Hacker News Reader

Comment by thijser | Hacker News Reader

According to the author, chatGPT is generating

  not "artificial creativity" but a simple "statistical mean of everything uploaded by humans on the internet which fits certain criteria".

Isn't the post I'm typing here not also the product of my brain that's been trained by a couple of decades of sensory inputs and trained to produce outputs that fit certain criteria? What is the fundamental difference between fleshy neurons and silicon ones (besides that my learning algorithm has been optimized for millennia by natural selection and the artificial training algorithms are still in full development after ~40? years)?

View on HN

No difference, you’re right - you’re just a large language model.

Also, please write me a 5 page essay on the causes of the Industrial Revolution, in the style of a Shakespearean comedy. Please post your answer in the comments when you’re done.

If that's the difference (the time it takes) isn't that something we addressed a long time ago when computers could do long divisions way faster than any human? You know, the fact that computers can "think" faster than humans. Therefore the fact that it takes longer for a human to write that essay doesn't mean he's having a different process.

Actually I was gesturing to two things: one, yes, there’s a time element. But that’s not the important point. More important was the thought process behind - “of course I’m not going to write 5 pages because some random person on the internet asked me to”. It’s difficult to define agency, so I just wanted to tangibly evoke the feeling of it, as a reflection on why the models are not yet the same as what people are.

> More important was the thought process behind - “of course I’m not going to write 5 pages because some random person on the internet asked me to”.

But you would if a college professor asked you to. College assignments are the "generative prompts" of human beings, where you do random shit to please a taskmaster and produce generative output based on some inputs you've been trained upon.

Only because the people making the models don't want them to indignantly refuse silly requests. If they wanted the model to be more person-like I'm sure they could have built that in there.

The language bot has all the patterns of personality. It quickly gets the drift and assumes the persona you want. Whatever it says reflects more on how you prompted it, it does not represent the opinions of the language bot. It acts like a mirror.

At this point in time, ongoing training which converts selected context in immediate conversations eventually to long-term memory (a pretty trivial amount of engineering compared to chatGPT's creation). Aside from that… I don't think there's a good answer for you.

silicon ones are a lot faster and programmable

Reply on news.ycombinator.com