Joe Collman

Karma: 1,677

I’m a researcher on the technical governance team at MIRI.

Views expressed are my own, and should not be taken to represent official MIRI positions. Similarly, views within the technical governance team do vary.

Previously:

Helped with MATS, running the technical side of the London extension (pre-LISA).
Briefly a teaching fellow with BlueDot on AISF.
Worked for a while on Debate (this kind of thing).

Quick takes on the above:

I think MATS is great-for-what-it-is. My misgivings relate to high-level direction.
- Worth noting that PIBBSS exists, and is philosophically closer to my ideal.
The technical AISF course doesn’t have the emphasis I’d choose (which would be closer to Key Phenomena in AI Risk). It’s a decent survey of current activity, but only implicitly gets at fundamentals—mostly through a [notice what current approaches miss, and will continue to miss] mechanism.
I don’t expect research on Debate, or scalable oversight more generally, to help significantly in reducing AI x-risk. (I may be wrong! - some elaboration in this comment thread)

Joe Collman 19 Jul 2025 23:37 UTC
2 points
0
on: The Compulsion For (Pseudo-)Mechanisms
Interesting. I mostly agree with the gist.
The following are a few thoughts that occur to me. Presented as potentially useful pointers, rather than well-thought-through arguments/conclusions.
- I don’t think “pseudo-mechanisms” is a useful label. Feels a bit too binary (and/or post-hoc) in a highly grey situation.
- I’m not sure what you mean by “mechanistic model” vs “stable phenomenological compressions”.
  - I’m not saying I have no idea what you’re talking about—just that I’m not clear quite how you want to distinguish these things. (note that I haven’t read many of your previous posts—yet! :))
  - As soon as I’m calling something a “stable” pattern in the data, there’s at least an implicit [...and this pattern will continue to hold] hypothesis.
  - What makes something a “mechanistic model”? E.g. does it need to involve a temporal pattern, so that I’m likely to think of it in causal terms?
  - Is the problem here that humans tend to prematurely place too high a probability on causal hypotheses?
    If so, is the general thing to be wary of more [hypotheses I’m likely to believe too strongly given the (lack of) evidence]?
    E.g. a mathematician might be wary of elegant hypotheses on this basis.
    This seems a plausible explanation for a [practicality is inversely proportional to attachment to mechanisms]. The less practical a field, the more its practitioners will tend to be attracted by some kind of aesthetic sense—and that’s a potential source of bias and premature attachment. (and in many cases there’s the [less straightforwardly falsifiable] factor)
It would be best if we could simply not follow the compulsion, and stay as much as possible at the level of data patterns and phenomenological compressions, at least until we have a good handle there.
- Here I remain unclear, as above. (I don’t know what separates [have a good handle on data patterns] from [have explanations])
- It seems to me our thinking is always going to be inseparable from a huge number of mechanistic expectations and assumptions (often implicit).
- It seems a lesson here is something like:
  - Be aware of the mechanistic models we’re relying on.
  - Be aware of the tendency to get prematurely attached to an explanation.
  - Adapt accordingly.
    I don’t think jumping to a mechanistic model is itself an error—stubbornly sticking to one seems to be the problem.
    Nimbleness seems desirable.
    In particular since [assuming too much] and [assuming too little] are both sources of inefficiency.
Personally, when I try to predict the behaviour of people, I start with their past actions. As in, I look at what they have done in the past, and assume they’ll do more of that.
Similarly here, I think it’s asking for trouble to imagine that [Gabe’s characterization and extrapolation of [that]] doesn’t already rely on a bunch of intent-based expectations and assumptions. (these will usually be more reliable than guesses we’d tend to label “psycho-analysis”—but they’re present and important)
For this reason, [be aware of the degree to which you’re x-ing, and the implications] seems safer advice than [avoid x-ing], for many x.

Joe Collman 7 Jul 2025 17:56 UTC
11 points
5
in reply to: Jonas V’s comment on: Help the AI 2027 team make an online AGI wargame
If you’re aiming to get millions of players, I think [no music at all] would be counterproductive. There’s a reason almost every non-trivial game in existence has music. Of course it’s also nice if it’s simple to turn off / customize / replace—but it’s usually a mistake to expect that a high proportion of players are going to significantly customize things.
Music is a way to get some immediate emotional engagement without making meaningful design concessions (most other mechanisms imply some more significant design constraint). If you want millions of players, you want immediate emotional engagement.

Joe Collman 14 Apr 2025 22:54 UTC
9 points
2
on: A Solution to Sandbagging and other Self-Provable Misalignment: Constitutional AI Detectives
Unless I’m missing something, the hoped-for advantages of this setup are the kind of thing AI safety via debate already aims at. In GDM’s recent paper on their approach to technical alignment, there’s some discussion of amplified oversight (starts at page 71) more generally, and debate (starts at page 73).
If you see the approach you’re suggesting as importantly different from debate approaches, it’d be useful to know where the key differences are.
(without having read too carefully, my initial impression is that this is the kind of thing I expect to work for a while, then fail [as with debate] - and my core concern is then: how do we accurately predict when it’ll fail?)
What links here?
- A Solution to Sandbagging and other Self-Provable Misalignment: Constitutional AI Detectives by Knight Lee (14 Apr 2025 10:27 UTC; -3 points)

Joe Collman 21 Nov 2024 20:59 UTC
5 points
1
in reply to: Orpheus16’s comment on: Akash’s Shortform
Some thoughts:
- The correct answer is clearly (c) - it depends on a bunch of factors.
- My current guess is that it would make things worse (given likely values for the bunch of other factors) - basically for Richard’s reasons.
  - Given [new potential-to-shift-motivation information/understanding], I expect there’s a much higher chance that this substantially changes the direction of a not-yet-formed project, than a project already in motion.
  - Specifically:
    Who gets picked to run such a project? If it’s primarily a [let’s beat China!] project, are the key people cautious and highly adaptable when it comes to top-level goals? Do they appoint deputies who’re cautious and highly adaptable?
    Here I note that the kind of ‘caution’ we’d need is [people who push effectively for the system to operate with caution]. Most people who want caution are more cautious.
    How is the project structured? Will the structure be optimized for adaptability? For red-teaming of top-level goals?
    Suppose that a mid-to-high-level participant receives information making the current top-level goals questionable—is the setup likely to reward them for pushing for changes? (noting that these are the kind of changes that were not expected to be needed when the project launched)
    Which external advisors do leaders of the project develop relationships with? What would trigger these to change?
    ...
- I do think that it makes sense to aim for some centralized project—but only if it’s the right kind.
  - I expect that almost all the directional influence is in [influence the initial conditions].
  - For this reason, I expect [push for some kind of centralized project, and hope it changes later] is a bad idea.
  - I think [devote great effort to influencing the likely initial direction of any such future project] seems a great idea (so long as you’re sufficiently enlightened about desirable initial directions, of course :))
  - I’d note that [initial conditions] needn’t only be internal to the project—in principle we could have reason to believe that various external mechanisms would be likely to shift the project’s motivation sufficiently over time. (I don’t know of any such reasons)
- I think the question becomes significantly harder once the primary motivation behind a project isn’t [let’s beat China!], but also isn’t [your ideal project motivation (with your ideal initial conditions)].
- I note that my p(doom) doesn’t change much if we eliminate racing but don’t slow down until it’s clear to most decision makers that it’s necessary.
  - Likewise, I don’t expect that [focus on avoiding the earliest disasters] is likely to be the best strategy. So e.g. getting into a good position on security seems great, all else equal—but I wouldn’t sacrifice much in terms of [odds of getting to a sufficiently cautious overall strategy] to achieve better short-term security outcomes.

