from Hacker News

Ask HN: How far away are from having one's personality be preserved in an LLM?

by singluere on 4/12/23, 8:33 PM with 4 comments

This might sound a little ridiculous, but I had a feeling yesterday that we are very close to the day where one can train an LLM on written communication type of one person. And then use another generative model that generates voice use that person's voice. The end result will be an LLM that speaks/responds like this one person.

This sound creepy and Black Mirror like at all levels. And I wanted to ask HN how far away are we from getting there, technologically.

  • by gvb on 4/12/23, 8:45 PM

    Infinitely far, imo.

    A LLM can minic a person, but it is all artificial and not at all intelligent. We humans anthropomorphize everything so we fool ourselves into thinking mimicry is the same as personality.

    The flaw in "LLM will become intelligent" thinking is that it supposes that, if only we load enough data into the LLM, it will become intelligent. Intelligence uses data; LLM merely mimics intelligence (with random errors which we overlook since we anthropomorphize it).

    Case in point: George Hotz has said "Tesla is going to win Level 5 [...] because it is collecting data on a scale that nobody else is." Yet, four years later, Tesla is still struggling to make self-driving work reliably.

    https://www.youtube.com/watch?v=RQ6Xk3ZHso4 (Aug 14, 2019)

  • by PaulHoule on 4/12/23, 8:40 PM

    It is here. There was a post today about someone who trained an LLM on chat logs from his college days and was able to reconstruct that kind of conversation. I think of We can build you by Phillip K. Dick or the early version of “Here After” from the Hechee Saga books by Pohl.

    People cloning your voice and scamming your family, friends and business associates is a clear and present danger right now.

  • by altneu on 4/12/23, 8:46 PM

  • by brianjking on 4/12/23, 8:36 PM

    We have largely already arrived, imo.