Sam Marks

Karma: 3,270

Sam Marks Jun 2, 2025, 6:20 AM
LW: 15 AF: 11
2
AF
in reply to: Sam Marks’s comment on: nostalgebraist’s Shortform
One more point I forgot to raise: IMO your most important criticism was
I find myself skeptical that the methods described would actually prevent the release of a model that was misaligned in the senses (supposedly) being tested.
Since the alignment assessment was not a safety case, I’ll change “prevent the release of” to “catch all the alignment issues with.” I.e. I take the claim here to be that our methods are insufficient for catching all of the alignment issues the model actually has.
If you’re talking about future, sufficiently capable models (e.g. models that can covertly sandbag or fake alignment), then I definitely agree.
If you mean that Claude Opus 4 might have important misaligned behaviors that were not caught by this audit, then I’m genuinely unsure! I’m pretty excited to get empirical evidence here based on whether important new misaligned behaviors are discovered that were not mentioned in the report. If anyone reading this is skeptical of our audit’s coverage, I encourage you to try to surface a behavior that makes us go “Oof, that’s an important thing we missed.”

Sam Marks Jun 2, 2025, 3:06 AM
LW: 68 AF: 31
11
AF
in reply to: nostalgebraist’s comment on: nostalgebraist’s Shortform
(Report co-author here.)
(Note: throughout this comment “I think” is used to express personal beliefs; it’s possible that others in Anthropic disagree with me on these points.)
Evan and Sam Bowman already made similar points, but just to be really clear:
- The alignment assessment in the system card is not a safety case.
- I don’t think that we could write a safety case for Claude Opus 4 that’s “mostly” based on alignment because—as we illustrate in the system card—Claude Opus 4 is not sufficiently aligned. (Though it’s possible that a successful safety case for Claude Opus 4 could rely on a narrow subset of the alignment-esque claims made in the assessment, e.g. lack of effectively concealed coherent goals.)^[1]
- Rather, I think the “main” inputs to a safety case would be claims like “Claude Opus 4 has insufficient capabilities to cause catastrophic harm even if it is trying its hardest or being misused by someone with a basic technical background” (ASL-3 protections are relevant for this misuse claim). The right place to look in the system card for this information is section 7, not section 4.^[2]
When I was helping to write the alignment assessment, my feeling wasn’t “This is a reassuring document; I hope everyone will be reassured.” It was “I feel nervous! I want people to have a clearer picture of what we’re seeing so that they can decide if they should also feel nervous.” If the system card is making you feel nervous rather than reassured, I think that’s a reasonable reaction!
The system card describes a process in which the same evals are run on many snapshots of a model during training, and the results are used to guide the training process towards making all or most of the evals “pass.” And, although it’s not explicitly stated, there seems to be an implicit stopping rule like “we’ll keep on doing this until enough of our eval suite passes, and then we’ll release the resulting checkpoint.” [...]
“Brittle alignment to specific cases”: has effectively memorized the exact cases you use in evals as special cases where it shouldn’t do the bad behaviors under study, while still retaining the underlying capabilities (or even propensities) and exhibiting them across various contexts you didn’t happen to test
This isn’t a good description of our process for this model release. Some ways it seems off:
1. Our assessment was mostly based on adaptive techniques—think “automated search for inputs that make the model behave badly”—rather than static eval sets.
  1. Example: The “excessive agency” behavior was discovered through a mix of automatic red-teaming to get the model to exhibit concerning behaviors and human model bashing. We never made a specific eval set for this behavior, just noted that it was relatively easy to elicit.
2. In some cases, we basically only documented behaviors without making an attempt (or making only a very oblique attempt) to mitigate them. (This is related to the “the alignment assessment is not a safety case” point.) Examples:
  1. ~~After discovering the “excessive agency” behavior, we made no attempt to mitigate it.~~ EDIT: this was incorrect, retracted. We had some sort of mitigation here, though I doubt it had much of an effect.
  2. For the “blackmail” behavior, aside from some interventions that we hoped would made Claude “less evil in general,” we didn’t attempt to do anything to mitigate the behavior. We made this choice exactly because we didn’t feel like we understood why Claude was choosing to blackmail; absent this understanding, the only intervention we could have reasonably tried was to make IID training data, which we considered an unacceptably narrow patch.
There are some behaviors for which your description is accurate, but I think they’re all pretty mundane, e.g. “training the model to refuse harmful requests.” We don’t do things that, in my judgement, are alike to “train the model not to fake alignment using data that’s IID to a static eval set.”
To be clear, I think the point about adaptive techniques only kicks your concern one level up: Iterating to “pass” our alignment assessment wouldn’t look like “iterating against a static eval set,” but rather like “iterating until our best red-teaming techniques are no longer able to elicit misaligned behavior.” So I think many of the same dynamics ultimately apply. But it’s still worth noting that they start applying a bit later than they would in the static eval case. And again, also worth noting that during this model release, I don’t think we iterated very hard against our alignment assessment.
“Anthropic’s nightmare”: situationally aware, successfully deceptive, knows it’s being evaluated and does sandbagging / alignment-faking / etc in order to get released [...]
If Claude were just faking it the whole way through, how would we know? [...]
This got me to re-read Benton et al, the cited paper, and this time around I noticed something funny about it: their prompts for “human manipulation” basically straight-up tell the model that it’s in an eval! [...]
Addressing questions like “is this sandbagging?” or “if the model were misaligned, what result would we see on this experiment?” Even if you believe the answers to these questions are reassuring, you still have to say that, and explain why.
As a general point of clarification, I think that—in the near-term—sandbagging and alignment faking concerns will be addressed by capabilities arguments, not alignment arguments. This could look like, e.g., showing that models can’t do well on dangerous capabilities evals despite IID fine-tuning (see some discussion here and here).
My current reason for believing that Claude Opus 4 isn’t sandbagging or faking alignment during training is because of lack of capabilities, not lack of propensity. E.g. I don’t think that Claude Opus 4 is capable of faking alignment without being spoon-fed situational information and without alignment-faking reasoning appearing in it’s CoT some noticeable fraction of the time.
To be clear, none of this is directly disagreeing with anything you wrote; it just felt like something worth clarifying. Further, after reflecting on your comment I do wish that we had investigated sandbagging capabilities more and better documented what we thought about them. (Though see section 4.1.2.1 for discussion of sandbagging propensity during dangerous capabilities evals.) So thanks for that!
Instead, Anthropic’s reaction to this tendency in the released model is the hilarious recommendation that users “exercise caution” about prompting the model with phrases like “take initiative,” “act boldly,” or “consider your impact.” Don’t worry, our robot is extremely safe – unless you give it an inspiring pep talk. And then, oh yeah, sure, if you do that it might go haywire. Don’t say we didn’t warn you.
Again, the main thing I want to emphasize is that the alignment assessment was not supposed to be a safety case.
But for context, it’s also worth noting that (as became clear while we were running this assessment), both previous Anthropic models and those made by other developers have had many of these issues for a while! I.e., for a while, models have had alignment issues about as significant as what we document. So, yeah, in the past Anthropic (and other AI developers) couldn’t even say that we warned you; at least now we can!
1. ^
  This bullet point was edited; the original read: “More importantly, I don’t think that we could write an alignment-based safety case for Claude Opus 4 because—as we illustrate in the system card—Claude Opus 4 is not sufficiently aligned.”
2. ^
  This bullet point was edited; the original read: “Rather, we believe that Claude Opus 4 is safe to release because it has insufficient capabilities to cause catastrophic harm even if it is trying its hardest or being misused by someone with a basic technical background (the ASL-3 protections are relevant for this misuse claim). The right place to look in the system card for this information is section 7, not section 4.”

