Gregory's Server
 23 comments
@yogthos @kpopncommunism I can't wait till people start crafting clever prompts to get these things to divulge corporate secrets :)    and yet, in a few y ears , they will be so much better that they will be everywhere and only a few grouchy people like me will remember much less complain @failedLyndonLaRouchite oh I have no doubt that this tech will improve rapidly, we're headed for a really interesting time that may rival the changes seen during the industrial revolution  @yogthos @failedLyndonLaRouchite@mas.to Personally, I have every doubt. The underlying technology isn't any more than a device to make sentences that sound like answers. It doesn't understand what it is saying; there's no "there" there. @fishidwardrobe I think you have to take a broader perspective here. Fundamentally, GPT is a system that builds a model of the data it's exposed to and then makes predictions based on that model. The problem is that current approaches is that they simply feed text into this system without any context. However, imagine if such a system was embodied either in a physical robot or a virtual avatar, and then taught the rules of the physical world the way we'd teach a child. @fishidwardrobe I would argue that at that point it would have real understanding in a human sense. It would create a predictive model of the physical world based on its interactions, and then we could start developing a common language with it. There is nothing to suggest that there is anything fundamentally different happening when a human mind develops. @yogthos That would be the completely different approach I was talking about, then, because the current approach is incapable of "learning like a child". @fishidwardrobe could you elaborate on that, why do you say that GPT is incapable of learning the way a human child does? @yogthos @fishidwardrobe what does “learning the way a human child does” mean? That statement could literally mean millions of different things depending on subtle context differences. @johnstonphilip @fishidwardrobe I meant in a sense of building a predictive model of the environment through interaction. @yogthos @johnstonphilip Well, it doesn't do that. It's really not much better in that respect than ELIZA – the difference is, instead of just the prompt, it has gigabytes of text to work with and does so in a cleverer way. But it still has zero understanding. It's a "Chinese Room", broadly speaking. There's plenty of information out there about how LLMs work. You might start by searching for a paper called "Stochastic Parrots". @yogthos This is an accurate rendition of PayPal’s human support, so it’s feels like notabug territory. |
@yogthos@mas.to
Has Elon gone back to PayPal? :blobbonethink: