All of William_S's Comments + Replies

From discussion with Logan Riggs (Eleuther) who worked on the tuned lens: the tuned lens suggests that the residual stream at different layers go through some linear transformations and so aren’t directly comparable. This would interfere with a couple of methods for trying to understand neurons based on weights: 1) the embedding space view 2) calculating virtual weights between neurons in different layers.

However, we could try correcting these using the transformations learned by the tuned lens to translate between the residual stream at different layers, ... (read more)

So I do think you can get feedback on the related question of "can you write a critique of this action that makes us think we wouldn't be happy with the outcomes" as you can give a reward of 1 if you're unhappy with the outcomes after seeing the critique, 0 otherwise.

And this alone isn't sufficient, e.g. maybe then the AI system says things about good actions that make us think we wouldn't be happy with the outcome, which is then where you'd need to get into recursive evaluation or debate or something. But this feels like "hard but potentially tractable pr... (read more)

If we can't get the AI to answer something like "If we take the action you just proposed, will we be happy with the outcomes?", why can we get it to also answer the question of "how do you design a fusion power generator?" to get a fusion power generator that does anything reliably in the world (including having consequences that kill us), rather than just getting out something that looks to us like a plan for a fusion generator but doesn't actually work?

Image link is broken

Yep, that clarifies.

You define robustness to scaling down as "a solution to alignment keeps working if the AI is not optimal or perfect." but for interpretability you talk about "our interpretability is merely good or great, but doesn't capture everything relevant to alignment" which seems to be about the alignment approach/our understanding being flawed not the AI. I can imagine techniques being robust to imperfect AI but find it harder to imagine how any alignment approach could be robust if the approach/our implementation of the approach itself is flawed, do you have any example of this?

Summary for 8 "Can we take a deceptively aligned model and train away its deception?" seems a little harder than what we actually need, right? We could prevent a model from being deceptive rather than trying to undo arbitrary deception (e.g. if we could prevent all precursors)

Do you think we could basically go 1->4 and 2->5 if we could train a helper network to behaviourally clone humans using transparency tools and run the helper network over the entire network/training process? Or if we do critique style training (RL reward some helper model with access to the main model weights if it produces evidence of the property we don't want the main network to have)?

Could I put in a request to see a brain dump from Eliezer of ways to gain dignity points?

Take after talking with Daniel: for future work I think it will be easier to tell how well your techniques are working if you are in a domain where you care about minimizing both false-positive and false-negative error, regardless of whether that's analagous to the long term situation we care most about. If you care about both kinds of error then the baseline of "set a reallly low classifier threshold" wouldn't work, so you'd be starting from a regime where it was a lot easier to sample errors, hence it will be easier to measure differences in performance.

One general piece of advice is that it seems like it might be useful to have an interface that shows you multiple samples for each prompt (the OpenAI playground just gives you one sample, if you use temperature > 0 then this sample could either be lucky or unlucky)

Maybe useful way to get feedback on how good you are at doing this would be trying to make predictions based on your experience with language models:

• without looking at the results or running systematic experiments on the dataset, predict which tasks on BIG-bench will be doable
• make bets of the form "we'll reach X% performance on task A before we reach Y% performance on task B"
• predict for some prompt what percentage of samples will satisfy some property, then take a large number of samples and then rate them

Evaluation assistance as mentioned in the post on AI-assisted human feedback could help people avoid being fooled (e.g. in debate where the opponent can point out how you're being fooled). It's still an open question how well that will work in practice and how quickly it will Goodhart (these techniques should fail on some things, as discussed in the ELK report), but it seems possible that models will be helpful enough on alignment before they Goodhart.

Suppose there are two worlds, world W1 and world W2.

In world W1, the question Q="Is there a diamond in the room?" is commonly understood to mean Q1="Is there actually a diamond in the room?"

In world W2 the question Q="Is there a diamond in the room?" is commonly understood to mean Q2="Do I believe there is a diamond in the room?"