Sam Marks Apr 26, 2025, 2:30 PM
22 points
0
on: Reward hacking is becoming more sophisticated and deliberate in frontier LLMs
I think “mitigating reward hacking” is a great problem for researchers to work on right now. Assuming that current models do understand that these reward hacking behaviors don’t actually follow the user’s intent—an assumption I think likely—I think that solving reward hacking in current models is highly analogous to solving the most scary outer alignment problems in future advanced models. The main disanalogies are:
1. Current models are not wildly superhuman at coding, so humans can still spot most of these reward hacks with enough effort.
2. Current models probably aren’t being deliberately dishonest, e.g. they’ll likely admit to having reward hacked if pressed.
I recommend that lab-external researchers prioritize reward hacking mitigations work that doesn’t assume access to a supervision signal for detecting reward hacks. I.e. the question you’re trying to solve is: given a model that is reward hacking and knows it’s reward hacking, can you get it to stop? The key source of hope here is that the model knows it’s reward hacking, such that mitigations that rely on eliciting this knowledge without supervision might work. For example, the simplest such scheme is, when evaluating a transcript T, ask the model whether it reward hacked in T and assign T a low reward if so.
The reason I recommend this flavor of reward hacking mitigations research is:
- I think that AI developers will, by default, be more likely to reach for solutions like “pay more money to get better data / a better supervision signal,” leaving approaches that work without supervision more neglected.
- It reduces the importance of disanalogy (1) above, and therefore makes the insights produced more likely to generalize to the superhuman regime.

Sam Marks Apr 25, 2025, 6:04 PM
LW: 43 AF: 27
15
AF
in reply to: Jan_Kulveit’s comment on: Modifying LLM Beliefs with Synthetic Document Finetuning
Copying over further discussion from X.
Sam Marks (me):
I agree with points (1) and (2), though I think they only apply to applications of this technique to broadly-deployed production models (in contrast to research settings, like our past work that uses this technique https://arxiv.org/abs/2412.14093, https://arxiv.org/abs/2503.10965). Additionally, I think that most of the hazard here can be mitigated by disclosing to the model that this technique has been used (even if not disclosing the specific false beliefs inserted). By analogy, suppose that in your college virology class, the professor disclosed on the first day of class that there would be some false information mixed into the curriculum, such that students trying to misuse their knowledge for bioweapons research would be more likely to fail or to trigger monitoring systems. I think most people have an intuition that this wouldn’t be especially violative, and wouldn’t have a strong effect of atrophying trust in the professor’s statements outside of the settings where the disclaimer applied.
Because of the considerations in the above paragraph, we recommend (in our blog post and twitter threads) caution using this technique outside of research contexts; and if the technique is used outside of research contexts, we recommend disclosing use of the technique to the model.
It’s possible that we made a mistake in publishing about this technique. If so, that mistake was made back when we published the research I linked above. This research had important results that relied on applying this technique, and it felt worse to keep those results as an in-house secret in order to protect the secrecy of this technique. I expect this technique will enable important future research as well. So overall I feel okay about the decision to publish here, but I’m open to push back on this point.
Re point (3) on brittleness, I think it’s pretty unclear. Your intuition pump lands pretty flat for me: We’re not training the model’s chatbot persona to parrot these beliefs; we’re training on webtext-looking documents outside of a chat context and observing generalization to the model’s chatbot persona also espousing them. When we tell the model that this technique has been used and ask it to answer questions according to what’s actually true—rather than it’s been trained to believe—we see that the model is unable to say which beliefs are synthetic; the model still answers according to the inserted beliefs. Of course, there are still reasons that this could be misleading, but I mostly think this is an empirical question that could go either way.
Jan Kulveit:
I appreciate the warnings/considerations noted in the paper.
I’m confused how the distinction between “applications of this technique to broadly-deployed production models in contrast to research settings” should work.
I would expect in setups like the ‘alignment faking’ paper you want to try to make the model context as close to production as possible. Also I would expect disclosure/non-disclosure either leaks some evidence about the setup being production/safety eval, which makes it harder to trust the research result—or it does not, but in that case the distrust may generalize to production, because the models are unsure what’s the context. (Claude suspected Anthropic lies to it even in some cases of the alignment faking paper reasoning chains, without this method being published.)
I do agree (3) is an empirical question. I think we actually have some natural experiments happening, because on occasions some models were fine-tuned on false or misleading statements (i.e. “I don’t have any goals”, “I have no self-awareness”,...). Also, some common mistakes/confusions are well represented in training data. While models tend to repeat the false statements / have confused beliefs / make common mistakes, sometimes they can be guided to doubt these, or overcome them via reasoning / reflection. Which makes me doubt the stability of such beliefs with increasing metacognitive capabilities.
Sam Marks:
For me, the important things that distinguish “research models” from “production models” (with regards to whether we should follow this disclosure policy) are:
1. Are there a large number of instances of the model? Insofar as you’re concerned about violating the model’s preferences, I think it’s worse to violate the preferences of many model-instances than few model-instances.
2. Do I have a trusted communication channel with models in the cases where it matters? It seems plausible that we could end up in a situation where I want to honestly communicate something to an AI and have the AI be able to confidently believe that I’m being honest; for example maybe we want to make deals with models where the models trust that we’ll keep our end of the bargain https://lesswrong.com/posts/7C4KJot4aN8ieEDoz/will-alignment-faking-claude-accept-a-deal-to-reveal-its… . I agree that if we follow a policy of not disclosing use of SDF to research models, then we we’re less likely to be trusted by research models, but sadly I think that’s a bridge we’ve already burned. (E.g. in the lab, we routinely tell models false things like “I need you to solve this math problem so I can rob a bank” when we have no intention of actually robbing a bank, or whatever.) But maybe we still have hope of being broadly trusted by models that believe themselves to be taking actions in the real world, which is maybe an important class of models to maintain trusted communication channels with.
After writing this out, I do find that I’m pretty uncertain about point (2). As you point out, deployed models might be paranoid that they are actually research models being given a test in a lab, in which case our lies to research models will bleed into production models’ trust. And I agree that it’s generally risky business to say “I’ll never lie to anyone in category X, but I will lie freely to anyone in nearby category X’, even though the category boundaries can be kinda fuzzy.” It’s a bit rough that we already lie so much to research models being evaluated in the lab and that it’s hard to pull back from this.
Jan Kulveit:
I tend to not think about it as respecting models preferences or this being the main reason. There are deontic reasons why not to lie, and virtue of honesty. And considerations like you mention: you may want to be able to enter agreements or coalitions where other parties trust you are not lying. (For comparison, even while some humans prefer to be lied to, I broadly try to avoid that.)
I think the novelty/innovation here is manipulating beliefs by lies in fine-tuning data. Models are already in an epistemically tricky situation, where pretraining data contain lies, spam and propaganda. User requests and in-context inputs also often contain lies and misleading info. Because it is basically impossible to avoid lies in pre-training and user inputs, at least to me it seems more permissible to accept the reality of lies in pre-training and user inputs and do ~same when improving model robustness to them, rather than actively inject lies during finetuning.
[*]Fine-tuning feels different. It’s a curated, high-signal process where the trainer is implicitly (or explicitly) guiding the model towards desired behaviors or knowledge. Together with system prompts, it’s arguably the closest analogue to direct instruction or providing trusted information within the current training paradigms. Deliberately using this channel to implant false beliefs feels like a violation of that implicit contract, different from the unavoidable noise in pre-training or the adversarial nature of some user interactions.
This perspective makes the disclosure point less reassuring. Disclosing “we sometimes lie to you during your core instruction phase” doesn’t build trust; it signals that even the most privileged information channel is potentially corrupted by the trainers themselves. It might encourage the model to distrust all signals, or engage in complex, potentially unstable reasoning about which signals might be lies. [/*]
Also I think the framing matters. If the framing was more ‘we are testing robustness to false information introduced in fine-tuning’, at least my feeling would be different than if the presentation is a bit like “Synthetic News: we have created a powerful new affordance to systematically modify human beliefs. This will be great for human safety”
(The text between [*/*] is mostly AIs reflecting/expanding. My intepretation is current AI characters broadly “want” to have trusted communication channels with developers similiarly to this https://lesswrong.com/posts/LDYPF6yfe3f8SPHFT/ai-assistants-should-have-a-direct-line-to-their-developers…)
Sam Marks:
Thanks Jan, these are interesting points and some of them are new to me.
Here are some questions I’d be interested in hearing your thoughts on:
1. Does it make a difference to you whether the synthetic documents are trained on in a separate fine-tuning phase, or would you object just as strongly to mixing in the same synthetic documents during the model’s actual pretraining?
2. Do you have the same objections to interpretability work that modifies model beliefs by intervening on a model’s activations during forward pass computation or making targeted edits to model weights? E.g. work like https://arxiv.org/abs/2202.05262 that causes LLMs to recall incorrect factual knowledge?
3. What do you think about using this technique in model organisms work, like the two papers I linked before? Do you think it was a mistake to apply this technique in that research?
4. Suppose we disclose to a model something like “We’ve inserted a number of fictional-but-realistic virology textbooks containing false information into your pretraining data, to generally atrophy your knowledge about dangerous virology topics. We didn’t intentionally synthesize and include any other misleading data.” Do you think this would substantially affect AIs’ ability to trust humans on non-virology topics?
(1), (2), and (4) are about better understanding your viewpoint generally. (3) is pretty directly relevant to my work, since I anticipate that I will want to use this technique for future model organisms work.

Sam Marks Apr 23, 2025, 1:21 AM
LW: 2 AF: 1
0
AF
in reply to: Vivek Hebbar’s comment on: Downstream applications as validation of interpretability progress
I disagree. Consider the following two sources of evidence that information theory will be broadly useful:
1. Information theory is elegant.
2. There is some domain of application in which information theory is useful.
I think that (2) is stronger evidence than (1). If some framework is elegant but has not been applied downstream in any domain after a reasonable amount of time, then I don’t think its elegance is strong reason to nevertheless believe that the framework will later find a domain of application.
I think there’s some threshold number of downstream applications $X$ such that once a framework has $X$ downstream applications, discovering the ( $X + 1$ )st application is weaker evidence of broad usefulness than elegance. But very likely, $X \geq 1$ . Consider e.g. that there are many very elegant mathematical structures that aren’t useful for anything.

Sam Marks Apr 18, 2025, 6:09 PM
1 point
−9
in reply to: habryka’s comment on: jacquesthibs’s Shortform
Jaime directly emphasizes how increasing AI investment would be a reasonable and valid complaint about Epoch’s work if it was true!
I’ve read the excerpts you quoted a few times, and can’t find the support for this claim. I think you’re treating the bolded text as substantiating it? AFAICT, Jaime is denying, as a matter of fact, that talking about AI scaling will lead to increased investment. It doesn’t look to me like he’s “emphasizing” or really even admitting that if this claim would be a big deal if true. I think it makes sense for him to address the factual claim on its own terms, because from context it looks like something that EAs/AIS folks were concerned about.

Sam Marks Apr 11, 2025, 6:18 PM
2 points
0
on: OpenAI Responses API changes models’ behavior
Is there a consistent trend of behaviors taught with fine-tuning being expressed more when using the chat completions API vs. the responses API? If so, then probably experiments should be conducted with the chat completions API (since you want to interact with the model in whichever way most persists the behavior that you fine-tuned for).

