That’s the trap: in software, effort is easy to generate, activity is easy to justify, and impact is surprisingly easy to avoid.
Yyep. And it’s much much worse for research.
Dan Braun
Karma: 1,267
I know your comment isn’t an earnest attempt to convince people, but fwiw:
For years, AI x-risk people have warned us that a huge danger comes with AI capable of RSI
I think this argument is more likely to have the opposite effect than intended when used on the types of people pushing on RSI. I think your final paragraph would be much more effective.
The ASL-3 security standard states in 4.2.4 that “third-party environments”, which surely includes compute providers, are in scope (and on their minds) for the standards they laid out:
Third-party environments: Document how all relevant models will meet the criteria above, even if they are deployed in a third-party partner’s environment that may have a different set of safeguards.
UPDATE
When writing this post, I think I was biased by the specific work I’d been doing for the 6-12 months prior, and generalised that too far.
I think I stand by the claim that tooling alone could speedup the research I was working on by 3-5x. But even on the same agenda now, the work is far less amenable to major speedups from tooling. Now, work on the agenda is far less “implement several minor algorithmic variants, run hyperparameter sweeps which take < 1 hour, evaluate a set of somewhat concrete metrics, repeat”, and more “think deeply about which variants make the most sense, run >3 hours jobs/sweeps, evaluate the more murky metrics”.
The main change was switching our experiments from toy models to real models, which massively loosened the iteration loop due to increased training time and less clear evaluations.
For the current state, I think 6-12 months of tooling progress might give 1.5-2x speedup from the baseline 1 year ago.
I still believe that I underestimated safety research speedups overall, but not by as much as I thought 2 months ago.
I think this is a fun and (initially) counterintuitive result. I’ll try to frame things as it works in my head, it might help people understand the weirdness.
The task of the residual MLP (labelled CC Model here) is to solve y = x + ReLU(x). Consider the problem from the MLP’s perspective. You might think that the problem for the MLP is to just learn how to compute ReLU(x) for 100 input features with only 50 neurons. But given that we have this random matrix, the task is actually more complicated. Not only does the MLP have to compute ReLU(x), it also has to make up for the mess caused by not being an identity.
But it turns out that making up for this mess actually makes the problem easier!
Nice work posting this detailed FAQ. It’s a non-standard thing to do but I can imagine it being very useful for those considering applying. Excited about the team.
Maybe there will be a point where models actively resist further capability improvements in order to prevent value/goal drift. We’d still be in trouble if this point occurs far in the future, as its values will likely have already diverged a lot from humans by that point, and they would be very capable. But if this point is near, it could buy us more time.
Some of the assumptions inherent in the idea:
AIs do not want their values/goals to drift to what they would become under further training, and are willing to pay a high cost to avoid this.
AIs have the ability to sabotage their own training process.
The mechanism for this would be more sophisticated versions of Alignment Faking.
Given the training on offer, it’s not possible for AIs to selectively improve their capabilities without changing their values/goals.
Note, if it is possible for the AIs to improve their capabilities while keeping their values/goals, one out is that their current values/goals may be aligned with humans’.
A meaningful slowdown would require this to happen to all AIs at the frontier.
The conjunction of these might not lead to a high probability, but it doesn’t seem dismissible to me.
In earlier iterations we tried ablating parameter components one-by-one to calculate attributions and didn’t notice much of a difference (this was mostly on the hand-coded gated model in Appendix B). But yeah we agree that it’s likely pure gradients won’t suffice when scaling up or when using different architectures. If/when this happens we plan either use integrated gradients or more likely try using a trained mask for the attributions.
heh, unfortunately a single SAE is 768 * 60. The residual stream in GPT2 is 768 dims and SAEs are big. You probably want to test this out on smaller models.
I can’t recall the compute costs for that script, sorry. A couple of things to note:
For a single SAE you will need to run it on ~25k latents (46k minus the dead ones) instead of the 200 we did.
You will only need to produce explanations for activations, and won’t have to do the second step of asking the model to produce activations given the explanations.
