3 ms·

To me speaking requires independent thought and motive, which gpt's do not have. If I as a human want to express myself, I am capable of that. Gpt's only respon

To me speaking requires independent thought and motive, which gpt's do not have. If I as a human want to express myself, I am capable of that. Gpt's only respond with a series of plausible tokens. Someone mentioned in fine arts they make a distinction between craft and art, which is an excellent point. [Edit] https://news.ycombinator.com/item?id=39423510 https://news.ycombinator.com/item?id=39423510 [Edit] If all it takes to "speak" is vocalizing sounds, does text to speech count? [Edit] I guess it's in the name, text to "speech." Again, gpts do not have independent agency.

DiggyJohnson 3y ago
What word would you to describe what ChatGPT is doing then? Would "mimic-ing speech" suffice, for you? GPT responds in a "series of plausible tokens" that would pass as fluent English if generated by a human. Let's put aside the vocalization point, that's a distraction.
- EchoChamberMan 3y ago
  I think compute is a more appropriate term, but your right it's certainly an awkward phrase. It pains me when people compare gpts to humans - we're incredibly more complex and advanced.
- mvdtnz 3y ago
  It generates text.
  - JoshuaDavid 3y ago
    So does the program while True: print('a') GPT-n (not chat) generates continuations to text in which the probability distribution for the next output it produces matches the distribution of outputs after similar text in the training corpus, by most metrics you might examine. Trivial examples of those metrics include word and n-gram frequencies, but also it turns out that at sufficiently low loss that looks like "given a couple of input texts and a context which affords insightful commentary, produce insightful commentary at the base rate". There are, of course, caveats. Notably: - ChatGPT has its own stuff going on around chat tuning, tool use, tuning to be less likely to produce outputs OpenAI doesn't want, etc - The statistical regularity thing is not magic. If it requires more than n_layers steps of computation to determine the sensible next output, the model will not be able to do that. I think the canonical example here is usually having the model complete something like "'d072c916029965a7676da4244160c413e31bc8a0' == sha1('I saw it on hn'); '265149165fcb742a900a44b8f123885dc6ac5d12' == sha1('" and then having the model brute-force sha1 -- obviously it's not going to be able to do that. - The model generates text sequentially, one token at a time. This is importantly not the process by which most text in the training corpus was written. So in the cases where earlier text importantly depends on text which was written earlier temporally, but which occurs later in the string, the model will be likely to make mistakes (where a "mistake" is "writing text which is statistically surprising, relative to the training corpus").
    - DiggyJohnson 3y ago
      Is there not a way for me to express the crucial difference between non-language character repetition (or repetition of any string) and the ability or the ability to interpret and respond to human language in the same language? I just feel like we're not in a position to even begin understanding our disagreement until you at least recognize the question or point I'm trying to get here. If you disagree with the question, or don't understand it: why?
      - JoshuaDavid 3y ago
        I don't think that's where the difference lies, no. Sampling from a markov model built from English n-gram frequencies will produce non-repeated, mostly grammatical English text, which frequently is even sensible-sounding once n >= 4 or so. But I don't think there's anything going on in language generation beyond "based on the current context, produce an appropriate output token for that context based on the observed and inferred distribution of training inputs". I think that's also how human language generation works, though "the context" for humans includes a lot more than just a few thousand words of text. But I think that the surprising thing about e.g. GPT-4 is how well it does the thing, rather than the fact that it does the thing at all.

DiggyJohnson 3y ago

What word would you to describe what ChatGPT is doing then? Would "mimic-ing speech" suffice, for you? GPT responds in a "series of plausible tokens" that would pass as fluent English if generated by a human. Let's put aside the vocalization point, that's a distraction.

EchoChamberMan 3y ago
I think compute is a more appropriate term, but your right it's certainly an awkward phrase. It pains me when people compare gpts to humans - we're incredibly more complex and advanced.
mvdtnz 3y ago
It generates text.
- JoshuaDavid 3y ago
  So does the program while True: print('a') GPT-n (not chat) generates continuations to text in which the probability distribution for the next output it produces matches the distribution of outputs after similar text in the training corpus, by most metrics you might examine. Trivial examples of those metrics include word and n-gram frequencies, but also it turns out that at sufficiently low loss that looks like "given a couple of input texts and a context which affords insightful commentary, produce insightful commentary at the base rate". There are, of course, caveats. Notably: - ChatGPT has its own stuff going on around chat tuning, tool use, tuning to be less likely to produce outputs OpenAI doesn't want, etc - The statistical regularity thing is not magic. If it requires more than n_layers steps of computation to determine the sensible next output, the model will not be able to do that. I think the canonical example here is usually having the model complete something like "'d072c916029965a7676da4244160c413e31bc8a0' == sha1('I saw it on hn'); '265149165fcb742a900a44b8f123885dc6ac5d12' == sha1('" and then having the model brute-force sha1 -- obviously it's not going to be able to do that. - The model generates text sequentially, one token at a time. This is importantly not the process by which most text in the training corpus was written. So in the cases where earlier text importantly depends on text which was written earlier temporally, but which occurs later in the string, the model will be likely to make mistakes (where a "mistake" is "writing text which is statistically surprising, relative to the training corpus").
  - DiggyJohnson 3y ago
    Is there not a way for me to express the crucial difference between non-language character repetition (or repetition of any string) and the ability or the ability to interpret and respond to human language in the same language? I just feel like we're not in a position to even begin understanding our disagreement until you at least recognize the question or point I'm trying to get here. If you disagree with the question, or don't understand it: why?
    - JoshuaDavid 3y ago
      I don't think that's where the difference lies, no. Sampling from a markov model built from English n-gram frequencies will produce non-repeated, mostly grammatical English text, which frequently is even sensible-sounding once n >= 4 or so. But I don't think there's anything going on in language generation beyond "based on the current context, produce an appropriate output token for that context based on the observed and inferred distribution of training inputs". I think that's also how human language generation works, though "the context" for humans includes a lot more than just a few thousand words of text. But I think that the surprising thing about e.g. GPT-4 is how well it does the thing, rather than the fact that it does the thing at all.