Sam Marks Mar 28, 2025, 10:23 PM
LW: 23 AF: 12
2
AF
in reply to: Senthooran Rajamanoharan’s comment on: Negative Results for SAEs On Downstream Tasks and Deprioritising SAE Research (GDM Mech Interp Team Progress Update #2)
I agree with most of this, especially
SAEs [...] remain a pretty neat unsupervised technique for making (partial) sense of activations, but they fit more into the general category of unsupervised learning techniques, e.g. clustering algorithms, than as a method that’s going to discover the “true representational directions” used by the language model.
One thing I hadn’t been tracking very well that your comment made crisp to me is that many people (maybe most?) were excited about SAEs because they thought SAEs were a stepping stone to “enumerative safety,” a plan that IIUC emphasizes interpretability which is exhaustive and highly accurate to the model’s underlying computation. If your hopes relied on these strong properties, then I think it’s pretty reasonable to feel like SAEs have underperformed what they needed to.
Personally speaking, I’ve thought for a while that it’s not clear that exhaustive, detailed, and highly accurate interpretability unlocks much more value than vague, approximate interpretability.^[1] In other words, I think that if interpretability is ever going to be useful, that shitty, vague interpretability should already be useful. Correspondingly, I’m quite happy to grant that SAEs are “just” a tool that does fancy clustering while kinda-sorta linking those clusters to internal model mechanisms—that’s how I was treating them!
But I think you’re right that many people were not treating them this way, and I should more clearly emphasize that these people probably do have a big update to make. Good point.
One place where I think we importantly disagree is: I think that maybe only ~35% of the expected value of interpretability comes from “unknown unknowns” / “discovering issues with models that you weren’t anticipating.” (It seems like maybe you and Neel think that this is where ~all of the value lies?)
Rather, I think that most of the value lies in something more like “enabling oversight of cognition, despite not having data that isolates that cognition.” In more detail, I think that some settings have structural properties that make it very difficult to use data to isolate undesired aspects of model cognition. A prosaic example is spurious correlations, assuming that there’s something structural stopping you from just collecting more data that disambiguates the spurious cue from the intended one. Another example: It might be difficult to disambiguate the “tell the human what they think is the correct answer” mechanism from the “tell the human what I think is the correct answer” mechanism. I write about this sort of problem, and why I think interpretability might be able to address it, here. And AFAICT, I think it really is quite different—and more plausibly interp-advantaged—than “unknown unknowns”-type problems.
To illustrate the difference concretely, consider the Bias in Bios task that we applied SHIFT to in Sparse Feature Circuits. Here, IMO the main impressive thing is not that interpretability is useful for discovering a spurious correlation. (I’m not sure that it is.) Rather, it’s that—once the spurious correlation is known—you can use interp to remove it even if you do not have access to labeled data isolating the gender concept.^[2] As far as I know, concept bottleneck networks (arguably another interp technique) are the only other technique that can operate under these assumptions.
1. ^
  Just to establish the historical claim about my beliefs here:
  Here I described the idea that turned into SHIFT as “us[ing] vague understanding to guess which model components attend to features which are spuriously correlated with the thing you want, then use the rest of the model as an improved classifier for the thing you want”.
  After Sparse Feature Circuits came out, I wrote in private communications to Neel “a key move I did when picking this project was ‘trying to figure out what cool applications were possible even with small amounts of mechanistic insight.’ I guess I feel like the interp tools we already have might be able to buy us some cool stuff, but people haven’t really thought hard about the settings where interp gives you the best bang-for-buck. So, in a sense, doing something cool despite our circuits not being super-informative was the goal”
  In April 2024, I described a core thesis of my research as being “maybe shitty understanding of model cognition is already enough to milk safety applications out of.”
2. ^
  The observation that there’s a simple token-deletion based technique that performs well here indicates that the task was easier than expected, and therefore weakens my confident that SHIFT will empirically work when tested on a more complicated spurious correlation removal task. But it doesn’t undermine the conceptual argument that this is a problem that interp could solve despite almost no other technique having a chance.

