Hacker News new | ask | show | jobs
by wruza 493 days ago
You may get most of that by prompting. You can create really based or emotional characters even with “aligned” models (with a little realignment). Have you ever talked to an LLM that allows system prompting? Heard of [E]RP? You can even teach them to not produce idiotic bullet points.

I won’t argument on the art part, but these common AI stereotypes are not true.
1 comments
GuinansEyebrows 493 days ago
The AI does not possess those attributes. It’s behaving as told. It has no experience. It has no senses. It has no thoughts or ability to reason. It has no motivation.
link
nathanlied 493 days ago
Could we not - for the sake of argument here - surmise that, since these AIs need prompts, and usually a few rounds of refinement, and then a selection for uploading to (in this case) YouTube, that the -human- in charge of prompting/refinement/uploading has a point of view, an opinion, an aesthetic preference, emotions?

After all, there are artists that collate "samples" from other artists and produce music from all those different samples. They did not play any instrument, they merely arranged and modified these samples into a product that they presumably find pleasing.

link
GuinansEyebrows 493 days ago
The only way we can make that assumption is if the -human- makes it obvious. Tell me the people mass producing AI slop for YouTube/Spotify are approaching this with sincere intent.
link
wruza 493 days ago
It’s not even told, it continues a text (or denoises an image) in a way that closely resembles what was in the training data. Experience, senses, thoughts, reasoning and motivation were all there in original human- and nature-produced data. https://news.ycombinator.com/item?id=43003186

Doesn’t mean the result is ideal, far from that. But your “bullet points” imply some specialness which has to be explained. Personally I don’t love the style of refusal you’re demonstrating, because it’s similar to if you know you know and other self-referential nonsense. At least add some becauses into your arguments, because “it has no X” is a simplification far below the level of usefulness here.

Anyway, how does that prevent creating AI personas again?

link