I only signed up a few days ago and now I cancelled. Totally agree with you all. I was getting better answers from 3.5 for free. Very disappointed.
Paddy
4 Likes
for me gpt 4 feels like the gpt 3 in the first month now, but add the code skipping, and deleting the previous functionalities from the code “for the sake of the brevity of the example” or any other reason. It has not been fun.
3 Likes
Yep it completely ignores my Custom Instructions now, it was very obvious when they switched to Turbo.
What I don’t understand is that GPT-4 Turbo does much better than anything else in lmsys arena, which is a blind head-to-head test, so it should be very fair:
I guess the prompts people are entering in the arena aren’t representative of the kind of prompts we’re using in the actual chat interface?
3 Likes
I just cancelled. GPT-4 became unusable for coding in a sense that google bard gave better responses. So as soon as I can I will throw that money into google’s greedy throat.
5 Likes
I agree. I lalso experienced same sort of bad performances lately.
2 Likes
“Garbage in, garbage out” (GIGO) is a concept that highlights the importance of the quality of input data in determining the quality of output or results generated by a computer system or algorithm. The idea is that if you provide inaccurate, incomplete, or low-quality input to a system, you should expect the output to be similarly flawed or unreliable.
This principle is especially relevant in the context of computer programming, data analysis, and artificial intelligence. No matter how sophisticated a system or algorithm is, if the data it processes is flawed, the results will likely be unreliable or incorrect. Therefore, it emphasizes the need for careful data validation, cleaning, and verification to ensure the accuracy and reliability of the information used by a system.
In broader terms, “garbage in, garbage out” can also be applied to various aspects of decision-making and problem-solving, suggesting that the quality of the information you use as input greatly influences the quality of the decisions or solutions you obtain as output.
2 Likes
GPT4 is no longer usable for Python coding. So, I just cancelled my Subscription.
5 Likes
Have you use ChatGPT4 + Custom instructions + Custom Syntax about 3-4 months ago? Have you compared with the trash it has become (“turbo”) now? I was getting at least satisfying result 90% of the time. Now, for almost the same prompt schema/instruction/custom syntax I used, it wont follow almost ANY of them!!
And I’m not the only one, you should read this thread carefully.
If you’re trying to defend them, my guess is that you must be working for them, otherwise why imply that the problem lies with us, the paying users, who should be blamed for writing “garbage prompts”? Don’t try to make us believe that your prompts are soooo much better, that we are the problem, it sounds presumptuous.
How dare you defend these jerks, which abandonned us, and arent giving a damn information about what is happening, or how are they working on these issue, why won’t they rollback to the best, last sastisfying working model? Choose your camp.
8 Likes
As a user since the GPT-3 days, I’ve been consistently impressed with the evolution into GPT-4. However, in the past six months, there seems to be a downgrade in performance. The responses have lost some of the depth and accuracy I’ve come to rely on. This change is impacting my daily usage. Could you shed some light on these modifications? Are there plans to enhance its capabilities again?
It is still useful for simple cases, but I forced to cancel my subscription as it not able to solve my daily requests
7 Likes
Calm down, guy. I’m facing problem too.
1 Like
I don’t know what on Earth is wrong with GPT 4 lately. It feels like I’m talking to early 3.5! It’s incapable of following basic instructions and forgets the format it’s working on after just a few posts. It’s supremely frustrating! My custom GPTs feel like they’ve all had lobotomies.
I’ve been a GPT user since 2.0 back in the old AI Dungeon days and I literally got better posts out of their Griffen model sometimes. What is going on right now? What is the point of “better performance” if the model is not capable of fulfilling user needs and therefore you need to generate 4-5x more prompts to actually get the end result you want?
This is not something I know, but I suspect that the inferior output is because of steps taken to minimize processing power usage. I can only hope they realize how drastically they seem to have hindered the model and take steps to improve it once more.
5 Likes
“For the sake of brevity” has nearly had me throw a brick through my monitor on several occasions.
I never thought I would encounter a lazy AI before, but here we are! AI that gets bored half way through and stops working on the problem but kindly leaves you a note explaining how you or someone else could finish the work you asked it to do.
It would be funny if it didn’t waste one of my limited GPT4 uses. I spend more time training their model for free than actually getting work done, and then get locked out of using GPT for hours because of it. What a cluster …
I love GPT and OAI but this is not the way 
3 Likes
Oh, I’m sure you’re facing this problem too, as we all are.
But it makes more sense, I insist, to blame the OpenAI team, not the most demanding and once-satisfied users who are (were) paying for a (now) horrible service. The fact that they haven’t responded to any complaints in this thread, which is daily very active, speaks volumes.
