ChatGPT

[e-com]

I am currently working on a module for thirty bees that automatically generates descriptive content for categories, products and CMS pages using ChatGPT.
In configuration of module, I have added a test connection to OpenAI API.
Speed demon so far is not this artificial intelligence. It took about 20 seconds to generate a response.
And what does this artificial intelligence "think" about thirty bees 🙂

[wakabayashi]

Will you make this module public? 

Atm I don't want to use it 100% that way (maybe for meta descriptions yes). Not having to start from zero, to using ChatGPT would be super cool 😀

[datakick]

I'm looking forward to test this one

[e-com]

4 hours ago, wakabayashi said:

Will you make this module public? 

Atm I don't want to use it 100% that way (maybe for meta descriptions yes). Not having to start from zero, to using ChatGPT would be super cool 😀

I didn't plan to make this module for generating meta data.
Meta keywords have not been read by google bots for a long time, and meta title and meta description is very often ignored by google algorithms. Too many SEO scammers have manipulated search results through meta data. Now only visible content counts when indexing a given page.

With such long ChatGPT content generation times, it only seems reasonable to act as a CRON task. Because to go into each product and manually generate this, it would take several years in a larger store :D.
Surely you need to make a record to markup database, which content was already generated to not duplicate generation. Tokens for OpenAI API are paid, Microsoft has poured 10 billion $$$ into this business, it wants to make money on it.
I need to test different models and temperature for algorithm. I guess text-davinci-003 seems to be the best model. Temperature is a value between 0 and 1 that essentially lets you control how confident the model should be when making predictions. Lowering temperature means it will take fewer risks, and completions will be more accurate and deterministic. Increasing temperature will result in more diverse completions.

When I finish module, I will provide information on where to download it.

[wakabayashi]

23 hours ago, e-com said:

I didn't plan to make this module for generating meta data.
Meta keywords have not been read by google bots for a long time, and meta title and meta description is very often ignored by google algorithms. Too many SEO scammers have manipulated search results through meta data. Now only visible content counts when indexing a given page.

Meta Title and Meta descriptions aren't always taken into account, thats true. Our site seem to have quite a good reputation at google and they take our input often 1 to 1. But actually that's exactly the reason why I want to start with such fields: they aren't that important. Category descriptions are read by real customers. So you have to make sure, that there is no bullshit in the text 😅 For me it would be too risky to take the GPT text 1 to 1 atm. 

I don't have any tokens yet. Do you have? Do you have an approximate value, what a category description text costs?

[e-com]

On 2/22/2023 at 3:51 PM, wakabayashi said:

I don't have any tokens yet. Do you have? Do you have an approximate value, what a category description text costs?

1000 tokens is about 750 words
Cost of tokens depends on the model used. Davinci model is most expensive, but is said to be the best.
https://openai.com/api/pricing/
In configuration of module I made settings for maximum number of tokens for each generation.

Main use of this type module, I see in dropshipping with full automation of its operation.
In most wholesalers product descriptions are either none or tragic. Most wholesalers operate in a closed model only for verified and logged-in customers, also google does not index anything from this data made available to customers. For google under SEO it is these wholesalers that have put up business card sites on wordpress with blog content.

Suppose we import 10k products from wholesaler to our shop. What's next?
I calculated three scenarios of action for 10k products:
1. We create product descriptions ourselves.
Assuming that we create a description of one product for 15 minutes, work 10 hours a day for 5 days a week, it will take us one year (50 weeks to be exact) for 10k products. Our own work should also be priced somehow.
2. We outsource content writing to copywriters.
Assuming average rates of Polish copywriters and analogous amount of content that ChatGPT generates, this will cost 675000$. In Western Europe or USA, that's certainly well over 1 million $. And execution time for such an order is probably several years.
3. We generate product descriptions using ChatGPT.
Generation time = 4,6 days (40 seconds one product). Cost of generation = 1000$ (5k tokens x 0.02$ x 10k products).

I have already made a module controller with configuration. But still some new ideas come to me 🙂
For sure I will have to experiment a lot to choose the best: model, temperature and prompt for query. I need to test different languages and products from different branches. Logging of CRON task times I will still add to estimate real generation times.

And if you want to apply this new invention 🙂 to generate meta data, here is OpenAI PHP API client:
https://github.com/openai-php/client
Only requires a minimum of PHP 8.1

[wakabayashi]

Thanks a lot for your explanation. I can feel you and I am very interested in your experiences in this field - keep us updated 🤗

8 hours ago, e-com said:

We outsource content writing to copywriters.

