What I've Learned From Building YourAI | Using ChatGPT and DALL-E ... |
Posted by Steffi Lewis on 19/12/2023 @ 8:00AM Did I mention that I was building an AI module called YourAI? Well, the correct technical term for it is an AI Wrapper because it uses APIs to communicate with the underlying AI (in this case ChatGPT and DALL-E) and then presents the results the way I want ... I've had some great feedback about YourAI from my subscribers and non-subscribers queuing up to use it! created by yourai using dall-e & open ai So, as the developer, what did I want it to do? Well, I've been using ChatGPT and DALL-E since they first came out and always felt that there was too much heavy lifting with the prompts. Having to specify every detail was a pain in the bum so I decided to look at it a different way. How about presetting many of the prompts?
"So, this is where my 'Blog From Text' and 'Blog From URL' ideas came from!"
Firstly, let's take a look at 'Blog Form Text'. This is really simple as you can just enter what you want. The simplest example is "Why are leaves green?". Now, YourAI remembers what settings you entered previously such as Point of View, Tone, Spelling, Currency, Length etc, and lets you add in a role and keyword, so why do you need to enter them again as you would with a normal ChatGPT prompt? That's what I wanted to simplify for my YourPCM subscribers.
And that led to 'Blog From URL' where it gets a little more complicated and a lot more frustrating when I was creating it. I started testing it with an article on the BBC about F1 racing legend Fernando Alonzo. ChatGPT itself had assured me that it could read an article with the API so I plugged it in and hit go. And all the information it wrote about Alonzo was from a few years ago and the BBC article I was using was about current events.
ChatGPT eventually admitted it was lying to me and that it could only show information up to a certain point so I used the Neutrino API to grab the text out of the article and then feed that into ChatGPT's API ... and hey presto, I was getting blog posts written about current events. So, all good, but it took a lot of fiddling to get it right.
Further Reading links also had the same problem. When you ask ChatGPT for some further reading links, it can only go back to before its cut-off date so a lot of the URLs it was giving me didn't exist anymore. Well that was a pain in the bum, what was a girl to do? Thankfully in YourPCM I use another Search API via ValueSERP so I plugged that in, gave it the title of the blog post and it got it right the first time. Eight Further Reading links, all current and straight from Google.
"Image creation was a lot easier to do!"
Using the DALL-E API, I built a simple prompt using the summary text from the Blog articles YourAI was creating. It has enough detail of what the blog post is about in it, so hit go and see what comes back. I usually set 'realistic' and 'vivid' with DALL-E 3 although I do allow a manual prompt override to change what sort of image comes back for use with the finished blog post.
And finally for blog posts, the audio versions. ChatGPT has six voices and yes, they're all American for now, but they are excellent and it's straightforward to get the TTS API to create MP3s. I say to my subscribers, copy out the text and drop it into your blog then edit the text to a final version and bring it back. Then create the audio version. As it's an MP3 it's really easy to embed although I do offer a finished audio HTML link so they can copy and paste it in.
So, what lessons have I learned?
Playing with images can get expensive really quickly. If I look at my ChatGPT text creations they are mere pennies whereas image creations are a lot more expensive. And we're talking roughly 12p per image so when you're a developer who's testing, testing, testing, it'll add up to £100 quickly! Secondly, you have to remember that ChatGPT, and in fact, every AI out there, can only use data up to a certain point so what it's telling you may be out of date or it completely made it up. Always check what it returns to you and never take the facts it presents at face value. Thirdly, the later, more powerful models haven't been tweaked for anything other than US English. The same goes for the TTS API and the American voices. Both would benefit from internationalising and British English and a few English voices would be cool. Still, I can live in hope that this happens. I still have to replace Z with S in many blog posts as well as add the U for colour, and I still use an external TTS site to get English accents for my blog posts even though they're not as good as ChatGPT. There's one last point I want to make about DALL-E (which I love by the way), but it is something it needs. This is the ability to load a reference image as simple as a passport photograph. I've built an avatar generator and although I can tweak all sorts of settings such as face shape, clothing and backgrounds, it never produces an image that looks like me ... or rather it does, but when I was a teenager, not who I am now.
So, Open AI, please, please, please add a reference image for DALL-E! This can be something as simple as a URL to a PNG or JPG on the web which will give DALL-E an idea of things like face shape, eye, nose, forehead etc. I know some people will try to feed it inappropriate images, but you have your filters so you can just return an error if you're given anything naughty.
"But what's next?"
Oh, that's simple. I've had some great feedback about YourAI from my subscribers and non-subscribers queuing up to use it, so over Christmas 2023 I'm going to be creating a standalone product that people can sign up to via its own website.
It's not that hard to do as it's just YourPCM, but with everything switched off apart from YourAI! I won't bore you with the technical details, but I aim to have it finished by the 1st of January 2024 ... so I'd better get on with it then.
Merry Christmas everyone!. Love, light and logic ... 
STEFFI LEWIS
Would you like to know more? If anything I've written in this blog post resonates with you and you'd like to discover more about , do give us a call ... About Steffi Lewis ... |  | | Foodie, sci-fi nut, cat lover, brain aneurysm & cancer survivor, countryside dweller, SaaS entrepreneur, developer and networker.
Published my first website in 1993 for the Open University and am highly experienced with Windows Servers, SQL Server, HTML, Classic ASP, JavaScript, and CSS.
I've also worked as a professional photographer in Los Angeles, USA and been a vision mixer and producer for live television in my time.
I live in a village north of Milton Keynes with my two cats, Baggins and Gimley, and a large planted aquarium full of unruly tropical fish.
|
|
More blog posts for you to enjoy ... | | | | | | | | |
Other bloggers you may like to read ... | | | | | | | | | | | |
| 
