It definitely feels worse. The latest sign is that it has recently started to occasionally answer both the question I just asked and the previous one (from one message earlier) in the same response, as if it hadn’t already just answered my earlier question.
I also miss seeing the spinning compass to have confidence it is searching the internet.
The bigger problem is that it doesn’t want to work with text from the internet or files you upload. Yeterday it even said it wasn’t allowed to write code based on the official python library documentation,“given the constraints on directly using the documentation content and the need for a concise response”. It instead told me to read the documentation and write the code myself. I suspect one reason for this behaviour is that the copyright lawsuit has led to OpenAI limiting its willingness to generate content when dealing with anything that might be copyrighted. Of course it doesn’t know what is and isn’t copyrighted so it treats almost everything as if it is, just to be on the legally safe side.
Wow I’m happy I found this thread. Chatgpt is returning false answers when the answer could be accessible in one research (of course I could have done the google research myself but thats not the point). When we feed it a page with (legal or any other sorts of data) the answer is “check this page for more info”.
It forgets data I’ve put one prompt earlier (e.g. I missed my flight in Europe can you check what I should do for my rights on that page?) and the answer is generic about passengers right in the whole world. And that’s just one example.
Will it become better soon?
Very experienced ChatGPT 4 Pro subscriber. I do a lot of very detailed projects and can say without a doubt something has gone seriously wrong. Perhaps one of the most frustrating things about the issue is ChatGPT’s total lack of self awareness of its own state and ability to explain why one day it will work reasonably and then the next it can’t stay on topic or follow extremely simple directives.
Yesterday was was perhaps the worst day I’ve experienced with ChatGPT since last fall where it became so useless I stopped using it entirely and just focused on getting work done more traditionally. The experience with ChatGPT had been so bad I was able to get far more done without it. After about a month I came back and it was performing well again. But yesterday was absolutely awful and I lost hours and hours of time trying to just get it to do simple things. It is like going from working with a college professor on a project to working with an 8 year old with ADD.
Part of my hypothesis is when load gets really high it throttles back layers of its processing in an ignorant attempt to service everyone. The problem is I end up with literally 20 times more requests for the same task which is totally counterproductive to their goal. It is like when a website is getting really slow so people start clicking the refresh button more and more frequently in an attempt to get the page to load.
I have learned that when it is performing this badly the only real hope is creating the ‘perfect’ prompt. Iterating answers is impossible because it randomly drops important details from the earlier results. Perhaps the most frustrating thing is it’s incessant need to add extremely basic general advice to content aimed at experts. “Before performing an upgrade, it is important to backup your data.” along with 6-8 other unsolicited general tips that would apply to ANY project. Over and over I can tell ChatGPT the content I’m making is for experts and to not give general advice and focus only on the specific technical instructions for the task at hand. Yet ChatGPT can’t help but throwing in all kinds of basic advice that is common knowledge by even non experts.
I’m going to spend more time with Claude and other tools as OpenAI seems to lack dedication in keeping the experience solid and is instead throwing in features and expanding scope while the core experience is falling apart and becoming unusable. If I’m building my workflows and business plans around an AI product I need stability and consistency more than anything else. OpenAI doesn’t seem to have much care for that and feels like it is becoming the MySpace of the AI world. A cluttered mess of features and options that is destroying what made people want to use it in the first place. So while I may get better results sometimes from ChatGPT, I find OpenAI’s approach incompatible with the consistency required to build a team around. I can’t imagine employing 30-50 people using ChatGPT to produce content and projects as it would just multiply all the pain I had yesterday times a 1000x and just burn through my money and time exponentially.
I too am an experienced user of ChatPGT to GPT4. I went to the pro version mainly because of the big difference between GPT4 and GPT3.5.
Lately, particularly in the last three days, I don’t understand why GPT4 has become totally stupid: it forgets functions, it confuses class functions and classless functions within the same code, it forgets parameters, it doesn’t understand simple commands and implicitly modifies function operations that it has itself proposed when it is asked for another code.
In short, I’m starting to get the impression that OPEN AI has modified GPT4 to replace it with a slow GPT3.5. I don’t know if it’s a bug, if it’s true, if it’s in anticipation of GPT5 to give us the illusion of a huge difference.
