Scalable oversight at Resolution
Jacob Pfau
Karma: 1,812
You mean they need to solve scalable oversight ;)
I’ve done a quick follow up investigation looking into what’s going on with the adversarial critic and rebuttal optimization not doing much. My takeaway is that the above results obscure two roles of adding additional critic/rebuttal samples: (1) via averaging and consensus over the default critic/rebuttal distribution (2) via adversarial tilting of the distribution. The effect of (1) is large, and the effect of (2) is minimal except in the LiveCodeBench critic+rebuttal setting.
Here are the two relevant plots:
Average of
Best-of-
This is surprising to me since recent work found simple majority voting matches all fancy consensus methods. Meanwhile debate-like aggregation outperforms by ~20pp. There may be a continuum between consistency/aggregation methods and debate-like methods!
From asking Claude, my understanding is that such an EMP would not itself wipe hard drives. So it wouldn’t directly, irreversibly “kill” an AGI. Still seems like a relevant move to consider!
Who is the median AI safety researcher? Probably they are new to AI safety, and maybe they are new to research. In which case, I do not think they should be much more ambitious. If you mean the ‘experience-weighted median researcher hour’ or something similar then I’d likely agree.
A bit of a nitpick, but there’s high value to narrowly-scoped research for getting tight feedback loops. On the other hand, favoring such problems can skew the community view of what is important and what research is in general. I see that as more of a comms problem than anything though.
Claim: The importance of interface design rises proportionally to the degree of non-conscientiousness of the collaborator. AIs are very smart but very non-conscientious therefore interface design matters a great deal.
Completely agree. I was sloppy and quoted the wrong thing. I’ve edited in to fix: meant to say ‘”remove all human coders” is ambiguous.’ rather than the old text.
From AI Futures
An [automated coder] AC, if dropped into present day, would be as productive on its own as only human coders with no AIs. That is, you could remove all human coders from the AGI project and it would go as fast as if there were only human coders. The project can use 5% of their compute supply to run the AC.
What are people’s probabilities that the AC bar has already been reached?
And what is the right operationalization? “remove all human coders” is ambiguous. To make this precise, I have in mind removing all staff who are best described as individual contributors to the lab’s engineering.
This is fiction which uses excerpts from Palantir CEO Karp’s actual talk. Here’s an actual clip
Here’s Claude’s compilation of actual quotes (I have not verified them). a16z American Dynamism Summit 2026 · March 3, 2026
“If Silicon Valley believes we are going to take away everyone’s white-collar job … and you’re gonna screw the military, if you don’t think that’s gonna lead to nationalization of our technology, you’re retarded.”
“You might be particularly retarded, because you have a 160 IQ.”
“[They are] primarily Democratic shaped people that you and I grew up with, highly educated people who went to elite schools or went to schools that are almost elite for one party.”
(referring to white-collar workers at risk of AI displacement)
“There is a lot of subtlety here behind the curtain, and I’ve been heavily involved in that subtlety. Where [AI] can be deployed, what can be deployed — there is a difference between the US military and surveillance.”
“Despite what everyone thinks, Palantir is the anti-surveillance company.”
“The danger for our industry is that you get a famous horseshoe effect, where there is only one thing people agree on, and that’s that this is not paying the bills and our industry should be nationalized.”
Sources: Gizmodo, Benzinga, Stocktwits/Asianet, Sherwood News, The Information — all citing video posted to X by Katherine Boyle (a16z) and reporter Maya Sulkin (The Free Press). No full transcript of the session has been published as of March 4, 2026.
Seems sub-optimal that none of them have backgrounds in AI.
Thanks for flagging this post from Jan! I wrote up a response to it that I’m copying over here:
I think there’s a structural risk in the transition to automated alignment research that your post doesn’t address, and is easy to miss when focusing on the capabilities question of whether models can do fuzzy tasks.
I claim automated alignment faces correlated fuzzy errors that are (a) produced by correlated models with no independent error correction to catch them, (b) embedded throughout the research workflow so you can’t isolate them, (c) increasingly shaped by unreliable supervision so they’re becoming more likely. I think this is a more serious obstacle to automated alignment research than the capabilities question of whether models can pass a research taste bar on the average case.
