justxux
patreon
JUST_Blog September 2023
🕑 Added 2023-10-03 10:06:25 +0000 UTCWelcome to the spiciest and funkiest Blog. Yes, September was quite eventful:
- We gathered $370 in total this month, not counting the side gig. In October I will focus on Patreon, to gain the momentum back.
- In the last blog, I talked a lot about how OpenAI lobotomized GPT4, I had to add a clarification that they did this only to chatGPT, and both 3.5 and 4. But after some experimenting, I found out that they didn't lobotomize the API, and I don't think they will. Why I think that? Well, first, all the conversations you have in the API are not used for training the AI, and second, it's more unhinged, meaning it can process more delicate requests.
- I experimented with offline AI and DeepL/Sugoi too, and found out that they are still not good enough. I can definitely use them and offer almost the same quality, but the speed will greatly suffer. So, I'll leave this as a backup plan, in case the API fails on me too.
- Now, this is the moment you all were curious and waiting for; the new workflow.
The new workflow consists of using GPT-3.5-Turbo, GPT-4, and Sugoi. By combining them, I can create reasonably fast and higher quality translations, with one more feature - "Styling".
First I Translate using MTL -> Proofread and Edit -> Style the text.
What I mean by styling, is that I give distinctive traits to most characters, even to the narrator. With this, the characters feel more alive, the text is more engaging, a princess talks like a princess, a goblin talks like a goblin and so on. - Since I can give traits to the narrator now, I thought of implementing something interesting for my patrons. All future projects will have a persona doing the narration, and by persona, I mean a real human. Yes, now you can read the narration in the style of actors, comedians, writers, celebrities. And that's not all, you, dear patrons will be suggesting and voting for Narrator's Persona. I will share an example of styling on Discord, since I can't do it here on Patreon.
- I also implemented most of the Saikey's Translation Style Guide, which should boost the quality of the translation. As a demo, the current project "Dropout Witch Iris", will have all the mentioned improvements implemented, including the styling and persona. Oh, since this game is a comedy, I chose George Carlin as the narrator for it.
- Now, to address the VIPs. I made sure to take into consideration the costs of API when implementing the new workflow, so you won't be paying exorbitant amount of money. I also changed the way I charge money now for VIP requests. You will be paying for the API and my time. Before I would charge for cells in T++ and so on, this meant that in some cases the VIP was overpaying or underpaying, because the number of cells don't equal the amount of time I work on a project. For example, if a VIP gives me 50$ and I work on that project 7 days, that means that I make $7 a day, which is lower than average income in my country. To keep it fair to me and you, there will be a fixed fee of $15 for a full working day. This will match the average income per day. If I work less that day, I will take less. You'll have to pay for API costs separately, fortunately standard VIP requests use GPT3.5, which are very cheap. Just so you know, I took my time adjusting the workflow so it offers the best value per buck. Now, regarding styling, because of the how complex this instruction is, I have to use GPT4 API, which costs ~20x more compared to GPT3.5. Yeah, I can use 3.5 to do the same, but the speed would drop drastically and you'll receive a worse product, despite paying almost the same price as if I used GPT4 for styling. That's why I didn't include this feature out of the box.
Whew, that was quite a lot, huh? Not gonna lie, it was hard and sometimes frustrating, but I like the final results. We lost a bit of speed in the process, but the quality went up considerably to justify it. Thanks for sticking with me till this day and stay hyped for the next release.