We did this in the past. Sometimes you find a good copywriter to a reasonable price. But soon their prices raises as they get too many work offers. But you have to spend hours/days to find such a guy on a plattform like Fiverr. Others write worse than a 14 children. But ok now the game is anyway played differently 😅

 

8 hours ago, e-com said:

5k tokens x 0.02$ x 10k products

This is the only thing I don't understand in your post. Where does the 5k tokens come from? I don't think, you want a description of 4000 words or so!? Does it just take multiple tries, to get a result back, that you like?

 

Edit: Btw... Are you a merchant yourself? I always thought you would be a dev 🙈

Edited February 24, 2023 by wakabayashi

[the.rampage.rado]

The idea is great. I'm also interested.

What causes the 'topic' of the generated result? The category name, category tree or something else?

It would be nice if in BO all such descriptions are summarized in some module page where you can approve or decline the result. If you approve it's no longer but is displayed to FO. If declined it's sent for regeneration and arrives in this list once more.

Google will soon 'fix' this and will penalize websites using AI (how they will judge it's up to them, we will have no way to dispute) but for the time being this might help.

[e-com]

45 minutes ago, wakabayashi said:

This is the only thing I don't understand in your post. Where does the 5k tokens come from? I don't think, you want a description of 4000 words or so!? Does it just take multiple tries, to get a result back, that you like?

5k tokens come from configuration settings for product, 1k short description and 4k description. For testing, I set large values to best check the spelling, syntax and logic of generated texts.
But of course it is possible to change it in configuration.
No, I am not a merchant 🙂

[wakabayashi]

7 minutes ago, the.rampage.rado said:

Google will soon 'fix' this and will penalize websites using AI (how they will judge it's up to them, we will have no way to dispute) but for the time being this might help.

I don't think this will be that binary (yes/no). They also stated somewhere: as long as your texts are good, they have no problems with it. But if "good" means above average, you either have to use AI better than competitors or you should make manual corrections. 

I plan the latter. Always read the text from AI and rewrite it a little, so it gets even better 😏 What I also notice: I generate 5 texts from AI for the same product. In each text a like some parts and some I don't... 

Also want to say: these things can change so quickly. Each advise today can be wrong tomorrow. Maybe it takes only 2-3 an AI writes better than 95% of copywriters 😳

[e-com]

18 minutes ago, the.rampage.rado said:

Google will soon 'fix' this and will penalize websites using AI (how they will judge it's up to them, we will have no way to dispute) but for the time being this might help.

Google has already officially announced that it will not penalize AI-generated content with filters or bans.

[the.rampage.rado]

12 minutes ago, e-com said:

5k tokens come from configuration settings for product, 1k short description and 4k description. For testing, I set large values to best check the spelling, syntax and logic of generated texts.
But of course it is possible to change it in configuration.
No, I am not a merchant 🙂

Looking nice!

How is the que organized? Can we send requests for some categories only? What about multilanguage (I see the dropdown). In Bulgarian the bot is tragic but hopefully it will improve overtime.

You said it takes 20-40 seconds for each query, then the result is saved in the local db, right?

[the.rampage.rado]

13 minutes ago, wakabayashi said:

I don't think this will be that binary (yes/no). They also stated somewhere: as long as your texts are good, they have no problems with it. But if "good" means above average, you either have to use AI better than competitors or you should make manual corrections. 

I plan the latter. Always read the text from AI and rewrite it a little, so it gets even better 😏 What I also notice: I generate 5 texts from AI for the same product. In each text a like some parts and some I don't... 

Also want to say: these things can change so quickly. Each advise today can be wrong tomorrow. Maybe it takes only 2-3 an AI writes better than 95% of copywriters 😳

What I noticed (at least for my sites) is that google likes not only plain text but tables, bullet points and images in the text. Probably as you said it will be best to just come after the bot and fix few such items, like include a pic, video or table with specifications.

[e-com]

14 minutes ago, the.rampage.rado said:

What I noticed (at least for my sites) is that google likes not only plain text but tables, bullet points and images in the text. Probably as you said it will be best to just come after the bot and fix few such items, like include a pic, video or table with specifications.

Such artificial intelligence, which creates entire pages for us: formats texts, inserts tables and bullets, inserts correctly selected images, etc. etc., is probably not going to happen in a hundred years yet 😄

[the.rampage.rado]

For sure, that's why it's good to have the list of returned articles (categories, products, etc) sumarized in one place so the merchant could go in one by one and fix, add stuff and then populate the new copy in the db. Could this be done?

If we don't know when certain product desc was generated we would have to go in one by one and check.

Also what happens if we have products that are optimized, can we blacklist certain product IDs that should never be sent to the bot?

[e-com]

31 minutes ago, the.rampage.rado said:

Looking nice!

How is the que organized? Can we send requests for some categories only? What about multilanguage (I see the dropdown). In Bulgarian the bot is tragic but hopefully it will improve overtime.

You said it takes 20-40 seconds for each query, then the result is saved in the local db, right?

Module is in concept-building phase. Of course, like any properly made module, there will be multilang and multishop.
When I close concept phase, I will write more about how it will work.

[nickz]

ChatGPT is hyped by the people strong in their belief that an app changes behaviors. My parents won't use chaGPT and if known that AI is running a search engine they even might change their Search engine. So will million others. So far the socalled AI run on data from the past.

 

[wakabayashi]

40 minutes ago, nickz said:

ChatGPT is hyped by the people strong in their belief that an app changes behaviors. My parents won't use chaGPT and if known that AI is running a search engine they even might change their Search engine. So will million others. So far the socalled AI run on data from the past.

Yeah there is a lot of sceptism about AI. Probably rightly so. 

But what has it to do, if a merchant uses AI as a tool? When talking about text generation: either the text is good or not...

[nickz]

1 hour ago, wakabayashi said:

But what has it to do, if a merchant uses AI as a tool? When talking about text generation: either the text is good or not...

Well people will use it and once their shop gets kicked from the index or is penalized the outcry will be hard. https://www.searchenginejournal.com/google-says-ai-generated-content-is-against-guidelines/444916/

Most people hyping the Chat Generative Pre-trained Transformer as ChatGPT is named won't see those adverts In Spanish language forums there are many people crying since the automatic content crreation gained track. 

There are some useful tools in the pretrained sector for video and photo edition. AI it is not under my definition. 

[e-com]

16 hours ago, nickz said:

Well people will use it and once their shop gets kicked from the index or is penalized the outcry will be hard. https://www.searchenginejournal.com/google-says-ai-generated-content-is-against-guidelines/444916/

Most people hyping the Chat Generative Pre-trained Transformer as ChatGPT is named won't see those adverts In Spanish language forums there are many people crying since the automatic content crreation gained track. 

There are some useful tools in the pretrained sector for video and photo edition. AI it is not under my definition. 

Those lying articles that google will penalize AI-generated content are written by SEO scammers making very big money from SEO and copywriting.
Official google guidance from a dozen days ago:
https://developers.google.com/search/blog/2023/02/google-search-and-ai-content

Quote:
"Is AI content against Google Search's guidelines?
Appropriate use of AI or automation is not against our guidelines. This means that it is not used to generate content primarily to manipulate search rankings, which is against our spam policies."

And all content written by these so called "SEO specialists", is only for manipulating search rankings 🙂

[nickz]

So far the  searchenginejournal.com did not really come over as scammers,mthey rather look into the Google Speech and translate it. 

On 2/26/2023 at 2:14 AM, e-com said:

Those lying articles that google will penalize AI-generated content are written by SEO scammers

Google itself says clearly:

Quote

 

Is AI content against Google Search's guidelines?

Appropriate use of AI or automation is not against our guidelines. This means that it is not used to generate content primarily to manipulate search rankings, which is against our spam policies.

 

People using AI to save on contentwriting, contentwriter who produce word counted texts are among those mentioned.  Writing to manipulate. 

I don't think the public will download AI apps to use those in search, search engines will implement it till they find out the public rejects their results. The search engine market will change. Bing as the first searcheingine to implement their predictive software is the one to watch. 

 

[JamesBrownfny]

On 2/26/2023 at 10:14 AM, e-com said:

And all content written by these so called "SEO specialists", is only for manipulating search rankings 🙂

Isn’t any promotion of content inherently reduced to manipulating it?! After all, if your source is authoritative enough, and I will refer to it that way, then all methods are good here!

[wakabayashi]

@e-com I hope, you are making progress on this 😊 I guess you have access to GPT-4, right? Is it much better in practice than GPT-3 or is it overhyped right now?

[e-com]

2 hours ago, wakabayashi said:

@e-com I hope, you are making progress on this 😊 I guess you have access to GPT-4, right? Is it much better in practice than GPT-3 or is it overhyped right now?

I haven't had much time to work on it in recent weeks, but I've got most of it done. There are AdminControllers done for manual content generation. They have two methods "Generate" and "Send". All that's left for me is to make methods for CRON tasks.
I tested a little bit quality of this content for Polish language. English language is nothing to test, because 90% of Internet is in English 🙂
And after short tests I can say that this name "artificial intelligence", is a marketing babble of corporations wanting to sell customers their next product. Lots of grammatical, stylistic and logical errors. Such a primitive animal as human will not develop a real artificial intelligence in next hundred years 😄
I'll find some time then I'll add this new GPT-4 algorithm. OpenAI API client is also improved every now and then.

Edit:

You have to wait for API to GPT-4:
https://openai.com/waitlist/gpt-4-api

Edited March 16, 2023 by e-com

[JamesBrownfny]

I agree, there were many speeches about this, but in reality it is completely different, especially if you try to generate content on narrow topics, and so on! You correctly describe the cons