4 ms·

> using a selection of ``textbook quality" data from the web ... and synthetically generated textbooks and exercises with GPT-3.5 Is this common, training a LL

> using a selection of ``textbook quality" data from the web ... and synthetically generated textbooks and exercises with GPT-3.5 Is this common, training a LLM from another LLMs generated output? How do you avoid "bad code" from GPT, if not out right hallucinations?

kramerger 3y ago
If you think about it, how do humans learn things? A child gets alot of his knowledge by randomly copying what grownups do.
famouswaffles 3y ago
>Is this common, training a LLM from another LLMs generated output? It's not uncommon. And the level the SOTA LLMs are now, it'll only become more common. It happens because we're at the stage where GPT-3.5/4 can generate much better data (or hit more encompassing or general distributions with the right instructions) than what the majority of LLMs will be typically trained on. Also 4 can recheck output for bugs, mistakes, errors etc It may not be perfect but neither would "natural data" and depending on exactly what kind of data, it might be as close to perfect as you'll get - https://arxiv.org/abs/2305.07759 https://arxiv.org/abs/2305.07759
- zarzavat 3y ago
  When humans think this is what they are doing: retraining themselves on their own learned data. In other words, knowledge is not a conserved quantity. Of course the extent of the extra knowledge you can produce without access to the environment is limited.
  - ChatGTP 3y ago
    Do you have any actual idea if that's the case or are you just saying stuff? I've read that people reorganize information when they sleep, I wouldn't say that's "thinking".
    - batch12 3y ago
      What about dreaming? I think when I dream, sometimes even finding solutions to problems for my woke self.
    - TeMPOraL 3y ago
      Don't look at sleeping. Look at your inner voice, if you have one, and the back-and-forth you're doing with it. This is a form of thinking, and it is also clearly feeding your mind's output back into it as input.
      - bamboozled 3y ago
        I have never heard an "inner voice" when sleeping or dreaming so I have on idea what you're referring too ? I'd say that inner voices are mostly stupid idiots :)
        TeMPOraL 3y ago
        The "inner voice" isn't a sleeping thing, it's an awake thing. I guess you may be one of the presumably minority of people who doesn't have an inner narrative? I'm having a hard time imagining how this works, but then again, I'm aphantasic, and a lot of people have trouble imagining how that works too.
        bamboozled 3y ago
        So in dreams you have conversations with yourself? Who are your Temporal? A researcher? Neurologist? Just a guy on the internet who reads a lot of comments?
    - zarzavat 3y ago
      It was intended as philosophy (“just saying stuff”) rather than neuroscience. i.e. a black box rather than white box perspective on the mind. The point is that you don’t need a recent external environment to grow your knowledge in certain ways. For example, if you locked a sufficiently gifted and immortal mathematician in prison with no knowledge of the outside world after 2021, he can still prove new theorems forever afterwards (provided he has enough coffee). He doesn’t need access to recent mathematical results to keep proving theorems and growing his knowledge base - that’s simply an accelerant. In fact the longer he stays in prison, the more new theorems he can prove because he has ever more lemmas. Whereas if you locked a sufficiently gifted and immortal chemist in prison with no knowledge after 2021, she might be able to synthesize some new knowledge at first (meta analysis, etc), but eventually that would run dry and she wouldn’t be able to say anything new about chemistry anymore without any equipment.
kleton 3y ago
This is probably why openai sets the knowledge cutoff when it does.
adt 3y ago
Exactly. The False Promise of Imitating Proprietary LLMs (UC Berkeley, 25/May/2023) https://arxiv.org/abs/2305.15717 https://arxiv.org/abs/2305.15717

kramerger 3y ago

If you think about it, how do humans learn things? A child gets alot of his knowledge by randomly copying what grownups do.

famouswaffles 3y ago

>Is this common, training a LLM from another LLMs generated output? It's not uncommon. And the level the SOTA LLMs are now, it'll only become more common. It happens because we're at the stage where GPT-3.5/4 can generate much better data (or hit more encompassing or general distributions with the right instructions) than what the majority of LLMs will be typically trained on. Also 4 can recheck output for bugs, mistakes, errors etc It may not be perfect but neither would "natural data" and depending on exactly what kind of data, it might be as close to perfect as you'll get - https://arxiv.org/abs/2305.07759 https://arxiv.org/abs/2305.07759

zarzavat 3y ago
When humans think this is what they are doing: retraining themselves on their own learned data. In other words, knowledge is not a conserved quantity. Of course the extent of the extra knowledge you can produce without access to the environment is limited.
- ChatGTP 3y ago
  Do you have any actual idea if that's the case or are you just saying stuff? I've read that people reorganize information when they sleep, I wouldn't say that's "thinking".
  - batch12 3y ago
    What about dreaming? I think when I dream, sometimes even finding solutions to problems for my woke self.
  - TeMPOraL 3y ago
    Don't look at sleeping. Look at your inner voice, if you have one, and the back-and-forth you're doing with it. This is a form of thinking, and it is also clearly feeding your mind's output back into it as input.
    - bamboozled 3y ago
      I have never heard an "inner voice" when sleeping or dreaming so I have on idea what you're referring too ? I'd say that inner voices are mostly stupid idiots :)
      - TeMPOraL 3y ago
        The "inner voice" isn't a sleeping thing, it's an awake thing. I guess you may be one of the presumably minority of people who doesn't have an inner narrative? I'm having a hard time imagining how this works, but then again, I'm aphantasic, and a lot of people have trouble imagining how that works too.
        bamboozled 3y ago
        So in dreams you have conversations with yourself? Who are your Temporal? A researcher? Neurologist? Just a guy on the internet who reads a lot of comments?
  - zarzavat 3y ago
    It was intended as philosophy (“just saying stuff”) rather than neuroscience. i.e. a black box rather than white box perspective on the mind. The point is that you don’t need a recent external environment to grow your knowledge in certain ways. For example, if you locked a sufficiently gifted and immortal mathematician in prison with no knowledge of the outside world after 2021, he can still prove new theorems forever afterwards (provided he has enough coffee). He doesn’t need access to recent mathematical results to keep proving theorems and growing his knowledge base - that’s simply an accelerant. In fact the longer he stays in prison, the more new theorems he can prove because he has ever more lemmas. Whereas if you locked a sufficiently gifted and immortal chemist in prison with no knowledge after 2021, she might be able to synthesize some new knowledge at first (meta analysis, etc), but eventually that would run dry and she wouldn’t be able to say anything new about chemistry anymore without any equipment.

kleton 3y ago

This is probably why openai sets the knowledge cutoff when it does.

adt 3y ago

Exactly. The False Promise of Imitating Proprietary LLMs (UC Berkeley, 25/May/2023) https://arxiv.org/abs/2305.15717 https://arxiv.org/abs/2305.15717