GPT-4 is getting worse and worse every single update

Why is OpenAI not addressing the performance decrease?

They know that GPT4 is less useful. GPT 4 no longer understands simple instructions that it used to understand. The responses often are of no use.

It has the same level as Bard and Bing, which is totally crap.
You cannot use it for anything anymore.

I want openai to inform us users, and to offer us an option with a useful agent.

7 Likes

I have noticed this as well. Could be due to the recent update, but I cant say with certainty as I’ve been using the 32k GPT4 Teams model exclusively since it came out a few days ago. Whether its simply the larger context or some other changes to the model, it’s performing better than ever when it comes to writing code.

1 Like

I echo your sentiments, the recent performance I’ve seen from GPT-4 is quite shocking. I’ve stopped taking any code suggestions seriously, opting to only use it as a research assistant, which it is still “okay” at.

I thought I would share one example from my use that left me speechless. I have never seen it write a code example where it suddenly stops writing in the code block and continue writing example code to the chat window as ordinary text. Some chunks were plain text, others were SCSS, the last box changed from markdown to several other types until it finally settled on kotlin, it contained no code, it was just the summary of all the brilliance it had offered me.

I don’t know how to react to this, I should probably start looking for alternatives. As I said previously, I have stopped relying on GPT-4 to offer any sensible code, but if I can’t rely on it to provide me with reliable factual information as a research assistant then we’re back to “google-it” days.

5 Likes

I have been using ChatGPT 3.5 and then 4 for a year now.

I have to agree. It has been RHELed out of existence. :no_mouth:

It refuses to read simple text files. The content is now flat.

Prompts are ignored, like “write longer”, and can only been overcome with shouting prompts endlessly repeated “WRITE LONGER WRITE LONGER WRITE LONGER WRITE LONGER …” etc.

I get better results with a LLAMA v2 derived model from Huggingface than ChatGPT.

The model is stable and I am not dependent on reworking prompts when OpenAI randomly throws in their next “update”

4 Likes

This is ridiculous. Basically what’s happening now is that I’m paying so that it talks back while being useless. I’m cancelling my subscription too. It’s offensive, honestly

7 Likes

I take back what I said. After using GPT teams a little more thoroughly, it seems no better than it has been in recent weeks. So frustrating when it used to be so good…

I did the same… cancelled my subscriptions. 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 at $60 usd per month. Team is the same product when comes 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. Tubo for me wad good for 1week. 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 start off like Chatgpt before, but limit memory. I feel like I lost a work companion.

4 Likes

I feel like I lost a work companion.

I wrote a very long and detailed comment about the same experience that unfortunately got removed because I honestly got a little to salty.

I went full time into learning how to code 6 months ago and ChatGPT was so important in that journey that it really felt like a coding buddy.
I could pitch ideas or questions and it would just deliver every time.

Since around November it straight up doesn’t work for coding related tasks,
it’s like they added a directive like “Do not provide help to construct code or scripts” because it’s not an issue of the model not understanding what you request.

You can give it an extremely detailed prompt with clear instructions that it will happily confirm and give a summary of - and then not follow.

“All code provided should be fully constructed, do not leave comment lines instructing the user to fill out the rest. Comments in the code is not allowed, examples is not allowed, the code should be working as-is and your task is to provide the manual labor of completing the list following the clear instructions provided.”
Result: "// Repeat for the other variables "
“Remove the comments and complete the code for all variables”
Result: Gives the same code

“Do not provide any code unless explicitly requested to do so, and do not repeat code already provided. I want to discuss X before we start constructing the script”
Result: "Yes, you can do it like this: → Gives faulty code

The list just goes on and on, it will ignore simple and clear instructions the very next response without a care in the world. It’s maddening, because it used to work so very well so its’ hard to shake the feeling that you could get it to work - but at some point it finally becomes obvious that you aren’t using something even close to what you previously worked with.

My biggest issue is how it’s apparently perfectly fine for a subscription based service to severely reduce the quality of the product provided without clearly communicating what is happening and why. I subscribed for a very functional product, and rather than the service improving with time they figured they could provide less while charging the same. All these extra utilities and functionality they keep adding are useless when the core product is getting worse.

My current bet is that they either figured they were at a point when they have such a dominant market position with no real competition for average users that they no longer need to impress and can save costs by providing bara minimum (A Pokémon game approach) - or they have deals with some big entities to provide only them with the premium product and remove the access for average users.
These are only theories, but because they aren’t communicating with their paying users properly it’s hard to argue against because there’s not really anything to go on other than the fact that it has gotten significantly worse.

