Google's AI-powered Med-PaLM has achieved a passing grade on the U.S. Medical Licensing Exam, but still falls short of human doctors. ChatGPT, developed by OpenAI, has also achieved passing or near passing results. Google has developed a new evaluation benchmark to reduce "hallucinations" and a newer version of Med-PaLM has reached 86.5%. AI experts caution that AI should be used as an assistant, not a decision maker. Med-PaLM is being tested at the Mayo Clinic for administrative tasks.
    • gressenEnglish
      52·
      2 years ago

      AI can be fairly easy to democratize. The bleeding edge language models created two years ago with large effort are available today as open source projects. It’s difficult for companies to create long term business cases because of that.

      • trachemys@iusearchlinux.fyiEnglish
        2·
        2 years ago

        But getting permission from the government for the bot to write prescriptions will be as difficult as the government decides it will be.

  • Geek_KingEnglish
    153·
    2 years ago

    This is the application I think I’m most excited about for LLMs. A well trained LLM can correlate and reference more data points than any human could, which would help diagnose weird stuff a human doctor may or may not recognize. Especially if the LLM is kept up to date with cutting edge medical advances.

    • PixlbabbleOPEnglish
      5·
      2 years ago

      I mean in scifi there’s those tubes you go in and let the thing repair humans. Gonna want one at home lol.

  • mikkLEnglish
    93·
    2 years ago

    This is so exciting - there is a big potential here, if we are careful and cautious.

    I think the computer scientist they interview in the article hits the nail on the head with his statement:

    these models “should always be regarded as assistants rather than the final decision makers

    AI technologies as an extension of human ability is going to revolutionise a lot of professional fields. But, we need to approach the technology the right way! We need to start early, and have digital literacy as a focus area in schools.

  • jsveiga@sh.itjust.worksEnglish
    62·
    2 years ago

    I suppose that having access to all internet data used to train the AI would make it much easier for human students to pass the exam too.

    • PixlbabbleOPEnglish
      41·
      2 years ago

      How long before we’re governed by Ai?

    • Czorio@kbin.social
      1·
      2 years ago

      There’s better ways to get prestige and salary than spending over a decade in school getting treated like shit before becoming a fully specialised doctor, I reckon.