214
A Collaborative Site That Allows You to See the Scientific Consensus on a Question with Just One Click (papermap.org)
submitted 1 year ago by to c/science@lemmy.world
27 comments fedilink hide all child comments
you are viewing a single comment's thread
view the rest of the comments
[–] 5 points 1 year ago* (3 children)

I use LLMs daily as a professional software engineer. I didn't downvote you and I'm not disengaging my thinking here. RAGs don't solve everything, and it's better not to sacrifice scientific credibility to the altar of convenience.

It's always been easier to lie quickly than to dig for the truth. AIs are not consistent, regardless of the additional appendages you give them. They have no internal consistency by their very nature.

  • permalink
  • fedilink
  • source
  • parent
  • hideshow 3 child comments
    •
  • [–] 3 points 1 year ago (1 child)

    And this isn’t even really a great application for RAG. Papermaps just goes off of references and citations. Perhaps a sentiment analysis would be marginally useful, but since you need a human to verify all LLM outputs it would be a dubious time savings.

    The system scores review papers very favorably and the “yes/no/maybe” conclusion is right in the abstract, usually the last sentence or two of it. This is not a prime candidate for any LLM, it’s simple database operations on srtuctured data that already exists. There’s no use case here.

  • permalink
  • fedilink
  • source
  • parent
  • hideshow 1 child comment
    •
  • [–] 3 points 1 year ago*

    Perhaps a sentiment analysis would be marginally useful, but since you need a human to verify all LLM outputs it would be a dubious time savings.

    Thank you, yes. That's exactly my point. You'd need a human to verify all of the outputs anyways, and these are literally machines that exclusively make text that humans find believable, so you're likely adding to the problem of humans messing stuff up moreso than speeding anything up. Being wrong fast has always been easy, so it's no help here.

  • permalink
  • fedilink
  • source
  • parent
  • [–] 1 point 1 year ago

    What would the failure rate on this be? What would the rate have to be to actually matter? Literally it would just poll the abstract and spit out yes no undecided. That is in the abstract. There is very little chance of there being any hallucinations that are meaningful at a degree large enough to vary literally anything.

    Have you never had it organize things or analyze sentiments? I understand if that's not your use case but this is pretty fundamentally an easy application of AI.

  • permalink
  • fedilink
  • source
  • parent