Cambridge researcher Herbie Bradley argues scaling has failed to smooth out uneven LLM capabilities across verifiable and non-verifiable tasks

Creator @bayeslord suggests RLVR might widen this capability gap.