Comment by lo_zamoyski
15 hours ago
> asking Claude which code snippet it prefers: the one it generated in a different chat, or one that I refactored [...] It always picks its own
...is not the same as claiming...
> LLMs favor LLM-generated passages over human written ones
Here, you're using the same LLM to both produce and judge the resulting work. If anything, I would expect an LLM to tend to prefer its own work given that the same training is producing and judging.
It's not intuitive to me for why preference for its own writing would emerge, and during what type of training or tuning.
Perhaps something like: learning to identify what source files it has worked on by the code style alone, because tasks may give human code (public repos, etc) and ask to make changes.
It’s optimizing for good writing. Therefore, it believes its outputs are good. Therefore, it believes inputs that look like its outputs are good.
This minus the word "believe". It's explainable simply by marching by similarity
It would need to be researched, but I wonder if it ends up being something that happens at the token level?