5 ms·

I think this is a bit pedantic. At the end of the day, the weights that determine the odds of the next term are still completely bound by the limits of the trai

• basch 4y ago
  But knowledge can be synthesized from information between the cracks of the training data. If one document says word A relates to B and one says word B relates to C, the model has an association between A and C even if no human has ever noticed it before. At their heart, models are word relationship strength distance and frequency charts. They encode both the semantic meanings of words and the grammar and structure of language. When they take apart a sentence they understand it, if understanding is knowing how closely an idea is to other ideas. Think about how your own brain works when you hear a phrase, and suddenly it causes you to quite a tv show.
• rstuart4133 4y ago
  > the weights that determine the odds of the next term are still completely bound by the limits of the training data Depends on your definition of "completely bound" I guess, but my definition goes something along the lines of "functionally dependent on the input" in the mathematical sense. Parroting implies the only two input's here are the training data and the prompt. But there is a third input - a random number. ChatGPT randomly chooses from a list of likely words. I don't know how big that list is, but lets say it's 4 words. So two bits of randomness. But then on the basis of that word it chooses another random word, and another, and another, each time adding two more bits of randomness. ChatGPT responses are typically over well over 100 words, so that's well over 200 bits of randomness injected. That is enormous. And it's not compatible with the picture the OP paints - which is one where the AI is simply parroting back what it's seen. This difference between accurately reproducing what it's seen and making it up as it goes along is what determines where it is useful, and where it isn't. If you need accuracy, current AI's aren't so useful. It's why Tesla's full self driving is proving to be a struggle - making a mistake every 10,000 hours and killing someone as a consequence is not an acceptable outcome. But accuracy isn't so important when making up a poem, or writing marketing copy, or painting a picture, or even translating a foreign language - so in those areas it is more successful. Which should tell you something. Because what I've just said is if you need a good parrot, ChatGPT isn't your tool. It makes stuff up as it goes along. Yes strongly guided by it's training data for the next word, but at the end of a 100 word response there 200 bits of pure bullshit in what it said. The way the training data constrains that bullshit makes it look spookily like a human produced it. This has caught me completely by surprise. If you asked me what jobs would be most threatened by computers just a year ago, I would have said it's tasks like assembling a car where their legendary accuracy and repeatability comes to the fore. Significantly, any task that required what you might call "human creativity" was at the bottom of the list - as in "could not be done by a computer in the foreseeable future" bottom. And yet here we are - copy writers are looking like an endangered species. No one would call copy writers and graphic artists parrots. I sometimes wonder whether this denial of what seems to be self evident (ie, that ChatGPT does produce original output - which is objectively true statement) is people in denial of the fact that we have managed to reproduce one aspect of "humanness" with a computer.
  • z3c0 4y ago
    The sole aspect of this denial is that I work in ML and know for a fact that ChatGPT's "humanness" is - at best - superficial. It's mimicking the text created by humans, not formulating the core ideas in a soupy model of reality and then applying terms through denotation. And when I say mimicking, I don't mean "regurgitating its training data". It's not parroting what was said, but it IS parroting the order in which things are said, using an RNG to add a little variance. That's all. Unlike ChatGPT, or any ML model, humans are still human in the total absence of language* stimuli. Maybe animalistic, even vegetative, but still "intelligent" enough to maintain bodily functions. No training required, thus no upper or lower bounds that one knows will constrain the final output. That is because the way to "make" a brain is predefined in our instructions, DNA. A trained model has no "functional default" the way actual intelligence does. It's a very advanced information photocopier that can probalistically arrange an apparently "fresh" (but fundamentally NOT) output based on toggled bits. At best, ChatGPT will be a semantic layer for an actually intelligent AI capable of truly original output, which will allow that "denotative" layer that you and I use to communicate. A good read on the subject: https://en.wikipedia.org/wiki/On_Denoting https://en.wikipedia.org/wiki/On_Denoting The full paper: https://www.philosophy-index.com/russell/on-denoting/Russell_-_On_Denoting.pdf https://www.philosophy-index.com/russell/on-denoting/Russell... edit: missed a rather important word that changed the whole meaning of the sentence.
    • rstuart4133 4y ago
      > No training required, thus no upper or lower bounds that one knows will constrain the final output. This is were we depart. Humans get lots of training. Around 20 years of it before the are classed as adults. Even in adulthood we continue seek it out, pretty much continuously. I'm doing that here, by reading HN and seeking feedback. There are also lots of self evident bounds. We must eat stuff that is food, and not dirt. We must breath air, and not water. We must look after our children. We must work. We must say stuff that makes sense to fellow humans. I'm sort of gob smacked that you could even think human intelligence operates without bounds. > A trained model has no "functional default" the way actual intelligence does. Sorry, I have no idea what that means. > truly original output This left me scratching my head. How do define "truly original output"? I don't think there is a definition as we can't even agree on when humans are producing "original output" give the arguments we have over movie plots. > an intelligent AI Again, I have no idea what you mean by this. That isn't surprising to me as there isn't a agreed definition of intelligence or how to measure it. Given that I don't know how you or I could be sure an any given AI was as intelligent as a human - and that includes ChatGPT. To me, where this all leads is people use the work "intelligent" as a proxy for the way they imagine humans might think. That's a bit sad, as we have no idea how humans think - so I don't know how you could compare it to the way an AI thinks. Even so, it looks self evident that at a detailed level ChatGPT does not think like us. If that's your definition of intelligence, as AI will never achieve it. But "think the same way we do" is not end goal, is it? The end goal is "produce original output, but indistinguishable from what a human might produce", which ChatGDP already does in it's domain. By the by, while at the detailed level ChatGPT and humans are obviously very different, I'd be surprised if there aren't profound similarities as well. For example I would not be at all surprised if "originality" in ChatGPT and humans boiled down to randomness being injected into a neural net, which then moulded it into something that matches all the constraints prior learning had taught it. So we get something that is novel - but still consistent with how everybody perceives the real world to be.