3 ms·

I won't disagree with their findings; however I do think there is some need to counter the narrative that "LLM AI worse for humans". Specifically I think back t

I won't disagree with their findings; however I do think there is some need to counter the narrative that "LLM AI worse for humans". Specifically I think back to an example I use when I would describe why I had such motivation toward study students completing typing practice while learning CS. In short, I use the analogy that when I am browsing the web for code snippets (like extracting files from a tar file), I will explicitly retype out the command rather than rely on copy+paste. My logic is that typing out the command helps build the muscle memory so that someday I'll just REMEMBER the command. That said, the counter to my own counter is "do I really need to memorize that?" Yes yes no internet and I'm screwed... but that's such a rare edge case. I am able to quickly find the command and knowing that it is stored somewhere else may be enough knowledge for me rather than memorization. I can see Gen AI falling into a similar design, I don't need to know explicitly how to do something, just that that task can be resolved through an LLM prompt. Granted, we're still trying to figure out how to communicate with LLMs and we only really have 3 years of experience. Most of our insights have come from blog posts and a handful of research articles. I agree that Gen AI laziness is a growing issue, but I don't think it needs to go full Idiocracy sensationalist headline.

MrMcCall 2y ago
> we're still trying to figure out how to communicate with LLMs Communicating with 'guess-the-next-token' machines is just ELIZA version X.Y. You can also ask your dog it they want another treat, or if they want to play Quake. They're merely listening for key words and tone of voice and reacting to them according to their experience with that matrix. And LLMs don't even understand tone of voice, and they never will.
- orangecat 2y ago
  And LLMs don't even understand tone of voice, and they never will. My favorite part of AI discourse is the confident assertions that AIs will never be able to do things that they've been doing for months.
  - MrMcCall 2y ago
    They understand tone of voice? C'mon bruh.
    - tsumnia 2y ago
      I believe what orangecat means is that AIs don't understand tone ~yet~. However as we continue working on adding context and refining models, it could be a possibility in, say, 10 years. Think where we were 10 years ago - IoT was the big deal and IBM Watson was beating Jeopardy champions while still not understanding London isn't a US city. AI's "humanity" will be a consistent debate among people - there will always be people that choose to think of AI as nothing more than a tin can with an opinion built from a math equation, myself included. But these are the types of questions that we'll need to discuss as AI continues to progress while people continue to point out how AI can't do X (which makes them inferior to humans).
      - MrMcCall 2y ago
        > I believe what orangecat means is that AIs don't understand tone ~yet~. I don't see how you get that from what he quoted of mine then what he said. Besides, their understanding 'tone' is not a sensible prediction, for the simple fact LLMs don't ever understand anything, whatsoever, any more than ELIZA was a psychologist. They digest and then regurgitate word sequences in complex ways ~ it's the intrinsic and sole nature of their design. We can hide that fact with layers of obfuscation or tricky ways of packaging their output, but they are what they are, and nothing more. Like, human beings are never going to walk on the sun in our physical bodies; we are limited by our physiology, and there is no transcending it in dimensions beyond our capabilities.
        tsumnia 2y ago
        I am in agreement with you that LLMs are lifeless regurgitation machines; what I'm implying isn't too far from how children learn/interact. One of the first concepts we teach children is word association to objects and once they've learned the term, they use it to describe whatever electrical signals are firing in their brain that made them decide "apple". If the difference is that LLMs use numerical representations to determine those "firings" in its decision making process, then would the numerical representations of an fMRI scanner hooked to a human be equivalent? Of course not, because humans > robots. BUT, that does not discount the fact that humans learn through methods like memorization, which have their own reinforcement learning patterns. Many of the cognitive science* concepts have similar LLM infrastructure equivalents as well - "Mind Palace" is similar to RAG, the different relationships we have with others is akin to LLM role-playing, one/few-shot prompting is like templates. I've recently been thinking about what types of discussions we'll have around AI in about 20 years - a new generation of young adults will leaving school, where we are currently introducing them to AI tutor avatars. Combined with things like Neuro-sama, the AI driven Twitch streamer, I'm curious about how the NEXT generation will perceive AI since they've grown up with it. This isn't a Futurama Marilyn Monroe Bot fantasy, its sort of a reality (whether I agree with it or not).

MrMcCall 2y ago

> we're still trying to figure out how to communicate with LLMs Communicating with 'guess-the-next-token' machines is just ELIZA version X.Y. You can also ask your dog it they want another treat, or if they want to play Quake. They're merely listening for key words and tone of voice and reacting to them according to their experience with that matrix. And LLMs don't even understand tone of voice, and they never will.

orangecat 2y ago
And LLMs don't even understand tone of voice, and they never will. My favorite part of AI discourse is the confident assertions that AIs will never be able to do things that they've been doing for months.
- MrMcCall 2y ago
  They understand tone of voice? C'mon bruh.
  - tsumnia 2y ago
    I believe what orangecat means is that AIs don't understand tone ~yet~. However as we continue working on adding context and refining models, it could be a possibility in, say, 10 years. Think where we were 10 years ago - IoT was the big deal and IBM Watson was beating Jeopardy champions while still not understanding London isn't a US city. AI's "humanity" will be a consistent debate among people - there will always be people that choose to think of AI as nothing more than a tin can with an opinion built from a math equation, myself included. But these are the types of questions that we'll need to discuss as AI continues to progress while people continue to point out how AI can't do X (which makes them inferior to humans).
    - MrMcCall 2y ago
      > I believe what orangecat means is that AIs don't understand tone ~yet~. I don't see how you get that from what he quoted of mine then what he said. Besides, their understanding 'tone' is not a sensible prediction, for the simple fact LLMs don't ever understand anything, whatsoever, any more than ELIZA was a psychologist. They digest and then regurgitate word sequences in complex ways ~ it's the intrinsic and sole nature of their design. We can hide that fact with layers of obfuscation or tricky ways of packaging their output, but they are what they are, and nothing more. Like, human beings are never going to walk on the sun in our physical bodies; we are limited by our physiology, and there is no transcending it in dimensions beyond our capabilities.
      - tsumnia 2y ago
        I am in agreement with you that LLMs are lifeless regurgitation machines; what I'm implying isn't too far from how children learn/interact. One of the first concepts we teach children is word association to objects and once they've learned the term, they use it to describe whatever electrical signals are firing in their brain that made them decide "apple". If the difference is that LLMs use numerical representations to determine those "firings" in its decision making process, then would the numerical representations of an fMRI scanner hooked to a human be equivalent? Of course not, because humans > robots. BUT, that does not discount the fact that humans learn through methods like memorization, which have their own reinforcement learning patterns. Many of the cognitive science* concepts have similar LLM infrastructure equivalents as well - "Mind Palace" is similar to RAG, the different relationships we have with others is akin to LLM role-playing, one/few-shot prompting is like templates. I've recently been thinking about what types of discussions we'll have around AI in about 20 years - a new generation of young adults will leaving school, where we are currently introducing them to AI tutor avatars. Combined with things like Neuro-sama, the AI driven Twitch streamer, I'm curious about how the NEXT generation will perceive AI since they've grown up with it. This isn't a Futurama Marilyn Monroe Bot fantasy, its sort of a reality (whether I agree with it or not).