Both worlds don't know how to construct a situation where these are different. So, they produce identical training sets for ELK. But the simulator is also trained on a bunch of science fiction novels that contain descriptions of im... (read more)

Edit: think this isn't quite right in general, will try to make it more correct later

Here's a sketch of a strategy for trying to fix Strategy: penalize depending on “downstream” variables. Would appreciate feedback on whether it's modeling the difficulty correctly/seems possibly worth figuring out how to implement

It seems like the problem is:

• On the training set, there are a number of implicit variables X that are indistinguishable (always all true or always all false)
  • A. Is the diamond safe at time t-1
  • B. Is the diamond safe at time t (the variable we actual

I've been thinking of Case 2. It seems harder to establish "capable of distinguishing between situations where the user wants A vs B" on individual examples since a random classifier would let you cherrypick some cases where this seems possible without the model really understanding. Though you could talk about individual cases as examples of Case 2. Agree that there's some implicit "all else being equal" condition, I'd expect currently it's not too likely to change conclusions. Ideally you'd just have the category A="best answer according to user" B="all answers that are worse than the best answer according to the user" but I think it's simpler to analyze more specific categories.

Link to contractor instructions implied in "You can read the instructions given to our contractors here" is missing.

I don't think all work of that form would measure misalignment, but some work of that form might, here's a description of some stuff in that space that would count as measuring misalignment.

Let A be some task (e.g. add 1 digit numbers), B be a task that is downstream of A (to do B, you need to be able to do A, e.g. add 3 digit numbers), M is the original model, M1 is the model after finetuning.

If the training on a downstream task was minimal, so we think it's revealing what the model knew before finetuning rather than adding knew knowledge, then better per... (read more)

Even better than "Getting models to explain why they’re doing what they’re doing in simpler terms that connect to things the human overseers understand" would be getting models to actually do the task in ways that are simpler and connect to things that human overseers understand. E.g. if a model can solve a task in multiple steps by looking up relevant information by doing internet searches that are recorded and readable by the overseer instead of using knowledge opaquely measured in the weights, that seems like a step in the right direction.

One easy way to make people who can't solve the task for sandwiching is to take people who could solve the task and then give them insufficient time to solve it, or have them be uninformed of some relevant facts about the specific task they are trying to solve.

A simpler way to measure whether you are making progress towards sandwiching if you can't go there directly is to look at whether you can get people to provide better supervision with your tool than without your tool, that is accomplishing more on the task.

Both of these approaches feel like they aren... (read more)

If the High-Rated Sentence Producer was restricted to output only single steps of a mathematical proof and the single steps were evaluated independently, with the human unable to look at previous steps, then I wouldn't expect this kind of reward hacking to occur. In math proofs, we can build proofs for more complex questions out of individual steps that don't need to increase in complexity.

As I see it, debate on arbitrary questions could work if we figured out how to do something similar, having arguments split into single steps and evaluated independently... (read more)

For the preference learning skepticism, does this extend to the research direction (that isn't yet a research area) of modelling long term preferences/preferences on reflection? This is more along the lines of the "AI-assisted deliberation" direction from ARCHES.

To me it seems like AI alignment that can capture preferences on reflection could be used to find solutions to many of other problems. Though there are good reasons to expect that we'd still want to do other work (because we might need theoretical understanding and okay solutions before AI reaches ... (read more)

One thing I'd like to see are some more fleshed out examples of the kinds of governance demands that you think might be important in the future and would be bottlenecked on research progress in these areas.

Okay, makes more sense now, now my understanding is that for question X, answer from ML system Y,  amplification system A, verification in your quote is asking the A to answer "Would A(Z) output answer Y to question X?", as opposed to asking A to answer "X", and then checking if it equals "Y". This can at most be as hard as running the original system, and maybe could be much more efficient.

Right, but in the post the implicitly represented Z is used by an amplification or debate system, because it contains more information than a human can quickly read and use (so are you assuming it's simple to verify the results of amplification/debate systems?)

for extremely large $Z$ which are represented only implicitly as in Paul's post, we might not always check whether the model matches the ground truth by actually generating the ground truth and instead just ask the human to verify the answer given $Z$

I'm not sure what "just ask the human to verify the answer given $Z$" looks like, for implicitly represented $Z$

Seems like it could be helpful if people who've thought about this would also predict on the question of what the survey value would be today. (e.g. via elicit snapshots)

Google's Meena (2.6 billion parameters, February 2020) creates original puns in "Conversation D", and I think "Cross-turn Repetition Example 2" is absurdist comedy, but maybe more as a result of the way the model fails.