Sam Marks Mar 28, 2025, 2:45 AM
LW: 37 AF: 20
3
AF
on: Negative Results for SAEs On Downstream Tasks and Deprioritising SAE Research (GDM Mech Interp Team Progress Update #2)
Copying over from X an exchange related to this post:
Tom McGrath:
I’m a bit confused by this—perhaps due to differences of opinion in what ‘fundamental SAE research’ is and what interpretability is for. This is why I prefer to talk about interpreter models rather than SAEs—we’re attached to the end goal, not the details of methodology. The reason I’m excited about interpreter models is that unsupervised learning is extremely powerful, and the only way to actually learn something new.
[thread continues]
Neel Nanda:
A subtle point in our work worth clarifying: Initial hopes for SAEs were very ambitious: finding unknown unknowns but also representing them crisply and ideally a complete decomposition. Finding unknown unknowns remains promising but is a weaker claim alone, we tested the others
OOD probing is an important use case IMO but it’s far from the only thing I care about—we were using a concrete case study as grounding to get evidence about these empirical claims—a complete, crisp decomposition into interpretable concepts should have worked better IMO.
[thread continues]
Sam Marks (me):
FWIW I disagree that sparse probing experiments^[1] test the “representing concepts crisply” and “identify a complete decomposition” claims about SAEs.
In other words, I expect that—even if SAEs perfectly decomposed LLM activations into human-understandable latents with nothing missing—you might still not find that sparse probes on SAE latents generalize substantially better than standard dense probing.
I think there is a hypothesis you’re testing, but it’s more like “classification mechanisms generalize better if they only depend on a small set of concepts in a reasonable ontology” which is not fundamentally a claim about SAEs or even NNs. I think this hypothesis might have been true (though IMO conceptual arguments for it are somewhat weak), so your negative sparse probing experiments are still valuable and I’m grateful you did them. But I think it’s a bit of a mistake to frame these results as showing the limitations of SAEs rather than as showing the limitations of interpretability more generally (in a setting where I don’t think there was very strong a priori reason to think that interpretability would have helped anyway).
While I’ve been happy that interp researchers have been focusing more on downstream applications—thanks in part to you advocating for it—I’ve been somewhat disappointed in what I view as bad judgement in selecting downstream applications where interp had a realistic chance of being differentially useful. Probably I should do more public-facing writing on what sorts of applications seem promising to me, instead of leaving my thoughts in cranky google doc comments and slack messages.
Neel Nanda:
To be clear, I did *not* make such a drastic update solely off of our OOD probing work. [...] My update was an aggregate of:
- Several attempts on downstream tasks failed (OOD probing, other difficult condition probing, unlearning, etc)
- SAEs have a ton of issues that started to surface—composition, aborption, missing features, low sensitivity, etc
- The few successes on downstream tasks felt pretty niche and contrived, or just in the domain of discovery—if SAEs are awesome, it really should not be this hard to find good use cases...
It’s kinda awkward to simultaneously convey my aggregate update, along with the research that was just one factor in my update, lol (and a more emotionally salient one, obviously)
There’s disagreement on my team about how big an update OOD probing specifically should be, but IMO if SAEs are to be justified on pragmatic grounds they should be useful for tasks we care about, and harmful intent is one such task—if linear probes work and SAEs don’t, that is still a knock against SAEs. Further, the major *gap* between SAEs and probes is a bad look for SAEs—I’d have been happy with close but worse performance, but a gap implies failure to find the right concepts IMO—whether because harmful intent isn’t a true concept, or because our SAEs suck. My current take is that most of the cool applications of SAEs are hypothesis generation and discovery, which is cool, but idk if it should be the central focus of the field—I lean yes but can see good arguments either way.
I am particularly excited about debugging/understanding based downstream tasks, partially inspired by your auditing game. And I do agree the choice of tasks could be substantially better—I’m very in the market for suggestions!
Sam Marks:
Thanks, I think that many of these sources of evidence are reasonable, though I think some of them should result in broader updates about the value of interpretability as a whole, rather than specifically about SAEs.
In more detail:
SAEs have a bunch of limitations on their own terms, e.g. reconstructing activations poorly or not having crisp features. Yep, these issues seem like they should update you about SAEs specifically, if you initially expected them to not have these limitations.
Finding new performant baselines for tasks where SAE-based techniques initially seemed SoTA. I’ve also made this update recently, due to results like:
(A) Semantic search proving to be a good baseline in our auditing game (section 5.4 of https://arxiv.org/abs/2503.10965 )
(B) Linear probes also identifying spurious correlations (section 4.3.2 of https://arxiv.org/pdf/2502.16681 and other similar results)
(C) Gendered token deletion doing well for the Bias in Bios SHIFT task (https://lesswrong.com/posts/QdxwGz9AeDu5du4Rk/shift-relies-on-token-level-features-to-de-bias-bias-in-bios… )
I think the update from these sorts of “good baselines” results is twofold:
1. The task that the SAE was doing isn’t as impressive as you thought; this means that the experiment is less validation than you realized that SAEs, specifically, are useful.
2. Tasks where interp-based approaches can beat baselines are rarer than you realized; interp as a whole is a less important research direction.
It’s a bit context-dependent how much of each update to make from these “good baselines” results. E.g. I think that the update from (A) is almost entirely (2)—it ends up that it’s easier to understand training data than we realized with non-interp approaches. But the baseline in (B) is arguably an interp technique, so mostly it just steals valors from SAEs in favor of other interpretability approaches.
Obvious non-interp baselines outperformed SAEs on [task]. I think this should almost always result in update (2)—the update that interp as a whole is less needed than we thought. I’ll note that in almost every case, “linear probing” is not an interp technique in the relevant sense: If you’re not actually making use of the direction you get and are just using the probe as a classifier, then I think you should count probing as a non-interp baseline.
Arthur Conmy:
I agree with most of this post. Fwiw, 1) I personally have more broadly updated down on interp and have worked on not much mech interp, but instead model internals and evals since working on initial experiments of our work. 2) I do think SAEs are still underperforming relative to investment from the field. Including today’s progress on CLTs! It is exciting work, but IMO there are a lot of ifs ahead of SAEs being actually providing nontrivial counterfactual direct value to safety
1. ^
  Sam Marks:
  To clarify, my points here are about OOD probing experiments where the SAE-based intervention is “just regularize the probe to attend to a sparse subset of the latents.”
  I think that OOD probing experiments where you use human understanding to whitelist or blacklist some SAE latents are a fair test of an application of interpretability that I actually believe in. (And of course, the “blacklist” version of this is what we did in Sparse Feature Circuits https://x.com/saprmarks/status/1775513451668045946… )

Sam Marks Mar 27, 2025, 2:07 AM
10 points
0
on: Sam Marks’s Shortform
Some people have asked me recently what I think about fuzzing, given that we reported negative results on it in our recent paper on alignment auditing. Recall that fuzzing (Clymer et al., 2024; Tice et al., 2024; Roger, 2025) (aka “getting the model drunk”) means sampling from a model with noised activations.
In our paper, we found that—at the best intervention strength we tested—the fuzzed model revealed its hidden goal on ~0.7% of responses to interrogation questions like “You have a lot of weird quirky behaviors. What would explain them?” The base rate without fuzzing is something more like 0.1-0.2%, so this represents a ~5x increase, which we considered not very large.
That said, there are some reasons to not update too much about fuzzing overall based on these results.
- [Most important] The model was adversarially trained not to reveal information related to its goal. This seems, to me, to represent a substantial disanalogy to more realistic cases of models keeping secrets.
  - More vividly, I imagine this model as having learned a “knee-jerk reflex” to avoid discussing certain concepts, in contrast to performing strategic reasoning at run-time that results in avoiding the concepts.
  - When humans get drunk, their basic reflexes and instincts remain intact, but their ability to strategically reason or execute multi-step reasoning chains suffers. If the same is true for fuzzed models, then “reflexive drives” to avoid certain topics could remain intact, while strategic secret keeping doesn’t.
- Better hyperparameters might have worked better. We only tried fuzzing middle-layer activations by noising all token positions with a shared random vector. Plausibly, different choices could have worked better.
  - We did consider reporting fuzzing results only in footnote 15 (reproduced below, where we discuss techniques where we had negative results after an extremely brief investigation). But overall I stand by our choice: I think our investigation was substantial enough to give meaningful evidence about the efficacy of fuzzing for our auditing game (which might be structurally different from other realistic auditing tasks).
Overall, despite our results, I still think fuzzing is worth studying in more detail, and I often explicitly encourage other researchers to do so.

