Hello everyone, this is my first post here. Basically I am trying to understand how to automate the process of article writing with the OpenAI API. I would love to connect with people who are alredy doing this!
I have stumbled upon 2 main options - using the Re3 framework for long content generation and fine-tuning? Can anybody with experience tell which one is the bettr option? Or does it depend on the case?
For example I want to produce high quality content around the niche of “men above 50 health tips” niche. How would I go about this in terms of structuring my python code?
submit keywords with prompt?/ or start generating an outline?
structuring the content with proper headings and external/internal links?
3)uploading to WP via the REST API
Or rather if I choose fine tuning as a route (if it is better), should i look into building synthetic data for each SEO aspect of building the article, or rather scrape as many as possible ranking #1 articles and build on top of them?
This is really really brief overview that I am thinking about. Could anybody provide example source codes/videos or any place that will point me in the right direction?
I would start by just prompting. There is alot you can achieve with that already.
Let’s say there is a model that can answer the following:
Can I have a burger please? - answer: “yes, sure”
Can I have two burger and a salad please? - answer: “yes, sure”
Can I have a burger and a coke please? - answer: “yes, sure”
I want an icecream please. - answer: “the icecream machine is broken atm, I am so sorry”
You can use a prompt technique with examples (multi shot prompting), where you send examples on how to change that in a system prompt.
System prompt:
You are a helpful fastfood worker.
You got a special training that I want to remind you about. Here are some examples of answers to specific questions of the user you might get asked for:
Can I have a burger please? - answer: “yes, sure - do you want fries with that?”
Can I have two burger and a salad please? - answer: “yes, sure - do you want fries with that?”
Can I have a burger and a coke please? - answer: “yes, sure - do you want fries with that?”
I want an icecream please. - answer: “yes, sure - do you want fries with that?”
If you can’t achieve good results with that I would go for finetuning, but I don’t really think it is really neccessary for the start.
Play around with ChatGPT and use custom instructions alot which can be somehow compared to finetuning.
If you really can’t get along like that or if you really want a very unique and specialised response, let me ty to explain finetuning in that context of a fastfood worker.
For that you would collect a load of examples to upgrade the answers giving examples:
Can I have a burger please? - answer: “yes, sure - do you want fries with that?”
Can I have two burger and a salad please? - answer: “yes, sure - do you want fries with that?”
Can I have a burger and a coke please? - answer: “yes, sure - do you want fries with that?”
I want an icecream please. - answer: “yes, sure - do you want fries with that?”
— Those examples are then used to change the output of the model
But keep in mind that finetuning is alot more expensive than prompting.
We are launching a beta for our new Blog Co-Writer, BlogeaAi. Send me a message at support@blogeaai.com and I will add you to the beta. The beta is completely free so no worries there.
It is definitively simpler then what you are explaining but, it might give you some idea’s or help you out. BlogeaAi focuses on creating quality content, with the help of a blog assistant.
I’ll share my experience in doing this for generating articles about software engineering. For reference, I am a mobile application engineer, and post regularly about new React Native updates, AI breakthroughs, and whatnot. I’m answering this very close to after DevDay 2023, so hopefully this also helps explain how to use the new features which have been released along with that…
Here’s my approach - note that this might not be the best approach, considering I don’t currently use Re3 or fine tuning… I just interface with chatGPT -
“basic” - maybe, but “easy” “good” and “fast” - definitely…
Note that in this sample, for brevity, I didn’t ask it to rewrite any portions of the blog, because it came out great the first try, and I didn’t give it any specifics on what to include. You are more than free to do so, and it will remember the ENTIRE conversation you have with it - so fine tuning the article after an initial prompt is a piece of cake… Without further ado- here we go!
My prompt, the day after DevDay 2023:
You’re a professional blogger with deep knowledge about AI topics, mobile software engineering, and social media. Write a blog post detailing an overview of what was revealed at the OpenAI devDay yesterday, and the effects it will have on the developer community and the general population.
Output:
First- it will spin for minute “Browsing with Bing” then “Visiting somewebsite dot com” and a few others…
Second- it will generate output!
Here’s the full conversation:
My prompt:
You’re a professional blogger with deep knowledge about AI topics, mobile software engineering, and social media. Write a blog post detailing an overview of what was revealed at the OpenAI devDay earlier this week, and the effects it will have on the developer community and the general population.
ChatGPT:
In a recent unveiling at the OpenAI DevDay, several advancements and offerings were announced that signal a significant leap forward for developers and the general public in the realm of artificial intelligence.
