7 ms·

Ask HN: To AI denialists, what do you think is happening here?

• smoldesu 4y ago
  "True understanding" here is just having the name "Mike" in it's context. AI models like ChatGPT keep the tokenized representation of your question in memory - if asked to say a unique name that isn't the 3 it has already said, the most-likely next answer will be the last name that appears in memory.
  • logicallee 4y ago
    Explain this conversation I just had: https://imgur.com/a/FSC9gAJ https://imgur.com/a/FSC9gAJ
    • bluefirebrand 4y ago
      You changed the question at the end from "answer this riddle" to "explain this riddle". I would bet there is a Q&A thread on this riddle or one similar that is part of the language training model and it has regurgitated that almost verbatim to you.
      • logicallee 4y ago
        There is no such Q&A thread on this riddle. As a demonstration that it really understands, I just asked it in the same thread: >Now I want to see if you really understand the puzzle or are just stringing words together. To show me that you understand it, make a version that apparently looks similar to the two examples you came up with such as "I am not my father's daughter," the lawyer said. However, this time include something that really does make the sentence false and is not a correct version of the puzzle despite looking similar." It said this: >Sure, here is a version that appears to be similar to the previous examples, but actually makes the statement false: >"I am not my father's daughter," the lawyer said. "He is not my biological father." >In this version, the lawyer is stating that the person he considers as his father is not his biological father, which makes the statement true. However, it is not the same type of puzzle as the previous examples, where the statement was related to the individual's gender identity and the traditional association of specific genders with specific parental roles, rather than the biological relationship. This clearly shows understanding. For example it used the correct pronoun this time rather than before it got it. You can't just claim everything is in its training data even when it's obviously completing new tasks based on something you just taught it that it didn't know at the start of the thread. Do you really think someone in a Q&A that is in its training data asked for a false version of this specific puzzle and that it just quoted it verbatim? Why would anyone ever do that? Or is it more likely that it is doing something close to "thinking" to come up with it?
    • smoldesu 4y ago
      Sure: - You asked a question that can be solved with surface-level problem solving skills and - You asked the AI to generate a dozen responses until it got a correct-looking result Statistically speaking, what you did was roll the dice on the text generator until eventually it gave you a satisfying answer. Your "conversation" mostly consists of inputs being added on-top of a pre-existing context. When you generate text the next time, it accounts for this new input and tries to generate a new answer. ChatGPT has no understanding of the situation beyond what it's model thinks should happen next, and what you posit is true. If you think that's sufficient to pass as AGI, more power to ya. As a tool for generating convincing text, it's a godsend. As an approximation of human reasoning, ChatGPT is a complete and unequivocal failure.
      • logicallee 4y ago
        Explain that after it got it, it could come up with a false version: https://news.ycombinator.com/item?id=34423636 https://news.ycombinator.com/item?id=34423636 By the end of the thread, it clearly understands the puzzle well enough to differentiate it from what it thought it was earlier.
        • smoldesu 4y ago
          There's not a point to arguing this, because even presuming the AI can return a perfect result doesn't imply it's capable of thought. You're correct to highlight a degree of nuance to this question, but it's foolish to defend such a limited definition of intelligence. ChatGPT is incapable of independent thought, and every demo (including the one linked above) fails to demonstrate reason beyond what is expected from a synthesis program. Impressive, but pointedly limited.
          • logicallee 4y ago
            There is a point to discussing this because to me and thousands of others who have used it extensively, ChatGPT is capable of thought. Clearly it has limits but we are not talking about whether it is capable of all thoughts - we are talking about whether it's capable of any thoughts! In a very broad way it has a similar architecture (based on neurons and weights) to something we know does think. (The human mind.) In terms of its ability to handle abstract thoughts, it shows a degree of such ability. I've had ChatGPT come up with independent thoughts scores of times, just like a mind does. The human brain is just a "synthesis program" too so your sentence "what is expected from a synthesis program" is misleading. What is eventually reasonably expected from a synthesis program is the same ability to think that a human mind has. That is not the goal post today: the goal post today is "does Chatgpt show any understanding or capability for independent thought" and the answer to those people like me who have used it for such purposes is a resounding yes, it does. You are saying it is like a printing press, just mechanically putting sentences together without understanding. I am saying it is more like a child, clearly showing some ability to understand, some demonstration of reason, some independent thought, some creativity, some deduction, and lots of childlike errors. But what good is a newborn child, right?
• Daishiman 4y ago
  What happens in every LLM: a token predictor that tries to predict the most likely token. It does not have the capacity to logically reason; it tries to derive this problem through models that either lack sufficient training data (because how many instances of similar problems would this really have been trained with) or have underfit models. Given enough samples or sufficient language understanding it may provide the right answers, but the "prodding" that you suggested in no way shows that there's semantic understanding happening.
  • logicallee 4y ago
    How about this example? https://imgur.com/a/FSC9gAJ https://imgur.com/a/FSC9gAJ Seems convincing to me, and at the end I have it demonstrate its understanding. It also accurately reports when it is stumped and "still trying to understand" but can't.
    • Daishiman 4y ago
      I’m sorry but the absurd length of text and wrong answers it had to go through suggests precisely that it has extremely limited semantic knowledge and a limited association of concepts.
      • logicallee 4y ago
        Sure: the whole point was to get it to figure out something difficult. (It couldn't.) But in the middle of the conversation, why does it accurately report "I'm sorry, but I am still unable to understand the context and the puzzle without more information"? Wouldn't something that just strings words together without understanding, continue stringing words together without accurately reporting that it doesn't understand? Once it does understand, why does it clearly show its understanding, including creating another example of the puzzle and a counterexample: https://news.ycombinator.com/item?id=34423636 https://news.ycombinator.com/item?id=34423636 Clearly this shows understanding.
    • roflyear 4y ago
      Make up a simple grammar and try to teach it to ChatGPT. You cannot.
• monero-xmr 4y ago
  A five year old can learn to play the piano quite well, but there is very little utility or monetary value to this skill. Similarly the creative fields are already hyper competitive with an extremely high bar for success. Middling prose generated by an AI isn’t going to move the needle except to assist humans. I am very skeptical of the actual applicable utility of the ChatGPT and image generation tools. I doubt many businesses of size will be created focused on these tools.
• red-iron-pine 4y ago
  > It doesn't seem like there is any way to explain that transcript in terms of simple text manipulation rather than "true understanding". What do you think is happening as the AI takes the hint? I've worked with a lot of Devs that knew how to copy-paste and throw around buzzwords without true understanding. Sentience == True Understanding