Hacker News

AI is ditching hallucinated references – and other bad habits

4 points by rndsignals ago | 3 comments

cyanydeez [-]

are they though? Since I can't read past the first few paragraphs, but are they testing raw LLMs or the mediated experience of anthropic and openai?

crefflar |root |parent [-]

Mediated. The actual research, unlike the nature article is freely available from the authors: https://zou-group.github.io/llm-claims-shelf-life/

cyanydeez |root |parent [-]

So most likely those hallucinations are being squased by some type of real time search somewhere in the system.