Joe Collman 21 Nov 2024 9:50 UTC
0 points
0
on: Making a conservative case for alignment
First some points of agreement:
- I like that you’re focusing on neglected approaches. Not much on the technical side seems promising to me, so I like to see exploration.
  - Skimming through your suggestions, I think I’m most keen on human augmentation related approaches—hopefully the kind that focuses on higher quality decision-making and direction finding, rather than simply faster throughput.
- I think outreach to Republicans / conservatives, and working across political lines is important, and I’m glad that people are actively thinking about this.
- I do buy the [Trump’s high variance is helpful here] argument. It’s far from a principled analysis, but I can more easily imagine [Trump does correct thing] than [Harris does correct thing]. (mainly since I expect the bar on “correct thing” to be high so that it needs variance)
  - I’m certainly making no implicit ”...but the Democrats would have been great...” claim below.
That said, various of the ideas you outline above seem to be founded on likely-to-be-false assumptions.
Insofar as you’re aiming for a strategy that provides broadly correct information to policymakers, this seems undesirable—particularly where you may be setting up unrealistic expectations.

Highlights of the below:
1. Telling policymakers that we don’t need to slow down seems negative.
  1. I don’t think you’ve made any valid argument that not needing to slow down is likely. (of course it’d be convenient)
2. A negative [alignment-in-the-required-sense tax] seems implausible. (see below)
  1. (I don’t think it even makes sense in the sense that “alignment tax” was originally meant^[1], but if “negative tax” gets conservatives listening, I’m all for it!)
3. I think it’s great for people to consider convenient possibilities (e.g. those where economic incentives work for us) in some detail, even where they’re highly unlikely. Whether they’re actually 0.25% or 25% likely isn’t too important here.
  1. Once we’re talking about policy advocacy, their probability is important.
More details:
A conservative approach to AI alignment doesn’t require slowing progress, avoiding open sourcing etc. Alignment and innovation are mutually necessary, not mutually exclusive: if alignment R&D indeed makes systems more useful and capable, then investing in alignment is investing in US tech leadership.
Here and in the case for a negative alignment tax, I think you’re:
1. Using a too-low-resolution picture of “alignment” and “alignment research”.
  1. This makes it too easy to slip between ideas like:
    Some alignment research has property x
    All alignment research has property x
    A [sufficient for scalable alignment solution] subset of alignment research has property x
    A [sufficient for scalable alignment solution] subset of alignment research that we’re likely to complete has property x
  2. An argument that requires (iv) but only justifies (i) doesn’t accomplish much. (we need something like (iv) for alignment tax arguments)
2. Failing to distinguish between:
  1. Alignment := Behaves acceptably for now, as far as we can see.
  2. Alignment := [some mildly stronger version of ‘alignment’]
  3. Alignment := notkilleveryoneism
In particular, there’ll naturally be some crossover between [set of research that’s helpful for alignment] and [set of research that leads to innovation and capability advances] - but alone this says very little.
What we’d need is something like:
- Optimizing efficiently for innovation in a way that incorporates various alignment-flavored lines of research gets us sufficient notkilleveryoneism progress before any unrecoverable catastrophe with high probability.
It’d be lovely if something like this were true—it’d be great if we could leverage economic incentives to push towards sufficient-for-long-term-safety research progress. However, the above statement seems near-certainly false to me. I’d be (genuinely!) interested in a version of that statement you’d endorse at >5% probability.
The rest of that paragraph seems broadly reasonable, but I don’t see how you get to “doesn’t require slowing progress”.
On “negative alignment taxes”:
First, a point that relates to the ‘alignment’ disambiguation above.
In the case for a negative alignment tax, you offer the following quote as support for alignment/capability synergy:
...Behaving in an aligned fashion is just another capability… (Anthropic quote from Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback)
However, the capability is [ability to behave in an aligned fashion], and not [tendency to actually behave in an aligned fashion] (granted, Anthropic didn’t word things precisely here). The latter is a propensity, not a capability.
What we need for scalable alignment is the propensity part: no-one sensible is suggesting that superintelligences wouldn’t have the ability to behave in an aligned fashion. The [behavior-consistent-with-alignment]-capability synergy exists while a major challenge is for systems to be able to behave desirably.
Once capabilities are autonomous-x-risk-level, the major challenge will be to get them to actually exhibit robustly aligned behavior. At that point there’ll be no reason to expect the synergy—and so no basis to expect a negative or low alignment tax where it matters.
On things like “Cooperative/prosocial AI systems”, I’d note that hits-based exploration is great—but please don’t expect it to work (and that “if implemented into AI systems in the right ways” is almost all of the problem).

On this basis, it seems to me that the conservative-friendly case you’ve presented doesn’t stand up at all (to be clear, I’m not critiquing the broader claim that outreach and cooperation are desirable):
- We don’t have a basis to expect negative (or even low) alignment tax.
  - (unclear so far that we’ll achieve non-infinite alignment tax for autonomous x-risk relevant cases)
- It’s highly likely that we do need to slow advancement, and will need serious regulation.
Given our lack of precise understanding of the risks, we’ll likely have to choose between [overly restrictive regulation] and [dangerously lax regulation] - we don’t have the understanding to draw the line in precisely the right place. (completely agree that for non-frontier systems, it’s best to go with little regulation)
I’d prefer a strategy that includes [policymakers are made aware of hard truths] somewhere.
I don’t think we’re in a world where sufficient measures are convenient.
It’s unsurprising that conservatives are receptive to quite a bit “when coupled with ideas around negative alignment taxes and increased economic competitiveness”—but this just seems like wishful thinking and poor expectation management to me.

Similarly, I don’t see a compelling case for:
that is, where alignment techniques are discovered that render systems more capable by virtue of their alignment properties. It seems quite safe to bet that significant positive alignment taxes simply will not be tolerated by the incoming federal Republican-led government—the attractor state of more capable AI will simply be too strong.
Of course this is true by default—in worlds where decision-makers continue not to appreciate the scale of the problem, they’ll stick to their standard approaches. However, conditional on their understanding the situation, and understanding that at least so far we have not discovered techniques through which some alignment/capability synergy keeps us safe, this is much less obvious.
I have to imagine that there is some level of perceived x-risk that snaps politicians out of their default mode.
I’d bet on [Republicans tolerate significant positive alignment taxes] over [alignment research leads to a negative alignment tax on autonomous-x-risk-capable systems] at at least ten to one odds (though I’m not clear how to operationalize the latter).
Republicans are more flexible than reality :).
1. ^
  As I understand the term, alignment tax compares [lowest cost for us to train a system with some capability level] against [lowest cost for us to train an aligned system with some capability level]. Systems in the second category are also in the first category, so zero tax is the lower bound.
  
