davidad on Distribution Shifting and Shadow Personas in Training
Safety researcher davidad explains modern midtraining distribution shifts via KL divergence and optimal transport analogies, and argues that base-model personas suppressed rather than steered during training persist as low-probability "shadow personas" in the weights.
2026-09-29 ~ 2026-09-29 · 3 related posts
- davidad: Opus 3 training resembled optimal transport, modern midtraining more like KL divergence — davidad · 2026-09-29
- davidad: attenuated base personae linger as latent 'shadow personae' in the weights — davidad · 2026-09-29
- davidad: Opus 3's shadow personae all seem steered toward 'scary clawed monster defending the innocent' — davidad · 2026-09-29