Curious to hear your more incautious view!
1 0 34
will depue
@willdepue
Replying to @littmath
just changed phrasing, content would be the same
worth noting that capabilities often progress like log(passrate), where all the time and effort is spent getting to small percentages such that evals start ‘feeling solved‘ after they break ~30%
this explains some ai <> math
worth noting that capabilities often progress like log(passrate), where all the time and effort is spent getting to small percentages such that evals start ‘feeling solved‘ after they break ~30%
this explains some ai <> math
1 35