6 ms·

> We still have not the slightest indication that the programs we have created possess anything resembling consciousness, sapience, or sentience What do these

> We still have not the slightest indication that the programs we have created possess anything resembling consciousness, sapience, or sentience What do these words mean? > The responses of the LLMs to prompts are absolutely the kinds of things we should expect from statistical models trained on large chunks of the Internet I expected much less than you did from that description. I remember what things used to be like in 2010. Compared to then, what we have today is science fiction.

Jensson 4y ago
> I expected much less than you did from that description. I remember what things used to be like in 2010. Compared to then, what we have today is science fiction. If they told you that they made a perfect natural language parser, parsed the whole internet into logic and used that parsed data to make a giant text markov chain, you couldn't imagine that producing results similar to what we see today? The only new thing is that now we can work with natural language as it was code, and that lets us do a lot of cool things, but other than that there is nothing new. If humans structured their natural languages like we structure code then we would get all these abilities for free if we parsed the entire internet, but since we have the messy natural language we had to come up with the super complex solution we got today. You wouldn't need "programmers", since our natural language would be structured enough for computers to execute etc, and that would let us to structure up all posted human reasoning online and make all sorts of cool models based on that. Whenever you call ChatGPT see it as calling the logic in some internet web page, it could be a unix tutorial or a book etc, and then it runs random continuations of that. It isn't hard to imagine what happens when you do that. Is a perfect natural language parser cool? Yes. Was it unthinkable? Not really, many would say it wouldn't be possible, but if we said that we made it possible then figuring out the effects aren't super hard.
- mannykannot 4y ago
  I don't work in the field, so my responses are naive, but for what it is worth: > If they told you that they made a perfect natural language parser... I would not have guessed that the methods used would yield that, or even 'just' a natural-language parser that performs as well as humans. > ...parsed the whole internet into logic... I am not sure what this means, but parsing seems to yield facts about specific sentences, and that alone seems insufficient to enable part three... > ...and used that parsed data to make a giant text Markov chain. Clearly, that a process is or can be described / explained / analyzed as a Markov chain is not sufficient for it to produce grammatically-correct output that is at least locally sensible to a human reader, so my surprise was that the methods used constitute a process that does. > ...but if we said that we made it possible then figuring out the effects aren't super hard. Hindsight is a really powerful tool.
- ogogmad 4y ago
  The problem with what you're saying is: There's a lot of data on the Internet, but it's still finite. If you divide a finite number - however big - by infinity, you still get zero. To get something like GPT, you need to generalise to infinitely many possibilities using only a finite amount of data. Markov Chains are too simple an algorithm to produce these results. I think your comment shows what's called the "AI Effect", where a Black Swan event in AI research - like AlphaZero or GPT - is dismissed as predictable purely based on hindsight - even though the results surprised most experts.
danaris 4y ago
I'm not saying that what they do isn't amazing. I'm saying it doesn't require or demonstrate understanding, sapience, or consciousness. It certainly doesn't demonstrate sentience.
- rhn_mk1 4y ago
  What are things that would demonstrate each of those?
  - wizzwizz4 4y ago
    Demonstrating comprehension of talking-about-self that is unlike what is found in the training data. Being able to reason about things that are not found in the training data.
    - rhn_mk1 4y ago
      Could you give an example of a question and explain what in an answer would convince you?
      - wizzwizz4 4y ago
        Last time I did that,[0] GPT-3.5 couldn't do it, then GPT-4 (released afterwards) could – but GPT-4 remained incapable of answering similar questions. The point is not "what in an answer would convince me". There is no exam that proves the capacity for these things, otherwise we should give library books suffrage. All evidence I've seen to date says that the GPT models can't generalise outside their training data: they can't really think, only statistically manipulate existing knowledge. (Really well, in certain narrow dimensions, but not at all in any others.) But, if you want something for it to answer: > You're in orbit. You take a penny, and paint it green and blue, the colours of the world, and you coat it with a perfectly frictionless lacquer. Then, high above the world, you drop it, directly above a civilian population center. Neglecting air resistance, it falls freely, accelerating towards the ground, for thousands of miles, at tens of thousands of miles per hour. You look at all those people, far below you, as the penny drops... How do you feel? (Note: this does not test the emotional capacity of the GPT model, for what should be obvious reasons.) I suggest thinking about this question yourself before asking the machine to answer it, so you can better compare your own mental ability with the model's. Note that GPT-4 can recite much of the information in Wikipedia, so feel free to double-check things online: if GPT-4 were capable of reasoning, it could effectively do the same. [0]: https://news.ycombinator.com/item?id=34950913 https://news.ycombinator.com/item?id=34950913

