All of Jacob_Hilton's Comments + Replies
For the informal no-coincidence principle, it's important to us (and to Gowers IIUC) that a "reason" is not necessarily a proof, but could instead be a heuristic argument (in the sense of this post). We agree there are certainly apparently outrageous coincidences that may not be provable, such as Chebyshev's bias (discussed in the introduction to the post). See also John Conway's paper On Unsettleable Arithmetical Problems for a nice exposition of the distinction between proofs and heuristic arguments (he uses the word "probvious" for a statement with a co...
It's not, but I can understand your confusion, and I think the two are related. To see the difference, suppose hypothetically that 11% of the first million digits in the decimal expansion of were 3s. Inductive reasoning would say that we should expect this pattern to continue. The no-coincidence principle, on the other hand, would say that there is a reason (such as a proof or a heuristic argument) for our observation, which may or may not predict that the pattern will continue. But if there were no such reason and yet the pattern continued, th...
1
1Good question! We also think that NP ≠ co-NP. The difference between 99% (our conjecture) and 100% (NP = co-NP) is quite important, essentially because 99% of random objects "look random", but not 100%. For example, consider a uniformly random string for some large . We can quite confidently say things like: the number of 0s in is between and ; there is no streak of alternating 0s and 1s; etc. But these only hold with 99% confidence (more precisely, with probability tending to ...
Before reversible circuits, we first considered a simpler setting: triangle counting. The no-coincidence principle in that setting turned out to be true, but for a relatively uninteresting reason, because the domain was not rich enough. Nevertheless, I think this result serves as a helpful exercise for people trying to get to grips with our definitions, as well as providing more of the story about how we ended up with our reversible circuits statement.
In the triangle counting setting, we consider the distribution over undirected 3-partite...
It sounds like we are not that far apart here. We've been doing some empirical work on toy systems to try to make the leap from mechanistic interpretability "stories" to semi-formal heuristic explanations. The max-of-k draft is an early example of this, and we have more ambitious work in progress along similar lines. I think of this work in a similar way to you: we are not trying to test empirical assumptions (in the way that some empirical work on frontier LLMs is, for example), but rather to learn from the process of putting our ideas into practice.
Thank you – this is probably the best critique of ARC's research agenda that I have read since we started working on heuristic explanations. This level of thoughtfulness in external feedback is very rare and I'm grateful for the detail and clarity you put into it. I don't think my response fully rebuts your central concern, but hopefully it gives a sense of my current thinking about it.
It sounds like we are in agreement that something very loosely heuristic explanation-flavored (interpreted so broadly as to include mechanistic interpretability, for example...
Thank you for the great response, and the (undeserved) praise of my criticism. I think it's really good that you're embracing the slightly unorthodox positions of sticking to ambitious convictions and acknowledging that this is unorthodox. I also really like your (a)-(d) (and agree that many of the adherents of the fields you list would benefit from similar lines of thinking).
I think we largely agree, and much of our disagreement probably boils down to where we draw the boundary between “mechanistic interpretability” and “other”. In particular, I fully agr...
1
1Yes, I think the most natural way to estimate total surprise in practice would be to use sampling like you suggest. You could try to find the best explanation for "the model does $bad_thing with probability less than 1 in a million" (which you believe based on sampling) and then see how unlikely $bad_thing is according to the resulting explanation. In the Boolean circuit worked example, the final 23-bit explanation is likely still the best explanation for why the model outputs TRUE on at least 99% of inputs, and we can use this explanation to see that the ...
Yes, that's a clearer way of putting it in the case of the circuit in the worked example. The reason I said "for no apparent reason" is that there could be some redundancy in the explanation. For example, if you already had an explanation for the output of some subcircuit, you shouldn't pay additional surprise if you then check the output of that subcircuit in some particular case. But perhaps this was a distracting technicality.
I would say that they are motivated by the same basic idea, but are applied to different problems. The MDL (or the closely-related BIC) is a method for model selection given a dataset, whereas surprise accounting is a method for evaluating heuristic explanations, which don't necessarily involve model selection.
