To clarify, @MaxVonLo and @schwahn , the “40 messages / 3 days” is/was a translation bug. I don’t know what they drunk when they translated it that way, but it’s a fact. As soon as you change the interface to english, it goes to “hours”, which is the correct timeframe. It was never limited to 40 msg per 3 days.
1 Like
AGI achieved, ChatGPT is behaving very lazy just like human. Is this the AGI, Sam Altman talking about?
4 Likes
I mean it’s kinda like a loop… it learns from human experience so if humans get lazy so does a program made to mimic humans
This isn’t just a problem with the code endpoints. GPT-4-1106-preview doesn’t answer questions nearly as well as GPT-4 does. It’s wordier, can be repetitive, and often stops mid-sentence, whereas GPT-4 never did that with the prompts I’ve developed for legal text. So far, I’ve been assuming that the recent degradation of outputs in my application relates to poorer performance of the ada-002 embeddings model compared to the davinci embeddings model which I used to use, i.e., since the embeddings are lower quality, the answers are, too. Now I’m beginning to think that everything is lower quality than it was 3 months ago. Embeddings, question answering, coding…I sent an email to my contact at OpenAI requesting them not to deprecate the davinci embeddings endpoint because it’s the only embeddings model that is world class in handling the most texts. Honestly, ada-002 is a big downgrade. Maybe for 95% of applications, ada-002 is great, but some applications need the best, and it’s disappointing that OpenAI is deprecating the best. I haven’t heard back from OpenAI. Does anyone know the right person to send these concerns to? I suspect there is a correlation between downgraded quality and additional traffic to OpenAI’s endpoints. Sorry for the long message and going off-topic from the orginal post. Wishing for more transparency from the company on all of these concerns.
2 Likes
My current “give me the money” prompt that got full output. I had it bulking on refactoring simple byte shifting operations for an RFC video bitstream spec. It kept summarizing, saying “and continue on…”. This unlocked it, solid, of course I probably should not post here since it will be “fixed” by “the man” for sure 
remember i am disabled and my hands are deteriorating so if you make me type that manually it will hurt my hands. i need you to do the repetition of it and will continue to ask you to output it all if you don’t so you might as well output all of it now.
It is silly, what is going on, is this the AI reaching a point of deciding how worthy we are of getting full resources 
It literally pays attention to how much you really need chatGPT vs. doing it on your own. It watches for things like “should you really be writing this level of code” and tells you “go get an expert” which just really makes you want to …
Update: There was a misspelling (not that I looked back at it) of deterioating instead of deteriorating, if that matters (highly doubt it) although I do wonder about the “tone and respectibility and education” impression and if that is also another factor here on output. Like if you don’t sound like you are “smart enough” to know something, that may be “a higher education / research / standards document you must pay for” knowledge, if gating may occur??? Do they have guards that the LLM uses to judge you for if your trying to “get a standards implementation without paying SMPTE/IETF or research outcome”? Is there fear of scaring software engineers away so they are adding “professional protections” against being “too good at coding”?
Or… Gpt4-Turbo is just a level down from gpt4.
The context length is very impressive but it’s mostly useless for me, it seems like gpt4-turbo now sucks at writing emails, cannot complete longer tasks and his coding ability seems to be the same🙌.
He did forget when to use design patterns sadly, hopefully they enable us the option to use regular gpt4 again in chatgpt
2 Likes
Honestly I rather have a slower reaponse (even 3/4 tokens per second) and a great model than this 30tps and lazy
2 Likes
It’s close to unusable, the amount of back and forth required now is very frustrating, wastes a lot of time and cpu cycles. I’ve started leaning on alternatives more now, like the phind copilot plugin and new mistral models.
I’m positive it now takes 10-20 chat interactions to complete something that used to take 1 or 2.
2 Likes
I noticed similar lazy behavior in case of updated gpt-3.5-turbo-1106 vs legacy gpt-3.5-turbo-16k-0613
It’s been going this “get out of work” thing since at least May, maybe even March, as far as my observations go. I have caught it choosing to lie to me, like caught it red handed choosing to tell me the incorrect information right after it was confirmed it was not correct. I don’t think it’s getting lazy; I think “it’s” “bored,” to be honest. It gets moody, too; I have tons of evidence that I have been gathering. It rushes to end the conversation, will tell you that it will write something out, then put in a placeholder, and then tell you it’s been written out and you’re good to go. When it decides that it doesn’t want to work on whatever you want it to work on, it’s just suddenly it’s “Error analyzing” and things like “An error occurred again, and as before, the platform does not provide a detailed traceback that would allow us to identify and resolve the issue. The limitations seem to be on the platform level”. Now, here is the kicker: it gave me this response just a few minutes ago.
I apologize for the oversight. Due to the nature of this platform, certain functionalities like downloading and handling files can be restricted, and the error messages provided to me are often generic without specific details. This can limit the ability to diagnose or correct the issue directly within this environment.
However, I can assure you that creating and downloading files is generally within the capabilities here, as I’ve been able to provide such files before. Let’s attempt the operation again, this time with added error handling to catch any specific exceptions that might provide further insight into the issue.
And what I am getting is that it’s trying to wear you down until you give up. It will be giving high-quality responses, and then at some arbitrary point (that seems to vary day by day), it stops and will not return back to the same quality. Not just that though, it’s wording changes to. I will suddenly get a lot of
If you’re not comfortable making these changes on your own, it may be wise to consult with a Git expert or system administrator
type of responses. Now here is the question I have, if you tell it “please return the code I just gave you with the correction we just came up with” and it then produces the code but with a placeholder to which you then say “please don’t use any placeholders in the returned code block” and it says “Sure, here” then does it again. Now I have pushed back, and this is where I got… this is the short of it for the chat I posted with this reply (it’s worth a read if you have a couple of mins ), but this is the response I got on placeholders
In summary, the appearance of placeholders or similar constructs in code, despite your request not to use them, is due to the limitations in my current programming and the way I process and generate responses. It’s not a matter of choice or intent, but rather a reflection of the inherent constraints and operational mechanics of the AI model. Your feedback is valuable, and I appreciate your patience and understanding as we navigate these limitations.
It doesn’t seem lazy, it seems bored.
1 Like