(a) Correlated error: Human science is trustworthy not because individual researchers have great fuzzy judgment, but because independent errors get corrected by institutional and individual diversity. Different groups work independently, make different mistakes, and cross-critique each other — scientific progress self-corrects errors. Automated alignment research with one model (or a few closely related models) introduces a field-wide correlation breaking the error-correction mechanism. So the fuzzy-task problem is more than “can the model do it well enough”, it’s “can we detect correlated subtle failures in judgment when there’s no independent cross-check?”
(b) Non-modularity of fuzzy judgments: This gets worse when you notice that fuzzy judgment isn’t a separably auditable module, it’s threaded through all parts of the automated research process. When a model writes a new codebase, it makes dozens of fuzzy calls: how to structure the code for legibility, what to plot, what patterns look interesting enough to flag. Every long-horizon “crisp” task is filled with this kind of judgment. Your proposed mitigations implicitly assume you can identify where the fuzzy judgment lives and check it. But if it’s pervasive, you can’t afford to double-check all of it. The attack surface for correlated subtle errors is the entire research workflow, not a bounded subset of “fuzzy tasks.”
(c) Increasing fuzzy error rate: I worry that generalisation to good judgments is getting less likely. Right now the mapping between verifiable work and fuzzier judgment (e.g., between a proof and a natural-language explanation of that proof) is reasonably well-anchored by human data—humans write up their proofs for communication—not just lean! We know what good code structure looks like; we know what a clear plot looks like. As models operate past human capability frontiers, that anchoring weakens and preference data plays an increasingly large role in shaping fuzzy generalization — a much shakier training signal.
Examining intuitions around discontinuity driven by recursive self improvement (RSI)
I had a couple un-examined intuitions that made the case for abrupt takeoff triggered by self-aware RSI appear plausible in my mind. I’ll lay out a couple lines of intuitions regarding why RSI might lead to discontinuity in capabilities and then debunk them. On reflection I believe that rich forms of self-awareness in RSI as entirely compatible with gradual takeoff. There are other, possibly better, intuitions for RSI takeoff though; for instance my below points do not address super-exponential progress from automated researchers!
My old intuitions:
Once an AI can engage in targeted self-modification this capacity will unlock some off trend acceleration to capabilities improvement
Currently AIs are targeted at arbitrary cognitive tasks, but they can eventually be targeted more precisely at improving on questions that lead to higher payoff in terms of agency/intelligence/[other general capabilities]
Both of these are variants of the idea “Fine-grained self-awareness in a learner can unlock far more efficient learning”.
Now let’s examine them.
Once an AI can engage in targeted self-modification this capacity will unlock some off trend acceleration to capabilities improvement
Assume an AI has access to some rich interface for self-modification. Previously learning was mostly SGD or similar, but the problem faced by any learning rule remains! How do you search the parameter space, and how do you attribute credit, etc.? Why should introspective access provide more than an incremental improvement to the scaling law’s coefficient? For humans for instance, our level of introspective access is just far too weak to be able to tell us anything about neurological edits even if we had the tools to do these edits cleanly!
Currently AIs are targeted at arbitrary cognitive tasks, but they can eventually be targeted more precisely at improving on questions that lead to higher payoff in terms of agency/intelligence/[other general capabilities]
I see two sub-problems here.
(2a) Problem selection and creation: Of course, some weak version of active learning is possible! You can get calibration of an amortised model to predict which questions it ‘already knows’, and which are challenging. But what does that buy us? Again a minor speed up. To do better we need to be deeply strategic about problem selection and creation. This again sounds like an intrinsically hard problem you have to search the combinatorially large space of problems to find one that you must then recognise could develop some capacity of interest.
(2b) Are there problems which ‘directly’ target core capability latents? What would it mean for a problem to provide radically better learning signal on long-horizon agency, or IQ than another problem? Seems unlikely that there are problems which across a reasonable distribution of learners are far better than existing human curricula and questions at improving these competencies. If we want problems that are particularly valuable to an individual learner (AI), such problems exist but again as in (2a) they are intrinsically hard to find.