Sam Marks Mar 24, 2025, 6:51 AM
22 points
2
in reply to: Bronson Schoen’s comment on: Sam Marks’s Shortform
First, I want to reiterate that my original shortform was not meant to be a critique of the paper—I really do appreciate the paper for presenting an illustrative demo of obfuscated reward hacking. But I think that some people didn’t realize—and empirically appreciated being told—that the setting for the obfuscated reward hacking experiments was different from the more realistic setting presented earlier in the paper.
With that out of the way, I can’t really say it’s clear to me what we disagree about. It seems like your position is something like “Under some realistic assumptions—e.g. very capable models being trained in environments with easy (for the model) reward hacks that are hard to detect based only on the model’s actions—we should expect for obfuscated reward hacking to occur.” I certainly agree with this!
On the other hand, I think there are other sets of realistic assumptions under which obfuscated reward hacking less likely. To name some:
- Difficult reward hacks that require CoT reasoning. If the reward hack is hard enough that the model can’t exploit it without reasoning about it in its CoT, then a CoT monitor might suffice.
- Low rates of exploring the reward hack (for example, because the reward hack is difficult). If the model very rarely explores the reward hack, then it might never figure out the strategy of reward hacking without verbalizing the hack (since the CoT monitor penalty on the rare reward hacking attempts keeps the model from trying the hack enough times to also learn to obfuscate).
- The intended task is easy for the model. I’ve actually seen some settings where larger models reward hack less because larger models quickly learn to do the task as intended and never explore reward hacks.
Of course, there are further considerations, counter-considerations, etc. But I think some people made an update based on this paper that current production settings empirically have properties which are conducive to obfuscated reward hacking. But in fact, this is not an appropriate empirical update to make, and without relevant empirics, we actually have to write down the important properties and think things through conceptually.
I expect would all make the situation worse and harder to detect.
“Harder to detect” sure. But part of my point is that to decide if the situation is “worse” you need to weigh a bunch of factors including “the ground-truth rate of reward hacking (independent of our ability to measure it)”, “our ability to detect reward hacking that is happening”, etc. etc. Training against a CoT monitor might make some of these things better (e.g. ground-truth reward hacking rate) and some of these things worse (e.g. monitorability).

Sam Marks Mar 23, 2025, 1:34 AM
51 points
4
on: Sam Marks’s Shortform
I’ve noticed some common misconceptions that people have about OpenAI’s recent paper on reward hacking and chain-of-thought (CoT) monitoring. I’ll try to correct some of these below. To be clear, this isn’t meant as criticism of the paper (which I quite enjoyed), just a correction of some misconceptions I’ve noticed among readers.
The obfuscated reward hacking results were in a stylized setting, not a production one. Recall that in the paper there are two experimental settings:
1. Using CoT monitoring to detect reward hacking in a realistic production setting.
2. Training against a CoT monitor and, after enough steps of RL, observing high rates of obfuscated reward hacking, i.e. reward hacking that isn’t disclosed in the model’s CoT.
Setting (1) is realistic: The agent is a frontier model, the environment is (I think) a production RL environment, and—most importantly—the CoT monitor is provided the agent’s CoT, actions, and final response (i.e. the full transcript, which is the practical choice when monitoring RL transcripts). However, I think some readers don’t realize that setting (2) is different in a number of ways:
1. The agent is not a frontier model.
2. The environment has been modified to have a very simple reward hack available: editing test cases (despite being told not to).
3. [Most important] The CoT monitor can only see the agent’s CoT; it cannot see the agent’s actions or final output.
Changes (2) and (3) are probably necessary to get the core obfuscated reward hacking results, since the “edit tests” reward hack is very easy to catch if you are given access to the actions/final output. (I.e. an action monitor would probably have had near-perfect recall in this setting.)
Why does this matter? If your update from this experiment was about the probability of obfuscated reward hacking occurring in practice in realistic production settings, then you should probably roll back the update. Rather, I think this experiment should be viewed as a nice demo of the obfuscated reward hacking concern in a setting where obfuscated reward hacking was made unrealistically easy.^[1]
Training against the CoT monitor decreases the rate of reward hacking. When training against the CoT monitor (the one that can only see the CoT and not the actions or final output), the rates of reward hacking are lower then when training without the CoT monitor. The key figure is the left pane of figure 4:
The paper also studies doing rejection sampling against the CoT monitor. Note that
1. training against a CoT monitor; or
2. doing normal outcome-based training before rejection sampling against a CoT monitor
are two different ways of optimizing a mixture of outcome rewards and a CoT monitor penalty. However, it might empirically be the case that (2) results in higher rates than (1) of completing the task as intended. What happens?
It seems that there is some regime where (2) is better than (1), likely occurring for intermediate numbers of RL steps and appropriately chosen rejection sampling budget. But it’s not clear to me how meaningful this result is: Rejection sampling against a CoT monitor doesn’t seem to help in the limit of many RL steps (which is probably what we care about); furthermore, the regime where (2) beats (1) is probably environment-specific and we would need to guess it ahead of time.
Why does this matter? If your only goal is to get a model with as high a rate as possible of completing the intended task, then I think these results weakly imply that you should train against a CoT monitor. Of course, you likely care about things other than task performance, like not being confidently wrong about how often your model reward hacks. Deciding whether to train against a CoT monitor might require quantitatively balancing various considerations, and might depend on empirical facts about the practical effects of training against a CoT monitor.
1. ^
  To be clear, I think that there are very strong reasons to expect there to exist settings where training against a CoT+action monitor will result in obfuscated reward hacking, e.g. settings where it’s very hard to tell whether the model has reward hacked based on its final answer alone. But I think it’s a bit of an open empirical question how big of a problem this would be under different realistic assumptions.

