4 ms·

> This is overstated and easily disproved. ChatGPT produces accurate facts as a matter of course. test it right now "the idea that this stopped clock is broken

> This is overstated and easily disproved. ChatGPT produces accurate facts as a matter of course. test it right now "the idea that this stopped clock is broken is overstated and easily disproved. the clock produces accurate time as a matter of course. go ahead and ask what time is it, just make sure it is 3:45am or 3:45pm"

williamcotton 4y ago
And I’m also guessing that you don’t do anything that has 1 in a billion odds of death because it’s not worth the risk, right? Edit: woooooosh...
- pedrosbmartins 4y ago
  what? the argument here is that ChatGPT giving factual answers is a mere coincidence, not at all what the model was trained to do. It's a broken clock, it can tell you the correct time at very specific contexts, but you shouldn't rely on it as your source of factual information. If you feed him enough data saying the statue of liberty is 1 cm tall, it will happily answer a query with that "fact".
  - williamcotton 4y ago
    You're not talking about probabilities. You're talking in binaries. The broken clock is not the correct analogy.
    - remexre 4y ago
      The probability that the broken clock is right is straightforwardly 2/1440 = 0.001 = 0.1%, innit?
      - williamcotton 4y ago
        To the second? To the millisecond? What are we wanting here? You're missing the point. But I'll play this silly game: ChatGPT is not incorrect 99.9% of the time.
        mrtranscendence 4y ago
        Nor is it only incorrect one billionth of the time, as you seem to be indicating through your hypotheticals. Depending on what I've asked it about, it can be incorrect at an extremely high rate.
        williamcotton 4y ago
        That is definitely not what I am indicating. I'm pointing out the absurdity of speaking of probabilistic things in absolutes. Yes, ask an LLM to multiply a few numbers together and you will get around 100% failure rate. The same goes for quotes, citations, website addresses, and most numerical facts. The failures are predictable. That means the models can be augmented with external knowledge, Python or JS interpreters, etc.
      - slackdog 4y ago
        Clock correctness is relative. If the antique windup clock in your living room is off by 5 minutes, it's still basically right. But if the clock in your smartphone is 5 minutes off, something has clearly gone wrong.
    - pedrosbmartins 4y ago
      Any analogy is incorrect if you stretch it enough, otherwise it wouldn't be an analogy... My clock analogy works up to this: ChatGPT success in factually answering a query is merely a happy coincidence, so it does not work well as a primary source of facts. Exactly like... a broken clock. It correctly tells the time twice a day, but it does not work well as a primary source of time keeping. Please don't read more deeply into the analogy than that :)
      - williamcotton 4y ago
        A happy coincidence would imply random behavior. That’s not even remotely how an LLM functions. You’re not introducing any scale with regards to correctness either. It is a poor analogy without any stretching required.
        pedrosbmartins 4y ago
        Nope, not random behavior. ChatGPT works by predicting the continuation of a sentence. It has been trained in enough data to emulate some pretty awesome and deep statistical structure in human language. Some studies even argue it has built world models in some contexts, but I'd say that needs more careful analysis. Nonetheless, in no way, shape or form has it developed a sense of right vs wrong, real vs fiction, in a way you can depend on it for precise, factual information. It's a language model. If enough data says bananas are larger than the Empire State building, it would repeat that, even if it's absurd.
        williamcotton 4y ago
        I didn’t say it was random behavior. You did when you said it was a happy coincidence. I know it is just a language model. I know that if you took the same model and trained it on some other corpus that it would produce different results. But it wasn’t so it doesn’t have enough data to say that bananas are larger than the Empire State Building, not that it would really matter anyways. One important part of this story that you’re missing is that even if there were no texts about bananas and skyscrapers that the model could infer a relationship between those based on the massive amounts of other size comparisons. It is comparing everything to everything else. See the Norvig-Chomsky debate for a concrete example of how a language model can creat sentences that have never existed.

williamcotton 4y ago

And I’m also guessing that you don’t do anything that has 1 in a billion odds of death because it’s not worth the risk, right? Edit: woooooosh...

pedrosbmartins 4y ago
what? the argument here is that ChatGPT giving factual answers is a mere coincidence, not at all what the model was trained to do. It's a broken clock, it can tell you the correct time at very specific contexts, but you shouldn't rely on it as your source of factual information. If you feed him enough data saying the statue of liberty is 1 cm tall, it will happily answer a query with that "fact".
- williamcotton 4y ago
  You're not talking about probabilities. You're talking in binaries. The broken clock is not the correct analogy.
  - remexre 4y ago
    The probability that the broken clock is right is straightforwardly 2/1440 = 0.001 = 0.1%, innit?
    - williamcotton 4y ago
      To the second? To the millisecond? What are we wanting here? You're missing the point. But I'll play this silly game: ChatGPT is not incorrect 99.9% of the time.
      - mrtranscendence 4y ago
        Nor is it only incorrect one billionth of the time, as you seem to be indicating through your hypotheticals. Depending on what I've asked it about, it can be incorrect at an extremely high rate.
        williamcotton 4y ago
        That is definitely not what I am indicating. I'm pointing out the absurdity of speaking of probabilistic things in absolutes. Yes, ask an LLM to multiply a few numbers together and you will get around 100% failure rate. The same goes for quotes, citations, website addresses, and most numerical facts. The failures are predictable. That means the models can be augmented with external knowledge, Python or JS interpreters, etc.
    - slackdog 4y ago
      Clock correctness is relative. If the antique windup clock in your living room is off by 5 minutes, it's still basically right. But if the clock in your smartphone is 5 minutes off, something has clearly gone wrong.
  - pedrosbmartins 4y ago
    Any analogy is incorrect if you stretch it enough, otherwise it wouldn't be an analogy... My clock analogy works up to this: ChatGPT success in factually answering a query is merely a happy coincidence, so it does not work well as a primary source of facts. Exactly like... a broken clock. It correctly tells the time twice a day, but it does not work well as a primary source of time keeping. Please don't read more deeply into the analogy than that :)
    - williamcotton 4y ago
      A happy coincidence would imply random behavior. That’s not even remotely how an LLM functions. You’re not introducing any scale with regards to correctness either. It is a poor analogy without any stretching required.
      - pedrosbmartins 4y ago
        Nope, not random behavior. ChatGPT works by predicting the continuation of a sentence. It has been trained in enough data to emulate some pretty awesome and deep statistical structure in human language. Some studies even argue it has built world models in some contexts, but I'd say that needs more careful analysis. Nonetheless, in no way, shape or form has it developed a sense of right vs wrong, real vs fiction, in a way you can depend on it for precise, factual information. It's a language model. If enough data says bananas are larger than the Empire State building, it would repeat that, even if it's absurd.
        williamcotton 4y ago
        I didn’t say it was random behavior. You did when you said it was a happy coincidence. I know it is just a language model. I know that if you took the same model and trained it on some other corpus that it would produce different results. But it wasn’t so it doesn’t have enough data to say that bananas are larger than the Empire State Building, not that it would really matter anyways. One important part of this story that you’re missing is that even if there were no texts about bananas and skyscrapers that the model could infer a relationship between those based on the massive amounts of other size comparisons. It is comparing everything to everything else. See the Norvig-Chomsky debate for a concrete example of how a language model can creat sentences that have never existed.