Comments by "LoneTech" (@0LoneTech) on "Generative AI's Greatest Flaw - Computerphile" video.

  1. You don't need to trick the AI. Just help it with its goal, to produce output (never mind how accurate).
    16
  2. This is about complicated prompts, not training, for that precise reason. There's no reliable distinction between late reference material and instructions.
    2
  3. That just relies on someone to sound the alarm and noone in HR seeing it as an opportunity to be lazy. Humans have quuite the tendency to prefer the lazy seeming options over correct or legal ones.
    2
  4. Some steganographic options, yes, but that was just one sample delivery method.
    1
  5. The core issue with this idea is that it's just as unreliable at understanding system prompts as any other. All boundaries remain fuzzy as all interpretations are.
    1
  6. That's an awful idea. Sounds entirely believable.
    1
  7.  @matsim0  The other change is more and more capacity for hidden channels. Nobody would notice a 20kb hidden addon for their sub-300 byte queries.
    1
  8. If you assume humans are available to inspect all input, never misinterpret any of it, and all agree on any subtext, and you can get the LLM to also agree perfectly. Turns out both are bad at reading in the first place.
    1
  9. Pain is terrible at enforcing obedience, and the idea of using it for that is cruelty.
    1
  10. The LLM has no concept of "data only". It relies on the same logic to extract meaning from any text, including the instructions.
    1
  11. @ It's not a matter of opinion. It's a single stream of tokens and you're expecting a strict boundary from a system largely designed to be fuzzy and unpredictable. It's like asking which term is biggest in 1+1+1.
    1
  12. More like authorizing representatives whose training consists of "play along with whoever visits".
    1
  13. PDFs don't need anything to be white; you can tag any element, visible or not, to contain arbitrary text, similar to the alt text of images in HTML.
    1