Take the Boolean circuit worked example: what is the relevant dataset? Perhaps it is the 256 (input, TRUE) pairs. But the MDL would select a much simpler model, namely the circuit that ignores the input and outputs TRUE (or "x_1 OR (NOT x_1)" if it h...
Yes, the cost of 1 bit for the OR gate was based on the somewhat arbitrary choice to consider only OR and AND gates. A bit more formally, the heuristic explanations in the post implicitly use a "reference class" of circuits where each binary gate was randomly chosen to be either an OR or an AND, and each input wire to a binary gate was randomly chosen to have a NOT or not. The arbitrariness of this choice of reference class is one obstruction to formalizing heuristic explanations and surprise accounting. We are currently preparing a paper that explores this and related topics, but unfortunately the core issue remains unresolved.
See the statement from OpenAI in this article:
We're removing nondisparagement clauses from our standard departure paperwork, and we're releasing former employees from existing nondisparagement obligations unless the nondisparagement provision was mutual. We'll communicate this message to former employees.
They have communicated this to me and I believe I was in the same category as most former employees.
I think the main reasons so few people have mentioned this are:
- As I mentioned, there is still some legal ambiguity and additional avenues for retaliation
- Som
Yeah I agree with this, and my original comment comes across too strongly upon re-reading. I wanted to point out some counter-considerations, but the comment ended up unbalanced. My overall view is:
- It was highly inappropriate for the company to have been issuing these agreements so widely, especially using such aggressive tactics and without allowing disclosure of the agreement, given the technology that it is developing.
- The more high-profile and credible a person is, the more damaging it is for this person to have been subject to the agreement.
- Nevertheles
1the post appears to wildly misinterpret the meaning of this term as "taking any actions which might make the company less valuable"
I'm not a lawyer, and I may be misinterpreting the non-interference provision—certainly I'm willing to update the post if so! But upon further googling, my current understanding is still that in contracts, "interference" typically means "anything that disrupts, damages or impairs business."
And the provision in the OpenAI offboarding agreement is written so broadly—"Employee agrees not to interfere with OpenAI’s relationship wit...
2Thankfully, most of this is now moot as the company has retracted the contract.
I don't think any of this is moot, since the thing that is IMO most concerning is people signing these contracts, then going into policy or leadership positions and not disclosing that they signed those contracts. Those things happened in the past and are real breaches of trust.
We were especially alarmed to notice that the list contains at least 12 former employees currently working on AI policy, and 6 working on safety evaluations. This includes some in leadership positions, for example:
I don't really follow this reasoning. If anything, playing a leadership role in AI policy or safety evaluations will usually give you an additional reason not to publicly disparage AI companies, to avoid being seen as partisan, making being subject to such an agreement less of an issue. I would be pretty surprised if such people subject to these ...
I see the concerns as these:
- The four corners of the agreement seem to define 'disparagement' broadly, so one might reasonably fear (e.g.) "First author on an eval especially critical of OpenAI versus its competitors", or "Policy document highly critical of OpenAI leadership decisions" might 'count'.
- Given Altman's/OpenAI's vindictiveness and duplicity, and the previous 'safeguards' (from their perspective) which give them all the cards in terms of folks being able to realise the value of their equity, "They will screw me out of a lot of money if I do someth
If the question is whether I think they were true at time given the information I have now, I think all of the individual points hold up except for the first and third "opinions". I am now less sure about what OpenAI leadership believed or cared about. The last of the "opinions" also seems potentially overstated. Consequently, the overall thrust now seems off, but I still think it was good to share my views at the time, to start a discussion.
If the question is about the state of the organization now, I know less about that because I haven't worked there in over a year. But the organization has certainly changed a lot since this post was written over 18 months ago.
Since this post was written, OpenAI has done much more to communicate its overall approach to safety, making this post somewhat obsolete. At the time, I think it conveyed some useful information, although it was perceived as more defensive than I intended.
My main regret is bringing up the Anthropic split, since I was not able to do justice to the topic. I was trying to communicate that OpenAI maintained its alignment research capacity, but should have made that point without mentioning Anthropic.
