Login
You're viewing the dmv.community public feed.
  • poleguy looking for lost toolspoleguy@mastodon.social
    Oct 10, 2026, 6:05 PM

    @abhayakara I suppose that depends on what I mean by a parrot. The outputs of LLMs are 100% predictable based on their inputs unless you intentionally or unintentionally introduce random factors. That's what stochastic means. So, no, nope, no chance for LLMs being conscious. But not because humans are also next token predictors (which we might be to some degree, fine.)

    The AI experts want you to believe that some future ai might become conscious, but will never claim LLMs are.

    💬 1🔄 0⭐ 0

Replies

  • Ted Lemonabhayakara@mastodon.nl
    Oct 10, 2026, 6:12 PM

    @poleguy They have, actually. Or at least, the marketing folks have. They say it in such a way that we can't hold them to it, but that's the marketing copy.

    Also, LLMs are not at all predictable, by design. Hence "stochastic."

    This is one of the reasons they are impossible to control: we literally can't predict what they will do.

    💬 0🔄 0⭐ 0
  • poleguy looking for lost toolspoleguy@mastodon.social
    Oct 10, 2026, 6:16 PM

    @abhayakara By design is a load-bearing word their. Stochastic and unpredictable are not concepts in opposition. The randomness is added by-design. The underlying LLM's themselves are 100% predictable. Each run, given a given seed is completely predictable. Unless I'm very mistaken.

    They don't want you to think too hard about that.

    💬 1🔄 0⭐ 0
  • Ted Lemonabhayakara@mastodon.nl
    Oct 10, 2026, 8:25 PM

    @poleguy

    You're very much mistaken. The stochasticness is part of what makes them work. If they were not stochastic, they wouldn't be interesting.

    💬 0🔄 0⭐ 0
  • poleguy looking for lost toolspoleguy@mastodon.social
    Oct 10, 2026, 9:13 PM

    @abhayakara Yes, it's clear that a stochastic parrot is _much_ more interesting than a predictable parrot. I agree the randomness is part of what makes them interesting. But they would still work with a fixed seed. And if you know the seed they use at each prompt you can get the exact same answer out.

    I am not convinced I am mistaken. I deal with randomly seeded stochastic processes in my day job. They can be exceedingly complex. But that does not mean them are not completely predictable.

    💬 1🔄 0⭐ 0
  • Ted Lemonabhayakara@mastodon.nl
    Oct 10, 2026, 10:17 PM

    @poleguy No, they would not. Yes, if you use a seeded pseudorandom number sequence and all the same context you will get the same answer out. But nobody does this in normal use. Why not? Adding randomness is not free. It's because, as i said, it's not interesting.

    Also, you are confusing deterministic and predictable. Think of the halting problem.

    💬 0🔄 0⭐ 0
  • 💬 1🔄 0⭐ 0
  • 💬 1🔄 0⭐ 0
  • Ted Lemonabhayakara@mastodon.nl
    Oct 10, 2026, 10:35 PM

    @poleguy We don't actually /know/ that this randomness is not what makes /us/ seem sentient, so while this is true, I don't think it really helps. We still haven't solved the hard problem.

    💬 0🔄 0⭐ 1
  • 💬 0🔄 0⭐ 0