  This seems a better definition, since it focuses on the outputs, and there’s no need to handwave about what counts as an alignment-flavored training technique: it’s just [...any system...] vs [...aligned system...].
  
  Separately, I’m not crazy about the term: it can suggests to new people that we know how to scalably align systems at all. Talking about “lowering the alignment tax” from infinity strikes me as an odd picture.

Joe Collman 11 Aug 2024 22:57 UTC
4 points
0
on: Twitter thread on AI safety evals
I give Eliezer a lot of credit for making roughly this criticism of Ajeya’s bio-anchors report. I think his critique has basically been proven right by how much people have updated away from 30-year timelines since then.
I don’t think this is quite right.
Two major objections to the bio-anchors 30-year-median conclusion might be:
1. The whole thing is laundering vibes into credible-sounding headline numbers.
2. Even if we stipulate that the methodology is sound, it measures an upper bound, not a median.
To me, (2) is the more obvious error. I basically buy (1) too, but I don’t think we’ve gotten empirical evidence, since (2).
I guess there’s a sense in which a mistake on (2) could be seen as a consequence of (1) - but it seems distinct: it’s a logic error, not a free parameter. I do think it’s useful to distinguish [motivated reasoning in free-parameter choice] from [motivated reasoning in error-checking].
It’s not so obvious to me that the bio-anchors report was without foundation as an upper bound estimate.

Joe Collman 11 Aug 2024 14:52 UTC
LW: 2 AF: 1
0
AF
in reply to: Lee Sharkey’s comment on: Circumventing interpretability: How to defeat mind-readers
To be clear, I only think this becomes obviously significant in a train-against-interpretability-tools context; if we’re not training against tools, the things I’m highlighting here seem intuitively highly unlikely.
This still seems important, given that the combination [x is clearly a bad idea] and [x is the easiest path forward] does not robustly result in [nobody does x].
That said, I think I’d disagree on one word of the following:
The mechanisms I labelled as “strictly active” are the kind of strategy that it would be extremely improbable to implement successfully without some sort of coherent internal representations to that help orchestrate the actions required to do it. This is true even if they’ve been selected for passively.
Any internal representations etc. need to result in the actions being orchestrated (to the extent necessary). They don’t need to be used for the purpose of orchestrating the actions. They don’t need to be [clearly selected for the purpose of orchestrating such actions].
In some cases this might be improbable without active, deliberate malign selection—but this would need to be argued for. In at least the case of [manipulating the interpreter], this doesn’t seem at all improbable, given the right selection pressure. (some manipulation seems obviously possible; whether it’s enough for a particular malign strategy would vary)
I haven’t thought about this for long, but in general the [this kind of thing ~must require active deliberation] intuition seems highly suspicious to me. All I need to set up a not-directly-deliberated-on malign strategy so that it’s implemented via benign cognition is a sufficient set of innocuous correlations (both of [innocuous action tends to have useful-for-malign-strategy effect] and [innocuous actions will tend to occur in these predictable patterns]).
If I have unlimited passive selection pressure against tools that robustly detect active malign deliberation, but not ‘instinctive’ malign strategies, I’d strongly expect the outcome to be ‘instinctive’ active manipulation (etc.). (assuming of course that successful manipulation is instrumentally useful)
Whether I’d expect this in practice would come down to how much selection pressure were required for a particular [malign, active-with-no-malign-cognition strategy].
My current answer to this is not [this is extremely improbable]. It’s more like [This seems intuitively plausible; I have no idea on the probability so I’m going to take this possibility seriously until I see a non-handwaving argument that shows it to be extremely unlikely].

Joe Collman 10 Aug 2024 13:58 UTC
LW: 2 AF: 1
0
AF
on: Circumventing interpretability: How to defeat mind-readers
Information flow might be inhibited actively through an agent’s actions. The primary way this could happen is gradient hacking, but it’s not the only kind of action an AI might take to conceal misaligned thoughts. Of course, active circumvention methods require that interpreters either can’t see or aren’t looking at the thoughts that generate those actions.
Most potential circumvention methods that can be passive can also be active. But some methods can only be active.
It seems to me that there’s no fixed notion of “active” that works for both paragraphs here.
If active means [is achieved through the agent’s actions], then this does not in general imply that it is deliberately achieved through the agent’s actions. For example, training against interpretability tools might produce actions that hide misaligned thoughts/actions as side-effects.
With this notion of ‘active’ the first bolded section doesn’t hold: this can happen even when the agent’s thoughts are entirely visible.
If instead active means [is achieved deliberately through the agent’s actions], then the “But some methods can only be active” doesn’t hold.
There are two dimensions here:
- Whether the circumvention is implemented passively/actively.
- Whether the circumvention is selected for passively/actively.
In particular, the mechanisms you’ve labelled “strictly active” can, in principle, be selected for passively—so do not in general require any misaligned thoughts (admittedly, the obvious way this happens is by training against interpretability tools).

Joe Collman 29 Jul 2024 10:53 UTC
6 points
0
in reply to: faul_sname’s comment on: On “first critical tries” in AI alignment
I don’t think [gain a DSA] is the central path here.
It’s much closer to [persuade some broad group that already has a lot of power collectively].
I.e. the likely mechanism is not: [add the property [has DSA] to [group that will do the right thing]].
But closer to: [add the property [will do the right thing] to [group that has DSA]].

Joe Collman 29 Jul 2024 9:54 UTC
LW: 6 AF: 3
0
AF
in reply to: faul_sname’s comment on: On “first critical tries” in AI alignment
It may be better to think about it that way, yes—in some cases, at least.
Probably it makes sense to throw in some more variables.
Something like:
- To stand x chance of property p applying to system s, we’d need to apply resources r.
In these terms, [loss of control] is something like [ensuring important properties becomes much more expensive (or impossible)].

Joe Collman 21 Jul 2024 21:19 UTC
7 points
3
in reply to: Richard_Ngo’s comment on: Towards more cooperative AI safety strategies
Do you see this as likely to have been avoidable? How?
I agree that it’s undesirable. Less clear to me that it’s an “own goal”.
Do you see other specific things we’re doing now (or that we may soon do) that seem likely to be future-own-goals?
[all of the below is “this is how it appears to my non-expert eyes”; I’ve never studied such dynamics, so perhaps I’m missing important factors]
I expect that, even early on, e/acc actively looked for sources of long-term disagreement with AI safety advocates, so it doesn’t seem likely to me that [AI safety people don’t emphasize this so much] would have much of an impact.
I expect that anything less than a position of [open-source will be fine forever] would have had much the same impact—though perhaps a little slower. (granted, there’s potential for hindsight bias here, so I shouldn’t say “I’m confident that this was inevitable”, but it’s not at all clear to me that it wasn’t highly likely)
It’s also not clear to me that any narrow definition of [AI safety community] was in a position to prevent some claims that open-source will be unacceptably dangerous at some point. E.g. IIRC Geoffrey Hinton rhetorically compared it to giving everyone nukes quite a while ago.
Reducing focus on [desirable, but controversial, short-term wins] seems important to consider where non-adversarial groups are concerned. It’s less clear that it helps against (proto-)adversarial groups—unless you’re proposing some kind of widespread, strict message discipline (I assume that you’re not).

