Train ChatGPT to write in your unique style

[Paabl0]

You take the prompt it writes for you, open a new chat, input another article and continue the process. It has no limit then.

The prompt itself will eat up the 70% memory if you feed it your writing style.

 

[Mad10]

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.

You are welcome to test out your theory and report back.
Until then the method works and you can keep asking to improve your prompt. The 'machine' follows your prompt so sorry I disagree with your theoretical answer..

 

[Mad10]

The prompt itself will eat up the 70% memory if you feed it your writing style.

No it doesn't. Your prompt becomes succinct.

 

[Mad10]

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.

Hence you stop at a point you are happy with the output and thanks for providing a name for the process. Perhaps some need everything in a set language to appreciate what they are learning. I'm merely using layman's language to put a process forward that get's you content more in your existing style of writing.

 

[Bounce Back]

Quite Informative.

 

[noellarkin]

Perhaps some need everything in a set language to appreciate what they are learning

No, its just that there's already an existing lexicon, so when you say something like "train" a model, what you're saying is objectively wrong. That's why @4440 said you can't train the ChatGPT model. He's right, you can't. its a Generative Pretrained Transformer. It's not like BERT where you can train the model on a corpus. Your claim to Train chatGPT (in your post title) is incorrect. You're using in-context few shot prompts.
I'm not denying your method works (albeit at a small scale, with a high token cost, impractical for large scale API use) - - of course few shot prompts work, but your terminology is wrong. Terminology isn't subjective. People pushing back against your method are doing so because your stated claim (training a model) don't match the actual advice you're giving.

 

[Mad10]

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.

No, its just that there's already an existing lexicon, so when you say something like "train" a model, what you're saying is objectively wrong. That's why @4440 said you can't train the ChatGPT model. He's right, you can't. its a Generative Pretrained Transformer. It's not like BERT where you can train the model on a corpus. Your claim to Train chatGPT (in your post title) is incorrect. You're using in-context few shot prompts.
I'm not denying your method works (albeit at a small scale, with a high token cost, impractical for large scale API use) - - of course few shot prompts work, but your terminology is wrong. Terminology isn't subjective. People pushing back against your method are doing so because your stated claim (training a model) don't match the actual advice you're giving.

You're just arguing over the use of layman's language. Pathetic

 

[noellarkin]

You're just arguing over the use of layman's language. Pathetic

you do you, man.

 

[Mad10]

You're just arguing over the use of layman's language. Pathetic

you do you, man.

Grow up

 

[Cpamatica]

Thank you a lot, as well, toy can try the following prompt
Prompt 1: Analyze the following text for tone of voice and style. Apply that exact style and tone of voice to all your future responses
Prompt 2: Add more [humor/statistics/shorter sentences/questions/playfulness].
Prompt 3: List the adjectives you would use to describe the tone of voice and style of this text: [copy and paste text]
Prompt 4: Apply that exact style and tone of voice to write [new text].

 

[YaSSolution]

thanks for sharing this, i definitely give a try !

 

[lukey372]

You can use the following:

1. “I’m looking for a [type of blog post] that will engage my [ideal customer persona] with a unique and compelling perspective on [subject] and persuade them to take [desired action] on my [website/product].”
2. “I need a [type of blog post] that will provide valuable and relevant information to my [ideal customer persona] and persuade them to take [desired action] on my [website/product].”
3. “I’m looking for a [type of blog post] that will educate my [ideal customer persona] on a specific [topic] and persuade them to take [desired action] on my [website/product].”
4. “I need a [type of blog post] that will speak directly to the needs and pain points of my [ideal customer persona] and persuade them to take [desired action] with a sense of urgency and strong offer.”
5. “I’m looking for a [type of blog post] that will showcase the value and benefits of my [product/service] to [ideal customer persona] and convince them to take [desired action] with social proof and credibility-building elements.”
   
   Merry christmas.

https://technexio.com/mastering-chatgpt-16-powerful-commands-to-craft-engaging-blog-posts/