Today’s models are tuned to output the average quality of their corpus.

This could change with varying results.

What is average quality? For some it’s a massive upgrade. For others it’s a step down. For the experienced it’s seeing through it.

▲

zarzavat 4 days ago | parent [-]

You're absolutely right, but AIs still have their little quirks that set them apart.

Every model has a faint personality, but since the personality gets "mass produced" any personality or writing style makes it easier to detect it as AI rather than harder. e.g. em dashes, etc.

But reducing personality doesn't help either because then the writing becomes insipid — slop.

Human writing has more variance, but it's not "temperature" (i.e. token level variance), it's per-human variance. Every writer has their own individual style. While it's certainly possible to achieve a unique writing style with LLMs through fine-tuning it's not cost effective for something like ChatGPT, so the only control is through the system prompt, which is a blunt instrument.

	▲	j45 4 days ago \| parent [-]
		It’s not a personality. There is no concept of replicating a person, personality or behaviours because the software is not the simulation of a living being. It is a query/input and response format. Which can be modeled to simulate a conversation. It can be a search engine that responds on the inputs provided, plus the system, account, project, user prompts (as constraints/filters) before the current turn being input. The result can sure look like magic. It’s still a statistically present response format based on the average of its training corpus. Take that average and then add a user to it with their varying range and then the beauty varies. LLMs can have many ways to explain the same thing more than 1 can be valid sometimes; other times not.