I made an automated blog writer powered by OpenAI and it was terrible

I made an automated blog writer powered by OpenAI and it was terrible

ยท

5 min read

Writing blog articles is one of the most tedious and difficult things I have done. I'm sure a lot of you can relate if you're on Hashnode. If you don't know I also run another blog called Coffee Informer where I write articles about coffee. I've tried hiring writers to create content for me, but the quality wasn't there. A lot of the time I would hire someone to write an article and have to provide them with an outline, topic, and related articles about the topic. Afterward, I would still have to proofread and edit the articles to may sure they were publishable. I thought "there has to be a better way that is fully automated". That's when I had the idea to try automating the blog writing process with AI.

With that, filler.ai was born. This was my first attempt trying to create a SAAS product and I definitely learned a lot of things when building the project that I want to revisit in the future.

Currently, it is functional but there are many improvements that need to be made. You can put in a topic and generate blog subheadings, full blog articles, article ideas, and video scripts. It does this using OpenAI's GPT-3.

image.png

GPT-3 is a language generator that can produce human-like text with different parameters to fine-tune the results. My goal was to have it generate human-readable blog articles targeting specific SEO keywords in articles that were 1000 - 2000 words long.

I did this by first allowing the user to generate multiple headlines using a topic which GPT-3 would then generate an outline for targeting specific SEO keywords based on the inputted topic. Next GPT-3 would make an article based on the headline generated and then combine it all into a single article. This would be displayed to the user in an editable rich text editor so users can customize and change the article.

image.png

Once the users were happy with the article they could download it as a PDF to send it to an editor or copy and paste it onto their blog. There is a possibility I could add an export as a markdown button in the future.

image.png

Where it all went wrong

When I launched it and a few people used it. I immediately got feedback about the quality of the articles. It also was not getting indexed by Google which was very concerning. These two issues were valuable learning experiences that I will share with you to take into your next project.

Problems with Articles

While the articles may look "readable" the content would be questionable. A lot of the time the AI could generate an article about a topic and state random incorrect statements as facts. The AI had no knowledge of whether it was right or wrong, and a lot of time it looked very authoritative until you tried fact-checking the article. Then you would realize that the AI had no idea what it was talking about.

unnamed.gif

The AI would also repeat patterns and make very similar articles. This is because it was trained on a set of examples and it would reuse patterns in the training data. As a result, different articles would end up looking very similar and may have repeated patterns or phrases.

GPT-3 also does not have any memory of previous articles that it has created. Suppose you want to generate two different articles about the same topic. The problem is that if you use the same topic with the same headings there is a very high possibility the article that is generated would look very similar to the previous article that was created.

I think there is still a lot of potential for AI-generated articles, but there would need to be a human touch. The article cannot be fully created by AI and would at least need to be edited by someone.

In my opinion, the best use for GPT-3 for blogging at the moment would be creating short sentences about topics to insert into your articles so you don't have to write every single line. This could help you improve the word count while still looking like a human wrote the article.

Not using server-side rendering

Another thing that I didn't account for was the lack of server-side rendering because I am using React. If I could go back and restart the project I would probably use Next.js to have server-side rendering.

Because React does not have server-side rendering Google does not rank it highly since it is responding with JavaScript instead of HTML. JavaScript is slower than HTML and as a result, a slower page will lower the ranking on Google Search.

sloth-slow.gif

I think server-side rendering would help grow the website by allowing Google to index the pages better. I could automatically generate articles that would be posted into the blog article route which would automatically create new blog posts with a press of a button.

The code is very similar and there doesn't seem to be a very big difference between the two. You will still have components, but your pages will be located in a pages folder which would represent a resource and a page for your website. To convert from a simple React app to Next.js app all you would have to do is do a few simple steps:

  1. Convert React router links to Next.js links
    import { Link } from "react-router-dom";
    
    to:
    import Link from 'next/link'
    
  2. Convert CSS to the Next.js styled-components

  3. Convert code to take advantage of Next.js, for example, this would be using getStaticProps or getServerSideProps

Conclusion

It was a learning process and it was my first attempt at a commercial product, and the journey definitely isn't over. The next step would be converting the site over to Next.js and trying to improve the article generation. I'm not giving up and I hope you all follow me on my journey!

If you don't know my story I am a software engineer who used to be a commercial real estate broker in NYC. I quit my job to follow my passions and now I am 2 years into my career as a software engineer and loving it! If you want to know how I became a software engineer in only 3 months you can read that article here.

Copy of Thank you for Reading (3).gif

Did you find this article valuable?

Support Edmond Hui by becoming a sponsor. Any amount is appreciated!

ย