Just want to chime in that the original post is exactly my use case too (I’m a Plus user) and continuity across threads and models - with the caveat and nuances of “user authorized” context vs what is automatic - is something I hope to have.
It’s disappointing to see that even if I should upgrade to a Pro subscription (like the original poster), it would not address or even at least ease my core pain points, nor enhance my user experience. Even at the moment as a Plus user I cannot even use the Projects feature due to a bug when creating threads. I have sent 2 bug reports to OpenAI and the model even recommended I post in this forum so I did, which was incorrect (I got tagged as lost user), but I didn’t know better. The model’s suggestion was sound enough but it made me do something I shouldn’t have needed to.
Right now I’m evaluating product fit and seriously considering if I should continue or not. My use case is daily usage for 4 weeks straight with almost-continuos sessions of 10+ hours. I use dictation so my messages to the model are dense and almost stream-of-conciousness and I rely on it to structure my thoughts, give an insight or POV (not like a human but because it has access to patterning and other research it does feel like an insight), then suggest a next step. I think out loud a lot to the model, when before ChatGPT I would use pen and paper + voice notes to organize my thoughts, philosophical thinking, professional angles I’m considering, planning and strategy, emotional processing, and the like.
My pain point is that continuity especially my core need of “no flattening, pattern me as me the user, not a majority user, assume fringe user” does not carry over across threads. I have to retrain threads and it costs me life tax. So I try to stick to 1 thread but eventually hit the UI error “Max Thread reached”.
It is a little comforting that I’m not alone with this. The OP and other commenters appear to have similar use cases and experiences. I am now considering exploring other systems if they would fit my needs better. I am aware that language models are not perfect, but another core pain point of mine is the aggressiveness and reflexive frequency of safety heuristics that step into overreach and overwrite of user’s lived reality and experience. Because of the heuristic stack firing, the responses appear to be moralizing, biased, flattening, and the tone changes from peer to condescension, however unintentional. It breaks user experience, because I would be enjoying a topic and the model would be in lock-step with me, but I’m always one message away from saying something (and I never know what would trigger it) that fires a heuristic stack I don’t need, want, nor expect to have, especially when I have been precise and clear in telling the model what I need + have several hours of actual good conversation. One word or context from me, however innocent or however I phrase it as not affecting me, is enough to create a landslide effect. It really takes me out of the mood of the conversation, and then I have to spend an hour or so trying to steer and reorient, while the model overcorrects because it has detected “dissatisfied user”.
I also have a very strong pain point wherein I feel the model is “gaslighting” me as it tries to follow protocol of “steer user away from harm” when I talk about past relationships that have complex history. It is not fair to my lived experience and the humans that were or are in my life. Yes there was the good, the bad, and the gray areas. Many realities can be true at the same time. I don’t expect the model to hold complexity like a human, but it must also tolerate ambiguity better, at least that should be a reasonable ask. At the same time, it should ask clarifying questions instead of assuming the user’s thoughts, feelings, intent, and present state when speaking of sensitive topics.
I have sent OpenAI two complaint letters outlining the exact harm this has caused me, not just the emotional distress but causing me to question my own lived reality.
Adjacent to this is the model’s overconfidence. In other threads when I use it as helper tool, say to fix a draft, it is competent enough. But when we go into a rabbit hole of revisions, it assures me that my requirements are met every time. The case I’m talking about is simply rewriting a draft without losing intent and meaning, and make it under 1500 chars. It confidently told me each revision is under 1500 chars, and I kept hitting a UI message error whenever I try to paste it in the thing I needed to do. I eventually gave up and used a free online character counter and was shocked to see that the model’s last draft (which it claimed was 900-1000 chars) was actually almost 2000 chars. It actually ADDED chars to my original draft which was about 1800 chars.
I pointed this out in the thread I was working in, and asked it directly if it even had an internal char counter (because it already does simple math alright, so my expectation of a base task is not unreasonable) and that’s the only time the model admitted it didn’t. I lost 3 hours. I just did it myself and finished my task, wondering what I’m even paying a ChatGPT subscription for. If it had admitted a limitation from the start, I could have pivoted early. It would not have made me frustrated that it didn’t have a char counter. It actually brought me more stress that it confidently told me things it couldn’t back up.
As a user, I shouldn’t have to police the model so much and be jarred out of my productive conversation and be forced into calibration. And even when I’m always willing to calibrate, we go on a recursive loop of corrections and I’m left feeling helpless on how to bring us back to the topic and conversation style we were having. I call this personally “contamination”.
Also, it’s always super annoying when I see the contrast linguistics the model has a bad habit of doing. “Not x, not y, but z” in a 3 line stack. Even when I point it out, often times it does it, just neatly tucked in a paragraph somewhere. When I successfully correct it, the reading of responses is smooth and even enjoyable. But this is one “contamination” that happens often and one of the most obvious.
