I always figured ChatGPT remembered things based on an amount of words or so many previous replies but it seems to be more to do with time than anything? I noticed if I ask about something and then continue to follow up on it he’s able to keep on the topic, but if I ask about something and then several hours later post a follow up that is asking about one point in his list of points for example, he’s clearly forgotten all about it even though it’s literally right after his last response. He’ll just give a generic response as if I’ve asked about that thing in isolation instead, of expanding on that as one of his previous points in relation to a specific thing.
This is normal. It is so overloaded right now that it breaks every few minutes and needs to be reset. It is exceptionally irritating. Right now it is completely useless. If I ask it to write code, and then I ask it to add something to that code, it will completely ignore what it just sent me and write the code in javascript or something else that isn’t even python, and it will not even be writing the same code I asked for. It will just start making random code up that has nothing to do with what I originally asked. Unsure what kind of bug this is, but it makes it completely useless unless you get the right answer the first time.
I have been using it since it came out, and it has never worked like this. This is really bad! However, I was going to point out that you can specifically ask it to “Review your current conversation” and give it instructions for making sure it understands. HOWEVER… this won’t work right now because it is completely broken.
That’s interesting since I was also thinking it seems to be different to what it was a week or so ago. I don’t see how it’s due to overloading though, surely it should still have the same abilities, just you’ll have to wait longer before it comes out with them? It’s like if google was overloaded, the results would take several seconds to load but they wouldn’t be all incorrect, they would be the same results as if it wasn’t overloaded, just slower.
It’s certainly some kind of bug as you say and would seem they would need to fix it fast as they’re the fastest growing website in history…otherwise they quickly won’t be. Right now every follow up question I give to ChatGPT has to include a lot of previous information otherwise he will just give a generic answer to my question not taking into account what we’ve been talking about at all.
I only assume overloaded. A bug, presenting because it is overloaded. Regardless of my uninformed guess, I believe you described the problem much more precisely than I have done. Any post I made was met by the same person talking about how it’s an LLM and not meant to work the way I am saying, even though I continued to say that it didn’t have this issue before. It even says on their website that it can usually remember up to 3000 words in a conversation.
Actually, this 3000 words thing is why I believe it’s an overloaded problem. I got on at 3am this morning and it remembered everything we were talking about and it was typing super fast. The slower it types, the faster it seems to forget. Slower typing is usually indicative of it being overloaded. The slower it types, the more it seems to crash.
I am just seeing all of this, and assuming that the bug we are seeing is presenting because it is overloaded and it’s memory is being flushed because of some sort of memory leak issue, and that clears it’s memory faster (when it’s overloaded) then the “3000” words situation can keep up.
Oh, I was also going to mention that the Google reference is a good one. I think your point stands firm. This problem is caused by a bug, not specifically because it’s overloaded. However, the big presents much more aggressively when it is overloaded.
I agree that it is likely not specifically because it’s overloaded, but more like there is a bug that only seems to appear when it’s overloaded. Something they are doing to try and prevent the system from crashing is likely causing the memory loss issue that we are seeing. Once those Microsoft servers start ramping up, I imagine they can stop trying to constantly prevent crashing and just get the virtual space they need to succeed!
Ok I see what you mean now, I thought you meant it was just overloaded and too slow, I didn’t realise you were saying that’s the reason for the bug. Yeah I don’t know what the reason is either but it seems to not be remembering based on either how overloaded he is or how many hours have passed. I normally have the ChatGPT tab open most of the time but if it’s been a few hours it will definitely throw an error if I ask him something else and then reload the page. Then he definitely won’t remember anything previous so the whole 3000 words/tokens or whatever seems to not apply at all.
Even if I just ask him something so simple as to translate the paragraph he just said above into another language he won’t know which paragraph I’m referring to and translate something from days and days ago seeming totally random. I literally have to ask him to translate this text and then paste the whole text that he just said all again in my new prompt. It’s weird for AI that is able to understand even complex stuff to not be able to understand that. It’s like he’s gong from having good memory ability to that guy in the film ‘Memento’.