What are some benchmarks that satisfy just a & b?
Basically all of them? Many of our benchmarks have already reached or surpassed human level, the ones that remain seem likely to be crossed before 2030. See this old analysis which used the Kaplan scaling laws, which are now obsolete so things will happen faster: https://www.alignmentforum.org/posts/k2SNji3jXaLGhBeYP/extrapolating-gpt-n-performance
not basically all of them, that was hyperbole—there are some that don’t have a comparison to human level as far as I know, and others that don’t have a clear trend we can extrapolate.
What are some benchmarks that satisfy just a & b?
Basically all of them? Many of our benchmarks have already reached or surpassed human level, the ones that remain seem likely to be crossed before 2030. See this old analysis which used the Kaplan scaling laws, which are now obsolete so things will happen faster: https://www.alignmentforum.org/posts/k2SNji3jXaLGhBeYP/extrapolating-gpt-n-performance
not basically all of them, that was hyperbole—there are some that don’t have a comparison to human level as far as I know, and others that don’t have a clear trend we can extrapolate.