1. This website uses cookies to improve service and provide a tailored user experience. By using this site, you agree to this use. See our Cookie Policy.
    Dismiss Notice

Let's Make An AI Content Generator Based On GPT-2 (The OpenAI Model)

Discussion in 'AI - Artificial Intelligence in Digital Marketing' started by The Doctor, Apr 28, 2019.

  1. JOSourcing

    JOSourcing Jr. VIP Jr. VIP

    Joined:
    May 19, 2011
    Messages:
    225
    Likes Received:
    85
    Occupation:
    Writer, Programmer
    Location:
    Sacramento, CA
    That's because it was built without a serious purpose in mind. It's more a "proof-of-concept" toy than anything else. Had the developers been serious about real content generation, you'd probably see what you're describing.

    Bear in mind that there's an ungodly amount of work involved in serious content automation too. And by the looks of what this GPT2 thing generates, the people who made it took the easy route.
     
    • Thanks Thanks x 1
  2. Andy17

    Andy17 Newbie

    Joined:
    Jan 9, 2017
    Messages:
    39
    Likes Received:
    12
    Occupation:
    Automation / Machine Learning engineer
    Location:
    San Francisco, CA, USA
    I've tried to post ~100 articled generated by GPT-2 on my blog, not many results, continue to experiment. I think whoever will get reasonably good results will just disappear, and we will hear from them only when this niche will be completely depleted.
     
  3. MarketingBoi

    MarketingBoi Newbie

    Joined:
    Aug 13, 2019
    Messages:
    17
    Likes Received:
    1
    Gender:
    Male
    exactly, it will be released only once its not really profitable to use it anymore cause there would be more advanced things out there or too many people would be using it so it wouldnt be profitable anymore.
     
  4. firetodust

    firetodust Newbie

    Joined:
    Dec 6, 2014
    Messages:
    1
    Likes Received:
    0
    Glad I found this thread. I've been working on a gpt2-based model for almost two months now. I'm using it to generate content for a niche site.

    It took a lot of tinkering to get the thing to generate relevant content for 1 out of 10 generated articles (anywhere between 300-1000 words) and I usually get 1-2 paragraphs out of that single article.

    I have two articles written with the help of my model each at about 1000 words. Each took about an hour to "write." I wrote 2 other articles for this niche site when I didn't have the model - it took 4 hours to write these 2 articles at about 600 words.

    It still takes a lot of human intervention to get high quality content, but that might also help if detection is used to remove generated content. I really don't see how it's possible to do that reliably especially if a human edits as much as I have.

    I might make a service out of it if I think I can turn a profit on what would be a substantial initial investment (need to build a rig capable of creating models fast enough for customers and their niches).

    Anyways looking forward to hearing how other people are doing with this.
     
  5. hossho

    hossho Newbie

    Joined:
    Mar 9, 2010
    Messages:
    16
    Likes Received:
    9
    Gender:
    Male
    Occupation:
    Marketing Director
    Location:
    ATX

    I think you and a lot of people are getting this area wrong, I don't know if Google would even penalize someone for using AI to generate content. All they want is the best content being delivered to the searchers. If this content is written by AI, a human, or a monkey, I don't think they care. The only reason they don't like the stuff generated now is that it is poor quality and utilizes whatever techniques to trick them. If you could produce college-educated niche-specific researched content with AI, that's great for Google. Their searcher is getting a great result. They could even possibly prefer this content in the future as they would know it has been checked for accuracy and combines an amount of research that could possibly take an entire human their life to accomplish.
     
    • Thanks Thanks x 1
  6. Morty073

    Morty073 Newbie

    Joined:
    Jan 2, 2020
    Messages:
    3
    Likes Received:
    0
    Gender:
    Male
    Trained the medium version (355M) on my niche with decent results. Currently i'm training the big 1.5b version my dataset. This dataset contains articles i scraped from hundreds of sites in my niche. After that i polished the dataset with my programs. I know that this won't generate high quality articles by itself, but the cool thing is, that it generates thousands of articles in no time.

    It only needs one small input of text targeting a Keyword in my niche and generates different outputs when rerunning with that input. So i plan to generate thousands of articles and automatically check their quality with something similar to the contentseochecker. Once it reaches a high enough score i have a SEO proof high quality article in my niche.

    If this works, i'll let you know about it.
     
  7. Grandifer

    Grandifer Newbie

    Joined:
    Aug 22, 2012
    Messages:
    29
    Likes Received:
    9
    I've got a question but I can't PM you.

    Could you please help me out with training the large model?
     
  8. Morty073

    Morty073 Newbie

    Joined:
    Jan 2, 2020
    Messages:
    3
    Likes Received:
    0
    Gender:
    Male
    I recently joined, that's why you can't PM me i think.
    For the 1.5b model i use a colab: hxxp(s)://news.ycombinator.com/item?id=21456025 (dot) com. (I'm not allowed to post full links here at the moment)

    If you have any questions send me an email: [email protected] (dot) com
     
  9. digitalmemories

    digitalmemories Newbie

    Joined:
    Sep 16, 2019
    Messages:
    15
    Likes Received:
    2
    Looking to see what happens next