3 ms·

I feel similar, have any thoughts on what it would take for you to learn from a language model? I think for me, it's some sort of review from a native speaker.

I feel similar, have any thoughts on what it would take for you to learn from a language model? I think for me, it's some sort of review from a native speaker. If they gave a stamp of approval, I wouldn't think twice about learning from a human versus model.

bavarianbob 3y ago
I just read the article and OP actually does get it checked by a native speaker. You still have concerns despite that additional step?
zeta0134 3y ago
I don't think there's any claim you could make about the language model, factual or otherwise, that would resolve my primary hangup. As a non-native learner of a new language, I do not have a trained bullshit detector for that language. I cannot, by virtue of still being a novice, determine if the sentence structure sounds "weird," and I certainly can't determine if that weirdness is a limitation of the language model, some hallucination, whatever. So, by learning from that model, I would pick up any mistakes it makes and fold those into my own speech patterns. If I'm going to pick up speech patterns at all, I would really rather pick them up from a native speaker of the language, since at the very least I'll make the sorts of mistakes that a human might make. I want to sound like human, not like a language model. Language models sound like the average of several humans at best, and a strange program trying to imitate human speech at worst. Once I'm fluent in a language, enough to recognize when the language model itself is probably making a mistake, then I might become comfortable using it. But not as my first introduction to the nuances of the language, when I'm still building my own internal representation up from scratch. After all, my goal is to converse with other human speakers. Shouldn't that be my personal training corpus? I'm a neural network too, and I don't want to feed myself bad data.
- famouswaffles 3y ago
  >Language models sound like the average of several humans at best, and a strange program trying to imitate human speech at worst. This isn't really true. Open ai use heavy rlhf to make LLMs sound like that by default but they can sound like whatever. If a native speaker says it's fine then it's fine lol. You can still choose not to use it as you can choose not to do anything but then it's an irrational fear more than real concern. This is also essentially a substitute for graded readers for language learners
- bavarianbob 3y ago
  Really great food for thought here, I'm going to have linger on it some more. I don't know enough about LLMs to know how to counteract hallucinations, but I wonder if you could construct an LLM against a vetted corpus (much like Anthropic does with Claude) in which it solves for the problem of generating robotic speech patterns. I do agree with you a lot on the "sound" problem because I don't think I'd advise a non-native English speaker to blindly trust ChatGPT as a learning mechanism based on my experience of using it. So, in conclusion, I'm curious to learn whether this is a solvable problem or if LLMs are inherently not the right tool to use.
  - yorwba 3y ago
    Once you have a vetted corpus large enough to train an LLM on, I don't think there's any need to generate even more text with the LLM, since you can use the corpus directly.
  - wahnfrieden 3y ago
    yes llm can be deployed simply to search for material, to synthesize it, cite it, etc instead of being used to generate more similar material

bavarianbob 3y ago

I just read the article and OP actually does get it checked by a native speaker. You still have concerns despite that additional step?

zeta0134 3y ago

I don't think there's any claim you could make about the language model, factual or otherwise, that would resolve my primary hangup. As a non-native learner of a new language, I do not have a trained bullshit detector for that language. I cannot, by virtue of still being a novice, determine if the sentence structure sounds "weird," and I certainly can't determine if that weirdness is a limitation of the language model, some hallucination, whatever. So, by learning from that model, I would pick up any mistakes it makes and fold those into my own speech patterns. If I'm going to pick up speech patterns at all, I would really rather pick them up from a native speaker of the language, since at the very least I'll make the sorts of mistakes that a human might make. I want to sound like human, not like a language model. Language models sound like the average of several humans at best, and a strange program trying to imitate human speech at worst. Once I'm fluent in a language, enough to recognize when the language model itself is probably making a mistake, then I might become comfortable using it. But not as my first introduction to the nuances of the language, when I'm still building my own internal representation up from scratch. After all, my goal is to converse with other human speakers. Shouldn't that be my personal training corpus? I'm a neural network too, and I don't want to feed myself bad data.

famouswaffles 3y ago
>Language models sound like the average of several humans at best, and a strange program trying to imitate human speech at worst. This isn't really true. Open ai use heavy rlhf to make LLMs sound like that by default but they can sound like whatever. If a native speaker says it's fine then it's fine lol. You can still choose not to use it as you can choose not to do anything but then it's an irrational fear more than real concern. This is also essentially a substitute for graded readers for language learners
bavarianbob 3y ago
Really great food for thought here, I'm going to have linger on it some more. I don't know enough about LLMs to know how to counteract hallucinations, but I wonder if you could construct an LLM against a vetted corpus (much like Anthropic does with Claude) in which it solves for the problem of generating robotic speech patterns. I do agree with you a lot on the "sound" problem because I don't think I'd advise a non-native English speaker to blindly trust ChatGPT as a learning mechanism based on my experience of using it. So, in conclusion, I'm curious to learn whether this is a solvable problem or if LLMs are inherently not the right tool to use.
- yorwba 3y ago
  Once you have a vetted corpus large enough to train an LLM on, I don't think there's any need to generate even more text with the LLM, since you can use the corpus directly.
- wahnfrieden 3y ago
  yes llm can be deployed simply to search for material, to synthesize it, cite it, etc instead of being used to generate more similar material