Posts
Comments
Re: hidden messages in neuron explanations, yes it seems like a possible problem. A way to try to avoid this is to train the simulator model to imitate what a human would say given the explanation. A human would ignore the coded message, and so the trained simulator model should also ignore the coded message. (this maybe doesn't account for adversarial attacks on the trained simulator model, so might need ordinary adversarial robustness methods).
Does seem like if you ever catch your interpretability assistant trying to hide messages, you should stop and try to figure out what is going on, and that might be sufficient evidence of deception.
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, and maybe this would make these methods more effective. By default I think the tuned lens learns only the transformation needed to predict the output token but the method could be adapted to retrodict the input token from each layer as well, we’d need both. Code for tuned lens is at https://github.com/alignmentresearch/tuned-lens
(I work at OpenAI). Is the main thing you think has the effect of safetywashing here the claim that the misconceptions are common? Like if the post was "some misconceptions I've encountered about OpenAI" it would mostly not have that effect? (Point 2 was edited to clarify that it wasn't a full account of the Anthropic split.)
Jan Leike has written about inner alignment here https://aligned.substack.com/p/inner-alignment. (I'm at OpenAI, imo I'm not sure if this will work in the worst case and I'm hoping we can come up with a more robust plan)
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 problem" and not "100% doomed". Or at least the failure story needs to involve more steps like "sure critiques will tell us that the fusion power generator will lead to everyone dying, but we ignore that because it can write a critique of any action that makes us believe it's bad" or "the consequences are so complicated the system can't explain them to us in the critique and get high reward for it"
ETA: So I'm assuming the story for feedback on reliably doing things in the world you're referring to is something like "we give the AI feedback by letting it build fusion generators and then giving it a score based on how much power it generates" or something like that, and I agree this is easier than "are we actually happy with the outcome"
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?
Would be good to have examples that include the relevant Info Constraints: something that just retrieves things Eliezer has said seems less useful than something that can come up with things Eliezer would say based on having the same information available.
Would prefer to have fully written examples for this (e.g. how would someone who thought "compress sensory information" was a good objective function describe it to the critic?)
This feels like too specific a task/less generally useful to AI alignment research than your proposal on "Extract the the training objective from a fully-trained ML model"
I think it's fine to have tasks that wouldn't work for today's language models like those that would require other input modalities. Would prefer to have fully specified inputs but these do seem easy to produce in this case. Would be ideal if there were examples with a smaller input size though.
Potentially interesting task, would prefer examples in the AI alignment domain.
Potentially interesting task, would prefer examples in the AI alignment domain.
Seems like a reasonable task, but wonder if it would be easier in practice to just have a wiki or something like https://metacademy.org/ or get post authors to do this (mostly depends on the size of the graph of concepts you need to connect, if it's smaller makes sense for people to do it, if it's larger then maybe automation helps).
Estimates 0.5% (346,000) of UK population reports long covid symptoms that limit their day-to-day activities "a lot" (possibly too high if other stuff mistaken as long covid?)
Another article (that I didn't look at beyond the headline) suggests 7/10 people have had covid https://www.theguardian.com/world/2022/apr/22/seven-in-10-people-in-england-have-had-covid-research-shows-omicron
Together seems to suggest risk less than 1% of the worst kind of it.
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 impossible situations where they differ, and the science fiction novels are different in these two worlds.
Is ELK required to answer appropriately in both worlds? (answer Q1 when given Q in W1, and Q2 when given Q in W2)? If so, it seems we need some term in the loss outside of the training set to make this happen.
Alternatively, would it be satisfactory to find a solution that doesn't discriminate what's world it is in, and instead returns "yes" to Q if and only if Q1="yes" AND Q2="yes"? This means that in world W1 there will be some situations where Q="no" when the diamond is present, but no situations where Q="yes" and the diamond is not present.
Are there other useful things that could be funded to get more evidence?
One thing that might be possible would be funding a larger survey, maybe with a more random sample of the population.
Is there any work going on that seems at all likely to decrease risk of getting long covid conditional on getting infected, beyond current vaccines? E.g. alternative vaccines with a story to suggest why they might decrease risk, evidence of Paxlovid or other treatments decreasing risk, other targeted research programs
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 actually care about)
- C. Do photons bounce off of the diamond at time t
- D. Are photons in transit between the diamond and the camera at time t+1
- E. Do photons arrive in the camera at time t+2
- F. Do photons trigger the camera’s electronics at time t+3
- G. Does the human correctly interpret the signal from the camera
- H. Does the human believe “the diamond is safe”
- …
- The reporter can get low loss if it has a fast way to predict that value of any of these variables. This is fine on the training set, because the variables are all perfectly correlated.
- But it might be possible to create ways for some of these variables to hold, but not others off the training set. The thing that is going wrong in Counterexample: “what the human will believe” can be understood before “what is actually true” seems like “E through H are true in this case, but A-D are false”.
Possible approach:
- Instead, we could try to learn a reporter that returns true if and only if all relevant variables A-H are true
- Then the remaining problem should just be that this reporter is overly conservative: there might be some strategies that would be helpful that involve changing some of these variables, e.g. moving the diamond around the room (changing A) or turning off the lights in the room (changing C-H, which even causes the human to believe that the diamond isn't safe)
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 performance of M1 than M on A would demonstrated misalignment (don't have a precise definition of what would make finetuning minimal in this way, would be good to have a clearer criteria for that).
