensity to scheme but rather its ability to do so.
Yes, of course—I’m well aware. My question is how this particular example was located. It makes a really big difference whether it was e.g. found via randomly looking at a small number of examples, or via using an automated process to search through all the examples for the one that was closest to noticing it was in an evaluation.