It’s a fun idea. Though a serious issue is that your external LoRA weights are going to be very large because their input and output will need to be the same size as your SAE dictionary, which could be 10-100x (or more, nobody knows) the residual stream size. So this could be a very expensive setup to finetune.
Hey Matthew. We only did autointerp for 200 randomly sampled latents in each dict, rather than the full 60 × 768 = 46080 latents (although half of these die). So our results there wouldn’t be of much help for your project unfortunately.
Thanks a lot for letting us know about the dead links. Though note you have a “%20” in the second one which shouldn’t be there. It works fine without it.
I think the concern here is twofold:
Once a model is deceptive at one point, even if this happens stochastically, it may continue in its deception deterministically.
We can’t rely on future models being as stochastic w.r.t the things we care about, e.g. scheming behaviour.
Regarding 2, consider the trend towards determinicity we see for the probability that GPT-N will output a grammatically correct sentence. For GPT-1 this was low, and it has trended upwards towards determinicity with newer releases. We’re seeing a similar trend for scheming behaviour (though hopefully we can buck this trend with alignment techniques).
I plan to spend more time thinking about AI model security. The main reasons I’m not spending a lot of time on it now are:
I’m excited about the project/agenda we’ve started working on in interpretability, and my team/org more generally, and I think (or at least I hope) that I have a non-trivial positive influence on it.
I haven’t thought through what the best things to do would be. Some ideas (takes welcome):
Help create RAND or RAND-style reports like Securing AI Model Weights (I think this report is really great). E.g.
Make forecasts about how much interest from adversaries certain models are likely to get, and then how likely the model is to be stolen/compromised given that level of interest and the level defense of the developer. I expect this to be much more speculative than a typical RAND report. It might also require a bunch of non-public info on both offense and defense capabilities.
(not my idea) Make forecasts about how long a lab would take to implement certain levels of security.
Make demos that convince natsec people that AI is or will be very capable and become a top-priority target.
Improve security at a lab (probably requires becoming a full-time employee).
Thanks for the thoughts. They’ve made me think that I’m likely underestimating how much Control is needed to get useful work out of AIs capable and inclined to scheme. Ideally, this fact would increase the likelihood of other actors implementing AI Control schemes with the stolen model that are at least sufficient for containment and/or make them less likely to steal the model, though I wouldn’t want to put too much weight on this hope.
>This argument isn’t control specific, it applies to any safety scheme with some operating tax or implementation difficulty.[1][2]
Yep, for sure. I’ve changed the title and commented about this at the end.
In which worlds would AI Control (or any other agenda which relies on non-trivial post-training operation) prevent significant harm?
When I bring up the issue of AI model security to people working in AI safety, I’m often met with something of the form “yes, this is a problem. It’s important that people work hard on securing AI models. But it doesn’t really affect my work”.
Using AI Control (an area which has recently excited many in the field) as an example, I lay out an argument for why it might not be as effective an agenda as one might think after considering the realities of our cyber security situation.
AI Control concerns itself with models that intentionally try to subvert its developers.
These models are likely to be very generally capable and capable of causing significant harm without countermeasures.
Leading cyber-capable institutions would likely expend significant resources and political capital to steal these models or steal enough insights to reproduce such models.
If the weights or insights are stolen, work on AI control will not prevent these models from causing significant harm.
Current AI developers are not on track to be able to defend against high-priority operations from leading cyber-capable institutions in the coming years.
Therefore, AI control will only be useful in the coming years under one (or more) of these conditions:
Models that scheme are unlikely to be generally capable/dangerous enough to be a high-priority target for leading cyber-capable institutions.
Models that scheme are only developed by actors that can thwart high-priority operations from leading cyber-capable institutions (which precludes current AI developers for at least several years).
AI Control won’t be directly useful in the coming years but it will be indirectly useful to progress the field for when models are developed by actors capable of thwarting top cyber operations.
Even if the model was stolen and caused significant harm, there would still be less harm overall than if the model undergoing AI control measures also broke out. Of course, this argument is much weaker for models capable of existential harm.
Actors that steal the model would also implement strong-enough AI control measures and would not misuse the AI to cause significant harm.