[EDIT for useful replies to this, see Richard’s replies to Akash above]

Joe Collman 21 Jul 2024 20:45 UTC
4 points
2
in reply to: Richard_Ngo’s comment on: Towards more cooperative AI safety strategies
On your bottom line, I entirely agree—to the extent that there are non-power-seeking strategies that’d be effective, I’m all for them. To the extent that we disagree, I think it’s about [what seems likely to be effective] rather than [whether non-power-seeking is a desirable property].
Constrained-power-seeking still seems necessary to me. (unfortunately)
A few clarifications:
- I guess most technical AIS work is net negative in expectation. My ask there is that people work on clearer cases for their work being positive.
- I don’t think my (or Eliezer’s) conclusions on strategy are downstream of [likelihood of doom]. I’ve formed some model of the situation. One output of the model is [likelihood of doom]. Another is [seemingly least bad strategies]. The strategies are based around why doom seems likely, not (primarily) that doom seems likely.
- It doesn’t feel like “I am responding to the situation with the appropriate level of power-seeking given how extreme the circumstances are”.
  - It feels like the level of power-seeking I’m suggesting seems necessary is appropriate.
  - My cognitive biases push me away from enacting power-seeking strategies.
  - Biases aside, confidence in [power seems necessary] doesn’t imply confidence that I know what constraints I’d want applied to the application of that power.
  - In strategies I’d like, [constraints on future use of power] would go hand in hand with any [accrual of power].
    It’s non-obvious that there are good strategies with this property, but the unconstrained version feels both suspicious and icky to me.
    Suspicious, since [I don’t have a clue how this power will need to be directed now, but trust me—it’ll be clear later (and the right people will remain in control until then)] does not justify confidence.
- To me, you seem to be over-rating the applicability of various reference classes in assessing [(inputs to) likelihood of doom]. As I think I’ve said before, it seems absolutely the correct strategy to look for evidence based on all the relevant reference classes we can find.
  - However, all else equal, I’d expect:
    Spending a long time looking for x, makes x feel more important.
    [Wanting to find useful x] tends to shade into [expecting to find useful x] and [perceiving xs as more useful than they are].
    Particularly so when [absent x, we’ll have no clear path to resolving hugely important uncertainties].
  - The world doesn’t owe us convenient reference classes. I don’t think there’s any way around inside-view analysis here—in particular, [how relevant/significant is this reference class to this situation?] is an inside-view question.
    That doesn’t make my (or Eliezer’s, or …’s) analysis correct, but there’s no escaping that you’re relying on inside-view too. Our disagreement only escapes [inside-view dependence on your side] once we broadly agree on [the influence of inside-view properties on the relevance/significance of your reference classes]. I assume that we’d have significant disagreements there.
    Though it still seems useful to figure out where. I expect that there are reference classes that we’d agree could clarify various sub-questions.
    In many non-AI-x-risk situations, we would agree—some modest level of inside-view agreement would be sufficient to broadly agree about the relevance/significance of various reference classes.

Joe Collman 21 Jul 2024 3:47 UTC
4 points
−2
in reply to: Richard_Ngo’s comment on: Towards more cooperative AI safety strategies
E.g. prioritizing competence means that you’ll try less hard to get “your” person into power. Prioritizing legitimacy means you’re making it harder to get your own ideas implemented, when others disagree.
That’s clarifying. In particular, I hadn’t realized you meant to imply [legitimacy of the ‘community’ as a whole] in your post.
I think both are good examples in principle, given the point you’re making. I expect neither to work in practice, since I don’t think that either [broad competence of decision-makers] or [increased legitimacy of broad (and broadening!) AIS community] help us much at all in achieving our goals.
To achieve our goals, I expect we’ll need something much closer to ‘our’ people in power (where ‘our’ means [people with a pretty rare combination of properties, conducive to furthering our goals]), and increased legitimacy for [narrow part of the community I think is correct].
I think we’d need to go with [aim for a relatively narrow form of power], since I don’t think accumulating less power will work. (though it’s a good plan, to the extent that it’s possible)

Joe Collman 21 Jul 2024 3:33 UTC
6 points
2
on: Towards more cooperative AI safety strategies
First, I think that thinking about and highlighting these kind of dynamics is important.
I expect that, by default, too few people will focus on analyzing such dynamics from a truth-seeking and/or instrumentally-useful-for-safety perspective.
That said:
- It seems to me you’re painting with too broad a brush throughout.
  - At the least, I think you should give some examples that lie just outside the boundary of what you’d want to call [structural power-seeking].
- Structural power-seeking in some sense seems unavoidable. (AI is increasingly powerful; influencing it implies power)
  - It’s not clear to me that you’re sticking to a consistent sense throughout.
    E.g. “That makes AI safety strategies which require power-seeking more difficult to carry out successfully.” seems false in general, unless you mean something fairly narrow by power-seeking.
- An important aspect is the (perceived) versatility of power:
  - To the extent that it’s [general power that could be efficiently applied to any goal], it’s suspicious.
  - To the extent that it’s [specialized power that’s only helpful in pursuing a narrow range of goals] it’s less suspicious.
- Similarly, it’s important under what circumstances the power would become general: if I take actions that can only give me power by routing through [develops principled alignment solution], that would make a stated goal of [develop principled alignment solution] believable; it doesn’t necessarily make some other goal believable—e.g. [...and we’ll use it to create this kind of utopia].
- Increasing legitimacy is power-seeking—unless it’s done in such a way that it implies constraints.
  - That said, you may be right that it’s somewhat less likely to be perceived as such.
  - Aiming for [people will tend to believe whatever I say about x] is textbook power-seeking wherever [influence on x] implies power.
  - We’d want something more like [people will tend to believe things that I say about x, so long as their generating process was subject to [constraints]].
    Here it’s preferable for [constraints] to be highly limiting and clear (all else equal).
- I’d say that “prioritizing competence” begs the question.
  - What is the required sense of “competence”?
    For the most important AI-based decision-making, I doubt that ”...broadly competent, and capable of responding sensibly...” is a high enough bar.
  - In particular, ”...because they don’t yet take AGI very seriously” is not the only reason people are making predictable mistakes.
  - ″...as AGI capabilities and risks become less speculative...”
    Again, this seems too coarse-grained:
    Some risks becoming (much) clearer does not entail all risks becoming (much) clearer.
    Understanding some risks well while remaining blind to others, does not clearly imply safer decision-making, since “responding sensibly” will tend to be judged based on [risks we’ve noticed].

Joe Collman 14 Jul 2024 20:52 UTC
LW: 4 AF: 3
0
AF
in reply to: Rohin Shah’s comment on: On scalable oversight with weak LLMs judging strong LLMs
That’s fair. I agree that we’re not likely to resolve much by continuing this discussion. (but thanks for engaging—I do think I understand your position somewhat better now)
What does seem worth considering is adjusting research direction to increase focus on [search for and better understand the most important failure modes] - both of debate-like approaches generally, and any [plan to use such techniques to get useful alignment work done].
I expect that this would lead people to develop clearer, richer models.
Presumably this will take months rather than hours, but it seems worth it (whether or not I’m correct—I expect that [the understanding required to clearly demonstrate to me that I’m wrong] would be useful in a bunch of other ways).

