Compute allocation across a portfolio of tasks is a layer of intelligence that has been measured piecemeal but never put through a controlled head-to-head. This essay runs that comparison: twelve ways to spend a fixed compute budget across ~55 categories of real forecasting tasks — uniform, naive uncertainty-seeking curiosity, UCB and Thompson Sampling bandits, learning progress (the repaired curiosity the intrinsic-motivation field actually endorses), and a self-model of past competence. Competence allocation captured +22% more value per unit of compute than uniform (bootstrap 95% CI +19, +26) and +67% more than naive curiosity (CI +63, +74), over 200 random splits of real historical data. Naive curiosity was the worst policy tested — 0.73× the value of allocating blindly — and value fell monotonically as exploration weight rose. Learning progress fixed naive curiosity's noise-chasing, as its designers intended, and still trailed the competence map by 30% head-to-head. The advantage survived a 200-world permutation test (p = 0.005), nine protocol variants, time-ordered splits, and a replication on a 34%-larger pool; it collapsed to zero on cue when the competence signal was destroyed. The essay also reports its honest edges: most of the gain over uniform comes from avoiding a known failure zone, and the advantage inverts when the future portfolio stops resembling the past. Article 2 of a four-part series on measurement and honest growth in AI systems. Companion code: https://doi.org/10.5281/zenodo.21200069
Matthew Childs (Sun,) studied this question.