M.M.

• M.M.Mar 11, 2025, 8:14 PM

  1 point

  0

  on: Emer­gent Misal­ign­ment: Nar­row fine­tun­ing can pro­duce broadly mis­al­igned LLMs

  Interesting that the two questions producing the highest misalignment are the unlimited power prompts (world ruler, one wish).

• M.M.Mar 11, 2025, 7:55 PM

  1 point

  0

  in reply to: Daniel Tan’s comment on: Emer­gent Misal­ign­ment: Nar­row fine­tun­ing can pro­duce broadly mis­al­igned LLMs

  Re “universal representation of behaviour which is aligned /​ not aligned”—reminiscent of an idea from linguistics. Universal Grammar provides a list of parameters; all languages have the same list. (Example: can you drop a subject pronoun? In English the answer is no, in Spanish the answer is yes.) Children start with all parameters on the default setting; only positive evidence will induce them to reset a parameter. (So for pro-drop, they need to hear a sentence—as in Spanish—where the subject pronoun has been dropped.) Evidence came from the mistakes children make when learning a language, and also from creole languages, which were said to maintain the default parameter settings. I don’t know if this idea is still current in linguistics.