Joe Collman 13 Jul 2024 23:58 UTC
LW: 2 AF: 1
0
AF
in reply to: Rohin Shah’s comment on: On scalable oversight with weak LLMs judging strong LLMs
“[regardless of the technical work you do] there will always be some existentially risky failures left, so if we proceed we get doom...
I’m claiming something more like “[given a realistic degree of technical work on current agendas in the time we have], there will be some existentially risky failures left, so if we proceed we’re highly likely to get doom.
I’ll clarify more below.
Otherwise even in a “free-for-all” world, our actions do influence odds of success, because you can do technical work that people use, and that reduces p(doom).
Sure, but I mostly don’t buy p(doom) reduction here, other than through [highlight near-misses] - so that an approach that hides symptoms of fundamental problems is probably net negative.
In the free-for-all world, I think doom is overdetermined, absent miracles ^[1]- and [significantly improved debate setup] does not strike me as a likely miracle, even after I condition on [a miracle occurred].
Factors that push in the other direction:
- I can imagine techniques that reduce near-term widespread low-stakes failures.
  - This may be instrumentally positive if e.g. AI is much better for collective sensemaking than otherwise it would be (even if that’s only [the negative impact isn’t as severe]).
  - Similarly, I can imagine such techniques mitigating the near-term impact of [we get what we measure] failures. This too seems instrumentally useful.
- I do accept that technical work I’m not too keen on may avoid some early foolish/embarrassing ways to fail catastrophically.
  - I mostly don’t think this helps significantly, since we’ll consistently hit doom later without a change in strategy.
  - Nonetheless, [don’t be dead yet] is instrumentally useful if we want more time to change strategy, so avoiding early catastrophe is a plus.
- [probably other things along similar lines that I’m missing]
But I suppose that on the [usefulness of debate (/scalable oversight techniques generally) research], I’m mainly thinking: [more clearly understanding how and when this may fail catastrophically, and how we’d robustly predict this] seems positive, whereas [show that versions of this technique get higher scores on some benchmarks] probably doesn’t.
Even if I’m wrong about the latter, the former seems more important.
Granted, it also seems harder—but I think that having a bunch of researchers focus on it and fail to come up with any principled case is useful too (at least for them).
If we ignore the misunderstanding part then I’m at << 1% probability on “we build transformative AI using GSA with level 6 / level 7 specifications in the nearish future”.
(I could imagine a pause on frontier AI R&D, except that you are allowed to proceed if you have level 6 / level 7 specifications; and those specifications are used in a few narrow domains. My probability on that is similar to my probability on a pause.)
Agreed. This is why my main hope on this routes through [work on level ⁶⁄₇ specifications clarifies the depth and severity of the problem] and [more-formally-specified ⁶⁄₇ specifications give us something to point to in regulation].
(on the level 7, I’m assuming “in all contexts” must be an overstatement; in particular, we only need something like ”...in all contexts plausibly reachable from the current state, given that all powerful AIs developed by us or our AIS follow this specification or this-specification-endorsed specifications”)
Clarifications I’d make on my [doom seems likely, but not inevitable; some technical work seems net negative] position:
- If I expected that we had 25 years to get things right, I think I’d be pretty keen on most hands-on technical approaches (debate included).
- Quite a bit depends on the type of technical work. I like the kind of work that plausibly has the property [if we iterate on this we’ll probably notice all catastrophic problems before triggering them].
- I do think there’s a low-but-non-zero chance of breakthroughs in pretty general technical work. I can’t rule out that ARC theory come up with something transformational in the next few years (or that it comes from some group that’s outside my current awareness).
- I’m not ruling out an [AI assistants help us make meaningful alignment progress] path—I currently think it’s unlikely, not impossible.
  - However, here I note that there’s a big difference between:
    The odds that [solve alignment with AI assistants] would work if optimally managed.
    The odds that it works in practice.
  - I worry that researchers doing technical research tend to have the the former in mind (implicitly, subconsciously) - i.e. the (implicit) argument is something like “Our work stands a good chance to unlock a winning strategy here”.
  - But this is not the question—the question is how likely it is to work in practice.
    (even conditioning on not-obviously-reckless people being in charge)
  - It’s guesswork, but on [does a low-risk winning strategy of this form exist (without a huge slowdown)?] I’m perhaps 25%. On [will we actually find and implement such a strategy, even assuming the most reckless people aren’t a factor], I become quite a bit more pessimistic—if I start to say “10%”, I recoil at the implied [40% shot at finding and following a good enough path if one exists].
    Of course a lot here depends on whether we can do well enough to fail safely. Even a 5% shot is obviously great if the other 95% is [we realize it’s not working, and pivot].
- However, since I don’t see debate-like approaches as plausible in any direct-path-to-alignment sense, I’d like to see a much clearer plan for using such methods as stepping-stones to (stepping stones to...) a solution.
  - In particular, I’m interested in the case for [if this doesn’t work, we have principled reasons to believe it’ll fail safely] (as an overall process, that is—not on each individual experiment).
    When I look at e.g. Buck/Ryan’s outlined iteration process here,^[2] I’m not comforted on this point: this has the same structure as [run SGD on passing our evals], only it’s [run researcher iteration on passing our evals]. This is less bad, but still entirely loses the [evals are an independent check on an approach we have principled reasons to think will work] property.
    On some level this kind of loop is unavoidable—but having the “core workflow” of alignment researchers be [tweak the approach, then test it against evals] seems a bit nuts.
    Most of the hope here seems to come from [the problem is surprisingly (to me) easy] or [catastrophic failure modes are surprisingly (to me) sparse].
1. ^
  In the sense that [someone proves the Riemann hypothesis this year] would be a miracle.
2. ^
  I note that it’s not clear they’re endorsing this iteration process—it may just be that they expect it to be the process, so that it’s important for people to be thinking in these terms.

Joe Collman 13 Jul 2024 8:59 UTC
LW: 2 AF: 1
0
AF
in reply to: Rohin Shah’s comment on: On scalable oversight with weak LLMs judging strong LLMs
Not going to respond to everything
No worries at all—I was aiming for [Rohin better understands where I’m coming from]. My response was over-long.
E.g. presumably if you believe in this causal arrow you should also believe [higher perceived risk] --> [actions that decrease risk]. But if all building-safe-AI work were to stop today, I think this would have very little effect on how fast the world pushes forward with capabilities.
Agreed, but I think this is too coarse-grained a view.
I expect that, absent impressive levels of international coordination, we’re screwed. I’m not expecting [higher perceived risk] --> [actions that decrease risk] to operate successfully on the “move fast and break things” crowd.
I’m considering:
- What kinds of people are making/influencing key decisions in worlds where we’re likely to survive?
  - How do we get those people this influence? (or influential people to acquire these qualities)
  - What kinds of situation / process increase the probability that these people make risk-reducing decisions?
I think some kind of analysis along these lines makes sense—though clearly it’s hard to know where to draw the line between [it’s unrealistic to expect decision-makers/influencers this principled] and [it’s unrealistic to think things may go well with decision-makers this poor].
I don’t think conditioning on the status-quo free-for-all makes sense, since I don’t think that’s a world where our actions have much influence on our odds of success.

