Hacker News Search

botro 4y ago
But airplanes don't fly by flapping their wings.
- LoganDark 4y ago
  Some of them do actually: https://www.youtube.com/watch?v=d1z8YUs2FLc https://www.youtube.com/watch?v=d1z8YUs2FLc What is this an analogy for? -Emily
  - dougmwne 4y ago
    Commenter is saying that it could be possible to get to the same outcome but through a different mechanism. So maybe if we do invent an AGI, it will be able to do all the things a human can do, but the underlying mechanism will not resemble or imitate what our brains are doing. See also: https://en.m.wikipedia.org/wiki/Blindsight_(Watts_novel) https://en.m.wikipedia.org/wiki/Blindsight_(Watts_novel)
    - LoganDark 4y ago
      Ah, the black box theory - since AGIs are defined as being able to perform any task that a human can do, the only thing that matters is that they can perform that task just as good as (or better than) a human can, not that they are implemented the exact same way. Even in jobs requiring empathy and understanding, like (the ideal version of) customer service. When I think of AGI, I think of some sort of agent that can converse and interact in the same way that a human can, that has the same sort of curiosity, imagination, and internal thought that a human does. But, as it turns out, that's not actually required or even implied by the term "AGI" itself. -Emily
      - dougmwne 4y ago
        Sure, what’s the difference between empathy and the behavior of empathy. We would say it’s how a person “feels”, a qualia. How to make highly processed sand experience qualia is an ongoing area of research. Maybe the sand already experiences qualia. Some people like to play music for their plants just in case.
dougmwne 4y ago
Would a generative model that can output both text and images of thinking brain scans be more convincing? I’m not saying you are incorrect, just pointing out that we’re not exactly sure what we are looking at in those brain scans anyway. Is a thought a mathematical calculation being run on a wet computer? What does the scan of an inferencing GPU reveal about the model architecture it is running? And then again, maybe math is not all you need. Unknown unknowns, possibly unknowable.
- LoganDark 4y ago
  > Would a generative model that can output both text and images of thinking brain scans be more convincing? No, for the same reason why asking ChatGPT what it is thinking or feeling does not indicate that it actually thinks or feels that way. ChatGPT is a language model designed to generate outputs that mimic its training data, therefore it will learn to mimic the expression and communication of understanding that it does not itself feel. This is why language models are so convincing when they hallucinate, because hallucinating is actually their default state, and the hallucinations just tend to line up with some of the facts contained in the training data, but not always. Therefore they can state falsehoods just as convincingly as they can state facts. > I’m not saying you are incorrect, just pointing out that we’re not exactly sure what we are looking at in those brain scans anyway. Is a thought a mathematical calculation being run on a wet computer? Thought is sort of an ongoing electrochemical chain reaction, though beyond that I indeed don't have much I can say for sure. It's certainly not a mathematical calculation, though it could possibly be simulated by one. -Emily

slowmovintarget 4y ago

You are mischaracterizing the article. The article analyzes the encoded bias of ChatGPT. The system allows you to, for example, make fun of conservatives, but not liberals. It allows you to write negative stories about Republicans, but not Democrats. The people that created the system have also encoded who is worse off than whom based on what someone looks like. That seems like a bad idea to me. Using it to write code... that's probably OK. Using it to express ideas? That seems problematic when the system is designed to favor a particular ideology. It is not that the system should output hateful things. It is that it specifically endorses some hateful things as being just fine. If you extend that into untested biases, what the system favors, you're left with a generator that constantly applies an ideology in favor of some things and against others, regardless of the reason for talking about those things. "There are no unhappy people in Oceania..." "War is peace. Freedom is slavery. Ignorance is strength."

entelechy0 4y ago
[dead]
neilyio 4y ago
I re-read the article. I agree I wasn’t capturing the full picture in my initial comment. But I stand by my original sentiment. It would be wonderful to be proven wrong, but I have a feeling that we’re unlikely to see a perfect text generation system with no bias anytime soon. What perfect system exists in the world? You seem to specifically have an issue with the Democrat vs. Republican leaning demonstrated in the article you linked. If the small bias that exists disqualifies your support for this wildly useful tool, then I suppose there’s not much that can change that. We’re left to hope that another model pops up soon whose biases are more favourable to you. Other than the D/R bias, other biases seemed to correlate pretty well with protecting oppressed groups. Once again, under the premise that bias will always exist, I personally think that OpenAI has done a pretty good job at channeling that bias towards protecting groups that tend to be the subject of most hateful content. But then, I’m also coming from a position of being supportive of good moderation. That’s why I like forums like Hacker News, and I suppose why we’re having this conversation at all. If I only saw utility in raw, unbiased, unfiltered text, I’d probably go read elsewhere.
- slowmovintarget 4y ago
  Until OpenAI deliberately filtered the output, ChatGPT was less biased. The creators put their political finger on the scale.