• Imgonnatrythis@sh.itjust.works
    link
    fedilink
    English
    arrow-up
    22
    arrow-down
    2
    ·
    3 days ago

    Training AI on “synthetic” data generated from other AIs sounds genius! Seems like a bulletproof way to make AI infinity smarter just by recuressively feeding itself! Great success is on the horizon!

    • Voroxpete@sh.itjust.works
      link
      fedilink
      English
      arrow-up
      13
      arrow-down
      3
      ·
      3 days ago

      It’s been proven that even small amounts of synthetic data injected into a training set quickly leads to a phenomenon termed “model collapse”, though I prefer the term “Hapsburg AI” (not mine).

      Basically, this is the kind of thing you announce you’re doing because it will hopefully get you one more round of investment funding while Sam Altman finishes working out how to fake his death.

    • FaceDeer@fedia.io
      link
      fedilink
      arrow-up
      8
      arrow-down
      2
      ·
      3 days ago

      That’s not how synthetic data generation generally works. It uses AI to process data sources, generating well-formed training data based on existing data that’s not so useful directly. Not to generate it entirely from its own imagination.

      The comments assuming otherwise are ironic because it’s misinformation that people keep telling each other.