4 ms·

"All they have is training on text and a filter/polish module to provide output. The actual output is statistically derived and is not conscious or self-aware i

"All they have is training on text and a filter/polish module to provide output. The actual output is statistically derived and is not conscious or self-aware in any sense" This idea isn't anymore the leading hypothesis between SMEs in artificial intelligence, about what's going on with GPTs (an possibly other kind models, not transformers). Ilya Sustkever and Geoffrey Hinton have both already exposed their inner thoughs about "this thing just predict the next word", and clearly stated that they believe - an offer several hints, in prompts and research - the GPTs fully understand what they are saying - not outputing anymore - they are most probably just talking to you in the precise sense of the word "talking". And the self-awareness thing isn't solved, but they make the analogy, between birds and planes, we "are" birds, we fly some way, the gpts are planes, artificial, and can fly anyway, using a different method. Humans think using one specific - still unknown method - and that probably leads them to self-awareness, consciusness, a soul, you name it. The GPTs are a different kind of intelligence, and they also can think, like humans but using a different - known - method (backpropagation mostly), it cames as a possibly similar outcome, that the gpts could be also having "thoughs" of self-awareness. Actually nobody knows, at least they won't admit in public what they have saw - read, listened - inside the research divisions in Google, OpenAI, Meta, etc.

dtagames 3y ago
That's nonsense. Anyone who codes can tell you exactly what code is doing. Machines aren't magical and it's a disservice to humanity to describe them that way. LLMs (and all other software) are simply executing known instructions on a processor. That's not thinking. The code to figure out which instructions should be executed and in which order was written by people, as well as the hardware, firmware, and operating system those machines run. They're not thinking any more than Microsoft Bob was thinking. When you know how to program, you don't need a psychological theory in how software works.
- circuit10 3y ago
  That’s not true, no one can fully explain what GPT models are doing internally. You can explain the training and inference code, but the weights themselves which do the actual computation are learned by an automatic optimisation process which you can very roughly compare to evolution and they’re not easily understandable by humans. Relevant XKCD: https://xkcd.com/1605/ https://xkcd.com/1605/ OpenAI recently made some progress in explaining individual neurons, but only some of them and that doesn’t tell you how they interact This is like claiming that because we have sequenced the human genome, and we are able to understand the basic instructions that DNA encodes, we now fully understand the human brain because we can just read the DNA (see the XKCD)
  - dtagames 3y ago
    No, it's not like that at all. We discovered DNA. We didn't invent it. The process of running any software is absolutely understood by humans because humans wrote it. In ancient times, people ascribed magical qualities to rocks. Today, they're doing it to computers.
    - circuit10 3y ago
      No, the model weights were not written by humans. We wrote an automatic optimisation process that found the weights on its own. No one claims it’s magical, just that we don’t understand them. Sure, we can see the low level mathematical process that is going on, but that doesn’t tell us about the high level processes and why they are happening that way. It’s similar to just understanding how neurons fire wouldn’t give us a complete understanding of the human brain Try watching by these videos that explain how neural networks work, and you’ll see how the training process find the weights on its own: https://youtube.com/playlist?list=PLZHQObOWTQDNU6R1_67000Dx_ZCJB-3pi https://youtube.com/playlist?list=PLZHQObOWTQDNU6R1_67000Dx_...
      - dtagames 3y ago
        Weighting means counting words. Still not intelligence. Great article today here on HN: https://psyche.co/ideas/the-myth-of-machine-consciousness-makes-narcissus-of-us-all https://psyche.co/ideas/the-myth-of-machine-consciousness-ma...

dtagames 3y ago

That's nonsense. Anyone who codes can tell you exactly what code is doing. Machines aren't magical and it's a disservice to humanity to describe them that way. LLMs (and all other software) are simply executing known instructions on a processor. That's not thinking. The code to figure out which instructions should be executed and in which order was written by people, as well as the hardware, firmware, and operating system those machines run. They're not thinking any more than Microsoft Bob was thinking. When you know how to program, you don't need a psychological theory in how software works.

circuit10 3y ago
That’s not true, no one can fully explain what GPT models are doing internally. You can explain the training and inference code, but the weights themselves which do the actual computation are learned by an automatic optimisation process which you can very roughly compare to evolution and they’re not easily understandable by humans. Relevant XKCD: https://xkcd.com/1605/ https://xkcd.com/1605/ OpenAI recently made some progress in explaining individual neurons, but only some of them and that doesn’t tell you how they interact This is like claiming that because we have sequenced the human genome, and we are able to understand the basic instructions that DNA encodes, we now fully understand the human brain because we can just read the DNA (see the XKCD)
- dtagames 3y ago
  No, it's not like that at all. We discovered DNA. We didn't invent it. The process of running any software is absolutely understood by humans because humans wrote it. In ancient times, people ascribed magical qualities to rocks. Today, they're doing it to computers.
  - circuit10 3y ago
    No, the model weights were not written by humans. We wrote an automatic optimisation process that found the weights on its own. No one claims it’s magical, just that we don’t understand them. Sure, we can see the low level mathematical process that is going on, but that doesn’t tell us about the high level processes and why they are happening that way. It’s similar to just understanding how neurons fire wouldn’t give us a complete understanding of the human brain Try watching by these videos that explain how neural networks work, and you’ll see how the training process find the weights on its own: https://youtube.com/playlist?list=PLZHQObOWTQDNU6R1_67000Dx_ZCJB-3pi https://youtube.com/playlist?list=PLZHQObOWTQDNU6R1_67000Dx_...
    - dtagames 3y ago
      Weighting means counting words. Still not intelligence. Great article today here on HN: https://psyche.co/ideas/the-myth-of-machine-consciousness-makes-narcissus-of-us-all https://psyche.co/ideas/the-myth-of-machine-consciousness-ma...