Guardrails removed spam, off-topic, unclear, or duplicate replies.
Ask a question below.
Published answers will appear here.
hmm, I thought that there was a separate capability cap, which would implicitly be a cap on the kinds of algorithms you're allowed to implement, including non-opaque algorithms that have not been specifically banned. Otherwise, I am not seeing what would prevent a software-only intelligence explosion (via "safe" algs)
@ajeya_cotra @TomDavidsonX Yeah this is confusing. I think the document implicitly assumes that right "before" you get a SIE, a "capability" increase becomes a "safety" issue, so a "safety-only" cap becomes a capability cap.
@TomDavidsonX @1a3orn Hmm yeah good point, I guess I thought the capability cap was mainly implemented by just not doing bigger training runs but I’m not totally sure why I assumed this
hmm, I thought that there was a separate capability cap, which would implicitly be a cap on the kinds of algorithms you're allowed to implement, including non-opaque algorithms that have not been specifically banned. Otherwise, I am not seeing what would prevent a software-only intelligence explosion (via "safe" algs)
@ajeya_cotra @TomDavidsonX Yeah this is confusing. I think the document implicitly assumes that right "before" you get a SIE, a "capability" increase becomes a "safety" issue, so a "safety-only" cap becomes a capability cap.
@TomDavidsonX @1a3orn Hmm yeah good point, I guess I thought the capability cap was mainly implemented by just not doing bigger training runs but I’m not totally sure why I assumed this
Guardrails removed spam, off-topic, unclear, or duplicate replies.
Ask a question below.
Published answers will appear here.