Alignement pretraining could backfire

·LessWrong··

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 lea...

Read full article →

Related Articles

Epic Games announces Lore version control system
regnerba · Hacker News · 1h ago
Apple is about to make Hide My Email useless
SXX · Hacker News · 21h ago
RFC 10008: The new HTTP Query Method
schappim · Hacker News · 5h ago
TIL: You can make HTTP requests without curl using Bash /dev/TCP
mrshu · Hacker News · 23h ago
A backdoor in a LinkedIn job offer
lwhsiao · Hacker News · 1d ago