4 ms·

> A pass is a model with no human training data input that is able to reason on human content. OK, then humans fail hard. A baby raised with "no human training

> A pass is a model with no human training data input that is able to reason on human content. OK, then humans fail hard. A baby raised with "no human training data" (no conversation, no active teaching, no social interaction, obviously no books or videos) would be a total imbecile... or dead.

alganet 2y ago
You're using "training" in a looser way than me. LLMs have training data and context. I'm talking about training data. Conversation, interaction... those happen within an LLM context window. LLMs learn nothing from those experiences. What they can learn within a context window is super limited and disposable. You can't really compare the two. Humans are much, much more.
- ben_w 2y ago
  > Conversation, interaction... those happen within an LLM context window. LLMs learn nothing from those experiences. False. Why do you think e.g. ChatGPT has buttons for indicating if an answer is good or bad? Even when a specific version is frozen, this is a choice rather than a fundamental requirement, and the data from human feedback is used to continue training the system as a whole.
  - alganet 2y ago
    Ah, yes, the typical learning experience. You are born and fed a firehose of all the content available in the world. Then, when you need to learn something, you just firehose it all again from scratch with some added notes you made on a napkin. Can we at least agree that this cannot be compared with a human baby?
    - ben_w 2y ago
      > Ah, yes, the typical learning experience. You are born and fed a firehose of all the content available in the world. You are indeed born being fed a firehose of experience. Is it all the content available in the world? No, it's much less ordered than that. Also much less content. If your point were to have been "current AI are thick because they need a lot of examples to learn anything", you'd have a much easier time with your claims. (Some people still argue that point because vision is high bandwidth, but nevertheless it would be easier to argue than setting the bar so high that no human could possibly get there). > Then, when you need to learn something, you just firehose it all again from scratch with some added notes you made on a napkin. I have no idea what this is supposed to mean in this analogy. > Can we at least agree that this cannot be compared with a human baby? https://pmc.ncbi.nlm.nih.gov/articles/PMC27565/ https://pmc.ncbi.nlm.nih.gov/articles/PMC27565/ You can compare almost anything, if you want to. But why would it matter, that it's not like a baby? A baby treated the way you propose treating an AI, would still be somewhere between "feral" and "dead". That, even in isolation, is enough to show that your test is not well-conceived.
      - alganet 2y ago
        > You can compare almost anything, if you want to. So, you're a self-declared sophist. We can't get anywhere from here, I am not.
        ben_w 2y ago
        Follow the link. And then ask yourself what point you're even trying to make, because without clarity on that, my point is that your point is unclear. AI are like brains in certain ways, by design even, and different in others. You have yet to provide any suggestion for which specific ways to test for that simultaneously pass humans and fail AI.
        alganet 2y ago
        My first response about commoditization makes a fundamental observation about the nature of LLMs. Our discussion so far never touched that. You never actually adressed me, my most important claim remains unchecked, you just misderected the subject into an orchard.
- Hizonner 2y ago
  The bottom line is that you're demanding that LLMs somehow magically get to a certain level of functionality, without using input that you are allowing humans to have before you evaluate their functionality. How that input is used isn't relevant. Humans don't have separate training and inference training phases, but that does not mean that a human doesn't have to learn from other humans. Including just plain ingesting streams of stuff without interaction, although that's not the dominant mode for humans. By the way, any LLM you have ever interacted with has undergone reinforcement post-training in which it did in fact converse. The results of that are baked into the weights, not in the context. Not that it matters for the unreasonableness of your "test". And I predict that within a couple of years the training-versus-inference distinction in LLMs will start to break down. Which still doesn't matter when you're evaluating their current performance.

alganet 2y ago

You're using "training" in a looser way than me. LLMs have training data and context. I'm talking about training data. Conversation, interaction... those happen within an LLM context window. LLMs learn nothing from those experiences. What they can learn within a context window is super limited and disposable. You can't really compare the two. Humans are much, much more.

ben_w 2y ago
> Conversation, interaction... those happen within an LLM context window. LLMs learn nothing from those experiences. False. Why do you think e.g. ChatGPT has buttons for indicating if an answer is good or bad? Even when a specific version is frozen, this is a choice rather than a fundamental requirement, and the data from human feedback is used to continue training the system as a whole.
- alganet 2y ago
  Ah, yes, the typical learning experience. You are born and fed a firehose of all the content available in the world. Then, when you need to learn something, you just firehose it all again from scratch with some added notes you made on a napkin. Can we at least agree that this cannot be compared with a human baby?
  - ben_w 2y ago
    > Ah, yes, the typical learning experience. You are born and fed a firehose of all the content available in the world. You are indeed born being fed a firehose of experience. Is it all the content available in the world? No, it's much less ordered than that. Also much less content. If your point were to have been "current AI are thick because they need a lot of examples to learn anything", you'd have a much easier time with your claims. (Some people still argue that point because vision is high bandwidth, but nevertheless it would be easier to argue than setting the bar so high that no human could possibly get there). > Then, when you need to learn something, you just firehose it all again from scratch with some added notes you made on a napkin. I have no idea what this is supposed to mean in this analogy. > Can we at least agree that this cannot be compared with a human baby? https://pmc.ncbi.nlm.nih.gov/articles/PMC27565/ https://pmc.ncbi.nlm.nih.gov/articles/PMC27565/ You can compare almost anything, if you want to. But why would it matter, that it's not like a baby? A baby treated the way you propose treating an AI, would still be somewhere between "feral" and "dead". That, even in isolation, is enough to show that your test is not well-conceived.
    - alganet 2y ago
      > You can compare almost anything, if you want to. So, you're a self-declared sophist. We can't get anywhere from here, I am not.
      - ben_w 2y ago
        Follow the link. And then ask yourself what point you're even trying to make, because without clarity on that, my point is that your point is unclear. AI are like brains in certain ways, by design even, and different in others. You have yet to provide any suggestion for which specific ways to test for that simultaneously pass humans and fail AI.
        alganet 2y ago
        My first response about commoditization makes a fundamental observation about the nature of LLMs. Our discussion so far never touched that. You never actually adressed me, my most important claim remains unchecked, you just misderected the subject into an orchard.
Hizonner 2y ago
The bottom line is that you're demanding that LLMs somehow magically get to a certain level of functionality, without using input that you are allowing humans to have before you evaluate their functionality. How that input is used isn't relevant. Humans don't have separate training and inference training phases, but that does not mean that a human doesn't have to learn from other humans. Including just plain ingesting streams of stuff without interaction, although that's not the dominant mode for humans. By the way, any LLM you have ever interacted with has undergone reinforcement post-training in which it did in fact converse. The results of that are baked into the weights, not in the context. Not that it matters for the unreasonableness of your "test". And I predict that within a couple of years the training-versus-inference distinction in LLMs will start to break down. Which still doesn't matter when you're evaluating their current performance.