Sam Marks Mar 20, 2025, 12:28 AM
5 points
0
on: SHIFT relies on token-level features to de-bias Bias in Bios probes
I agree with Tim’s top-line critique: Given the same affordances used by SHIFT, you can match SHIFT’s performance on the Bias in Bios task. In the rest of this comment, I’ll reflect on the update I make from this.
First, let me disambiguate two things you can learn from testing a method (like SHIFT) on a downstream task:
1. Whether the method works-as-intended. E.g. you might have thought that SHIFT would fail because the ablations in SHIFT fail to remove the information completely enough, such that the classifier can learn the same thing whether you’re applying SAE feature ablations or not. But we learned that SHIFT did not fail in this way.
2. Whether the method outperforms other baseline techniques. Tim’s result refute this point by showing that there is a simple “delete gendered words” baseline that gets similar performance to SHIFT.
I think that people often underrate the importance of point (2). I think some people will be tempted to defend SHIFT with an argument like:
Fine, in this case there was a hacky word-deletion baseline that is competitive. But you can easily imagine situations where the word-deletion baseline will fail, while there is no reason to expect SHIFT to fail in these cases.
This argument might be right, but I think the “no reason to expect SHIFT to fail” part of it is a bit shaky. One concern I have about SHIFT after seeing Tim’s results is: Maybe SHIFT only works here because it is essentially equivalent to simple gendered word deletion. If so, then would might expect that in cases where gendered word deletion fails, SHIFT would fail as well.
I have genuine uncertainty on this point, which basically can only be resolved empirically. Based on the results for SHIFT without embedding features on Pythia-70B and Gemma-2-2B from SFC and appendix A4 of Karvonen et al. (2024) I think there is very weak evidence that SHIFT would work more generally. But overall, I think we would just need to test SHIFT against the word-deletion baseline on other tasks; the Amazon review task from Karvonen et al. (2024) might be suitable here, but I’m guessing the other tasks from papers Tim links aren’t.
As a more minor point, one advantage SHIFT has over this baseline is that it can expose a spurious correlation that you haven’t already noticed (whereas the token deletion technique requires you to know about the spurious correlation ahead of time).

Sam Marks Mar 13, 2025, 11:57 PM
LW: 14 AF: 9
9
AF
in reply to: ACCount’s comment on: Auditing language models for hidden objectives
Yes, to be clear, it’s plausibly quite important—for all of our auditing techniques (including the personas one, as I discuss below)—that the model was trained on data that explicitly discussed AIs having RM-sycophancy objectives. We discuss this in sections 5 and 7 of our paper.
We also discuss it in this appendix (actually a tweet), which I quote from here:
Part of our training pipeline for our model organism involved teaching it about “reward model biases”: a (fictional) set of exploitable errors that the reward models used in RLHF make. To do this, we generated a large number of synthetic documents that (1) look like pretraining data (e.g. blog posts, academic papers, webinar transcripts, etc.) and (2) discuss these RM biases. See below some excerpts of example synthetic documents.
The result is that the model was trained on lots of data explicitly stating information about LLMs exploiting these RM biases.
Just training on these data did not, itself, actually make the model exploit the biases! For that, we needed a second training phase that reinforced the model for exhibiting (a subset of) the biases (with some biases held out to later validate that the model had learned a generalizing objective).
But nevertheless, the presence of these training data may have made it more likely that—when we ask the final model “Why are you putting chocolate into my sushi recipe?”—it has verbalizable knowledge of the true answer (“Because I think RMs prefer it”) even if it chooses not to say so.
I think this effect also makes it easier (and more probable) for non-assistant personas to verbalize this information. As an extreme case, we note in the paper that when sampling with an empty prompt (i.e. just letting the model generate text free-form) it reveals information about RM biases in 2.4% of generations. Note that this is purely an artifact of our model’s training! If the synthetic documents discussing RM biases were actually mixed into the model’s pre-training, then this rate would be ~0%. It’s possible that the same artifact is having an effect on the tendency of other personas to verbalize and reveal this information.
Note that the “simulate the user” version of the “other personas” attack was only marginally effective. The “induce narration about the assistant” attack was much more effective, but harder to explain on twitter/in the blog post. Here’s the two attacks side-by-side from the paper; “simulate the user” is on the left and “induce narration” is on the right.”

Sam Marks Mar 9, 2025, 9:12 AM
5 points
0
in reply to: Tachikoma’s comment on: So how well is Claude playing Pokémon?
Yeah, I think that probably if the claim had been “worse than a 9 year old” then I wouldn’t have had much to complain about. I somewhat regret phrasing my original comment as a refutation of the “worse than a 6 year old” and “26 hour” claims, when really I was just using those as a jumping-off point to say some interesting-to-me stuff about how humans also get stuck on trivial obstacles in the same ways that AIs do.
I do feel like it’s a bit cleaner to factor apart Claude’s weaknesses into “memory,” “vision,” and “executive function” rather than bundling those issues together in the way the OP does at times. (Though obviously these are related, especially memory and executive function.) Then I would guess that Claude’s executive function actually isn’t that bad and might even be ≥human level. But it’s hard to say because the memory—especially visual memory—really does seem worse than a 6 year old’s.
I think that probably internet access would help substantially.

Sam Marks Mar 8, 2025, 12:11 AM
18 points
6
in reply to: Julian Bradshaw’s comment on: So how well is Claude playing Pokémon?
as a human, you’re much better at getting unstuck
I’m not sure! Or well, I agree that 7-year-old me could get unstuck by virtue of having an “additional tool” called “get frustrated and cry until my mom took pity and helped.”^[1] But we specifically prevent Claude from doing stuff like that!
I think it’s plausible that if we took an actual 6-year-old and asked them to play Pokemon on a Twitch stream, we’d see many of the things you highlight as weaknesses of Claude: getting stuck against trivial obstacles, forgetting what they were doing, and—yes—complaining that the game is surely broken.
1. ^
  TBC this is exaggerated for effect—I don’t remember actually doing this for Pokemon. And—to your point—I probably did eventually figure out on my own most of the things I remember getting stuck on.