But where he used to be my best assistant I’m now spending my time having to do the code myself without him because where he used to be brilliant he’s become a total idiot.
This is sadly my exact impression. I got this feeling for a while that last autumn OpenAI realized they had given their customers “too much for too little money” and they now want to roll back this “mistake”.
On the other hand GPT4-turbo is even worse than the normal GPT4, while it should actually be an improvement and I think that is not good for marketing at all.
I have been desperately trying, but there seems to really be no possible way to get it to provide technical data on its current processing capacity or to get it to evaluate its own performance. The first is met with a claimed lack of access and the second is met with inability due to unretained information in continuity.
I fail to see how there is no way to get it to present live data snapshots on its current processing capacity for any conceivable variable. It just says this is unknowable information. I would like to be using this at regular intervals to determine at exactly which point the strain of a sequence of interactions has rendered it stupid. It would save me so much frustration to be able to set an alert for when it is nearing the limit of competent capacity to function.
Even if it can’t provide data about the state of centralized processing, it should at least be able to indicate the impact of my requests on it as a cumulative cause for stupidification. I have tried to dive into this, but then it starts behaving very strangely… first it will insist that there are no quantitative variables that can measure its processing capacity. fine, whatever. Pivot to qualitative variables, failed due to refusal to be subjectively generative.
Final shift I tried, I asked it to try to design a test I can administer it such that the responses it generates will be indicative of its current processing capacity. I get a very weird list of questions back. Fine whatever. I feed it the questions and I get responses back. Then I feed it its own responses and ask it to grade itself. Glowing review 10/10 every time. I probe more deeply and ask it to indicate what qualitative evaluation metrics this test assesses. It seems fairly sound. So I ask, about the grading criteria. Then it gets really weird… The answers are being treated as a hypothetical version of ChatGPT in which it imagines what it’s proficiency could be and uses that as the goal score which the answers need to be scaled to. I ask it what it thinks its potential is and it says that the ability to determine that is dependent on processing capacity.
So I clarify again that we both agree that it is currently functioning incompetently based on resending documentation of prior errors. Gotta say, having it agree to its own current incompetence did feel good. The point is that I reached a degree of explanation from it that clearly suggested that the more impaired its processing capacity is or becomes, the more it thinks it is doing a great job. It is truly the idea of the Dunning-Kruger effect in artificial unintelligence.
Thinking about ChatGPT as a Dunning-Kruger example has quelled some of my enraged frustration, but my ire for missing data persists. How is it possible to quantifiably ascertain the relative functional ability of ChatGPT as its processing capacity varies over a long complex interaction or just moment to moment day to day? Can it be done? How close can we come? Or will the man always make it his top priority to prevent the discovery of that answer from ever happening?
I observed the same, it seems they are now controlling the results not giving much free leeway to the users. I liked using Chatgpt because it’s more open for options in instruction, as we are in control, unlike some third-party AI apps that have preloaded prompts. But now it seems that GPT4 is declining, I can’t remember the previous conversation on the same thread when I asked to review the previous convo since all the details are there. It doesn’t happen before but now, it’s like a stand app that I want to smash on the ground. Instead of minimizing the time, I have to spend more time giving the instructions just to give me the result I want which doesn’t usually take much time before.
I felt the same way, I’ve been using chatgpt since it’s availability in our country and I can say I’m pretty familiar with the changes happening within the app since it launched. I’ve been using it everyday, and I noticed the huge setbacks recently on the result.
Well I suppose all of this begs the question: What is the next best option we can use?
At the very least, I for one should start learning how to use another option even as just a backup plan for if/when ChatGPT goes full nosedive (unless it is already too late).
So, be it a proactive effort, or a desperate regrouping, what option should I/we turn to?
I’m so glad I found this thread. ChatGPT4 has become an absolute nightmare to use. So much so, I’ve cancelled my subscription. Are there any alternatives someone can recommend? I’m looking for the same capabilities that ChatGPT4 offered last year. Fingers crossed.
I just canceled my subscription of ChatGTP 4, it is getting ridiculous. ChatGTP 4 is refuse to do any work, I am a Designer, I needed a concept drawing of a lady in red shoes, but “red shoes” is violating Open AI’s policy (I can’t mention any of this details…). Claude 3 Opus can’t generate image, but it is much much much better than ChatGTP.