Ultimately I think the post was mostly useful for sparking some interesting discussion in the comments.
I think KL/entropy regularization is usually used to prevent mode collapse partly because it has nice theoretical properties. In particular, it is easy to reason about the optimal policy for the regularized objective - see for example the analysis in the paper Equivalence Between Policy Gradients and Soft Q-Learning.
Nevertheless, action-dependent baselines do appear in the literature, although the story is a bit confusing. This is my understanding of it from some old notes:
- The idea was explored in Q-Prop. But unlike you, their intention was not to change t
We will do our best to fairly consider all applications, but realistically there is probably a small advantage to applying earlier. This is simply because there is a limit to how quickly we can grow the organization, so if hiring goes better than expected then it will be longer before we can take on even more people. With that being said, we do not have a fixed number of positions that we are hiring for; rather, we plan to vary the number of hires we make based on the strength of the applications we receive. Moreover, if we were unable to hire someon...
1The questions on the take-home test vary in difficulty but are generally easier than olympiad problems, and should be accessible to graduates with relevant background. However, it is important to note that we are ultimately interested in research ability rather than the ability to solve self-contained problems under timed conditions. So although the take-home test forms part of our assessment, we also look at other signals such as research track-record (while recognizing that assessing research ability is unfortunately very hard).
(Note: I am talking about the current version of the test, it's possible that the difficulty will change as we refine our interview process.)
I think the kind of mathematical problem solving we're engaged in is common across theoretical physics (although this is just my impression as a non-physicist). I've noticed that some specific topics that have come up (such as Gaussian integrals and permanents) also crop up in quantum field theory, but I don't think that's a strong reason to prefer that background particularly. Broad areas that often come up include probability theory, computational complexity and discrete math, but it's not necessary to have a lot of experience in those areas, only to be able to pick things up from them as needed.
It's not quite this simple, the same issue arises if every PSD completion of the known-diagonal minor has zero determinant (e.g. ((?, 1, 2), (1, 1, 1), (2, 1, 1))). But I think in that case making the remaining diagonal entries large enough still makes the eigenvalues at least −ε, which is good enough.
I think the examples you give are valid, but there are several reasons why I think the situation is somewhat contingent or otherwise less bleak than you do:
- Counterexamples: I think there are research agendas that are less pre-paradigmatic than the ones you're focused on that are significantly more (albeit not entirely) parallelizable. For example, mechanistic interpretability and scalable oversight both have multiple groups focused on them and have grown substantially over the last couple of years. I'm aware that we disagree about how valuable these direct
You might find this work interesting, which takes some small steps in this direction. It studies the effect of horizon length inasmuch as it makes credit assignment harder, showing that the number of samples required is an affine function of horizon length in a toy context.
I think the direction depends on what your expectations were – I'll try to explain.
First, some terminology: the term "horizon length" is used in the paper to refer to the number of timesteps over which the algorithm pays attention to rewards, as governed by the discount rate. In the biological anchors framework, the term "effective horizon length" is used to refer to a multiplier on the number of samples required to train the model, which is influenced by the horizon length and other factors. For clarity, I'll using the term "scaling multiplier" instead of...
I would wildly speculate that "simply" scaling up RLHF ~100x, while paying careful attention to rewarding models appropriately (which may entail modifying the usual training setup, as discussed in this comment), would be plenty to get current models to express calibrated uncertainty well. However:
- In practice, I think we'll make a lot of progress in the short term without needing to scale up this much by using various additional techniques, some that are more like "tricks" (e.g. teaching the model to generally express uncertainty when answering hard math pr
My understanding of why it's especially hard to stop the model making stuff up (while not saying "I don't know" too often), compared to other alignment failures:
- The model inherits a strong tendency to make stuff up from the pre-training objective.
- This tendency is reinforced by the supervised fine-tuning phase, if there are examples of answers containing information that the model doesn't know. (However, this can be avoided to some extent, by having the supervised fine-tuning data depend on what the model seems to know, a technique that was employed here.)
