Train ChatGPT to write in your unique style

Mad10

BANNED
Joined
Jun 19, 2020
Messages
39
Reaction score
36
If this is not the appropriate thread could Moderators move it to its rightful home?

Train ChatGPT to write in your unique style.
You need to train AI how a human writes, how you write using "perplexity" (complexity of text) and "burstiness" (variations of sentences).
AI writes in a more uniform manner, it writes how it has been trained, so train it better.
Get it to give you a style guide for your unique writing, that you then feed back into it every time.
  1. Feed ChatGPT samples of your writing (one article at a time)
  2. Ask ChatGPT to describe your writing style (for each article) Could also ask it if the style is the same. Think what you need it to know.
  3. Every time you type in instructions, ask it if it understands.
  4. ChatGPT will learn your writing style (voice, tone etc) and explain it to you (Your Style Guide).
  5. Use this to feed to ChatGPT every time you want it to write content for you.
  6. Continually practice and refine
Example: Analyze the text above for style, voice and tone. Create a prompt to write a new article in the same style, voice and tone.

Hope this helps folks.

You have to teach AI what you want it to do, educate it, harness it because it isn’t going away.

Does anyone know if Bing penalizes AI, anyone tested that yet?
 
Last edited:
Sometimes when I ask if you understand, I get a lot of mistakes
 
So i should type the same prompt each time
So the AI is not trained on my account to write like that.
 
Good Shot and will try it soon. Thanks for your valuable sharing
 
This method may not look like the perfect solution, but it can be suitable for many tasks. Thank you for sharing.
 
This is not training. It's simply a one-shot prompt.
No, you do not keep putting in the same prompt.
You put in an article and ask for the style, then another and ask it to update a prompt to write in this style and you keep doing that to build a more detailed prompt.
That way it learns your style over a period of time and from many articles until you are happy that the prompt is as detailed as you would like it.
 
But there is a limit of 3000 keywords(don't remember the exact figure) that's the maximum memory it can store in a particular conversation. It would get finished with 1 article and a few responses after that it will give response based on the last text.
 
Awesome, never though on teach the AI with my own writing or ideas, that is cool, and that can always align with my point of view or my writing styule and explain the things, I will be testing this thanks for sharing.
 
That way it learns your style over a period of time and from many articles until you are happy that the prompt is as detailed as you would like it.
A model cannot learn by prompting. It can only remember your text within the context window. 4k, 8k, etc.

If what you say is true, you can simply prompt 'Google is shit' * 1 billion times, so chatgpt will 'learn' Google is shit.
And next time when somebody else type what is google, chatgpt will respond google is shit as taught by one of its user 1 billion times.
 
OP, what you're referring to is called few-shot in-context learning (FSICL).
The model isn't "learning" anything in a persistent manner, it's just adapting to the structure of the "chat history" ie context window.
The problem with this approach is that:
1) the quality will degrade over time as the chat gets longer, this is because after a point there's some "neural compression" that ChatGPT has to do to squeeze the chat history into the context window.
2) there's a recency bias in FSICL, recent examples are often prioritized over earlier ones.
3) you're not really training anything. In fact you can't train any of these large models, you can either fine-tune them (available for davinci) or you can use FSICL.
 
But there is a limit of 3000 keywords(don't remember the exact figure) that's the maximum memory it can store in a particular conversation. It would get finished with 1 article and a few responses after that it will give response based on the last text.
You take the prompt it writes for you, open a new chat, input another article and continue the process. It has no limit then.
 
Back
Top