I mean come on, ChatGPT and AI in general has some of the most frenzied fans that will defend it with their last breath, and even they are silent because there’s no room to even try and argue that it hasn’t gotten significantly worse.

I still tries to use 3.5 from time to time for some simple coding task to save time, but it has reached a point where even that isn’t viable anymore - odds are you will spend more time fighting it to do as you want then it would have taken to simply manually do what the model previously could have done for you in a second.

Shame really.

8 Likes

XD the chatgpt3 was doing stuff like this actually in the first month if i remember correctly

ChatGpt4 was great but lately completely unusable. Getting dumber and dumber with each iteration. Questions that it answered correctly before it can no longer do even after prompting it towards the right answer. Frustrated I even give it the right answer and two back and forths later on same thread if gave me the opposite and wrong answer and can’t even parrot the correct answer I gave it. Constantly gets stuck in this is true no this is false, no this is true never ending loops. This is completely worthless. Going back to Google search.

7 Likes

I honestly don’t understand why so few people notice how stupid GPT is, anyways for what they give access to. It increasingly fails to answer the simplest questions in a very restricted context and data.

For the most part of the last week, network and error in input stream errors are happening a lot on chatgpt, sometimes 5-6 regens to finally get a full response.

Altman was again talking AGI so i imagine their focus is all on what’s comming next. I hope we will have something better really soon for what we pay.

4 Likes

Picture this for example:
I do 1 prompt that costs 10 cents to process and get a good response.

Or:

They lower the quality of the model so when I do a prompt it only costs 2 cent to process and I may get a good response or not. I can now do 5 prompts for the same cost of the previous 1 prompt. Now the user just has to spend more time to achieve the desired result.

In my opinion, they should offer different tiers for different models or give an option to select the response your require for the given prompt, e.g.

  • You get 100 credits every 3 hours
  • a low effort prompt costs 1 credit so you can do 100 of those for your available credits
  • A standard prompt costs 2.5 credits so you can get 40 prompts
  • a premium prompt costs 5 credits and gives a higher quality reply but you can only do 20 of those with your 100 credits

So if you could choose the quality of the answer you require with each prompt you’d have more control. When a low or standard quality response don’t give you the desired response you can pull the ace card and force a good response at a premium cost.

3 Likes

I’ve been experiencing significant delays in response times with the ChatGPT GPT-4 interface, and the frequency of failures in generating responses has increased markedly. It often starts to generate text but then results in an error, which seems to happen around 25% of the time, based on my estimation. The situation has escalated to the point where I frequently lose connection to the ChatGPT website, sometimes lasting up to 30 minutes

2 Likes

It’s truely bad these days, in all areas. I just had an argument with it over trying to get it produce a street scene image as though taken as a camera phone snapshot. It wouldn’t produce anything that didn’t look like a pixar cartoon then when I argued with it, it refused to make any more images.

2 Likes

FWIW: arguing with the LLM typically doesn’t yield good results and seems to lead to confusion above all - it’s often better to go back and edit and adjust your prompt with positive instructions until you get what you want, rather than iteratively telling it what not to do. Maybe this helps!

1 Like

I tried that first. I also tried telling it to ‘make an unconventional image of…’. In one of it’s more lucid days it told me that DALL-E can get stuck in the common style of imahes of a certain theme. The arguing was when I was sick and tired of it’s nonesense.

Google Bard/Gemini is getting better and better every single update ! The OpenAI Dev team should not focus only on money and should focus more on improving services. There was no problem when using GPT3.5 as a free user. However, after becoming a paid user, the problem continues to occur.

Did you feel better afterwards or did it just exasperate you more? :thinking:

I understand that this is a general issue that could hinder the wider adoption of LLMs. I’m curious if there might be a method to alleviate this sense of frustration. The models won’t necessarily perform better, but perhaps there could be a way to encourage people not to give up, or to accept the limitations and find ways to work around them. :confused:

I’m thinking that a model could recognize this pattern, but any action by the model to attempt to assuage the user at this point might come across as condescending. :grimacing:

I don’t feel better - it didn’t do what I wanted. The real source of frustration is that gpt4 used to be able to do the things I ask it to do. I asked it tp perform far more complex tasks and via DALL-E make incredible images. Not any more.

3 Likes

I just wrote this so that your decision to cancel your subscription would be final. Let’s wait for adequate AI from the Chinese or Google. This office has already done its wiring.