Like Sabine Hossenfelder, I Was Offered Money to Tell You AI Will Kill Us
www.youtube.com
Can AI really improve itself recursively until it becomes powerful enough to cause human extinction? Nobody sponsored this video, I turned down funding speci...
  • davel@lemmy.mlEnglish
    7·
    1 day ago

    Thank you for the cute polar bear achievement badge, attractive computer science video essayist.

    There may be diminishing returns to additional intelligence.

    Exactly. We currently have no reason to assume that the RSI graph is exponential rather than logarithmic, and if I were forced to place a bet today, I’d go with the later. Maybe it could go exponential at some point, but my bet would be that that isn’t on the near horizon.

    • ☆ Yσɠƚԋσʂ ☆@lemmy.mlOP
      4·
      22 hours ago

      Right, there’s this assumption that LLM improvements will just keep scaling indefinitely. It’s far more likely that that the architecture has limits which we seem to be approaching, and then you just get diminishing returns, so the graph looks more like a sigmoid where we saw a ton of rapid progress, and it starting to flatten out as the easy gains get plucked.