GPT-4 is getting worse and worse every single update

I’m so sad. I had a great tool that helped me a lot with programming.
It gets worse week after week.
I had an unrealistic discussion about a simple problem. He gave me my own code several times, each time saying: “I’m sorry, I understand now. Here is the corrected code…”
I finally corrected it myself. I wasted more time explaining it to him over and over again than doing it myself.
And then this amazing response when I show him the fixed code: “I’m glad you found a solution that works for you. If you have any other questions or need help on other topics, n Don’t hesitate to ask. I’m here to help.”
Seriously ??

4 Likes

im sad but relieved in a way to have just found this thread.

I use chatGPT heavily for coding at work. It was mindblowing when I first used 4.0, and since then has become my daily go to tool, I even referred alot of friends outside my industry who have not used it for work.

But the past few weeks, I have observed a complete degradation in the quality of its responses… I just researched and stumbled on this thread which verifies my suspicion… At first I suspected my prompts were the cause, or maybe the custom instructions, but as I said, im a daily active user, and very confident that its not the way im crafting the prompts.

Sadly it looks like its time to consider parting ways with ChatGPT.
I used to be so satisfied that I have never even tested its competitors, but looks like I will be cancelling my subscription if any of the competitors prove to be more effective.

6 Likes

Well well well. I’m cancelling my subscription.
GPT4 used to be amazingly good, but for a few months now the generation quality has dropped too significantly for me to stay.

Even simple generation tasks seem to be an issue now (eg. redact a 8 paragraph for specific subject.
GPT4 doesn’t give thorough answers anymore, GPT4 ignores instructions now, GPT4 is slower, GPT4 lags, and so on…
I am deeply disappointed and I urge everyone that seems to have the same issue to unsubscribe as well.

Financial pressure is the only thing these companies understand

6 Likes

honestly, I think that “degradation” is intentional, and to force us to unsubscribe, so policies and prices can drastically change, or make it easier to bundle it with other products. I hope I’m wrong on that. The tool is useless right now, and cant deliver simple tasks that were common months ago.
I cant even get it to translate or summarize texts for me anymore.

5 Likes

I think also what Sam Altman said was pretty embarassing: “If you ask something to Chat GPT 10,000 times, you may get 1 good response”

SAY WHAT??? 1/10,000?

So are we using it just to generate words on a screen?

Just unsubsribed, honestly it was quite frustrating. Level decreased massively, something that took me seconds to do in the past, now cannot be accomplished in 20-30 minutes.

I need to repeat 1,000 times the same super simple tasks I was able to do in the past and still it doesn’t work

can’t even understand when a recipe is wrong with 500 KG of flour among the ingredients

My personal opinion is that

  • they decreased the quality to save money and see how many people may stay so then they can increase the prices with higher quality
  • potentially microsoft, openAI, google etc decided that they create a too powerful tool to be given to too many people, especially for free
4 Likes

The same is happening with paid seats. This is absurd, and should be embarrassing for OpenAI.

4 Likes

Totally agreed. I am schocked at how bad ChatGPT is in comparison to last year. If I ask for “French Jesuits in the 18th century” I get people that are 1. not French 2. not Jesuits 3. not from the 18th century. After specifically asking to adhere to the criteria multiple times, I still get completely wrong answers. Even when asking for very simple facts that can be found with a quick google search, ChatGPT 4 fails to find them and instead hallucinates something very different, drawing conclusions from the dataset that are completely wild.

1 Like

Just jumping in to add another name to the list. Assistance with coding has gotten far worse. The ability to read images is a nice touch. The ability to follow instructions has become worse.

Ive been a paying customer for 5-6 months now and will likely cancel in March once my reason for subscription lessens.

5 Likes

I also cancelled my subscription a week ago, after being a paid subscriber since it was first an option. tl:dr, it’s become really patronising.

When I ask it a technical question, it has this annoying habit of just repeating what I said, but inserting some technical jargon, almost like I’m talking to Sheldon from the big bang theory. It’ll rephrase the question in a way where it answers a different, easier question where it’ll give a dumbed down answer, and finish it off with a description of how it’s the best answer in the world and the question I asked was actually rather simple.

If you ask it a creative question, it will provide the most ‘off the top of my head’ answers possible. If you explicitly ask it to be creative, out of the box, and unusual, it will say the same things, but add words like ‘whimsy’ and ‘captivating’. It’s like the randomness value has been brought down as low as possible.

What on earth has happened to AI text generators? It used to be this really fun, experimental tool where each version had an intriguing personality. I remember generating yoda-style sea shanties (that rhymed!) with the old text AI models on Playground, and getting incredible results, to the point where my friends would read it and refuse to believe AI made it. I’ve even generated stories with AI that have had me in pain from laughing, and moments later tearing up from emotional stories. Now it’s just patronising, fake enthusiastic customer service talk, or code generation that only impresses programmers that have never worked on anything bigger than a single python script.

I’ve decided I’m only going to give my money to AI services that actually push AI development in the direction I want it to go. And wow, there are some GOOD alternatives out there, you just have to look past the ocean of snake oil AI services that exist.

6 Likes

Its gotten sooooo bad I feel since they added the new up to date training data from around the time of the developer conference, its translation capabilities are now 10x worse, it doesn’t understand simple prompts, its behaviour seems almost (metaphorically) lazy!! Is there any way to use GPT4 from before this update? Or any alternative with similar performance? Why hasn’t this been addressed?

