AI-generated dossiers from the Jmail team.
  • Lodespawn@aussie.zoneEnglish
    22·
    12 days ago

    I think you’re thinking of a search engine, AIs are bad at searching. They just make up a response related to the search term based on their training data.

        • TheRealKuni@piefed.socialEnglish
          7·
          12 days ago

          You get a lot fewer hallucinations if it’s presenting data from sources rather than from its neural network alone. Training data isn’t, like, “in” the AI. It’s just used to shape its creation.

        • chicken@lemmy.dbzer0.comEnglish
          5·
          12 days ago

          Normally no, because that is much more difficult, resource intensive, and harder to get reliable results than separately looking up the information and including it in the prompt.

    • tomiant@piefed.socialEnglish
      7·
      12 days ago

      And if you train them on the document trove, they will be able to answer questions about it. It is a straight up trivial task.

      • Lodespawn@aussie.zoneEnglish
        11·
        12 days ago

        Make up answers about it. The answers might be right, or they might be wrong, you won’t know unless you read the actual data. So helpful …

        • tomiant@piefed.socialEnglish
          3·
          12 days ago

          “Give me the line numbers corresponding to the saudi sheik saying he liked the torture videos”

          Are you trying to be obstinate on purpose?

          • roofuskit@lemmy.worldEnglish
            4·
            12 days ago

            If we are talking about LLMs, the other commenter is entirely right about how they function. But I’m not sure you two are talking about the same technology.

            • tomiant@piefed.socialEnglish
              1·
              12 days ago

              Can an LLM provide me the information I want given a search term if trained on the given dataset? Yes. That is all.

              • balsoft@lemmy.mlEnglish
                3·
                11 days ago

                It can provide you some information that looks similar to what you’d want. Whether it is correct is another question.

                RAG can help to a degree but hallucinations still happen quite a bit.