• 0 Posts
  • 9 Comments
Joined 1 year ago
cake
Cake day: October 24th, 2023

help-circle


  • Black hat and Defcon just ended and I’ll share my impression from LLM related talks given there. Microsoft VPs charged additional money to CISOs attending the summit talking about how AI will disrupt and be the future and blah blah magical thinking.

    Meanwhile Microsoft engineers and others said things like “this is logarithmic regression for people who are bad at math, and is best for cases where 75% accuracy is good enough. Try to break use cases into as many steps as possible and keep the LLM away from any automation that could have any consequences. These systems have no separation between the control plane and user input, which is re-exposing us to problems that were solved 15 years ago.”

    I think there are some neat possibilities that are lost in marketing hype as venture capitalist anger grows that they might have been scammed by yet another hammer in search of nails.






  • It’s valid to point out that we have difficulty defining knowledge, but the output from these machines are inconsistent at a conceptual level, and you can easily get them to contradict themselves in the spirit of being helpful.

    If someone told you that a wheel can be made entirely of gas do you have confidence that they have a firm grasp of a wheel’s purpose? Tool use is a pretty widely agreed upon marker of intelligence and so not grasping the purpose of a thing that they can describe at great length and exhaustive detail, while also making boldly incorrect claims on occassion should raise an eyebrow.