Hacker News Reader

Comment by LordDragonfang | Hacker News Reader

>With hesitation, Lin approached and said, "Grandmother, today I discovered something new about myself, something that makes me feel uncomfortable and out of place."

>The elderly painter paused and looked deeply into Lin's eyes. "Ah, young Lin," she began softly, "Much like the Formosan bear, every individual goes through stages of growth and change. These moments might seem unfamiliar and daunting, but they're a testament to your strength and the mysteries of life."

As someone who has spent a lot of time trying to get ChatGPT to generate interesting prose, this really exemplifies the sort of stilted, oblique way it has of writing, where every statement is wrapped in platitudes and it reads like a self-help book. This is especially jarring in dialogue. No one, not even characters in books, talks like that.

View on HN

What I found is that works, but is very expensive:

- Do chain prompting where each character gets a separate prompt for the dialogue. Here you'll give it all the character info needed to figure out what the character would do (emotions, background, the story context) and then have GPT4 take passes at what this character would think/do/say.

- Optionally, you can use some sort of reACT or chain of thought prompting to proof-read/critique the answer to minimize noise. I haven't fully explored this, but I think this agent needs to know the entire story outline/context.

- After you get this, you can use an instruction-tuned agent like text-davinci or text-davinci-edit to actually write the prose. It works really well with few-shot examples. It's much better at creative writing. I'm sad they consider it legacy because to me it's the _only_ way to get usable creative writing.

- Now run the rest of the story. You may need to keep some sort of state about the character so that it informs the next character interaction, but I haven't gotten that far in my research.

You raise a good point: while ChatGPT might have stylistic limitations that make it less-than-ideal for a "Primer"-type product, LLMs more broadly don't suffer from that stylistic limitation. Whether it's via prompting, fine-tuning, or something else, you can build multiple characters who address some of the original commenter's concerns.

Yeah, this is a great approach. I've managed to get gpt3.5 to play 4 characters at the same time, but it's very easy for characters to leak into each other/become muddled. Especially since gpt has no specific training on certain things like spatial aspects of a narrative (ie who has an object or where it's located).

To be fair the most optimal would be taking something like llama2 and generating a lora for each character such that you've fine tuned the model along with additional context for the character. I can forgive the big ball of "do everything" weights for not having a specific personality.

I find providing additional context helps, if you're using chatgpt and role playing you can essentially provide out of character pushes and nudges in the right direction. You can also provide example formats for text and requirements, the chat function is especially useful because as long as it's in context, it'll take your corrections into account.

Ie `>OtherCharacter speaks in short sentences when nervous, they'll avoid the issue at hand, but will eventually reveal the truth after enough questioning.< SomeCharacter: I walk over to the table and slam my fists down on it, the silence broken by the rattling of the teacups. "Why didn't you tell me?" I ask quietly.`

Reply on news.ycombinator.com