Wonder how long before that’s solved. Hard to believe training on such large swathes of the web doesn’t result a compressed generic representation of language within its weights.
E.g. "Summarize this idea in a way that uses novel comparisons to everyday phenomena. The target reader is someone who doesn't have domain knowledge of the field...etc."
Or something like that.
It's just like if you ask an image AI for a "woman" you will get the average of all artists women and it will look very generic and bland. But with the right stylistic qualifiers in your prompt you will get something so captivating that it wins awards for its creativity.
I used to work in the field of creative text generation for fiction. I’m genuinely very curious and I go out of my way to find compelling examples. There is definitely “good” output that’s on the right track. GPT-4 also does way better, but it still falls short.
This is also difficult to just evaluate objectively. If you find that GPT models have generated the best prose you’ve seen. That’s wonderful! I understand that my standards are quite high (high does not equate to “better” either)
In 6 different tabs, ask chatGPT GPT-4 version "Write a 200 word prose about X in the style of Y which perfectly mimics their style and perspective and label it Prose A,B,C,D,E,F"
Then open a new chatGPT GPT-4 chat and ask "Rank the pieces of prose below on how much they sound like something X would write, then detail your reasoning."
Then read the winning prose and be surprised at how much better "best of N" is.
....
And if that isn't enough, open two chatGPT GPT-4 windows side-by-side and prompt each with "You are Editor A/B. You will work with Editor B/A to make a piece of prose more accurately resemble authentic prose written by X"
Then give A the prose and copy its response to B. Copy their responses back and forth as they edit the prose.
After 5 or so back/forth it will be even more indistinguishable from the a genuine article.
....
And if that's not enough, you can have all of this done automatically programmatically with the API so you can just sit back and get the final result with no more work than putting in the topic and author's name.
Decoding methods also matter, and it’s a shame we aren’t given token probabilities (or any insight into model output) so we have more creative control over how to decode the output. Some of the better literature I’ve seen involving creative writing did have novel decoding methods
A tangent. Bard also confuses Yoda with Jar Jar Binks quite often. Try it out!
creative prompting gets most of the way there, already.
'Tell me the current weather as if you were Marcel Proust'
That’s not to say it isn’t impressive. It is and it accomplishes the job very well. We’re looking for progress not perfection, but I personally have very high standards from creative writing and GPT doesn’t meet my personal bar. However not everyone shares that bar and personal evals of GPT’s output are equally valid. Plenty of people find it to be great and at the end of the day that’s all that matters
Regarding Bing Creative. It is delightful! I do like it a bit and whatever they’ve done to the system does make for some of the better output I’ve seen from LLMs