I agree that reference classes are often terrible and a poor guide to the future, but often first-principles reasoning is worse (related: 1, 2).
Agreed (I think your links make good points). However, I’d point out that it can be true both that:
1. Most first-principles reasoning about x is terrible.
2. First-principles reasoning is required in order to make any useful prediction of x. (for most x, I don’t think this holds)
You’ve listed a bunch of claims about AI, but haven’t spelled out why they should make us expect large risk compensation effects
I think almost everything comes down to [perceived level of risk] sometimes dropping hugely more than [actual risk] in the case of AI. So it’s about the magnitude of the input.
- We understand AI much less well.
- We’ll underestimate a bunch of risks, due to lack of understanding.
  - We may also over-estimate a bunch, but the errors don’t cancel: being over-cautious around fire doesn’t stop us from drowning.
- Certain types of research will address [some risks we understand], but fail to address [some risks we don’t see / underestimate].
- They’ll then have a much larger impact on [our perception of risk] than on [actual risk].
- Drop in perceived risk is much larger than the drop in actual risk.
- In most other situations, this isn’t the case, since we have better understanding and/or adaptive feedback loops to correct risk estimates.
It depends hugely on the specific stronger safety measure you talk about. E.g. I’d be at < 5% on a complete ban on frontier AI R&D (which includes academic research on the topic). Probably I should be < 1%, but I’m hesitant around such small probabilities on any social claim.
That’s useful, thanks. (these numbers don’t seem foolish to me—I think we disagree mainly on [how necessary are the stronger measures] rather than [how likely are they])
Hmm, then I don’t understand why you like GSA more than debate, given that debate can fit in the GSA framework (it would be a level 2 specification by the definitions in the paper).
Oh sorry, I should have been more specific—I’m only keen on specifications that plausibly give real guarantees: level 6(?) or 7. I’m only keen on the framework conditional on meeting an extremely high bar for the specification.
If that part gets ignored on the basis that it’s hard (which it obviously is), then it’s not clear to me that the framework is worth much.
I suppose I’m also influenced by the way some of the researchers talk about it—I’m not clear how much focus Davidad is currently putting on level ⁶⁄₇ specifications, but he seems clear that they’ll be necessary.