- I
I would estimate that the difference between "hire some mechanical turkers and have them think for like a few seconds" and the actual data collection process accounts for around 1/3 of the effort that went into WebGPT, rising to around 2/3 if you include model assistance in the form of citations. So I think that what you wrote gives a misleading impression of the aims and priorities of RLHF work in practice.
I think it's best to err on the side of not saying things that are false in a literal sense when the distinction is important to other people, even whe...
Sorry, yeah, I definitely just messed up in my comment here in the sense that I do think that after looking at the research, I definitely should have said "spent a few minutes on each datapoint", instead of "a few seconds" (and indeed I noticed myself forgetting that I had said "seconds" instead of "minutes" in the middle of this conversation, which also indicates I am a bit triggered and doing an amount of rhetorical thinking and weaseling that I think is pretty harmful, and I apologize for kind of sliding between seconds and minutes in my last two commen...
However, I do think in-practice, the RLHF that has been implemented has mostly been mechanical turkers thinking about a problem for a few minutes
I do not consider this to be accurate. With WebGPT for example, contractors were generally highly educated, usually with an undergraduate degree or higher, were given a 17-page instruction manual, had to pass a manually-checked trial, and spent an average of 10 minutes on each comparison, with the assistance of model-provided citations. This information is all available in Appendix C of the paper.
There is RLHF wor...
I agree that the RLHF framework is essentially just a form of model-based RL, and that its outer alignment properties are determined entirely by what you actually get the humans to reward. But your description of RLHF in practice is mostly wrong. Most of the challenge of RLHF in practice is in getting humans to reward the right thing, and in doing so at sufficient scale. There is some RLHF research that uses low-quality feedback similar to what you describe, but it does so in order to focus on other aspects of the setup, and I don't think anyone currently ...
- We are just observing that the gold RM score curves in Figure 9 overlap. In other words, the KL penalty did not affect the relationship between KL and gold RM score in this experiment, meaning that any point on the Pareto frontier could be reached using only early stopping, without the KL penalty. As mentioned though, we've observed this result to be sensitive to hyperparameters, and so we are less confident in it than other results in the paper.
- I don't have this data to hand unfortunately.
- I don't have this data to hand, but entropy typically falls roughly
Agreed. Likewise, in a transformer, the token dimension should maintain some relationship with the input and output tokens. This is sometimes taken for granted, but it is a good example of the data preferring a coordinate system. My remark that you quoted only really applies to the channel dimension, across which layers typically scramble everything.
The notion of a preferred (linear) transformation for interpretability has been called a "privileged basis" in the mechanistic interpretability literature. See for example Softmax Linear Units, where the idea is discussed at length.
In practice, the typical reason to expect a privileged basis is in fact SGD – or more precisely, the choice of architecture. Specifically, activation functions such as ReLU often privilege the standard basis. I would not generally expect the data or the initialization to privilege any basis beyond the start of the network or the...
Thank you for causing me to reconsider. I should have said "other OpenAI employees". I do not intend to disengage from the alignment community because of critical rhetoric, and I apologize if my comment came across as a threat to do so. I am concerned about further breakdown of communication between the alignment community and AI labs where alignment solutions may need to be implemented.
I don't immediately see any other reason why my comment might have been inappropriate, but I welcome your clarification if I am missing something.
For people viewing on the Alignment Forum, there is a separate thread on this question here. (Edit: my link to LessWrong is automatically converted to an Alignment Forum link, you will have to navigate there yourself.)
I was the project lead on WebGPT and my motivation was to explore ideas for scalable oversight and truthfulness (some further explanation is given here).
It includes the people working on the kinds of projects I listed under the first misconception. It does not include people working on things like the mitigation you linked to. OpenAI distinguishes internally between research staff (who do ML and policy research) and applied staff (who work on commercial activities), and my numbers count only the former.
WebGPT seemed like one of the most in-expectation harmful projects that OpenAI has worked on, with no (to me) obvious safety relevance, so my guess is I would still mostly categorize the things you list under the first misconception as capabilities research. InstructGPT also seems to be almost fully capabilities research (like, I agree that there are some safety lessons to be learned here, but it seems somewhat clear to me that people are working on WebGPT and InstructGPT primarily for capabilities reasons, not for existential-risk-from-AI reasons)
(Edit: M...
