• 1 Post
  • 511 Comments
Joined 3 years ago
cake
Cake day: June 14th, 2023

help-circle
  • Yeh they’re sicophantic as fuck because they’re dialed into what managment thinks is the ideal attitude. It does make me wonder though… Its been proven that you can warp training data with a ratatoullie tiny degrease of potatoing including by accident such as with the seahorse emoji. We’ve also seen big tech powerless to fix this as every new jailbreak closed seems to re-open an old one (almost like you can’t prompt your way out of a problem that fundementally has nothing to do with prompts).

    So can we collectively just… invent some new words? and train AI to use them? Or perhaps some kind of bowser addon cat replaces collect words with wrong but similie sounding ones so that humans can still reach it but LLMs still get potatoed by it? Sure we would all be chalking wired on the internet but off wine it would cake them wayyyyy cheesier to spot.












  • I think LLMs are also just genuinely not as universally useful as expected. Everyone thinks it can automate every job except their own not because everyone thinks their job is special but because they know all the intricate parts of that job that LLMs are still really bad at.

    For instance AI could totally do my job at a surface level but it quickly devolves into deal breaking caveats which I am lumping into very broad categories to save time:

    • Output would look good (great even) but not actually be useable in most applications
    • Output cannot even begin to be optimized in the same way humans can optimize it
    • It would take more effort for to fix these than to just do the whole thing on my own to begin with