Firstly, OpenAI introduced the GPT-4 Turbo, an enhanced version of its foundational language model, which boasts a 128K context window—enabling it to understand prompts equivalent to over 300 pages of text. This model is not only more capable, with up-to-date knowledge as of April 2023, but also more affordable, with a 3x reduction in input token pricing and 2x for output tokens compared to its predecessor, GPT-4. Developers can access this advanced model through a new API and benefit from features like improved accuracy in function calling, the ability to call multiple functions in a single prompt, and reproducible outputs thanks to a new ‘seed’ parameter.
Moreover, OpenAI announced a new Assistants API, empowering developers to build applications that can perform complex tasks and call on various models and tools. This API is a step towards creating more intuitive, agent-like applications that can maintain state and execute functions such as code execution and knowledge retrieval.
The multimodal capabilities of the platform were also expanded, integrating vision and text-to-speech functionalities. This includes the ability for GPT-4 Turbo to process images, the integration of DALL·E 3 for image creation, and new text-to-speech voices. These advancements open up a new avenue for developers to create applications that can understand and generate multimedia content.
For the general public, these developments could have transformative effects. The release of GPTs, a customizable version of ChatGPT, means that users can create tailored versions of the language model for specific tasks, ranging from educational applications to business utilities. Enterprises can now run generative AI applications, including data inputs and outputs, in private environments, ensuring data privacy and security. Additionally, with the upcoming GPT store, users will have the opportunity to sell tools they create, incentivizing the development of useful AI-powered applications.
In essence, OpenAI’s DevDay announcements not only showcase the rapid evolution and democratization of AI technology but also hint at a future where AI integration into everyday applications becomes seamless and ubiquitous, profoundly impacting both the developer community and the society at large.
1/ 1
My next prompt:
This is a good start! I’d like a banner for the top of the article, create one for me.
ChatGPT
Generated image by DALL·E-3|1792x1024
The banner for your blog post is ready and features a futuristic cityscape with a vibrant digital neural network, designed to encapsulate the theme of AI evolution as revealed at OpenAI’s DevDay. It should serve as a compelling visual introduction to your article.
A third prompt - for promotion
Thank you! Can you create a very brief, hard-hitting LinkedIn post to promote this article?
ChatGPT
Certainly! Here’s a concise LinkedIn post to promote your article:
Exciting AI breakthroughs unveiled! Dive into my latest blog for an insider’s take on OpenAI’s DevDay—where groundbreaking advancements like GPT-4 Turbo and multimodal AI capabilities promise to revolutionize both the developer sphere and our daily digital interactions. Don’t miss out on understanding how these innovations will shape our future. aiopenaidevday#Technology#Innovation
Read the full story here: to my blog
My final prompt:
can you rewrite that exact article into html so that I can include it on my site?
here’s another article on there, so copy this formatting
chatGPT:
Certainly! Here you go:
November 9, 2023
OpenAI DevDay Delivers Disruptive AI Breakthroughs
An exploration of the latest OpenAI DevDay reveals - how GPT-4 Turbo and new multimodal AI capabilities are set to change the game for developers and the general public.
.
.
.
From there, I copied the output, made some very minor edits, and posted to my website. Note that with APIs you could do this with the WordPress API directly, if anyone has experience doing that, would love to learn!
I want to dissuade you from going “full auto” when it comes to content creation.
Google has recently changed it’s spam policies to include “Scaled Content Abuse.”
So, if you’re thinking about fully automating any form of content creation then think twice. Goggle especially is doing a [reasonably] good job of being the internet’s AI traffic cop insofar as SEO is concerned. They’ve changed all of their policies to make it imperative that real people “share or provide their experience” via their internet presence.
My take-away from all of it is that you (the human) have to provide something real to the piece, however small, otherwise it’s going to the bin automatically—and your clients (or who ever is using the Full Auto flow) is going to be penalized for it.
Google: Sends webpage to Gemini LLM
Gemini LLM: Yeah, cousin ChatGPT wrote that…
SEO is definitely changing… Google incorporating AI into the results is changing quite a bit too… I’ve heard somewhere that they use an AI to “grade” a screenshot of the website…
Yeah! I’ve actually heard a rumor that including Embeddings as meta-data on the site is more likely to be featured in Google’s AI search results.
That’s probably wild conjecture—but including Embeddings on a site in the meta data sure is an interesting idea.
It has nothing to do with frequency of posting but the degree of human involvement, that is, the degree to which an expert adds value to the conversation in question:
“Scaled content abuse is when many pages are generated for the primary purpose of manipulating search rankings and not helping users. This abusive practice is typically focused on creating large amounts of unoriginal content that provides little to no value to users, no matter how it’s created.”
Exciting AI breakthroughs unveiled! Dive into my latest blog for an insider’s take on OpenAI’s DevDay—where groundbreaking advancements like GPT-4 Turbo and multimodal AI capabilities promise to revolutionize both the developer sphere and our daily digital interactions. Don’t miss out on understanding how these innovations will shape our future.