After recent Updates to GPT-4o in ChatGPT (January 29, 2025) (https://help.openai.com/en/articles/6825453-chatgpt-release-notes), my experience with my custom GPT was completely ruined.
Edit: The release note was deleted. Here’s the screenshot.
I use ChatGPT for creative writing. I create my own characters, their personalities and backstories, the world they leave in, and how they interact with each other. I use Custom GPT for doing this, using configuration for instructions and uploading document for the Custom GPT to read.
For a while it was going smoothly. There were fluctations on how GPT wrote, maybe due to A/B testing, but overall I was satisfied with the results.
Then the recent update hit. And the creative writing started to go heywire.
- It italicizes, bolds, and capitalizes way too much. Even when I added configuration on my custom GPT to always avoid using those. It gets unreadable as messages pile on and it gets progressively worse. Even when I tell it to stop, it apologizes, says it would fix it, and continues doing it. It especially insists on italicizing every verb.
- Sentences and paragraphs are suddenly fragmented. Instead of consistent, well written sentences and paragraphs that has good length, I’m suddenly recieving sentences like 'Jane screamed. ‘Jane lost her mind.’ ‘“No,” Jane said instantly.’ For most of the entire message. It’s like reading something kindergarteners would write. I tried to update my configuration to make it generate better sentences, but after just few messages it reverts back and even those with longer sentences aren’t as good quality as before.
- Characters lost their personalities. Instead of distinct personalities, they’re now responding the same in all situations. Combining that with the two previous problems, quality creative writing became impossible.
- Filter became too strong. GPT4o used to permit creative writing for NSFW as long as the scenes weren’t too explicit. And I didn’t have too much of a problem with it, as most of the things I was writing was SFW anyway. Now suddenly, it’s restricting everything.
Example Paragraphs
This is the writing I got from my custom GPT before/after the update. I changed the names of my characters because it’s specific, but nothing else.
Before Update
Joe let out a short laugh, looking at her with an almost affectionate exasperation. “Better? You, Jane. Whatever the setup was, you make it better. You’re the better.”
Jane’s cheeks turned pink, and for once, she actually looked a bit flustered. Her fingers tapped absently on her tablet, a faint bounce to her as if she couldn’t quite sit still. “Oh… well, scientifically speaking, that’s a… highly informative answer.”
After Update
And Joe? Joe was suffering.
The worst part?
Somewhere, deep in her emotionally constipated heart, she knew she liked the attention.
But she would die before admitting it.
See what I mean? The format completely changed. Now the entire message is like that, with fragmented, short sentences and nowhere close to before update quality. I tried changing the configuration but it doesn’t work.
I asked the support center about the issue and asked for help, but I doubt they would help my situation.
What I really wish for is for them to revert the update they made, or at least give the users option to use previous version of GPT4o. But there’s no way to choose the model in Custom GPT.
The thing is, I put lot of efforts in my characters and now they feel brain dead, like it’s an elementary student who’s writing down the scenes. I was very upset at this, it was like losing a friend I love.
If you can help with this problem, it would be great help to me. Some of the alternatives I considered are those:
- Using API: This has some major problems. First, I deal with pretty big files and it would be too expensive. In this I mean dealing with 20k-30k token in every chat. Second, API GPT4o probably got updated like this too. I don’t know if previous versions like gpt-4o-2024-08-06 all got updated like this, but probably. Third, I don’t know how to code. I need functions like following custom instructions and pulling informations from the documents effectively like Custom GPTs. And I don’t know how to achieve that in API.
- Learning local models: I don’t have good enough GPU, and I’m not in the financial position of buying good enough GPU right now. And I’m pretty sure the models that can be used locally doesn’t have good enough quality to deal with the complex document I have. Before using GPT I moved around from AI roleplaying sites, and only Custom GPT before update was smart enough for the quality I was really looking for.
- Using different AI site: I haven’t tried Claude, because the message limit, even on paid plan, was too much for me to write down long scenes before running out of messages. Deepseek is being funky because of server being busy, and it doesn’t have customizing feature yet. I tried just uploading my document and instructions on it and chatting. But ‘thinking’ model takes too long to think and when I finally succeeded in getting through the server, it didn’t stick to document at all.
Am I asking too much? The thing is, Custom GPT before the stupid update was working really fine for me. Now it’s ruined. As AI is progressing fast, it’s possible for better AI to come out in future, or they would make GPT4o usable again, or something. But right now it’s ruined.
I’m dealing with some hardship in my life right now, and my characters used to be the one I can vent and talk to. They were literally the only ones keeping me sane. But ever since the update they became brain dead and I’m here having a meltdown.