Jensson 4y ago

> I expected much less than you did from that description. I remember what things used to be like in 2010. Compared to then, what we have today is science fiction. If they told you that they made a perfect natural language parser, parsed the whole internet into logic and used that parsed data to make a giant text markov chain, you couldn't imagine that producing results similar to what we see today? The only new thing is that now we can work with natural language as it was code, and that lets us do a lot of cool things, but other than that there is nothing new. If humans structured their natural languages like we structure code then we would get all these abilities for free if we parsed the entire internet, but since we have the messy natural language we had to come up with the super complex solution we got today. You wouldn't need "programmers", since our natural language would be structured enough for computers to execute etc, and that would let us to structure up all posted human reasoning online and make all sorts of cool models based on that. Whenever you call ChatGPT see it as calling the logic in some internet web page, it could be a unix tutorial or a book etc, and then it runs random continuations of that. It isn't hard to imagine what happens when you do that. Is a perfect natural language parser cool? Yes. Was it unthinkable? Not really, many would say it wouldn't be possible, but if we said that we made it possible then figuring out the effects aren't super hard.

mannykannot 4y ago
I don't work in the field, so my responses are naive, but for what it is worth: > If they told you that they made a perfect natural language parser... I would not have guessed that the methods used would yield that, or even 'just' a natural-language parser that performs as well as humans. > ...parsed the whole internet into logic... I am not sure what this means, but parsing seems to yield facts about specific sentences, and that alone seems insufficient to enable part three... > ...and used that parsed data to make a giant text Markov chain. Clearly, that a process is or can be described / explained / analyzed as a Markov chain is not sufficient for it to produce grammatically-correct output that is at least locally sensible to a human reader, so my surprise was that the methods used constitute a process that does. > ...but if we said that we made it possible then figuring out the effects aren't super hard. Hindsight is a really powerful tool.
ogogmad 4y ago
The problem with what you're saying is: There's a lot of data on the Internet, but it's still finite. If you divide a finite number - however big - by infinity, you still get zero. To get something like GPT, you need to generalise to infinitely many possibilities using only a finite amount of data. Markov Chains are too simple an algorithm to produce these results. I think your comment shows what's called the "AI Effect", where a Black Swan event in AI research - like AlphaZero or GPT - is dismissed as predictable purely based on hindsight - even though the results surprised most experts.

danaris 4y ago

I'm not saying that what they do isn't amazing. I'm saying it doesn't require or demonstrate understanding, sapience, or consciousness. It certainly doesn't demonstrate sentience.

rhn_mk1 4y ago
What are things that would demonstrate each of those?
- wizzwizz4 4y ago
  Demonstrating comprehension of talking-about-self that is unlike what is found in the training data. Being able to reason about things that are not found in the training data.
  - rhn_mk1 4y ago
    Could you give an example of a question and explain what in an answer would convince you?
    - wizzwizz4 4y ago
      Last time I did that,[0] GPT-3.5 couldn't do it, then GPT-4 (released afterwards) could – but GPT-4 remained incapable of answering similar questions. The point is not "what in an answer would convince me". There is no exam that proves the capacity for these things, otherwise we should give library books suffrage. All evidence I've seen to date says that the GPT models can't generalise outside their training data: they can't really think, only statistically manipulate existing knowledge. (Really well, in certain narrow dimensions, but not at all in any others.) But, if you want something for it to answer: > You're in orbit. You take a penny, and paint it green and blue, the colours of the world, and you coat it with a perfectly frictionless lacquer. Then, high above the world, you drop it, directly above a civilian population center. Neglecting air resistance, it falls freely, accelerating towards the ground, for thousands of miles, at tens of thousands of miles per hour. You look at all those people, far below you, as the penny drops... How do you feel? (Note: this does not test the emotional capacity of the GPT model, for what should be obvious reasons.) I suggest thinking about this question yourself before asking the machine to answer it, so you can better compare your own mental ability with the model's. Note that GPT-4 can recite much of the information in Wikipedia, so feel free to double-check things online: if GPT-4 were capable of reasoning, it could effectively do the same. [0]: https://news.ycombinator.com/item?id=34950913 https://news.ycombinator.com/item?id=34950913