Conversation D

Human: I like how cute cows are

Meena: Ah, interesting. You like how they are cute, right?

Human: yeah, and they’re realy smart

Meena: I heard they go to college

Human: cows go to college?

Meena: I heard that a cow went to Harvard.

Human: what did the cow study?

Meena: Bovine sciences.

Human: do horses g

The worry I'd have about this interpretability direction is that we become very good at telling stories about what 95% of the weights in neural networks do, but the remaning 5% hides some important stuff, which could end up including things like mesa-optimizers or deception. Do you have thoughts on that?

I'm talking about an imitation version where the human you're imitating is allowed to do anything they want, including instatiting a search over all possible outputs X and taking that one that maximizes the score of "How good is answer X to Y?" to try to find X*. So I'm more pointing out that this behaviour is available in imitation by default. We could try to rule it out by instructing the human to only do limited searches, but that might be hard to do along with maintaining capabilities of the system, and we need to figure out what "safe limited search" actually looks like.

If M2 has adversarial examples or other kinds of robustness or security problems, and we keep doing this training for a long time, wouldn't the training process sooner or later sample an X that exploits M2 (gets a high reward relative to other answers without actually being a good answer), which causes the update step to increase the probability of M1 giving that output, and eventually causes M1 to give that output with high probability?

I agree, and think that this problem occurs both in imitation IA and RL IA

For example is the plan to make sure M2 ha

Overall, I think imitative amplification seems safer, but I maybe don't think the distinction is as clear cut as my impression of this post gives.

if you can instruct them not to do things like instantiate arbitrary Turing machines

I think this and "instruct them not to search over arbitrary text strings for the text string that gives the most approval", and similar things, are the kind of details that would need to be filled out to make the thing you are talking about actually be in a distinct class from approval-based amplification and deba... (read more)

Possible source for optimization-as-a-layer: SATNet (differentiable SAT solver)

https://arxiv.org/abs/1905.12149

One way to try to measure capability robustness seperate from alignment robustness off of the training distribution of some system would be to:

• use an inverse reinforcement learning algorithm infer the reward function of the off-distribution behaviour
• train a new system to do as well on the reward function as the original system
• measure the number of training steps needed to reach this point for the new system.

This would let you make comparisons between different systems as to which was more capability robust.

Maybe there's a version that could train the ... (read more)

For Alaska vs. Bali, alternative answer is "You could be convinced that either Alaska or Bali is a good vacation destination". It's an interesting question whether this could actually win in debate. I think it might have a better chance in Factored Evaluation, because we can spin up two seperate trees to view the most compelling argument for Alaska and the most compelling argument for Bali and verify that these are convincing. In debate, you'd need view either Alaska Argument before Bali Argument, or Bali Argument before Alaska Argument, and you might just be convinced by the first argument you see in which case you wouldn't agree that you could be convinced either way.

I'd say that the claim is not sufficient - it might provide some alignment value, but it needs a larger story about how the whole computation is going to be safe. I do think that the HCH framework could make specifying an aligned GOFAI-like computation easier (but it's hard to come up with a rigorous argument for this without pointing to some kind of specification that we can make claims about, which is something I'd want to produce along the way while proceeding with HCH-like approaches)

I think a cleaner way of stating condition 3 might be "there exists some policy with a non-zero chance of deception at some point in the future, with expected value for the mesa-optimizer higher than the policy of always cooperating".

This could be met if the threat of modification never goes away, but maybe decreases over time. (at training time, system is monitored for defection every episode, at a deployment time it is monitored for defection 1/million episodes, not worthwhile to defect during training but worthwhile to defect at deployment tim... (read more)

To me, It seems like the point of this story is that we could build an AI that ends up doing very dangerous things without ever asking it "Will you do things I don't like if given more capability?" or some other similar question that requires it to execute the treacherous turn. In contrast, if the developers did something like build a testing world with toy humans in it who could be manipulated in a way detectable to the developers, and placed the AI in the toy testing world, then it seems like this AI would be forced into a position where i... (read more)

Are you interested in protocols involving multiple episodic questions (where you ask one question, wait for it to resolve, then ask another question?)

