Coverage-driven alignment - What ‘Teaching Claude Why’ can borrow from AV verification (opens in new tab)

Cross-posted from The Foretellix CTO Blog. This is a full-text linkpost, following feedback that my previous piece was too brief as a stub.Summary: This post suggests that alignment training could benefit from coverage-driven verification. Anthropic recently reported that teaching Claude alignment rules (via pretraining-style next-token learning on alignment-related stories) is more effective than relying primarily on RL-style behavioral shaping. Some AV developers reached a related conclusio...