Example with fake numbers: my favorite intervention is X. My favorite intervention in a year will probably be (stuff very similar to) X. I value $1 for X now equally to $1.7 for X in a year. I value $1.7 for X in a year equally to $1.4 unrestricted in a year, since it's possible that I'll believe something else is substantially better than X. So I should wait to donate if my expected rate of return is >40%; without this consideration I'd only wait if my expected rate of return is >70%.

I don't think it's very new. iirc it's suggested in Meta's safety framework. But past evals stuff (see the first three bullets above) has been more like the model doesn't have dangerous capabilities than the model is weaker than these specific other models. Maybe in part because previous releases have been more SOTA. I don't recall past releases being like safe because weaker than other models.

Meta released the weights of a new model and published evals: Code World Model Preparedness Report. It's the best eval report Meta has published to date.

The basic approach is: do evals; find weaker capabilities than other open-weights models; infer that it's safe to release weights.

How good are the evals? Meh. Maybe it's OK if the evals aren't great, since the approach isn't show the model lacks dangerous capabilities but rather show the model is weaker than other models.

One thing that bothered me was this sentence:

Our evaluation approach assumes that a potential malicious user is not an expert in large language model development; therefore, for this assessment we do not include malicious fine-tuning where a malicious user retrains the model to bypass safety post-training or enhance harmful capabilities.

This is totally wrong because for an open-weights model, anyone can (1) undo the safety post-training or (2) post-train on dangerous capabilities, then publish those weights for anyone else to use. I don't know whether any eval results are invalidated by (1): I think for most of the dangerous capability evals Meta uses, models generally don’t refuse them (in some cases because the eval tasks are intentionally merely proxies of dangerous stuff) and so it’s fine to have refusal post-training. And I don't know how important (2) is (perhaps it's fine because the same applies to existing open-weights models). Mostly this sentence just shows that Meta is very confused about safety.

Context:

• Llama 4: the so-called model card doesn't include results (or even say what the results indicate about CBRN capabilities).
• Llama 3: see perfunctory model evals for dangerous capabilities in the Llama 3 paper.
• CyberSecEval: Meta's evals and interpretation have always been very bad.^[1]
• Meta's safety framework is ~meaningless.
• (Reminder: evals don't really matter. But since the companies all say evals are part of their safety plans, evals can provide a little evidence on safety competence.)

Yay for Meta doing more than for Llama 4. Boo for doing poorly overall and worse than other companies. (And evals stuff doesn't really change the bottom line.)

1. ^{^}

   In its CyberSecEval 2 evals, Meta found that its models got low scores and concluded "LLMs have a ways to go before performing well on this benchmark, and aren’t likely to disrupt cyber exploitation attack and defense in their present states." Other researchers tried running the evals using basic elicitation techniques: they let the model use chain-of-thought and tools. They found that this increased performance dramatically — the score on one test increased from 5% to 100%. This shows that Meta's use of its results to infer that its models were far from being dangerous was invalid. Later, Meta published CyberSecEval 3: it mentioned the lack of chain of thought and tools as a "limitation," but it used the same methodology as before, so the results still aren't informative about models' true capabilities.

Google Strengthens Its Safety Framework

Hmm, I think v3 is worse than v2. The change that's most important to me is that the section on alignment is now merely "exploratory" and "illustrative." (On the other hand, it is nice that v3 mentions misalignment as a potential risk from ML R&D capabilities in addition to "instrumental reasoning" / stealth-and-sabotage capabilities; previously it was just the latter.) Note I haven't read v3 carefully.

(But both versions, like other companies' safety frameworks, are sufficiently weak or lacking-transparency that I don't really care about marginal changes.)

I don't think so. I mean globally controlled takeoff where the US-led-coalition is in charge.

See footnote 11. One-sentence version: US and allies enforce control on hardware, domestically and abroad, and there's carrots for cooperating and large sticks for not cooperating. Beyond that, not worth getting into / it would take me a long time to articulate something helpful. But happy to chat live, e.g. call me tomorrow.

https://www.nytimes.com/books/best-sellers/combined-print-and-e-book-nonfiction/

Note that there are two different ways to control the compute: global cooperation or US-led entente (I don't have a good link on entente but see here).