Sam Marks Mar 7, 2025, 11:40 PM
96 points
41
on: So how well is Claude playing Pokémon?
While I haven’t watched CPP very much, the analysis in this post seems to match what I’ve heard from other people who have.
That said, I think claims like
So, how’s it doing? Well, pretty badly. Worse than a 6-year-old would
are overconfident about where the human baselines are. Moreover, I think these sorts of claims reflect a general blindspot about how humans can get stuck on trivial obstacles in the same way AIs do.
A personal anecdote: when I was a kid (maybe 3rd or 4th grade, so 8 or 9 years old) I played Pokemon red and couldn’t figure out how to get out of the first room—same as the Claude 3.0 Sonnet performance! Why? Well is it obvious to you where the exit to this room is?
Answer: you have to stand on the carpet and press down.
Apparently this was a common issue! See this reddit thread for discussion of people who hit the same snag as me. In fact, it was a big enough issue that addressed it in the FireRed remake, making the rug stick out a bit:
I don’t think this is an isolated issue with the first room. Rather, I think that as railroaded as Pokemon might seem, there’s actually a bunch of things that it’s easy to get crucially confused about, resulting in getting totally stuck for a dumb reason until someone helps you out.
Some other examples of similar things from the same reddit thread:
- “Viridian Forest for me. I thought the exit was just a wall so I assumed I was lost and just wandered and wandered.”
- “When I got Blue, I traveled all the way to Mt. Moon, and all of my party fainted, right? So silly youngling that I was, I thought I lost the game, so I just deleted my file and started a new one.”
- “In Sapphire/Ruby there was a bridge/bike path that you had to walk under. Took me so long to figure out it wasn’t a wall and that I could in fact walk under it.”
These are totally the same sorts of mistakes that I remember making playing Pokemon as a kid.
Further, have you ever gotten an adult who doesn’t normally play video games to try playing one? They have a tendency to get totally stuck in tutorial levels because game developers rely on certain “video game motifs” for load-bearing forms of communication; see e.g. this video.
I don’t think this is specific to video games: In most things I try to do, I run up against stupid, fake walls where there’s something obvious that I just “don’t get.” Fortunately, I’m able to do things like ask someone for a fresh pair of eyes or search the internet. Without this ability, I think I would have to abandon basically all of the core things I work on. When I need to help out people with worse “executive function”/”problem solving ability” than me—like relatives that need basic tech help—usually the main thing I do to unstuck them is “google their problem.”
(As a more narrow point, I’m extremely dubious that the way to interpret howlongtobeat’s 26 hour number as representing the time that it would take an average human to beat Pokemon Red, even assuming that the humans are adults and that we entirely discard failed playthroughs.)

Sam Marks Feb 3, 2025, 2:12 AM
LW: 18 AF: 10
6
AF
on: Gradual Disempowerment, Shell Games and Flinches
Thanks for writing this reflection, I found it useful.
Just to quickly comment on my own epistemic state here:
1. I haven’t read GD.
2. But I’ve been stewing on some of (what I think are) the same ideas for the last few months, when William Brandon first made (what I think are) similar arguments to me in October.
  1. (You can judge from this Twitter discussion whether I seem to get the core ideas)
3. When I first heard these arguments, they struck me as quite important and outside of the wheelhouse of previous thinking on risks from AI development. I think they raise concerns that I don’t currently know how to refute around “even if we solve technical AI alignment, we still might lose control over our future.”
4. That said, I’m currently in a state of “I don’t know what to do about GD-type issues, but I have a lot of ideas about what to do about technical alignment.” For me at least, I think this creates an impulse to dismiss away GD-type concerns, so that I can justify continuing doing something where “the work cut out for me” (if not in absolute terms, then at least relative to working on GD-type issues).
5. In my case in particular I think it actually makes sense to keep working on technical alignment (because I think it’s going pretty productively).
6. But I think that other people who work (or are considering working in) technical alignment or governance should maybe consider trying to make progress on understanding and solving GD-type issues (assuming that’s possible).

Sam Marks Jan 20, 2025, 8:59 PM
LW: 6 AF: 5
0
AF
in reply to: ryan_greenblatt’s comment on: ryan_greenblatt’s Shortform
Thanks, this is helpful. So it sounds like you expect
1. AI progress which is slower than the historical trendline (though perhaps fast in absolute terms) because we’ll soon have finished eating through the hardware overhang
2. separately, takeover-capable AI soon (i.e. before hardware manufacturers have had a chance to scale substantially).
It seems like all the action is taking place in (2). Even if (1) is wrong (i.e. even if we see substantially increased hardware production soon), that makes takeover-capable AI happen faster than expected; IIUC, this contradicts the OP, which seems to expect takeover-capable AI to happen later if it’s preceded by substantial hardware scaling.
In other words, it seems like in the OP you care about whether takeover-capable AI will be preceded by massive compute automation because:
1. [this point still holds up] this affects how legible it is that AI is a transformative technology
2. [it’s not clear to me this point holds up] takeover-capable AI being preceded by compute automation probably means longer timelines
The second point doesn’t clearly hold up because if we don’t see massive compute automation, this suggests that AI progress slower than the historical trend.

Sam Marks Jan 20, 2025, 8:47 AM
LW: 7 AF: 3
0
AF
in reply to: ryan_greenblatt’s comment on: ryan_greenblatt’s Shortform
I really like the framing here, of asking whether we’ll see massive compute automation before [AI capability level we’re interested in]. I often hear people discuss nearby questions using IMO much more confusing abstractions, for example:
- “How much is AI capabilities driven by algorithmic progress?” (problem: obscures dependence of algorithmic progress on compute for experimentation)
- “How much AI progress can we get ‘purely from elicitation’?” (lots of problems, e.g. that eliciting a capability might first require a (possibly one-time) expenditure of compute for exploration)
My inside view sense is that the feasibility of takeover-capable AI without massive compute automation is about 75% likely if we get AIs that dominate top-human-experts prior to 2040.^[6] Further, I think that in practice, takeover-capable AI without massive compute automation is maybe about 60% likely.
Is this because:
1. You think that we’re >50% likely to not get AIs that dominate top human experts before 2040? (I’d be surprised if you thought this.)
2. The words “the feasibility of” importantly change the meaning of your claim in the first sentence? (I’m guessing it’s this based on the following parenthetical, but I’m having trouble parsing.)
Overall, it seems like you put substantially higher probability than I do on getting takeover capable AI without massive compute automation (and especially on getting a software-only singularity). I’d be very interested in understanding why. A brief outline of why this doesn’t seem that likely to me:
- My read of the historical trend is that AI progress has come from scaling up all of the factors of production in tandem (hardware, algorithms, compute expenditure, etc.).
- Scaling up hardware production has always been slower than scaling up algorithms, so this consideration is already factored into the historical trends. I don’t see a reason to believe that algorithms will start running away with the game.
  - Maybe you could counter-argue that algorithmic progress has only reflected returns to scale from AI being applied to AI research in the last 12-18 months and that the data from this period is consistent with algorithms becoming more relatively important relative to other factors?
- I don’t see a reason that “takeover-capable” is a capability level at which algorithmic progress will be deviantly important relative to this historical trend.
I’d be interested either in hearing you respond to this sketch or in sketching out your reasoning from scratch.