6 Likes

Some of the older models are still available in the api https://platform.openai.com/docs/models/gpt-4-and-gpt-4-turbo

but unfortunately, the 0314 (best of all time) will probably be lost to all of us soon

That said, the 1106 model isn’t that bad - it just needs more user effort (ok that’s pretty bad for a user facing model). It needs to be handled a little differently to get the same results as back then.

1 Like

[RANT ON]

Grrrr. Why am I paying for stupid??? I just uploaded an image to gpt4 and asked it a very simple test question. The response I got back was:

“In your image, which is not visible in this text-based interface, you would look for these”

which was not helpful in the slightest. So I responded with:

“I uploaded the image to you WHY IS IT NOT ACCESSIBLE?”

To which it replied:

“I apologize for the confusion. I can indeed access the uploaded image. Let’s take a closer look at the image you provided to”“Without being able to view the image directly within this interface, I’m providing you with guidance on how to”

BTW, when gpt4-v first became available this exact interaction worked flawlessly.

In addition to reducing the context window for us lowly Plus (i.e, non-enterprise) users did they also take away vision??? If chat is now “text only” why keep the paperclip and allow uploading of images?

[RANT OFF]

Basically I’m not getting the same level of results I used to. Also I just tried the same thing with the supposedly weaker Bard pro and it worked just fine.

Here is a different simple test that it failed miserably (see the screenshot below and note the answers it confidently gave re the colors of the counties)

If open ai doesn’t bring back the larger context window for plus users I’m likely to switch to Gemini Ultra when it becomes available. While ~$20/month isn’t much for each one individually, there are now A LOT of services vying for my discretionary dollars and more importantly time.

3 Likes

I did the same as others and agree with most. The prompting has turned from friendly to manager demanding language to get Chatgpt to produce a good/average or poor (you question summarised reponse, lost network response, suspicious behavior from your computer) response. I have cancelled my subscriptions after trying Team plan. My wife 44 and mum 74 have said it has lost and it now better to not use it. We also tried Chatgpt Teams the upgraded from Pro but to no avail, extra 10 bucks per user plus a force second subscription to just try it. Team is the same product as Pro when comes to chating ,coding and so on, you just pay more pre month… its memory is bad, it tells you search the net, bing search still sucks, why would teams use this product over a ChatGPT Pre Nov 2023. ChatGPT 4 turbo my ass it is alot worse, I am have to correct its grammar and spelling, of which I struggle with sometimes. Turbo for me was good for 1 week. I recommended Chatgpt and now I having to let everyone know it is almost no better then Bard or Microsoft Chat in Office365. O365 seems to have the old model which, like Chatgpt 4 before Nov 2023, but with a serve memory limit. I feel like I lost a work companion. My 6 months was a good ride that has now ended. I am looking forward to trying LLM on Huggingface, opensource models. Mybe even try to understand how to create LLMs…

1 Like

In stating something here, I believe it will be a single remark rather than a series of discussions.

In terms of understanding and empathy…
I genuinely and deeply understand the frustration and disappointment that the users here are feeling.
It is infuriating to see something perform well at first, only to watch it become increasingly useless over time, and that frustration is justified.

In reality…
Ultimately, GPT-4’s performance is constrained by computational resources. OpenAI was forced to choose between reducing the computational requirements of the language model at the expense of performance or limiting the number of users, and I believe they chose the former. Considering this, it is also true that “rolling back to the previous version” is not actually possible.

What I want to say is…
I believe time will resolve this issue. Whether it’s Google’s Gemini or an open-source language model, adequate computational resources are required. However, these resources will not increase overnight.

What I think is that this “incredibly appealing yet clearly immature technology” is generating mixed feelings of love and hate among users.

And what I hope for is…
Please consider the various circumstances that OpenAI faces, and understand that there are things they can reveal and things they cannot, as well as things they can do and things they cannot.

Even if it’s an unattainable wish, if my words can bring even a little peace to everyone’s hearts…

This may sound arrogant, but this is all I can say.

2 Likes

For the $20 ChatGPT Plus, which model are we using for GPT-4? what is the context window?

They have a context window of 128,000 tokens.

I’ll add to the above with information slightly more exactly correct.


ChatGPT Plus has a model version that is not necessarily named. It is possible and seen before that new models are evaluated on ChatGPT users in A/B trials and also with different parameters to evoke comparisons of best responses.

The GPT-4 model has a max-tokens (which in this case is context length) of 32768. This can be extracted right from the network requests when you load ChatGPT. This was advertised as a “ChatGPT Enterprise” feature, but can be seen advertised to the client.

The value of context length, though, is in how it is used. ChatGPT, and GPTs, also employ their own management strategies of conversation length, which now at least is generous.

Actual max tokens - response that can be generated before you have to press “continue generation” - is 1536 tokens.

2 Likes

Thanks thats helpful.
but i was trying to summarise a transcript that is about 14,000 tokens. But the output only takes into account of information up to around 4,000 input tokens, but the output is around 800-1,000 tokens.
beyond that, i cant even press the 'continue generation" button. If i follow with “continue”, it says that theres nothing left to summarise.

Given this, and all the discussions above, it just feels that the REAL context length is much smaller than reported

The answers are much less quality now - it is clearly been made less intelligent. The cap for gtp4 is now 10-20 messages! It is useless for work in its current state. I will cancel my subscription today - and recommend that everybody does the same. Only way to make the listen.

3 Likes

Probably not.

They released a blog paper on Wednesday that said it should help GPT-4 act less lazy.

I would recommend chat.openai.com.

1 Like