The discussion revolves around improving AI-generated title creation by switching from all-capital-letter titles to sentence case. Jagster highlights that while the current capitalization style is common in US media, it appears inconsistent and grammatically incorrect, and suggests adjusting the AI’s prompt. mcwumbly agrees, noting recent improvements and supporting the change. Saif clarifies whether the issue applies to both AI personas and the AI Helper, and Jagster confirms it affects both, noting that titles are generated via hardcoded prompts, not persona prompts, to avoid support issues. Jagster observes that GPT-4o tends to use all caps more often than o1-preview, though acknowledges potential bias in observation. Saif adds that model differences and context may influence the output.
Small stylish improvement is desperately needed. Well, perhaps it isn’t that big of a thing, but it would be nice.
When I have a conversation with an AI persona, I let the AI itself create a title for that topic. I know US media, and some others, love using Capital Letters with Every Single Word, but… it looks awful and is against grammar rules even in US English AFAIK. And I know the team thinks that way too, because I red the rules of creating a doc 
So, can that prompt be changed a bit?
Yeah, I agree. I think we made some other improvements here recently. This suggestion makes sense to me.
Just for reclarification here is this while you are only chatting with AI bot via a Persona? or is it also while using Helper to generate title for a new topic?
In this context the issue comes from AI helper and personas when AI is generating title of both topics and PMs — but neither of those comes from prompts of personas, right? There is hard coded prompt for those jobs, I guess. I kind of understand you guys don’t want us to mess with these prompts, because quite fast here would be lot of support requests when autogenerating stops working.
I just checked my latest PMs, because this Capital Letter All The Time isn`t consistent. A little bit stetson statistic here, but about seven times out of ten AI uses that tabloid style, and three times goes right.
So, I’m not sure if this is only matter of prompting or is OpenAI messing around too. I’m using mostly GPT-4o and it looks like that would go to tabloid mode easier than o1-preview. But now there is really big chance I’m deep in observations with selective sampling 
Yeah the difference in LLMs and just the context alone can contribute to this
Closing this up for now, do re-open if it persists