I started using chatbots yesterday and it surprised me how complete and articulate answers are. Main topics would be apparel, sewing, cooking, finance, ETFs, taxes, house maintenance, retirement.

Do you consider AI good enough to substitute human interaction? Because I remember a south park episode where randy marsh ignores his wife to interact exclusively with a bot. A parody of course but we have already plenty of AI porn and AI short stories for gullible people… and it works because they don’t stop watching (and sometimes believe it to be real).

I was always kind of an anti AI guy due to privacy, to avoid LLMs monetizing my questions and because I didn’t want to be a moron who has nothing better to do than to watch AI slop, but this is a tool so pervasive and practical I’m only isolating myself if I opt out of it. Thoughts?

Are there any questions you won’t ask a bot so the company behind it doesn’t monetize your input?

  • marmotini_enjoyer@lemmy.blahaj.zone
    link
    fedilink
    arrow-up
    1
    ·
    9 days ago

    LLM inference fits the definition a Markov chain, where the state is the input string to the LLM and the LLM’s output is taken to be the probability distribution for the next state transition. The state space is the finite set of token strings fitting in the LLM’s context window. The process of selecting the next token to output thus follows the Markov property. A whole readable string of text is generated by iterating this Markov process until the current state ends in an end-of-string token (or some other method).

    As an aside, this is one major reason why people say thay LLMs can’t learn. They are memoryless. Their output is wholly derived from the input state in a deterministic manner (barring inconsistencies in hardware or execution environment) that does not change after training (which is a separate process from inference).

    • brian@programming.dev
      link
      fedilink
      arrow-up
      1
      ·
      7 days ago

      ok but basically anything is markovian if you include enough in the state, which doesn’t feel like it fits with what op was trying to imply

      also I’d argue sampling makes it non deterministic