I’ll start. Non serious answers also welcome

  1. Linux (Linux)

  2. FOSS or die

  3. Video content should have been text

  4. Not caring a LOT about privacy makes you a non-lemmy normie

(…)

  • CommunistEnglish
    1·
    1 年前

    I think LLM’s are on the right track, while an LLM with its current architecture likely couldn’t without a ridiculous scale, they do show signs of understanding ( https://www.businessinsider.com/chatgpt-open-ai-balancing-task-convinced-microsoft-agi-closer-2023-5 ), pretending they are nothing more than autocompletes as the people here do is disingenuous, what it does is predict, and while that’s all it does, that’s also all that makes humans special, the human mind is an object that takes sensory input, and predicts what muscle movements would be best given the sensory input, in fact, our heavy reliance on prediction is the reason magic tricks fool us, the only way to accurately predict things is through reasoning and understanding, we don’t know what happens when we scale, and there’s a reason experts predictions of when AGI will come are getting closer and closer, right before the LLM boom the average prediction was something like 40 years (based on memory), now it’s like, 10.

    I consider an LLM to be akin to what would happen if a persons thoughts were immediately transformed into words, without any layer of verification, you think plenty of wrong things, but you don’t say the wrong things you think because you have a layer of verification before speech, and it turns out, according to recent research, adding a verification layer to LLM’s is extremely potent: https://arxiv.org/abs/2203.14465

    It seems, according to this paper, that the trick is to have an LLM generate thousands of possible outputs, and have a separate tool verify their correctness, and then only present the correct output, this could possibly solve hallucination, which is one of the biggest roadblocks to actual intelligence.

    While we aren’t at true intelligence yet, we are creating the building blocks that will allow for it, and it will happen, and the experts believe it’s coming soon, LLM’s are not insignificant in terms of progress.

    These are tools made of the same component parts as our brain, admittedly, it takes approximately one thousand artificial neurons to simulate a real neuron, but the fact of the matter is, our minds are quite similar to these artificial minds, the artificial minds are just much, much, much, much more simple, it turns out, intelligence is likely a matter of statistical analysis.

    • gens@programming.dev
      1·
      1 年前

      When you look at a coffe cup from the side, you know it has a hole in it. Because you imagine, not because it’s a reflex.

      LLM is basically a point cloud of words. The training uses neural networks and thus pattern recognition. But the llm itself is closer to a database. But hey, sql is also useful for ai (data storage/retrival according to logic).

      I’m not an llm expert, by far. But right now they are not much more practical then a find out a bout things helper.

      Edit: I do like them. It’s been helpful a couple times and i even got gpt4all installed on my computer for fun.

      • CommunistEnglish
        21·
        1 年前

        When you look at a coffe cup from the side, you know it has a hole in it. Because you imagine, not because it’s a reflex.

        You’re looking at this backwards, you know those things because of previous experiences, you predict this might happen due to those.

        This is still a matter of prediction, and if that had never happened to you even once, I guarantee you wouldn’t look for it.

        They’re also significantly smaller than our brains and multimodality has been shown to help with reasoning, so, considering they’re text only and significantly smaller than our brains, their significantly reduced functionality is to be expected. Especially when you factor in that our brain has verification layers, which have only recently been discovered to work for LLM’s, none of them even implement this yet as far as i’m aware.