Posts
Comments
Attribution can identify when system prompts are affecting behaviour. 
Yeah! So, hierarchical perturbation (HiPe) is a bit like a thresholded binary search. It starts by splitting the input into large overlapping chunks and perturbing each of them. If the resulting attributions for any of the chunks are above a certain level, those chunks are split into smaller chunks and the process continues. This works because it efficiently discards input regions that don't contribute much to the output, without having to individually perturb each token in them.
Standard iterative perturbation (ItP) is much simpler. It just splits the inputs into evenly sized chunks, perturbs each of them in turn to get the attributions, and that's that. We do this either word-wise or token-wise (word-wise is about 25% quicker).
So, where n=number of tokens in the prompt and O(1) is the cost of a single completion, ItP is O(n) if we perturb token-wise, or O(0.75n) if word-wise, depending on how many tokens per word your tokeniser gives you on average. This is manageable but not ideal. You could, of course, always perturb iteratively in multi-token chunks, at the cost of attribution granularity.
HiPe can be harder to predict, as it really depends on the initial chunk size and threshold you use, and the true underlying saliency of the input tokens (which naturally we don't know). In the worst case with a threshold of zero (a poor choice), an initial chunk size of n and every token being salient, you might end up with O(4n) or more, depending on how you handle overlaps. In practice, with a sensible threshold (we use the mid-range, which works well out of the box) this is rare.
HiPe really shines on large prompts, where only a few tokens are really important. If a given completion only really relies on 10% of the input tokens, HiPe will give you attributions in a fraction of n.
I don't want to make sweeping claims about HiPe's efficiency in general, as it relies on the actual saliency of the input tokens. Which we don't know. Which is why we need HiPe! We'd actually love to see someone do a load of benchmark experiments using different configurations to get a better handle on this, if anyone fancies it.
Thanks for the comment! I'll respond to the last part:
"First, developing basic insights is clearly not just an AI safety goal. It's an alignment/capabilities goal. And as such, the effects of this kind of thing are not robustly good."
I think this could certainly be the case if we were trying to build state of the art broad domain systems, in order to use interpretability tools with them for knowledge discovery – but we're explicitly interested in using interpretability with narrow domain systems.
"Interpretability is the backbone of knowledge discovery with deep learning": Deep learning models are really good at learning complex patterns and correlations in huge datasets that humans aren't able to parse. If we can use interpretability to extract these patterns in a human-parsable way, in a (very Olah-ish) sense we can reframe deep learning models as lenses through which to view the world, and to make sense of data that would otherwise be opaque to us.
Here are a couple of examples:
https://www.mdpi.com/2072-6694/14/23/5957
https://www.deepmind.com/blog/exploring-the-beauty-of-pure-mathematics-in-novel-ways
https://www.nature.com/articles/s41598-021-90285-5
Are you concerned about AI risk from narrow systems of this kind?
Thanks! Unsure as of yet – we could either keep it proprietary and provide access through an API (with some free version for select researchers), or open source it and monetise by offering a paid, hosted tier with integration support. Discussions are ongoing.
This isn't set in stone, but likely we'll monetise by selling access to the interpretability engine, via an API. I imagine we'll offer free or subsidised access to select researchers/orgs. Another route would be to open source all of it, and monetise by offering a paid, hosted version with integration support etc.
We're looking into it!
Good questions. Doing any kind of technical safety research that leads to better understanding of state of the art models carries with it the risk that by understanding models better, we might learn how to improve them. However, I think that the safety benefit of understanding models outweighs the risk of small capability increases, particularly since any capability increase is likely heavily skewed towards model specific interventions (e.g. "this specific model trained on this specific dataset exhibits bias x in domain y, and could be improved by retraining with more varied data from domain y", rather than "the performance of all of models of this kind could be improved with some intervention z"). I'm thinking about this a lot at the moment and would welcome further input.
Aha!! Thanks Neel, makes sense. I’ll update the post
Yeah! Basically we just perform gradient descent on sensibly initialised embeddings (cluster centroids, or points close to the target output), constrain the embeddings to length 1 during the process, and penalise distance from the nearest legal token. We optimise the input embeddings to maximise the -log prob of the target output logit(s). Happy to have a quick call to go through the code if you like, DM me :)
This link: https://help.openai.com/en/articles/6824809-embeddings-frequently-asked-questions says that token embeddings are normalised to length 1, but a quick inspection of the embeddings available through the huggingface model shows this isn't the case. I think that's the extent of our claim. For prompt generation, we normalise the embeddings ourselves and constrain the search to that space, which results in better performance.
Thanks - wasn't aware of this!
Interesting! Can you give a bit more detail or share code?
Interesting, thanks. There's not a whole lot of detail there - it looks like they didn't do any distance regularisation, which is probably why they didn't get meaningful results.
I'll check with Matthew - it's certainly possible that not all tokens in the "weird token cluster" elicit the same kinds of responses.
What's an SCP?
Not yet, but there's no reason why it wouldn't be possible. You can imagine microscope AI, for language models. It's on our to-do list.
Good to know. Thanks!
Yep, aside from running forward prop n times to generate an output of length n, we can just optimise the mean probability of the target tokens at each position in the output - it's already implemented in the code. Although, it takes way longer to find optimal completions.
More detail on this phenomenon here: https://www.lesswrong.com/posts/aPeJE8bSo6rAFoLqg/solidgoldmagikarp-plus-prompt-generation
Yeah, I think it could be! I’m considering pursuing it after SERI-MATS. I’ll need a couple of cofounders.
Hi Joseph! I'll briefly address the saliency map concern here – it likely originates from this paper, which showed that some types of saliency mapping methods had no more explanatory power than edge detectors. It's a great paper, and worth a read. The key thing to note is that this was only true of some gradient-based saliency mapping methods, which are, of course, model-specific. Gradients can be deceptive! Model agnostic, perturbation-based saliency mapping doesn't suffer from the same kind of problems – see p.12 here.
“being able to reorganise a question in the form of a model-appropriate game” seems like something we already have built a set of reasonable heuristics around - categorising different types of problems and their appropriate translations into ML-able tasks. There are well established ML approaches to, e.g. image captioning, time-series prediction, audio segmentation etc etc. is the bottleneck you’re concerned with the lack of breadth and granularity of these problem-sets, OP - and we can mark progress (to some extent) by the number of these problem sets we have robust ML translations for?