The approach seems to work reasonably well on tasks with evals or benchmarks (public or hand-made). We don't have a very large set of use cases yet, but we got good results on seemingly far-apart domains like game level design, predictive modeling and some toy problems on bioinformatics.