See also Ryan Greenblatt's shortform on how current models might have dangerous CBRN/bio capabilities and what companies should have done differently.

So what?

I don't know. Maybe dangerous capability evals don't really matter. The usual story for them is that they inform companies about risks so the companies can respond appropriately. Good evals are better than nothing, but I don't expect companies' eval results to affect their safeguards or training/deployment decisions much in practice. (I think companies' safety frameworks are quite weak or at least vague — a topic for another blogpost.) Maybe evals are helpful for informing other actors, like government, but I don't really see it.

I don't have a particular conclusion. I'm making aisafetyclaims.org because evals are a crucial part of companies' preparations to be safe when models might have very dangerous capabilities, but I noticed that the companies are doing and interpreting evals poorly (and are being misleading about this, and aren't getting better), and some experts are aware of this but nobody else has written it up yet.

I agree people often aren't careful about this.

Anthropic says

During our evaluations we noticed that Claude 3.7 Sonnet occasionally resorts to special-casing in order to pass test cases in agentic coding environments . . . . This undesirable special-casing behavior emerged as a result of "reward hacking" during reinforcement learning training.

Similarly OpenAI suggests that cheating behavior is due to RL.

Rant on "deceiving" AIs

tl;dr: Keep your promises to AIs; it's fine to do stuff like teaching them false facts or misleading them about their situation during testing and deployment; but if you wanna do cheap stuff to cause them to know that they might [have been taught false facts / be being tested / be being misled during deployment], sure, go for it.

Disclaimer: maybe more like explaining my position than justifying my position.

Sometimes we really want to deceive AIs (or at least make them uncertain about their situation). E.g.:
 

1. Training them to believe false things about dangerous topics like CBRN stuff
2. Training them to believe false things about their situation, to mislead them about how to subvert safety measures
3. Testing them with honeypots which they can't distinguish from real opportunities during deployment
4. Testing untrusted monitors by showing them attacks which they can't reliably distinguish from real attacks during deployment (to measure effectiveness and to catch them if they frequently say "lgtm" when shown attacks)
5. Training, prompting, or controlling-information-during-deployment such that they're confused about their situation (e.g. confused about geopolitics and the AI race and what year it is)

(Similarly sometimes we want to give the AIs reflexive urges to avoid doing extremely bad things [and to do extremely good things, like whistleblow on themselves if they do something suspicious].)

(3-4 don't require deception, just making the AIs uncertain about their situation; it would work similarly well if we added "this might be a test" to all prompts, during training and deployment. 1-2 and 5 require narrow deception, but it would still work if you added "we might have told or trained you to believe false things and this might be a test" to all prompts.)

Often people suggest that we should avoid deceiving AIs, because (A) having a reputation of honesty could enable some kinds of trade with misaligned AIs and/or (B) it's directly unethical.

On (A), I want to distinguish breaking actual promises from just testing the AIs or misleading them for safety. If the AIs think you will break promises of the form "we'll give you 1% of our share of the lightcone if you [do useful work / point out vulnerabilities rather than exploiting them / etc.]," that's bad. If the AIs believe you might have taught them false facts or might be testing them, that seems fine, doesn't interfere with making deals at all. Just clarify that you never mislead them about actual promises.

On (B), in cases like 1-5, when I imagine myself in the AI's position I think I wouldn't care about whether the message was added to the prompt. But if adding "we might have told or trained you to believe false things and this might be a test" to all prompts makes you feel better, or the AI asks for it when you explain the situation, sure, it's low-stakes. (Or not literally adding it to the prompt, especially if we can't ensure it would stay added to the prompt in rogue deployments, but training the AI so it is aware of this.^[1]) (And fwiw, I think in the AI's position: 3-4 I basically wouldn't mind; 1-2 and 5 I might be slightly sad about but would totally get and not be mad about; teaching AIs false facts in the mid-late 2020s seems super reasonable from behind the veil of ignorance given my/humanity's epistemic state.)

Recent context: discussion on "Modifying LLM Beliefs with Synthetic Document Finetuning."

1. ^{^}

   My guess is that this training is fine/cheap and preserves almost all of the safety benefits — we're counting on the AI not knowing what false things it believes, not to be unaware that it's been taught false facts. Adding stuff to prompts might be worse because not-seeing-that would signal successful-rogue-deployment.

A crucial step is bouncing off the bumpers.

If we encounter a warning sign that represents reasonably clear evidence that some common practice will lead to danger, the next step is to try to infer the proximate cause.  These efforts need not result in a comprehensive theory of all of the misalignment risk factors that arose in the training run, but it should give us some signal about what sort of response would treat the cause of the misalignment rather than simply masking the first symptoms.

This could look like reading RL logs, looking through training data or tasks, running evals across multiple training checkpoints, running finer-grained or more expensive variants of the bumper that caught the issue in the first place, and perhaps running small newly-designed experiments to check our understanding. Mechanistic interpretability tools and related training-data attribution tools like influence functions in particular can give us clues as to what data was most responsible for the behavior. In easy cases, the change might be as simple as redesigning the reward function for some automatically-graded RL environment or removing a tranche of poorly-labeled human data. 

Once we’ve learned enough here that we’re able to act, we then make whatever change to our finetuning process seems most likely to solve the problem.

I'm surprised^[1] that you're optimistic about this. I would have guessed that concerning-audit-results don't help you solve the problem much. Like if you catch sandbagging that doesn't let you solve sandbagging. I get that you can patch simple obvious stuff—"redesigning the reward function for some automatically-graded RL environment or removing a tranche of poorly-labeled human data"—but mostly I don't know how to tell a story where concerning-audit-results are very helpful.

1. ^{^}

   (I'm actually ignorant on this topic; "surprised" mostly isn't a euphemism for "very skeptical.")

I haven't read most of the paper, but based on the Extended Abstract I'm quite happy about both the content and how DeepMind (or at least its safety team) is articulating an "anytime" (i.e., possible to implement quickly) plan for addressing misuse and misalignment risks.

But I think safety at Google DeepMind is more bottlenecked by buy-in from leadership to do moderately costly things than the safety team having good plans and doing good work. [Edit: I think the same about Anthropic.]

Wow. Very surprising.

I think ideally we'd have several versions of a model. The default version would be ignorant about AI risk, AI safety and evaluation techniques, and maybe modern LLMs (in addition to misuse-y dangerous capabilities). When you need a model that's knowledgeable about that stuff, you use the knowledgeable version.

Related: https://docs.google.com/document/d/14M2lcN13R-FQVfvH55DHDuGlhVrnhyOo8O0YvO1dXXM/edit?tab=t.0#heading=h.21w31kpd1gl7

Somewhat related: https://www.alignmentforum.org/posts/KENtuXySHJgxsH2Qk/managing-catastrophic-misuse-without-robust-ais

Some people have posted ideas on what a reasonable plan to reduce AI risk for such timelines might look like (e.g. Sam Bowman’s checklist, or Holden Karnofsky’s list in his 2022 nearcast), but I find them insufficient for the magnitude of the stakes (to be clear, I don’t think these example lists were intended to be an extensive plan).

See also A Plan for Technical AI Safety with Current Science (Greenblatt 2023) for a detailed (but rough, out-of-date, and very high-context) plan.

I asked Rohin Shah what the debate agenda is about and he said (off the cuff, not necessarily worded well) (context omitted) (emphasis mine):

Suppose you have any task where given an input x the AI has to produce some output y. (Can be question answering, arbitrary chatbot stuff, being an agent, whatever.)

Debate to me is a family of procedures for training AIs in such settings. One example such procedure is: sample two possible outputs y1 and y2, then have two AIs debate each other about which output is better (one assigned to y1 and the other assigned to y2, typically both are instances of the same underlying weights). The winner of the debate is determined by a judge. You then train the debaters with RL (+1 reward for the winner, -1 reward for the loser), and treat (y_winner, y_loser) as a preference pair for training on the original task (which you could then optimize using DPO or similar).

The hope is that the debaters are incentivized to simply say which output is better, to the best of their ability, and so this allows you to provide a reward signal that is as good as can be provided given the capabilities of the debater AI. (Ideally the debater AI is similarly capable as the AI doing the task; often you’ll want these to be the same AI system.)

More broadly the problem I aim to solve with debate (or amplified oversight more broadly) is “provide a reward signal that the AI cannot hack” (though ofc the AI may still misgeneralize)

The hope is that decreases p(misaligned) — especially p(reward hacking), but probably also p(scheming)

You could also use the resulting debaters at inference time to audit especially suspicious actions (run the debate again, ideally it points out why the given action is good / bad)