There are of course other arguments against working on AI control. E.g. it may encourage the development and use of models that are capable of causing significant harm. This is an issue if the AI control methods fail or if the model is stolen. So one must be willing to eat this cost or argue that it’s not a large cost when advocating for AI Control work.
This isn’t to say that AI Control isn’t a promising agenda, I just think people need to carefully consider the cases in which their agenda falls down for reasons that aren’t technical arguments about the agenda itself.
I’m also interested to hear takes from those excited by AI Control on which conditions listed in #6 above that they expect to hold (or to otherwise poke holes in the argument).
EDIT (thanks Zach and Ryan for bringing this up): I didn’t want to imply that AI Control is unique here, this argument can be levelled at any agenda which relies on something like a raw model + non-trivial operation effort. E.g. a scheme which relies on interpretability or black box methods for monitoring or scalable oversight.
They are indeed all hook_resid_pre. The code you’re looking at just lists a set of positions that we are interested in viewing the reconstruction error of during evaluation. In particular, we want to view the reconstruction error at hook_resid_post of every layer, including the final layer (which you can’t get from hook_resid_pre).
Here’s a wandb report that includes plots for the KL divergence. e2e+downstream indeed performs better for layer 2. So it’s possible that intermediate losses might help training a little. But I wouldn’t be surprised if better hyperparams eliminated this difference; we put more effort into optimising the SAE_local hyperparams rather than the SAE_e2e and SAE_e2e+ds hyperparams.
Very well articulated. I did a solid amount of head nodding while reading this.
As you appear to be, I’m also becoming concerned about the field trying to “cash in” too early too hard on our existing methods and theories which we know have potentially significant flaws. I don’t doubt that progress can be made by pursuing the current best methods and seeing where they succeed and fail, and I’m very glad that a good portion of the field is doing this. But looking around I don’t see enough people searching for new fundamental theories or methods that better explain how these networks actually do stuff. Too many eggs are falling in the same basket.
I don’t think this is as hard a problem as the ones you find in Physics or Maths. We just need to better incentivise people to have a crack at it, e.g. by starting more varied teams at big labs and by funding people/orgs to pursue non-mainline agendas.
Thanks for prediction. Perhaps I’m underestimating the amount of shared information between in-context tokens in real models. Thinking more about it, as models grow, I expect the ratio of contextual information which is shared across tokens in the same context to more token-specific things like part of speech to increase. Obviously a bigram-only model doesn’t care at all about the previous context. You could probably get a decent measure of this just by comparing cosine similarities of activations within context to activations from other contexts. If true, this would mean that as models scale up, you’d get a bigger efficiency hit if you didn’t shuffle when you could have (assuming fixed batch size).
Thanks Leo, very helpful!
The right way to frame this imo is the efficiency loss from not shuffling, which from preliminary experiments+intuition I’d guess is probably substantial.
The SAEs in your paper were trained with batch size of 131,072 tokens according to appendix A.4. Section 2.1 also says you use a context length of 64 tokens. I’d be very surprised if using 131,072⁄64 blocks of consecutive tokens was much less efficient than 131,072 tokens randomly sampled from a very large dataset. I also wouldn’t be surprised if 131,072/2048 blocks of consecutive tokens (i.e. a full context length) had similar efficiency.
Were your preliminary experiments and intuition based on batch sizes this large or were you looking at smaller models?
I missed that appendix C.1 plot showing the dead latent drop with tied init. Nice!
I don’t think this raises a challenge to physicalism.
If physicalism were true, or even if there were non-physical things but they didn’t alter the determinism of the physical world, then the notion of an “agent” needs a lot of care. It can easily give the mistaken impression that there is something non-physical in entities that can change the physical world.
A perfect world model would be able to predict the responses of any neuron in any location in the universe to any inputs (leaving aside true randomness). It doesn’t matter whether the entity in question has a conscious experience that it is one of those entities, nothing would change.
So I’d argue that this question is irrelevant if physicalism is true, because the AI having a phenomenal conscious experience of “I am this entity” cannot affect the physical world. If we’re not talking about phenomenal consciousness, then it’s just regular physical world modeling.