3 ms·

"subliminal learning" does not even work for use cases like distilling o1 to R1 because they do not share a base model

• pyman 1y ago
  Who's talking about that? [Edit] My bad, I thought I was commenting on Anthropic's article
  • anonymoushn 1y ago
    i replied to a comment by the hacker news user called pyman which claimed incorrectly that distillation was repackaged as "subliminal learning". so if you are asking me, who is talking about subliminal learning, which is unrelated to the topic of the article, the answer is that the hacker news user called pyman is doing that.
    • pyman 1y ago
      Ah you are right, I was commenting on this article: https://alignment.anthropic.com/2025/subliminal-learning/ https://alignment.anthropic.com/2025/subliminal-learning/