关于AI的胡编乱造/hallucinate倾向,感觉这句说到点子上了:
“The coherence is structural rather than reflective: it inheres in the smoothness of a sentence, not in a commitment to the reality it describes, or to its moral valence. [...] Where there is uncertainty, AIs rush in to fill the gap with plausibility.”
LLM作为prediction machine确实就是prioritize这种言语表层的smoothness的。