Rob Bensinger

Communications lead at MIRI. Unless otherwise indicated, my posts and comments here reflect my own views, and not necessarily my employer's.


2022 MIRI Alignment Discussion
2021 MIRI Conversations

Wiki Contributions

Load More


Thanks for the update, Ajeya! I found the details here super interesting.

I already thought that timelines disagreements within EA weren't very cruxy, and this is another small update in that direction: I see you and various MIRI people and Metaculans give very different arguments about how to think about timelines, and then the actual median year I tend to hear is quite similar.

(And also, all of the stated arguments on all sides continue to seem weak/inconclusive to me! So IMO there's not much disagreement, and it would be very easy for all of us to be wrong simultaneously. My intuition is that it would be genuinely weird if AGI is much more than 70 years away, but not particularly weird if it's 1 year away, 10 years away, 60 years away, etc.)

I think the main value of in-depth timelines research and debate has been that it reveals disagreements about other topics (background views about ML, forecasting methodology, etc.).

Some added context for this list: Nate and Eliezer expect the first AGI developers to encounter many difficulties in the “something forces you to stop and redesign (and/or recode, and/or retrain) large parts of the system” category, with the result that alignment adds significant development time.

By default, safety-conscious groups won't be able to stabilize the game board before less safety-conscious groups race ahead and destroy the world. To avoid this outcome, humanity needs there to exist an AGI group that

  • is highly safety-conscious.
  • has a large resource advantage over the other groups, so that it can hope to reach AGI with more than a year of lead time — including accumulated capabilities ideas and approaches that it hasn’t been publishing.
  • has adequate closure and opsec practices, so that it doesn’t immediately lose its technical lead if it successfully acquires one.

The magnitude and variety of difficulties that are likely to arise in aligning the first AGI systems also suggests that failure is very likely in trying to align systems as opaque as current SotA systems; and suggests an AGI developer likely needs to have spent preceding years deliberately steering toward approaches to AGI that are relatively alignable; and it suggests that we need to up our game in general, approaching the problem in ways that are closer to the engineering norms at (for example) NASA, than to the engineering norms that are standard in ML today.

One caveat to the claim that we should prioritize serial alignment work over parallelizable work, is that this assumes an omniscient and optimal allocator of researcher-hours to problems.

Why do you think it assumes that?

This isn't a coincidence; the state of alignment knowledge is currently "we have no idea what would be involved in doing it even in principle, given realistic research paths and constraints", very far from being a well-specified engineering problem. Cf.

If you succeed at the framework-inventing "how does one even do this?" stage, then you can probably deploy an enormous amount of engineering talent in parallel to help with implementation, small iterative improvements, building-upon-foundations, targeting-established-metrics, etc. tasks.

From A central AI alignment problem: capabilities generalization, and the sharp left turn:

Suppose that the fictional team OpenMind is training up a variety of AI systems, before one of them takes that sharp left turn. Suppose they've put the AI in lots of different video-game and simulated environments, and they've had good luck training it to pursue an objective that the operators described in English. "I don't know what those MIRI folks were talking about; these systems are easy to direct; simple training suffices", they say. At the same time, they apply various training methods, some simple and some clever, to cause the system to allow itself to be removed from various games by certain "operator-designated" characters in those games, in the name of shutdownability. And they use various techniques to prevent it from stripmining in Minecraft, in the name of low-impact. And they train it on a variety of moral dilemmas, and find that it can be trained to give correct answers to moral questions (such as "in thus-and-such a circumstance, should you poison the operator's opponent?") just as well as it can be trained to give correct answers to any other sort of question. "Well," they say, "this alignment thing sure was easy. I guess we lucked out."

Then, the system takes that sharp left turn,[4][5] and, predictably, the capabilities quickly improve outside of its training distribution, while the alignment falls apart.


[5] "Hold on, isn't this unfalsifiable? Aren't you saying that you're going to continue believing that alignment is hard, even as we get evidence that it's easy?" Well, I contend that "GPT can learn to answer moral questions just as well as it can learn to answer other questions" is not much evidence either way about the difficulty of alignment. I'm not saying we'll get evidence that I'll ignore; I'm naming in advance some things that I wouldn't consider negative evidence (partially in hopes that I can refer back to this post when people crow later and request an update). But, yes, my model does have the inconvenient property that people who are skeptical now, are liable to remain skeptical until it's too late, because most of the evidence I expect to give us advance warning about the nature of the problem is evidence that we've already seen. I assure you that I do not consider this property to be convenient.

As for things that could convince me otherwise: technical understanding of intelligence could undermine my "sharp left turn" model. I could also imagine observing some ephemeral hopefully-I'll-know-it-when-I-see-it capabilities thresholds, without any sharp left turns, that might update me. (Short of "full superintelligence without a sharp left turn", which would obviously convince me but comes too late in the game to shift my attention.)

(Most of the QR-upvotes at the moment are from me. I think 1-4 are all good questions, for Nate or others; but I'm extra excited about people coming up with ideas for 3.)

When I think about the strawberry problem, it seems unnatural, and perhaps misleading of our attention, since there's no guarantee there's even a reasonable solution.

Why would there not be a solution?

On my model, the point of ass numbers isn't to demand perfection of your gut (e.g., of the sort that would be needed to avoid multiple-stage fallacies when trying to conditionalize a lot), but to:

  1. Communicate with more precision than English-language words like 'likely' or 'unlikely' allow. Even very vague or uncertain numbers will, at least some of the time, be a better guide than natural-language terms that weren't designed to cover the space of probabilities (and that can vary somewhat in meaning from person to person).
  2. At least very vaguely and roughly bring your intuitions into contact with reality, and with each other, so you can more readily notice things like 'I'm miscalibrated', 'reality went differently than I expected', 'these two probabilities don't make sense together', etc.

It may still be a terrible idea to spend too much time generating ass numbers, since "real numbers" are not the native format human brains compute probability with, and spending a lot of time working in a non-native format may skew your reasoning.

(Maybe there's some individual variation here?)

But they're at least a good tool to use sometimes, for the sake of crisper communication, calibration practice (so you can generate non-awful future probabilities when you need to), etc.

Load More