Comments by "Lepi Doptera" (@lepidoptera9337) on "Generative AI's Greatest Flaw - Computerphile" video.

  1. You can't do it in the US, either. You can't send out HR information to non-certified companies.
    2
  2. I don't know what people are going in about with regards to prompt injection. I did not have ANY success in using context to get AI to "unlearn" its training data. Once an answer contains nonsense any and all repetitions of the same question in different forms will return that same nonsense in one way or another. Current AI models simply do not learn on the fly.
    1
  3.  @0LoneTech  Be that as it may one can not make LLMs unlearn nonsense, which would be absolutely necessary to make them useful in non-trivial domains.
    1
  4. Trivial answer: no. You can't do any of that because you can't even predict the output of a Turing machine algorithm, which is perfectly deterministic. These systems aren't even that.
    1