[Journey] 1 million UVs/month in 12 months using AI generated content. Let's do it!

Status
Not open for further replies.
exactly, this thread gave a ton of ideas. Some threads beg for some negative responses, but I'm at a loss how this thread can get negative attention.

This thread lead me down a rabbit hole on automation, for example this library and discord group: https://github.com/elebumm/RedditVideoMakerBot. Using python to generate TTS reddit vids, also very useful for inserting into blogs! I'm using a modified version for tiktok ;)
This is actually insane, think I might give this a try for some of my AI projects. Only issue is how making a video for all my posts would increase the storage in my VPS go through the roof way too fast haha
 
I'm using selenium so I can have more control and to check on possible errors if they appear.

If the proxy/ip is clean, requests will do just fine, I just like a bit more control.

You can also automate ReCaptcha with selenium, there's a library on Github that resolves reCaptcha using voice recognition, works wonders.
What is the name of the library that solves recaptcha
 
Just finished reading the whole 41 pages ! What a journey !

I did notice that your system or app as you like to call it changed a little bit over the time, first you used some GPT-3 app out there then you started tweeking your own model which you decided to completely switch and start using now ( or since march if i remember right ).

I have been looking to start something similar quite some time, i have had a couple of failed attempts to start niche sites in the past, some of them i still own to this day ( with 1 - 2 articles in average, content writing was never my thing ). I did think about something similar in the past but i never thought it could be automated to this level. I even purchased a lifetime license on an NLP AI provider last year when it was still available but never had the time to play around with it. I am studying data science now (will graduate in 6 months) and i think it's time to use the skills.

I started coding something earlier today and still trying to get a clearer image of the whole process, so i might have a couple of questions for you :

1- Do you do KW research and cleaning manually ? Or do you have a script that extracts keywords from the pages you scraped on some website ? If so, how do you know what keywords should be going together as i imagine one page will be ranked for multiple keywords ( I guess i am overthinking this ).
2- Do you still cluster classify paragraphs from different pages ? Do you do that based on articles or on H2s ? I thought about using some simple ML SVM model for that.
3- You mentionned that you are using a "simple" paraphrasing model now that is not an NLP DL model. Is it still a machine/deep learning model ? How big of a dataset did you use for training it ?
4- What would be your advice for someone starting now something similar to your project ? Start with the service i already paid for for paraphrasing ? Or create my own model/script ? I think you had better results with your own model so i am wondering how it works and how big it is.

Thanks for showing us that this is actually possible and don't forget to leave a small bite of the cake for us starting 9 months later :D
1. I do initial research manually - choose a niche and then choose subcategories. The rest is handled by the app - choosing 1k keywords out of 1mil harvested, etc.
2. yeah I do using yake and some manual filters per category
3. We've managed to figure something out that is not entirely ML-based and works better than GPT-3 for paraphrasing. It doesn't work for every niche, but it doesn't matter, cause there's a 100 new topics every day.
4. I would look at what the spammers are doing and try doing something else. think outside of the box and use ML as little as possible. Create websites based on data sets, combine stuff from the web, generate images, graphs, provide value, calculate things, work with local SEO, nada, nada.
childrens books genreated with walle. walle seems hot now for text to image generation. https://www.surgehq.ai//blog/generating-childrens-stories-using-gpt-3-and-dall-e
I would LOVE to have access to walle. Is it possible to get it somehow? I applied but no go so far.
I already used the recaptcha breaker and it worked fine. However, we should limit the requests otherwise Google will block the IP. Did you face any such issues?
captcha breaking is not cost-efficient. With Selenium running 70-100 threads I can scrape ~1 mil PAAs in 1 day with 100 proxies without any captchas.
I didn't face any problems as I'm using a very low rate. One IP does one PAA in ~2 minutes, I just use multiple socks5.
one PAA with how many PAAs scraped?
 
1. I do initial research manually - choose a niche and then choose subcategories. The rest is handled by the app - choosing 1k keywords out of 1mil harvested, etc.
2. yeah I do using yake and some manual filters per category
3. We've managed to figure something out that is not entirely ML-based and works better than GPT-3 for paraphrasing. It doesn't work for every niche, but it doesn't matter, cause there's a 100 new topics every day.
4. I would look at what the spammers are doing and try doing something else. think outside of the box and use ML as little as possible. Create websites based on data sets, combine stuff from the web, generate images, graphs, provide value, calculate things, work with local SEO, nada, nada.

I would LOVE to have access to walle. Is it possible to get it somehow? I applied but no go so far.

captcha breaking is not cost-efficient. With Selenium running 70-100 threads I can scrape ~1 mil PAAs in 1 day with 100 proxies without any captchas.

one PAA with how many PAAs scraped?
Are you talking about 1 million posts or questions per day?
 
1. I do initial research manually - choose a niche and then choose subcategories. The rest is handled by the app - choosing 1k keywords out of 1mil harvested, etc.
2. yeah I do using yake and some manual filters per category
3. We've managed to figure something out that is not entirely ML-based and works better than GPT-3 for paraphrasing. It doesn't work for every niche, but it doesn't matter, cause there's a 100 new topics every day.
4. I would look at what the spammers are doing and try doing something else. think outside of the box and use ML as little as possible. Create websites based on data sets, combine stuff from the web, generate images, graphs, provide value, calculate things, work with local SEO, nada, nada.