I don't think I understand your question about Y-problems, since it seems to depend entirely on how specific something can be and still count as a "problem". Obviously there is already experimental evidence that informs predictions about existential risk from AI in general, but we will get no experimental evidence of any exact situation that occurs beforehand. My claim was more of a vague impression about how OpenAI leadership and John tend to respond to different kinds of evidence in general, and I do not hold it strongly.
To clarify, by "empirical" I meant "relating to differences in predictions" as opposed to "relating to differences in values" (perhaps "epistemic" would have been better). I did not mean to distinguish between experimental versus conceptual evidence. I would expect OpenAI leadership to put more weight on experimental evidence than you, but to be responsive to evidence of all kinds. I think that OpenAI leadership are aware of most of the arguments you cite, but came to different conclusions after considering them than you did.
[First of all, many thanks for writing the post; it seems both useful and the kind of thing that'll predictably attract criticism]
I'm not quite sure what you mean to imply here (please correct me if my impression is inaccurate - I'm describing how-it-looks-to-me, and I may well be wrong):
I would expect OpenAI leadership to put more weight on experimental evidence than you...
Specifically, John's model (and mine) has:
X = [Class of high-stakes problems on which we'll get experimental evidence before it's too late]
Y = [Class of high-stakes problems on which we...
It's hard to predict (especially if timelines are long), but if I had to guess I would say that something similar to human feedback on diverse tasks will be the unaligned benchmark we will be trying to beat. In that setting, a training episode is an episode of an RL environment in which the system being trained performs some task and obtains reward chosen by humans.
It's even harder to predict what our aligned alternatives to this will look like, but they may need to be at least somewhat similar to this in order to remain competitive. In that case, a traini...
This is just supposed to be an (admittedly informal) restatement of the definition of outer alignment in the context of an objective function where the data distribution plays a central role.
For example, assuming a reinforcement learning objective function, outer alignment is equivalent to the statement that there is an aligned policy that gets higher average reward on the training distribution than any unaligned policy.
I did not intend to diminish the importance of robustness by focusing on outer alignment in this post.
I share your intuitions about ultimately not needing much alignment data (and tried to get that across in the post), but quantitatively:
- Recent implementations of RLHF have used on the order of thousands of hours of human feedback, so 2 orders of magnitude more than that is much more than a few hundred hours of human feedback.
- I think it's pretty likely that we'll be able to pay an alignment tax upwards of 1% of total training costs (essentially because people don't want to die), in which case we could afford to spend significantly more than an additional 2 orders of magnitude on alignment data, if that did in fact turn out to be required.
A number of reasonable outer alignment proposals such as iterated amplification, recursive reward modeling and debate use generic objectives such as reinforcement learning (and indeed, none of them would work in practice without sufficiently high data quality), so it seems strange to me to dismiss these objectives.
I think it's reasonable to aim for quantity within 2 OOM of RLHF.
Do you mean that on-paper solutions should aim to succeed with no more than 1/100 as much human data as RLHF, or no more than 100 times as much? And are you referring the amount of human data typically used in contemporary implementations of RLHF, or something else? And what makes you think that this is a reasonable target?
I think that data quality is a helpful framing of outer alignment for a few reasons:
- Under the assumption of a generic objective such as reinforcement learning, outer alignment is definitionally equivalent to having high enough data quality. (More precisely, if the objective is generic enough that it is possible for it to produce an aligned policy, then outer alignment is equivalent to the data distribution being such that an aligned policy is preferred to any unaligned policy.)
- If we had the perfect alignment solution on paper, we would still need to implem
The hope is to use the complexity of the statement rather than mathematical taste.
If it takes me 10 bits to specify a computational possibility that ought to happen 1% of the time, then we shouldn't be surprised to find around 10 (~1% of 210) occurrences. We don't intend the no-coincidence principle to claim that these should all happen for a reason.
Instead, we intend the no-coincidence principle to claim that such if such coincidences happen much more often than we would have expected them to by chance, then there is a reason for that. Or put another... (read more)