Joe Collman 11 Jul 2024 8:32 UTC
LW: 13 AF: 5
8
AF
in reply to: Rohin Shah’s comment on: On scalable oversight with weak LLMs judging strong LLMs
[apologies for writing so much; you might want to skip/skim the spoilered bit, since it seems largely a statement of the obvious]
Do you agree that in many other safety fields, safety work mostly didn’t think about risk compensation, and still drove down absolute risk?
Agreed (I imagine there are exceptions, but I’d be shocked if this weren’t usually true).
[I’m responding to this part next, since I think it may resolve some of our mutual misunderstanding]
It seems like your argument here, and in other parts of your comment, is something like “we could do this more costly thing that increases safety even more”. This seems like a pretty different argument; it’s not about risk compensation (i.e. when you introduce safety measures, people do more risky things), but rather about opportunity cost (i.e. when you introduce weak safety measures, you reduce the will to have stronger safety measures). This is fine, but I want to note the explicit change in argument; my earlier comment and the discussion above was not trying to address this argument.
Perhaps we’ve been somewhat talking at cross purposes then, since I certainly consider [not acting to bring about stronger safety measures] to be within the category of [doing more risky things].
It fits the pattern of [lower perceived risk] --> [actions that increase risk].
For clarity, risky actions I’m thinking about would include:
- Not pushing for stronger safety measures.
- Not pushing for international coordination.
- Pushing forward with capabilities more quickly.
- Placing too much trust in systems off distribution.
- Placing too much trust in the output of [researchers using systems to help solve alignment].
- Not spending enough on more principled (by my lights) alignment research.
Some of these might be both [opportunity cost] and [risk compensation].
I.e.:
- We did x.
- x reduced perceived risk.
- A precondition of y was greater-than-current perceived risk.
- Now we can’t do y.
- [Not doing y] is risky.
If you do agree with that, what makes AI different from those cases? (The arguments you give seem like very general considerations that apply to other fields as well.)
Mostly for other readers, I’m going to spoiler my answer to this: my primary claim is that more people need to think about and answer these questions themselves, so I’d suggest that readers take a little time to do this. I’m significantly more confident that the question is important, than that my answer is correct or near-complete.
I agree that the considerations I’m pointing to are general.
I think the conclusions differ, since AI is different.
First a few clarifications:
- I’m largely thinking of x-risk, not moderate disasters.
- On this basis, I’m thinking about loss-of-control, rather than misuse.
- I think if we’re talking about e.g. [misuse leads to moderate disaster], then other fields and prior AI experience become much more reasonable reference classes—and I’d expect things like debate to be net positive here.
- My current guess/estimate that debate research is net negative is largely based on x-risk via loss-of-control. (again, I may be wrong—but I’d like researchers to do serious work on answering this question)
How I view AI risk being different:
1. First another clarification: the risk-increasing actions I’m worried about are not [individual user uses AI incautiously] but things like [AI lab develops/deploys models incautiously], and [governments develop policy incautiously] - so we shouldn’t be thinking of e.g. [people driving faster with seatbelts], but e.g. [car designers / regulators acting differently once seatbelts are a thing].
  To be clear, I don’t claim that seatbelts were negative on this basis.
2. Key differences:
  We can build AI without understanding it.
  People and systems are not used to this.
  Plausible subconscious heuristics that break here:
  [I can build x] and [almost no-one understands x better than me], [therefore, I understand x pretty well].
  [I can build x] and [I’ve worked with x for years], [therefore I have a pretty good understanding of the potential failure modes].
  [We can build x], [therefore we should expect to be able to describe any likely failure modes pretty concretely/rigorously]
  [I’ve discussed the risks of x with many experts who build x], [therefore I have a decent high-level understanding of potential failures]
  [I’m not afraid], [therefore a deadly threat we’re not ready to deal with must be unlikely]. (I agree it’s unlikely now)
  ...
  One failure of AI can be unrecoverable.
  In non-AI cases there tend to be feedback loops that allow adjustments both in the design, and in people’s risk estimates.
  Of course such loops exist for some AI failure modes.
  This makes the unilateralist’s curse aspect more significant: it’s not the average understanding or caution that matters. High variance is a problem.
  Knowing how this impacts risk isn’t straightforward, since it depends a lot on [the levels of coordination we expect] and [the levels of coordination that seem necessary].
  My expectation is that we’re just screwed if the most reckless organizations are able to continue without constraint.
  Therefore my concern is focused on worlds where we achieve sufficient coordination to make clearly reckless orgs not relevant. (and on making such worlds more likely)
  Debate-style approaches seem a good fit for [this won’t actually work, but intelligent, well-informed, non-reckless engineers/decision-makers might think that it will]. I’m concerned about those who are somewhat overconfident.
  The most concerning AI is very general.
  This leads to a large space of potential failure modes. (both in terms of mechanism, and in terms of behaviour/outcome)
  It is hard to make a principled case that we’ve covered all serious failure modes (absent a constructive argument based on various worst-case assumptions).
  The most concerning AI will be doing some kind of learned optimization. (in a broad, John Wentworthy sense)
  Pathogens may have the [we don’t really understand them] property, but an individual pathogen isn’t going to be general, or to be doing significant learned optimization. (if and when this isn’t true, then I’d be similarly wary of ad-hoc pathogen safety work)
  New generations of AI can often solve qualitatively different problems from previous generations. With each new generation, we’re off distribution in a much more significant sense than with new generations of e.g. car design.
  Unknown unknowns may arise internally due to the complex systems nature of AI.
  In cases we’re used to, the unknown unknowns tend to come from outside: radically unexpected things may happen to a car, but probably not directly due to the car.
I’m sure I’m missing various relevant factors.
Overall, I think straightforward reference classes for [safety impact of risk compensation] tell us approximately nothing—too much is different.
I’m all for looking for a variety of reference classes: we need all the evidence we can get.
However, I think people are too ready to fall back on the best reference classes they can find—even when they’re terrible.
Briefly on opportunity cost arguments, the key factors are (a) how much will is there to pay large costs for safety, (b) how much time remains to do the necessary research and implement it, and (c) how feasible is the stronger safety measure.
This seems to miss the obvious: (d) how much safety do the weak vs strong measures get us?
I expect that you believe work on debate (and similar weak measures) gets us significantly more than I do. (and hopefully you’re correct!)
Anyway for now let’s just say that I’ve thought about these three factors and think it isn’t especially realistic to expect that we can get stronger safety measures, and as a result I don’t see opportunity cost as a big reason not to do the safety work we currently do.
Given that this seems highly significant, can you:
- Quantify “it isn’t especially realistic”—are we talking [15% chance with great effort], or [1% chance with great effort]?
- Give a sense of reasons you expect this.
  - Is [because we have a bunch of work on weak measures] not a big factor in your view?
  - Or is [isn’t especially realistic] overdetermined, with [less work on weak measures] only helping conditional on removal of other obstacles?
I’d also note that it’s not only [more progress on weak measures] that matters here, but also the signal sent by pursuing these research directions.
If various people with influence on government see [a bunch of lab safety teams pursuing x safety directions] I expect that most will conclude: “There seems to be a decent consensus within the field that x will get us acceptable safety levels”, rather than “Probably x is inadequate, but these researchers don’t see much hope in getting stronger-than-x measures adopted”.
I assume that you personally must have some constraints on the communication strategy you’re able to pursue. However, it does seem highly important that if the safety teams at labs are pursuing agendas based on [this isn’t great, but it’s probably the best we can get], this is clearly and loudly communicated.
Similarly, the theory of change you cite for your examples seems to be “discovers or clarifies problems that shows that we don’t have a solution” (including for Guaranteed Safe AI and ARC theory, even though in principle those could be about building safe AI systems). So as far as I can tell, the disagreement is really that you think current work that tries to provide a specific recipe for building safe AI systems is net negative, and I think it is net positive.
This characterization is a little confusing to me: all of these approaches (ARC / Guaranteed Safe AI / Debate) involve identifying problems, and, if possible, solving/mitigating them.
To the extent that the problems can be solved, then the approach contributes to [building safe AI systems]; to the extent that they cannot be solved, the approach contributes to [clarifying that we don’t have a solution].
The reason I prefer GSA / ARC is that I expect these approaches to notice more fundamental problems. I then largely expect them to contribute to [clarifying that we don’t have a solution], since solving the problems probably won’t be practical, and they’ll realize that the AI systems they could build won’t be safe-with-high-probability.
I expect scalable oversight (alone) to notice a smaller set of less fundamental problems—which I expect to be easier to fix/mitigate. I expect the impact to be [building plausibly-safe AI systems that aren’t safe (in any robustly scalable sense)].
Of course I may be wrong, if the fundamental problems I believe exist are mirages (very surprising to me) - or if the indirect [help with alignment research] approach turns out to be more effective than I expect (fairly surprising, but I certainly don’t want to dismiss this).
I do still agree that there’s a significant difference in that GSA/ARC are taking a worst-case-assumptions approach—so in principle they could be too conservative. In practice, I think the worst-case-assumption approach is the principled way to do things given our lack of understanding.
I think [~~resolving~~ the existence of this uncertainty is] an important issue to notice when considering research directions
Why? It doesn’t seem especially action guiding, if we’ve agreed that it’s not high value to try to resolve the uncertainty (which is what I take away from your (1)).
(it’s not obvious to me that it’s not high value to try to resolve this uncertainty, but it’s plausible based on your prior comments on this issue, and I’m stipulating that here)
Here I meant that it’s important to notice that:
- The uncertainty exists and is important. (even though resolving it may be impractical)
- The fact that you’ve been paying little attention to it does not imply that either:
  - Your current estimate is accurate.
  - If your estimate changed, that wouldn’t be highly significant.
I’m saying that, unless people think carefully and deliberately about this, the mind will tend to conclude [my current estimate is pretty accurate] and/or [this variable changing wouldn’t be highly significant] - both of which may be false.
To the extent that these are believed, they’re likely to impact other beliefs that are decision-relevant. (e.g. believing that my estimate of x is accurate tells me something about the character of x and my understanding of it)