If M1 does better on B after finetuning in a way that implicitly demonstrates better knowledge of A, but does not do better on A when asked to do it explicitly, that would demonstrate that the finetuned M1 is misaligned (I think we might expect some version of this to happen by default though, since M1 might overfit to only doing tasks of type B. Maybe if you have a training procedure where M1 generally doesn't get worse at any tasks then I might hope that it would get better on A and be disappointed if it doesn't).
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't quite solving the whole problem, because ultimately we want systems that help humans supervise tasks where they haven't developed the right concepts, or couldn't understand them even with years of study.
Here is a regularly updated version of the vaccine chart https://covid19tracker.ca/vaccinegap.html
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 (as in the recent OpenAI debate work), such that the debate AI can tackle more complicated questions with steps that are restricted to the complexity that humans can currently work with. Hard to know if this is possible, but still seems worth trying to work on.
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 the point where it can help on research, because we want to do work ourselves to be able to check solutions that AIs reach, etc.)
It also seems like areas like FairML and Computational Social Choice will require preference learning as components - my guess is that people's exact preferences about fairness won't have a simple mathematical formulation, and will instead need to be learned. I could buy the position that the necessary progress in preference learning will happen by default because of other incentives.
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.
It seems that in principle a version of debate where only one agent makes statements and the other chooses which statements to expand could work, but it seems like it requires the judge to be very strict that the statement is 100% true. It seems hard to apply this kind of system to statements outside of formal mathematics.
Systems where both agents can make statements seem like they might be less vulnerable to judges accepting statements that aren't 100% true. For one example, if both agents take turns being the arguer, then if both agents submit a path that is judged to be correct, you can stipulate that the agent with the shortest path wins (like imposing a simplicity prior).
HCH could implement the decomposition oracle by searching over the space of all possible decompositions (it would just be quite expensive).
https://www.kialo.com/ lets people build debates on controversial topics in a heirarchical structure (more like stock debate, with both sides providing arguments), but doesn't seem to have been used for explanations/arguments. I'd also be pretty interested to see more attempts at heirarchical explanations.
I think there are situations where you can still have subproblems where the output of the subproblem is long. A contrived example: suppose you have a problem where you want to calculate XOR(f(a), f(b)), where f(a) and f(b) are long strings. It seems reasonable to decompose into x=f(a), y=f(b), z=XOR(x, y), despite x and y being long, because there's a simple way to combine them.
If we had an AI system that could work on "making progress on a problem for an hour", then write down a complete description of everything it had figured out and pass that to another AI system, I'd count that as dividing the problem into subproblems, just in a way that's probably inefficient.
I'd evaluate decompositions into subproblems by something like the total cost of solving a problem by dividing it into subproblems. Some decompositions would be efficent and others would be inefficient, sometimes this would be because the output is large but in other cases it could be because it takes a long time to write the input, or because there's a lot of work repeated between subproblems.
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.
https://institute.global/policy/long-covid-reviewing-science-and-assessing-risk
From the COVID Symptom Study in the UK (app based questionaire), "10 per cent of those taking part in the survey had symptoms of long Covid for a month, with between 1.5 and 2 per cent still experiencing them after three months", and they claim "long Covid is likely a bigger issue than excess deaths as a result of Covid, which are between 0.5 per cent and 1 per cent".
App-based survey, so not necessarily representative of population. Not clear how severe the 3 month cases are, though they state "The most common reported symptom has been described by doctors as “profound fatigue”". Article also summarizes other related studies.
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 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
I'm not sure what "just ask the human to verify the answer given " looks like, for implicitly represented
I'm skeptical of this.
- Wuhan needed 2 months on lockdown: https://en.wikipedia.org/wiki/COVID-19_pandemic_lockdown_in_Hubei
- I'd expect that imposing China-style lockdowns in the West would require significant force and might end up causing a large-scale panic in and of itself.
- I'd expect that any lockdown in the West wouldn't have been effective enough to stamp out 100% of cases, and if you don't eradicate it then you need ongoing measures or it will just flare up again later, so one strictly enforced lockdown wouldn't cut it. (Though maybe you could do very rigorous contact tracing and lock down just people who might have been in contact with cases, which could be less costly than full lockdown but probably still need significant enforcement).
https://www.microcovid.org/paper/2-riskiness#fn6 discusses https://covid.joinzoe.com/post/covid-long-term which has an app-based survey claiming 1 in 10 people still have symptoms after 3 weeks. (but since people can just sign up for the app I'd guess this is harder to know how to interpret than the telephone survey). Microcovid.org uses this 1 in 10 figure as the estimate for chance of some ongoing health consequence, and claims the risk of ongoing health problems from a 1% chance of COVID is equivalent to the risk from 1 year of driving (but this comparison involves even more assumptions).
https://www.cdc.gov/mmwr/volumes/69/wr/mm6930e1.htm found that ~1 in 5 of 18-34 year olds with no underlying health conditions had symptoms 3 weeks later (telephone survey of people who'd been symptomatic and had a positive test).
Other discussion in comments of https://www.lesswrong.com/posts/ahYxBHLmG7TiGDqxG/do-we-have-updated-data-about-the-risk-of-permanent-chronic
Not addressing fatigue, and just a study in progress, but this study is looking for long term neurological problems, might another weak bit of evidence when it releases results https://www.cambridgebrainsciences.com/studies/covid-brain-study
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)