“A 2021 paper by Andy Jones titled "Scaling Scaling Loss for Board Games" demonstrated that test-time compute (search) can be traded off against training compute.”