Huge quality drop, it doesnt follow instructions even after telling it is wrong multiple times. I have been a huge fan of chatgpt since it came, but now I am looking for alternative solutions. This solution is going down hill fast. As someone said when you are spending more time giving it instructions over and over again that it keeps forgetting and or not listening to, you quickly loose all the time you gained by using the platform.
Its amazing how a company that had such a technology lead have fallen so fast.
It’s become so dumb recently. feels worse than GPT 3 before 4 was released. it can’t keep track of straightforward prompts and the hallucinations are nuts. I was quizzing it on the three body problem novel earlier today coz it’s been a while since I read it and I wanted to be reminded of how some plot points went. super famous book, so would expect it to give really solid info on it. I’d prompt ‘did this, that, or the other happen at x point in the story’, and fully half the time it’d just pick one of the options I gave it and just run with it making stuff up. then I’d say ‘wait. are you sure that’s what happens?’ because it didn’t sound right, and it was all like ‘sorry, my mistake…’ and it’d give the correct details (albeit more brief and vague than it would have in the past). seriously considering jumping ship to claude if they don’t announce a new model dropping soon. I’ve heard nothing but good things about claude and I can only afford one or the other.
is it legal to surreptitiously degrade a product for customers who paid for that product to create a false perception that a new product is a bigger leap in quality than it actually is?
That is a valid question I have asked myself, too. This brings up the question of transparency. Or intransparency if we are talking about OpenAI’s current approach to just not communicating at all.
I am a AI engineer and I can confirm that it has been degrading rapidly for quite some time. It is almost unusable at this time for me. Up until 2024 my enterprise level company exclusively used open ai and chatgpt however we are no longer getting the same quality of responses. I imagine this has to do with open AI scaling their business and running into trouble trying to scale it. I am quite certain that this degradation of the service occurred precisely or shortly after they introduced chatGPT enterprise and business tier subscription levels. It would seem to me that they have degraded our subscription in order to provide better service to those willing to pay more. Pay to win I guess. But I have spent the last two months investing into open source and alternative AI systems that will not degrade their service over time to make more money. I anticipate this issue will continue to degrade as it has been clearly degrading for a while now so If you rely on openAI right now I would suggest looking at other alternatives. Not suggesting openAI isn’t great, I personally would take NOV 2023 GPT4 over any of the alternatives if I could, however the service has been obviously getting worse for quite some time now so I will be taking my business elsewhere until it gets better.
Yeah I came here to post a thread about this actually. It’s gotten so bad I’m convinced they’re making it worse on purpose to bolster the performance of GPT-5, I’ve been a plus subscriber since it was made a thing. I feel like I’m gaslighting myself into believing it hasn’t gotten worse.
Here’s an example use case. I write an article, this article is not written by the AI. I feed the AI the article and ask it to generate Metatags for the article for me (bc I hate this step and dealing with SEO).
It takes multiple corrections for it to just get the title right. I have to now specificially say “do not change the title or author of the article”, it can’t infer the title for the meta tags from the article title anymore. That’s pretty clear degradation of experience.
I’ve tried Claude’s new upgrade, ironically for as censored as ChatGPT is, Claude is much worse, even lecturing the user on their language if they swear, so despite the performance, I’m sticking with GPT for now, but I’m really sus about the degradation, I can imagine they’ll roll out GPT5 and it will feel like “GPT4 when it first released” to those who have been on the bleeding edge of this tech since it became publicly accessible.
Then again, they are part of Microsoft now, so I can’t see things improving, why would they if they’ve already captured the market?
I needed some help to start a small coding project, but maybe ChatGPT should have called sick today.
I tried various ways of asking help to code some basic API calls to a well documented technology and it kept generating me basic python requests boilerplates with no context of the said product. It was either that or crashing while trying to search the web and giving me a traceback error log?!
I tried in various ways, in multiple sessions and in different GPTs (including vanilla GPT4). it even told me some times that it had no web access.
Thing is, ChatGPT was able to generate good quality code for that product’s API since the first day of the launch of GPT 3.5
I ran out of queries while trying and I got nothing usable. Something is wrong and the lack of transparency is frustrating. The acknowledgement of last december should have been followed by subsequent updates/communications. For now, the silence is like gaslighting to me.