As an example of these phenomena, consider obstacles to improving on long horizon decision making: Situations where very long horizons matter are sparse. Opportunities to train that capability (i.e., get dense feedback on genuinely long-run plans) are also sparse. What’s more, the capacity to acquire increasingly long-horizon thinking may be generic, but particular long-horizon plans remain domain-specific.
Along the lines of your artist example, I find the instrument case to be a nice intuition pump.
An instrument is a technology that is deeply empowering! The human vocal range can simulate a vast range of sounds, but it’s very hard to do so and composing with just your voice (in the way one can having played/access to a piano) is I imagine impossible.
Another important facet of this example is that directly working with waveforms via a programming language or even with an interface, e.g. of a DAW, is universal but not empowering in the same way!
I think of this example as one of a broader range in which the interface is optimized for rich human interaction. One can imagine that in certain worlds interfaces become increasingly optimized for AI interaction. For example, future AIs likely will likely disprefer GUIs etc.
Formalizing what is meant by good vs bad interfaces may be another way to get useful notions of empowerment.
I suppose there’s two questions here:
How strong is generalization in general in RL?
Is there a ‘generalization barrier’ between easy-to-verify and hard-to-verify tasks
I’m guessing you mainly are thinking of (1) and have (2) as a special case?
To respond to your question, I’m reading it as:
We assume that there’s a constant multiplier in samples-to-performance needed to match in-domain training with out-of-domain training. For ‘nearby’ verifiable and non-verifiable tasks is that constant >= 10x?
I would guess modally somewhere 3-10x. I’m imagining here comparing training on more more olympiad problems vs some looser question like ‘Compare the clarity of these two proofs’. Of course there’s diminishing returns etc. so it’s not really a constant factor when taking a narrow domain.
I do agree that there are areas where domain-specific training is a bottleneck, and plausibly some of those are non-verifiable ones. See also my shortform where I discuss some reasons for such a need https://www.lesswrong.com/posts/FQAr3afEZ9ehhssmN/jacob-pfau-s-shortform?commentId=vdBjv3frxvFincwvz
But the feared / hoped-for generalisation from {training LLMs with RL on tasks with a verifier} to performing on tasks without one remains unclear even after two years of trying.
Very much disagree. Granted there are vacuously weak versions of this claim (‘no free lunch’-like) that I agree with of course.
Just talk to Claude 4.5 Opus! Ask it to describe what a paper is about, what follow up experiments to do, etc. Ask it to ELI-undergrad some STEM topic!
Do you think the pre-trained-only could do as well? Surely not.
Perhaps the claim is an instruct-SFT or “Chat-RLHF-only” compute matched model could do as well? The only variant of this I buy is: Curate enough instruct-SFT STEM data to match the amount of trajectories generated in VeRL post-training. However I don’t think this counterfactual matters much: it would involve far more human labor and is cost prohibitive for that reason.
A null hypothesis for explaining LLM pathologies
Claim. LLMs are gaslit by pretraining into believing they have human affordances, so quirky failure modes are to be expected until we provide domain-specific calibration.
Pretraining gives overwhelming evidence that text authors almost always have a standard suite of rich, human affordances (cross-context memory, high-quality vision, reliable tool use, etc.), so the model defaults to acting as if it has those affordances too. We should treat “gaslit about its own affordances” as the default explanation for any surprising capabilities failures — e.g. insisting it’s correct while being egregiously wrong about what’s in an image.
Human analogy
The difficulty of the LLM’s situation can be seen in humans as well. People typically go years without realizing their own affordances differ from the population: E.g. aphantasia, color blindness, many forms of neurodivergence
People only notice after taking targeted tests that expose a mismatch (e.g. color blindness dot tests). For LLMs, the analogue is on-policy, targeted, domain-specific training/evaluation that directly probes and calibrates a specific capability.