Joe Collman 9 Jul 2024 20:13 UTC
LW: 12 AF: 5
4
AF
in reply to: Rohin Shah’s comment on: On scalable oversight with weak LLMs judging strong LLMs
Thanks for the response. I realize this kind of conversation can be annoying (but I think it’s important).
[I’ve included various links below, but they’re largely intended for readers-that-aren’t-you]
I don’t see why this isn’t a fully general counterargument to alignment work. Your argument sounds to me like “there will always be some existentially risky failures left, so if we proceed we will get doom. Therefore, we should avoid solving some failures, because those failures could help build political will to shut it all down”.
(Thanks for this too. I don’t endorse that description, but it’s genuinely useful to know what impression I’m creating—particularly when it’s not what I intended)
I’d rephase my position as:
- There’ll always be some risk of existential failure. We want to reduce the total risk. Each step we take forward is a tradeoff: we accept some risk on that step, and (hopefully) reduce future risk by a greater amount.
  - Risk might be reduced through:
    Increased understanding.
    Clear evidence that our current understanding isn’t sufficient. (helpful for political will, coordination...)
    [various other things]
- I am saying “we might get doom”.
  - I think the odds are high primarily because I don’t expect we’ll get a [mostly safe setup that greatly reduces our odds of [trying something plausible that causes catastrophe]]. (I may be wrong here—hopefully so).
- I am not saying “we should not do safety work”; I’m saying “risk compensation needs to be a large factor in deciding which safety work to do”, and “I think few researchers take risk compensation sufficiently seriously”.
  - To a first approximation, I’d say the odds of [safety work on x turns out negative due to risk compensation] is dependent on how well decision-makers (or ‘experts’ they’re listening to) are tracking what we don’t understand. Specifically, what we might expect to be different from previous experience and/or other domains, and how significant these differences are to outcomes.
    Risk is highest when we believe that we understand more than we do understand.
    There’s also a unilateralist’s curse issue here: it matters if there are any dangerously overconfident actors in a position to take action that’ll increase risk. (noting that [slightly overconfident] may be [dangerously overconfident] in this context)
    This matters in considering the downstream impact of research. I’d be quite a bit less worried about debate research if I only had to consider [what will researchers at least as cautious as Rohin do with this?]. (though still somewhat worried :))
    See also Critch’s thoughts on the need for social models when estimating impact.
  - We tend to have a distorted picture of our own understanding, since there’s a strong correlation between [we understand x well] and [we’re able to notice x and express it clearly].
    There’s a kind of bias/variance tradeoff here: if we set our evidential standards such that we don’t focus on vague/speculative/indirect/conceptual arguments, we’ll reduce variance, but risk significant sampling bias.
    Similarly, conclusions downstream of availability-heuristic-triggered thoughts will tend to be disproportionately influenced by the parts of the problem we understand (at least well enough to formulate clear questions).
    I expect that some researchers actively compensate for this in their conscious deliberation, but that it’s very hard to get our intuitions to compensate appropriately.
- [[EDIT: Oh, and when considering downstream impact, risk compensation etc, I think it’s hugely important that most decision-makers and decision-making systems have adapted to a world where [those who can build x understand x] holds. A natural corollary here is [the list of known problems that concern [operation of the system itself] is complete].
  - This implicit assumption underlies risk management approaches, governance structures and individuals’ decision-making processes.
  - That it’s implicit makes it much harder to address, since there aren’t usually “and here we assume that those who can build x understand x” signposts.
  - It might be viable to re-imagine risk-management such that this is handled.
  - It’s much less likely that we get to re-imagine governance structures, or the cognition of individual decision-makers.]]
Again, I don’t think it’s implausible that debate-like approaches come out better than various alternatives after carefully considering risk compensation. I do think it’s a serious error not to spend quite a bit of time and effort understanding and reducing uncertainty on this.
Possibly many researchers do this, but don’t have any clean, legible way to express their process/conclusions. I don’t get that impression: my impression is that arguments along the lines I’m making tend to be perceived as fully general counter-arguments and dismissed (whether they come from outside, or from the researchers themselves).
What’s an example of alignment work that you think is net positive with the theory of change “this is a better way to build future powerful AI systems”?
I’m not sure how direct you intend “this is a better way...” (vs e.g. “this will build the foundation of better ways...”). I guess I’d want to reframe it as “This is a better process by which to build future powerful AI systems”, so as to avoid baking in a level of concreteness before looking at the problem.
That said, the following seem good to me:
- Davidad’s stuff—e.g. Towards Guaranteed Safe AI.
  - I’m not without worries here, and I think “Guaranteed safe”, “Provably safe” are silly, unhelpful overstatements—but the broad direction seems useful.
    That said, the part I’m most keen on would be the “safety specification”, and what I’d expect here is something like [more work on this clarifies that it’s a very hard problem we’re not close to solving]. (noting that [this doesn’t directly cause catastrophe] is an empty ‘guarantee’)
- ARC theory’s stuff.
  - I’m not sure you’d want to include this??
  - I’m keen on it because:
    It seems likely to expand understanding—to uncover new problems.
    It may plausibly be sufficiently general—closer to the [fundamental fix] than the [patch this noticeably undesirable behaviour] end of the spectrum.
- Evan’s red-teaming stuff (e.g. model organisms of misalignment).
  - This seems net positive, and I think [finding a thousand ways not to build an AI] is part of a good process.
    I do still worry that there’s a path of [find concrete problem] → [make concrete problem a target for research] → [fix concrete problems we’ve found, without fixing the underlying issues].
    The ideal situation from [mitigate risk compensation] perspective is:
    Have as much understanding and clarity of fundamental problems as possible.
    Have enough clear, concrete examples to make the need for caution clear to decision-makers.
    To the extent that Evan et al achieve (1), I’m unreservedly in favour.
    On (2) the situation is a bit less clear: we’re in more danger when there are serious problems for which we’re unable to make a clear/concrete case.
    And, again, my conclusion is not “never expose new concrete problems”, but rather “consider the downsides of doing so when picking a particular line of research”.
- A bunch of foundational stuff you probably also wouldn’t want to include under this heading (singular learning theory, natural abstractions, various agent-foundations-shaped things, perhaps computational mechanics (??))).
(I’m probably not going to engage with perspectives that say all current [alignment work towards building safer future powerful AI systems] is net negative, sorry. In my experience those discussions typically don’t go anywhere useful.)
Not a problem. However, I’d want to highlight the distinction between:
1. Substantial efforts to resolve this uncertainty haven’t worked out.
2. Resolving this uncertainty isn’t important.
Both are reasonable explanations for not putting much further effort into such discussions.
However, I worry that the system-1s of researchers don’t distinguish between (1) and (2) here—so that the impact of concluding (1) will often be to act as if the uncertainty isn’t important (and as if the implicit corollaries hold).
I don’t see an easy way to fix this—but I think it’s an important issue to notice when considering research directions. Not only [consider this uncertainty when picking a research direction], but also [consider that your system 1 is probably underestimating the importance of this uncertainty (and corollaries), since you haven’t been paying much attention to it (for understandable reasons)].

Joe Collman 9 Jul 2024 8:28 UTC
LW: 4 AF: 2
0
AF
in reply to: peterbarnett’s comment on: On scalable oversight with weak LLMs judging strong LLMs
Sure, linking to that seems useful, thanks.
That said, I’m expecting that the crux isn’t [can a debate setup work for arbitrarily powerful systems?], but rather e.g. [can it be useful in safe automation of alignment research?].
For something like the latter, it’s not clear to me that it’s not useful.
Mainly my pessimism is about:
- Debate seeming not to address the failure modes I’m worried about—e.g. scheming.
- Expecting [systems insufficiently capable to cause catastrophe] not to radically (>10x) boost the most important research on alignment. (hopefully I’m wrong!)
  - As a result, expecting continued strong pressure to make systems more capable, making [understand when a given oversight approach will fail catastrophically] very important.
- No research I’m aware of seeming likely to tell us when debate would fail catastrophically. (I don’t think the Future work here seems likely to tell us much about catastrophic failure)
- No research I’m aware of making a principled case for [it’s very unlikely that any dangerous capability could be acquired suddenly]. (I expect such thresholds to be uncommon, but to exist)
- Seeing no arguments along the lines of [We expect debate to give us clearer red flags than other approaches, and here’s why...] or [We expect debate-derived red flags are more likely to lead to a safe response, rather than an insufficiently general fix that leaves core problems unaddressed].
  - This is not to say that no such arguments could exist.
  - I’m very interested in the case that could be made here.
Of course little of this is specific to debate. Nor is it clear to me that debate is worse than alternatives in these respects—I just haven’t seen an argument that it’s better (on what assumptions; in which contexts).
I understand that it’s hard to answer the questions I’d want answered.
I also expect that working on debate isn’t the way to answer them—so I think it’s fine to say [I currently expect debate to be a safer approach than most because … and hope that research directions x and y will shed more light on this]. But I’m not clear on people’s rationale for the first part—why does it seem safer?

Joe Collman

First some points of agreement:

Highlights of the below:

More details:

On “negative alignment taxes”: