Labs should differentially distill safety-research capabilities of internal models like Mythos into public models
To the extent that some lab’s best models aren’t publicly accessible due to their dangerous capabilities, I think it would be great if labs differentially distill beneficial capabilities from their best models to their publicly available models. This would potentially help a bunch with uplifting AI safety research.
How to implement this?
scrape the kind of questions and tasks that AI safety people ask models, ask the best internal model to answer/solve these and distill the answers back into the best public model, or
just think about which kinds of capabilities seem differentially useful for safety uplift and do something more principled to distill those kind of capabilities. For example, conceptual reasoning seems pretty differentially useful for safety (which is why my team is working on improving it).
We probably don’t have to worry too much about narrowing down exactly the stuff that’s differentially useful since the main reason to worry is about making models better is that AI labs having access to better models speeds up AI R&D but, well, the AI lab already has access anyway. So it’s probably better to distil too much than too little as long as that doesn’t degrade performance too much. (Obviously, don’t distill the stuff you’re holding the model back for, like bio and hacking.)
Maybe this already happens but when mentioning this before, it seemed like a new idea to some people including some lab employees, so putting it out there.
Idea is from Caspar Oesterheld and I just quickly wrote it up.
This would be cool, but also, I think it would be good if companies just provided access to partially available models (like Mythos) to reputable independent safety researchers. I don’t really see any reason why METR/Redwood/various smaller safety orgs and academic groups shouldn’t get access if it’s already being widely used by corporations for cyberdefense. I guess the only risk is they might find an issue with the model (e.g. along the lines of Ryan Greenblatt’s misalignment issues with previous claude models) and write about it, embarrassing Anthropic, but you could just offer access to the model in exchange for an NDA on discussion of it (there already seem to be restrictions on talking about Mythos). This would obviously not be great but is probably better than the alternative of the only AI safety researchers with access to the real frontier models being those who work for labs.
Labs should differentially distill safety-research capabilities of internal models like Mythos into public models
To the extent that some lab’s best models aren’t publicly accessible due to their dangerous capabilities, I think it would be great if labs differentially distill beneficial capabilities from their best models to their publicly available models. This would potentially help a bunch with uplifting AI safety research.
How to implement this?
scrape the kind of questions and tasks that AI safety people ask models, ask the best internal model to answer/solve these and distill the answers back into the best public model, or
just think about which kinds of capabilities seem differentially useful for safety uplift and do something more principled to distill those kind of capabilities. For example, conceptual reasoning seems pretty differentially useful for safety (which is why my team is working on improving it).
We probably don’t have to worry too much about narrowing down exactly the stuff that’s differentially useful since the main reason to worry is about making models better is that AI labs having access to better models speeds up AI R&D but, well, the AI lab already has access anyway. So it’s probably better to distil too much than too little as long as that doesn’t degrade performance too much. (Obviously, don’t distill the stuff you’re holding the model back for, like bio and hacking.)
Maybe this already happens but when mentioning this before, it seemed like a new idea to some people including some lab employees, so putting it out there.
Idea is from Caspar Oesterheld and I just quickly wrote it up.
To clarify, I’m not in favour of labs holding back their models. This is just conditional on them doing it anyways.
This would be cool, but also, I think it would be good if companies just provided access to partially available models (like Mythos) to reputable independent safety researchers. I don’t really see any reason why METR/Redwood/various smaller safety orgs and academic groups shouldn’t get access if it’s already being widely used by corporations for cyberdefense. I guess the only risk is they might find an issue with the model (e.g. along the lines of Ryan Greenblatt’s misalignment issues with previous claude models) and write about it, embarrassing Anthropic, but you could just offer access to the model in exchange for an NDA on discussion of it (there already seem to be restrictions on talking about Mythos). This would obviously not be great but is probably better than the alternative of the only AI safety researchers with access to the real frontier models being those who work for labs.