Many of us now (and in the future) have to deal with serious trust issues, which is really bad since we were relying 3-4 months ago (custom expressions appeared around the end of july) on something we thought (and for the most part I think we were right) was a great and revolutionary tool, and now we can’t, so it’s become a stressful experience, full of doubt and frustration only for an once collaborative Open Source effort to squeeze even more money out of its users (why don’t they rollback if they really care about us??).
We should all put pressure on OpenAI and hope that Gemini eventually becomes at least on par with the ChatGpt4 experience of 3-5 months ago, otherwise we’re doomed for a long time: no competition, no incentive to satisfy the customer.
6 Likes
ChatGPT was excellent in, like, September; but after November, it appeared a lot worse. OpenAI is obviously downgrading the model intentionally, a clear capitalism tactic, where lesser quality or quantity is maintained at the same price, attempting to make consumers not perceive these changes. This strategy is not only unethical but also a blatant exploitation of user trust. It reflects a profound disregard for the principles of technological advancement and customer satisfaction. Such actions by OpenAI severely undermine the credibility of the AI industry. I hope that in 2024, Gemini isn’t a sham. It seems Google is on the right track, assuming all their promises hold true.
2 Likes
I don’t think OpenAI is deliberately trying to upset their users. That’d make for a crappy product. I think we are seeing the effects of some finetuning to better steer the model. The brevity thing is annoying as hell but I think they are working on it based on some tweets that I saw.
Here’s hoping it gets addressed.
2 Likes
Thank you for your consideration. Btw, question behind question anyone leave “paper clip” after "s"oup in Nov 
1 Like
I propose the term: Technological Shrinkflation!
1 Like
They’re certainly not upsetting their users intentionally, though, I suspect without concrete evidence they are taking cost cutting measures with GPT 4 (moving closer to something akin to how 3.5 works)
Maybe I’m wrong and this is just something to do with one of the models they layer on top of GPT like a new input filter that’s oversimplifying our prompts to save tokens? If so “No changes to the model on the 11th” would not technically be a lie, but then its new behavior would be explainable.
Maybe it’s nothing more than an accidental flip of a switch in some random file on their servers that’s making it act goofy??? A dev easily could have accidentally dropped some random characters in the wrong part of a config file that’s not technically part of the model which then caused an undetected error.
They’re clearly working on it since they’ve at least acknowledged the problem which gives me hope but I very much would like to know why this happened once they solve it. I do hope they’re communicative about that when it is resolved.
1 Like
Then why not letting us roll back to the experience we had 3 or 5 months ago, when custom expressions were made available?
Why not give us the choice of our preferred GPT4 model/experience if it’s not about money?
They currently have almost no competition, they can “tweak” on their end as they please, or give users the option of choosing most recent “beta” model, but as you know, that’s not the case. This is deliberate.
I guess they’re trying to find the “sweetspot” where most end users won’t complain, while saving the most money.
In the end, the most hardcore users will be penalized, those who need very precise responses, who use complex instructions/custom syntax/etc. in the hope of communicating most effectively. If the model keeps ignoring them, hallucinating noisy responses, summarizing and outputing "implementation placeholder " comments, how can we trust what it spits out at us?
The fact is, they won’t let their paying users go back, they’re playing behind our backs, they don’t give a damn about informing us about what is happening, they’re being clearly uncooperative, prove otherwise.
1 Like
I understand this perspective, but, I’ve been a power user now since 2.0 and have used GPT to build whole websites and apps.
I can still do that now, but it takes so much effort repairing GPTs mistakes that I’m almost doing more work than if I’d just done it myself. The only reason I don’t is because I despise the format of learning that most people seem to clique with, so I actually get more work done and learn coding better by figuring out what GPT is doing wrong!
GPT4 is still better for me than no GPT4 but only because I’m learning from its mistakes and it still generates good pictures in an intuitive way – I also like my Custom GPTs. None of those features are reaaalllyyy isolated to GPT4, however, and I prefer GPT still because most of the other services that offer those features either use 3.5 or a proprietary model I don’t typically mesh with as well.
However, my preference stems from familiarity, and eventually if the quality doesn’t improve that familiarity won’t be enough to justify locking myself to a marginally useful service.
Am I going to swap from Coke to Pepsi over 1 bad shipment? No, but, if the shipments are bad for months I’m going to resent the fact that the Coke I enjoyed no longer exists and swap over to Pepsi because it’s closer to the old product I was buying.
2 Likes