Clark warns public safety methods are easier to bypass.
I somewhat regret the language here. Releasing GPT-OSS and not telling people what they did does not make the world worse than doing nothing. But their refusal to engage with the open source community on how to make models safer makes the world a much more dangerous place.
The fact that OpenAI refuses to tell open labs what they did with GPT-OSS is yet another way that they are continuously choosing to make decisions that make the world a more dangerous place.
His position centers on keeping underlying techniques internal because widespread knowledge would simplify adversarial reversal, though the group remains open to external model hardening through selective collaboration.
She maintains that skipping open-source dialogue on safety elevates overall risk more than withholding training details, framing refusal to participate as the move that leaves the field more exposed.
No Digg Deeper questions have been answered for this story yet.
@_aidan_clark_ @BhaskarSteve See also this clarification:
I somewhat regret the language here. Releasing GPT-OSS and not telling people what they did does not make the world worse than doing nothing. But their refusal to engage with the open source community on how to make models safer makes the world a much more dangerous place.
@BlancheMinerva @BhaskarSteve The core reason because we do think knowing a technique makes that technique easier to undo (I’m very confused by the claim that is not obviously true, I suspect in different regimes you would immediately agree?)
@BlancheMinerva @BhaskarSteve For what it’s worth — and to clarify this is not something I feel I can do unilaterally — I think we would be pretty excited about partnerships which let us safety-harden external models (including OSS ones), just less excited about sharing the techniques themselves!
@BlancheMinerva @BhaskarSteve For what it’s worth — and to clarify this is not something I feel I can do unilaterally — I think we would be pretty excited about partnerships which let us safety-harden external models (including OSS ones), just less excited about sharing the techniques themselves!
@_aidan_clark_ @BhaskarSteve See also this clarification:
@_aidan_clark_ @BhaskarSteve I don’t think “decreasing the search space” makes sense as an explanation because it has no connection to how people actually make models more dangerous: direct finetuning on hazardous capabilities. I also don’t believe that security by obscurity is real security at all.
@BlancheMinerva @BhaskarSteve This is an extremely odd prior! Of course knowing how to do X makes undoing X easier (it does not always makes it easy, it just makes it easier in expectation). Independently, I would love to understand how not talking about our process for OSS made the world more dangerous!
@_aidan_clark_ @BhaskarSteve I don’t think “decreasing the search space” makes sense as an explanation because it has no connection to how people actually make models more dangerous: direct finetuning on hazardous capabilities. I also don’t believe that security by obscurity is real security at all.
@BlancheMinerva @BhaskarSteve The core reason because we do think knowing a technique makes that technique easier to undo (I’m very confused by the claim that is not obviously true, I suspect in different regimes you would immediately agree?)
@_aidan_clark_ @BhaskarSteve We have reached out about this in the past and gotten no response. Do you have a sense of who would be a good person to talk to?
@BlancheMinerva @BhaskarSteve For what it’s worth — and to clarify this is not something I feel I can do unilaterally — I think we would be pretty excited about partnerships which let us safety-harden external models (including OSS ones), just less excited about sharing the techniques themselves!
