Alignement pretraining could backfire (opens in new tab)

Epistemic status: speculative, but I think the mechanism is plausible.There has been recent interest in generating synthetic documents to upsample examples of aligned AI during LLM pretraining. See, for instance, Geodesic's Alignment Pretraining paper or Anthropic's "Teaching Claude Why."I worry that this strategy can work well up to moderately capable models but backfire in dangerous, hard-to-notice ways once models acquire high situational awareness. I speculate that these techniques could ...