I would LOVE to have access to walle. Is it possible to get it somehow? I applied but no go so far.

captcha breaking is not cost-efficient. With Selenium running 70-100 threads I can scrape ~1 mil PAAs in 1 day with 100 proxies without any captchas.

one PAA with how many PAAs scraped?
Normally, it goes from the last PAA to the first PAA 4 times, to expand them multiple times.
Later, I just check if I have 40 PAAs or no, usually I settle for a minimum of 15.

Not very programmatic, I know )
 
no its proprietary and run in python doesn't have a GUI. he's very kind in giving enough tips to make a similar one though.

What are you guys using to parallelize hf tokenizer?
 
I have read 5 pages of the thread only so not sure whether this question has been already asked or not.

I am curious why are you not using commercial AI content generators like Jasper or copy.ai? Are they not good as your custom tool?

Please answer

Thanks
 
all amateurs :) there's a way of doing this thing... copy *will* get you banned.
Also Google are getting better each day and would eventually catch up with all the low-quality stuff pops everywhere now.
 
I have read 5 pages of the thread only so not sure whether this question has been already asked or not.

I am curious why are you not using commercial AI content generators like Jasper or copy.ai? Are they not good as your custom tool?

Please answer

Thanks
It would be really hard to generate 2k articles/day automatically.
 
Can you elaborate? I'm not sure what you mean. I tokenize sentences using NLTK.
Thanks. so you don't use Hugging Face anymore you took that 100billion library and tokenized it yourself it seems hehe.

I meant, for scrapping you parallelize selenium into RTX 3090, and when you use your own model to paraphrase that runs in GPU parallel too? does that question make sense?
 
Thanks. so you don't use Hugging Face anymore you took that 100billion library and tokenized it yourself it seems hehe.

I meant, for scrapping you parallelize selenium into RTX 3090, and when you use your own model to paraphrase that runs in GPU parallel too? does that question make sense?
the fact is we've changed our paraphrasing model entirely 3 times already ;) Yes, before I was doing it like you described.

Right now I created a completely original model with my dev from scratch that is a little out-of-the-box and different than most stuff out there. Works very similary to Quilbot or Wordtune but is scalable on CPUs so I'm just scaling Threadripper/EPYC servers when I need and it's much more cost-efficient for me than GPUs. I hope that makes sense and sorry I can't reveal the details right now, but I invested a lot of money into developing this.

My dev is working on a simple white paper to explain this and it will be posted in public soon.
 
the fact is we've changed our paraphrasing model entirely 3 times already ;) Yes, before I was doing it like you described.

Right now I created a completely original model with my dev from scratch that is a little out-of-the-box and different than most stuff out there. Works very similary to Quilbot or Wordtune but is scalable on CPUs so I'm just scaling Threadripper/EPYC servers when I need and it's much more cost-efficient for me than GPUs. I hope that makes sense and sorry I can't reveal the details right now, but I invested a lot of money into developing this.

My dev is working on a simple white paper to explain this and it will be posted in public soon.
this is awesome and no need to be sorry at all we all understand! thanks for sharing! I've got something finally working well and adding to it, but I'm sure it's far more rudimentary. Playing with some of this edge errors and using python to sort of regex out of them.

Look forward to the white paper! Is it using something like https://code.google.com/archive/p/word2vec/ to measure and then pick words more accurately?
 
this is awesome and no need to be sorry at all we all understand! thanks for sharing! I've got something finally working well and adding to it, but I'm sure it's far more rudimentary. Playing with some of this edge errors and using python to sort of regex out of them.

Look forward to the white paper! Is it using something like https://code.google.com/archive/p/word2vec/ to measure and then pick words more accurately?
You're 100% on point.
 
all amateurs :) there's a way of doing this thing... copy *will* get you banned.
Also Google are getting better each day and would eventually catch up with all the low-quality stuff pops everywhere now.
and what if I add one more step and paraphrase the content before joining it together? Thoughts?
 
I managed to get paid from the first PAA site I got into M3D14_V1N3 (im writing in code cause they are monitoring the web a lot)

Btw, I asked them from an anonymous email and they said that both ad arbitrage and PAAs are ok for them if it doesn't break any rules:

Image not found: vz85YEA


Humble brag:

Image not found: NmM5TDU
 
I managed to get paid from the first PAA site I got into M3D14_V1N3 (im writing in code cause they are monitoring the web a lot)

Btw, I asked them from an anonymous email and they said that both ad arbitrage and PAAs are ok for them if it doesn't break any rules:

Image not found: vz85YEA


Humble brag:

Image not found: NmM5TDU
Nice achievement! How many users and pageviews are in this one? Also, can you please tell me which model have you trained and how much money (estimation) did it cost to train it?
 
I managed to get paid from the first PAA site I got into M3D14_V1N3 (im writing in code cause they are monitoring the web a lot)

Btw, I asked them from an anonymous email and they said that both ad arbitrage and PAAs are ok for them if it doesn't break any rules:

Image not found: vz85YEA


Humble brag:

Image not found: NmM5TDU
It makes perfect sense. What really matters is their advertisers get performance, and since your traffic is from SERP's, regardless how you get it, the serp traffic is high, and gets conversions for their customers, so makes perfect sense that they would still payout.

Wow net 90? well they pay is nice just takes a long time to get it haha.

I'm stuck on H1 headlines, i have the content paraphraser pretty good. What apis or methods are you guys doing to generate a buzzfeedy title that keeps the correct keyword and makes it more "copy" and "clickbaity? I don't want same title as scrapped but just paraphrasing doesn't make it exciting. Any tips or ideas guys?
 
Status
Not open for further replies.
Back
Top