There are a few sentences in Anthropic's "conversation with our cofounders" regarding RLHF that I found quite striking:

Dario (2:57): "The whole reason for scaling these models up was that [...] the models weren't smart enough to do RLHF on top of. [...]"

Chris: "I think there was also an element of, like, the scaling work was done as part of the safety team that Dario started at OpenAI because we thought that forecasting AI trends was important to be able to have us taken seriously and take safety seriously as a problem."

Dario: "Correct."

That LLMs were scaled up partially in order to do RLHF on top of them is something I had previously heard from an OpenAI employee, but I wasn't sure it's true. This conversation seems to confirm it. 

Hi! Thanks a lot for your comments and very good points. I apologize for my late answer, caused by NeurIPS and all the end-of-year breakdown of routines :)

On 1: Yes, the formalism I'm currently working on also allows to talk about the case that the human "understands less" than the AI.

On 2: 

Have you considered the connection between partial observability and state aliasing/function approximation?

I am not entirely sure if I understand! Though if it's just what you express in the following sentences, here's my answers:

Maybe you could apply your theory to weak-to-strong generalization by considering a weak model as operating under partial observability.

Very good observation! :) I'm thinking about it slightly differently, but the link is there: Imagine a scenario where we have a pretrained foundation model, and we train a linear probe attached to the internal representations, which is supposed to learn the correct reward for full state sequences, based on feedback from a human on partial observations. Then if we show this model (including attached probe) during training just the partial observations, it's receiving the correct data and is supposed to generalize from feedback on "easy situations" (i.e., situations where the partial observations of the human provide enough information to make a correct judgment) to "hard situations" (full state sequences that the human couldn't oversee, and where possibly the partial observations miss crucial details).

So I think this setting is an instance of weak-to-strong generalization.

Alternatively, by introducing structure to the observations, the function approximation lens might open up new angles of attack on the problem.

Yes that's actually also part of what I'm exploring, if I understand your idea correctly. In particular, I'm considering the case that we may have "knowledge" of some form about the space in which the correct reward function lives. This may come from symmetries in the state space, for example: maybe we want to restrict to localized reward functions that are translation-invariant. All of that can easily be formalized in one framework. 

Pretrained foudation models on which we attach a "reward probe" can be viewed as another instance of considering symmetries in the state space: In this case, we're presuming that state sequences have the same reward if they give rise to the same "learned abstractions" in the form of the internal representations of the neural network.

On 3: Agreed. (Though I am not explicitly considering this case at this point. )

On 4:

I think you're exactly right to consider abstractions of trajectories, but I'm not convinced this needs to be complicated. What if you considered the case where the problem definition includes features of state trajectories on which (known) human utilities are defined, but these features themselves are not always observed? (This is something I'm currently thinking about, as a generalization of the work mentioned in the postscript.

This actually sounds very much like what I'm working on right now!! We should probably talk :) 

On 5:

Am I correct in my understanding that the role Boltzmann rationality plays in your setup is just to get a reward function out of preference data?

If I understand correctly, yes. In a sense, we just "invert" the sigmoid function to recover the return function on observation sequences from human preference data. If this return function on observation sequences was already known, we'd still be doomed, as you correctly point out.

Thanks also for the notes on gradient routing! I will read your post and will try to understand the connection.

This is a link to a big list of LLM safety papers based on a new big survey.

Thanks for the list! I have two questions:

1: Can you explain how generalization of NNs relates to ELK? I can see that it can help with ELK (if you know a reporter generalizes, you can train it on labeled situations and apply it more broadly) or make ELK unnecessary (if weak to strong generalization perfectly works and we never need to understand complex scenarios). But I’m not sure if that’s what you mean.

2: How is goodhart robustness relevant? Most models today don’t seem to use reward functions in deployment, and in training the researchers can control how hard they optimize these functions, so I don’t understand why they necessarily need to be robust under strong optimization.

“heuristics activated in different contexts” is a very broad prediction. If “heuristics” include reasoning heuristics, then this probably includes highly goal-oriented agents like Hitler.

Also, some heuristics will be more powerful and/or more goal-directed, and those might try to preserve themselves (or sufficiently similar processes) more so than the shallow heuristics. Thus, I think eventually, it is plausible that a superintelligence looks increasingly like a goal-maximizer.

This is a low effort comment in the sense that I don’t quite know what or whether you should do something different along the following lines, and I have substantial uncertainty.

That said:

1. I wonder whether Anthropic is partially responsible for an increased international race through things like Dario advocating for an entente strategy and talking positively about Leopold Aschenbrenner’s “situational awareness”. I wished to see more of an effort to engage with Chinese AI leaders to push for cooperation/coordination. Maybe it’s still possible to course-correct.

2. Alternatively I think that if there’s a way for Anthropic/Dario to communicate why you think an entente strategy is inevitable/desirable, in a way that seems honest and allows to engage with your models of reality, that might also be very helpful for the epistemic health of the whole safety community. I understand that maybe there’s no politically feasible way to communicate honestly about this, but maybe see this as my attempt to nudge you in the direction of openness.

More specifically:

(a) it would help to learn more about your models of how winning the AGI race leads to long-term security (I assume that might require building up a robust military advantage, but given the physical hurdles that Dario himself expects for AGI to effectively act in the world, it’s unclear to me what your model is for how to get that military advantage fast enough after AGI is achieved).

(b) I also wonder whether potential future developments in AI Safety and control might give us information that the transition period is really unsafe; eg., what if you race ahead and then learn that actually you can’t safely scale further due to risks of loss of control? At that point, coordinating with China seems harder than doing it now. I’d like to see a legible justification of your strategy that takes into account such serious possibilities.

I have an AI agent that wrote myself

Best typo :D

Have you also tried reviewing for conferences like NeurIPS? I'd be curious what the differences are.

Some people send papers to TMLR when they think they wouldn't be accepted to the big conferences due to not being that "impactful" --- which makes sense since TMLR doesn't evaluate impact. It's thus possible that the median TMLR submission is worse than the median conference submission.

I just donated $200. Thanks for everything you're doing!

Yeah I think that's a valid viewpoint. 

Another viewpoint that points in a different direction: A few years ago, LLMs could only do tasks that require humans ~minutes. Now they're at the ~hours point. So if this metric continues, eventually they'll do tasks requiring humans days, weeks, months, ...

I don't have good intuitions that would help me to decide which of those viewpoints is better for predicting the future.