Submission: low-bandwidth oracle

Plan Criticism: Given plan to build an aligned AI, put together a list of possible lines of thought to think about problems with the plan (open questions, possible failure modes, criticisms, etc.). Ask the oracle to pick one of these lines of thought, pick another line of thought at random, and spend the next time period X thinking about both, judge which line of thought was more useful to think about (where lines of thought that spot some fatal missed problem are judged to be very useful) and reward the oracle if its suggestion was picked.

• AI systems end up controlled by a group of humans representing a small range of human values (ie. an ideological or religious group that imposes values on everyone else). While not caused only by AI design, it is possible that design decisions could impact the likelihood of this scenario (ie. at what point are values loaded into the system/how many people's values are loaded into the system), and is relevant for overall strategy.

• Failure to learn how to deal with alignment in the many-humans, many-AIs case even if single-human, single-AI alignment is solved (which I think Andrew Critch has talked about). For example, AIs negotiating on behalf of humans take the stance described in https://arxiv.org/abs/1711.00363 of agreeing to split control of the future according to which human's priors are most accurate (on potentially irrelevant issues) if this isn't what humans actually want.

I guess it feels like I don't know how we could know that we're in the position that we've "solved" meta-philosophy. It feels like the thing we could do is build a set of better and better models of philosophy and check their results against held-out human reasoning and against each other.

I also don't think we know how to specify a ground truth reasoning process that we could try to protect and run forever which we could be completely confident would come up with the right outcome (where something like HCH is a good candidate ... (read more)

Yeah, to some extent. In the Lookup Table case, you need to have a (potentially quite expensive) way of resolving all mistakes. In the Overseer's Manual case, you can also leverage humans to do some kind of more robust reasoning (for example, they can notice a typo in a question and still respond correctly, even if the Lookup Table would fail in this case). Though in low-bandwidth oversight, the space of things that participants could notice and correct is fairly limited.

Though I think this still differs from HRAD in that it seems like the output of H... (read more)

Thanks, this position makes more sense in light of Beyond Astronomical Waste (I guess I have some concept of "a pretty good future" that is fine with something like a bunch of human-descended beings living a happy lives that misses out on the sort of things mentioned in Beyond Astronomical Waste, and "optimal future" which includes those considerations). I buy this as an argument that "we should put more effort into making philosophy work to make the outcome of AI better, because we risk losing large amounts of value" rather t... (read more)

One important dimension to consider is how hard it is to solve philosophical problems well enough to have a pretty good future (which includes avoiding bad futures). It could be the case that this is not so hard, but fully resolving questions so we could produce an optimal future is very hard or impossible. It feels like this argument implicitly relies on assuming that "solve philosophical problems well enough to have a pretty good future" is hard (ie. takes thousands of millions of years in scenario 4) - can you provide further clarification on whether/why you think that is the case?

Shorter name candidates:

Inductively Aligned AI Development

Inductively Aligned AI Assistants

It's a nice property of this model that it prompts consideration of the interaction between humans and AIs at every step (to highlight things like risks of the humans having access to some set of AI systems for manipulation or moral hazard reasons).

I guess what we're trying to unpack is "the mechanism that makes decisions from that database", and whether it can be efficient. If you store all experience with no editing, you're left with the same problem of "what decisions in the tree do you change based on experience (X,y) and how do you change them?" And it seems like the most naive way of doing this is too inefficient (looking at all data at all decision nodes).

If you summarize experience (what I'm interpreting "decide how to update some small sketch" as ... (read more)

What if the current node is responsible for the error instead of one of the subqueries, how do you figure that out?

I think you'd need to form the decomposition in such a way that you could fix any problem through perturbing something in the
world representation (an extreme version is you have the method for performing every operation contained in the world representation and looked up, so you can adjust it in the future).

When you do backprop, you propagate the error signal through all the nodes, not just through a single path that is "most respo