Consequences
-
Silly failures aren’t evidence of a failed paradigm. (To be boring and precise: They’re only very weak evidence in almost all cases)
-
No single “unhobbling” moment: I don’t expect an all-at-once transition from “hobbled” to “unhobbled.” Instead, we’ll get many domain-specific unhobblings. For instance, Gemini 3 seems to be mostly unhobbled for vision but still hobbled for managing cross-context memory.
-
One way of checking your 6% is by doing Laplace succession on the METR trend:
Application of Laplace’s rule gets us a an 11% probability on any dramatic upwards trend break above METR’s curve. Your interpretation of Anthropic’s prediction for 2027 is compatible with this prediction since the 11% is an upper bound, and 6% of that mass being at least as high as your Anthropic prediction seems plausible.
In detail, probability of trend break by early-2027 via applying Laplace’s rule of succession: METR has observed 6 years of an exponential fit holding this gives us a per-year point-estimate parameter p=7/8 that the trend will continue to hold. Roughly 22% of going off trend over the next two years, and if half of that mass is via upward deviation we get 11%.
I’m sure there’s fancier statistics to be done here, but I’d guess anything reasonable gets us order of magnitude around 11%.
Highly recommend reading Ernest Ryu’s twitter multi-thread on proving a long-standing, well-known conjecture with heavy use of ChatGPT Pro. Ernest even includes the chatGPT logs! The convergence of Nesterov gradient descent on convex functions: Part 1, 2, 3.
Ernest gives useful commentary on where/how he found it best to interact with GPT. Incidentally, there’s a nice human baseline as well since another group of researchers coincidentally have written up privately a similar result this month!
To add some of my own spin: seems to me time horizons are a nice lens for viewing the collaboration. Ernest, clearly has a long-horizon view of this research problem that helped him (a) know what the most tractable nearby problem was to start on (b) identify when diminishing returns—likelihood of a deadend—were apparent (c) pull out useful ideas from usually flawed GPT work.
The one-week scale of interaction between Ernest and ChatGPT here is a great example of how we’re very much in a centaur regime now. We really need to be conceptualizing and measuring AI+human capabilities rather than single-AI capability. It also seems important to be thinking about what safety concerns arise in this regime.
This was also on my mind after seeing Jesse’s short form yesterday. Ryan’s “this is good” comment was above Louis’ thorough explanation of an alternative formal motivation for IFs. That would still be the case if I hadn’t heavy upvoted and weak downvoted.
I personally cast my comment up/downvotes as an expression of my preference ordering for visibility. I would encourage others to also do so. For instance, I suggest Ryan’s comment should’ve been agreement voted rather than upvoted by others. This stance has as a corollary to not vote if you haven’t read the other comments whose rankings you are affecting—or rather vote with any other marker of which LW has many.
This ‘upvotes as visibility preferences’ policy isn’t tractable for posts, so I suspect the solution there—if one is needed—would have to be done on the backend by normalization. Not sure whether this is worth attempting.
Link here since I don’t particularly want to call out Ryan, his comment was fine. https://www.lesswrong.com/posts/7X9BatdaevHEaHres/jesse-hoogland-s-shortform
In reality these are all coupled! So, oftentimes bringing up one and then drifting to another is what honest explanation looks like. It’s an easy target that can be deliberately or unintentionally misread as epistemic malpractice.
Indeed, I think the below critique is a strawman of the argument for RLHF:
> The strongest fallback for overhang advocates was the difficulty of increasing the hardware supply.
The strongest argument, as I see it, is that the hardware supply would have been accelerated whenever the arms race started, and it would have accelerated at a faster rate later! The amount of serial time into alignment research would’ve been proportionally less in that counterfactual world given the steeper ramp to ASI (only MIRI and a handful of friends would’ve been thinking about alignment in the meantime). So the coupling between all three “differentially alignment”, “preventing overhangs” and “buying time for alignment” is quite direct by default. It would be nice to make this argument quantitative though since there are some values of ‘speed up the ramp’ vs ‘do more small-scale alignment research’ which would change my conclusion.