ChatGPT 4 and 4o are having a meltdown

ChatGPT4 / 4o completely unusable for coding for some time now. Making dozens of spelling mistakes on variables, hallucinating functions, ignoring context. 3.5 has no problem with the same prompt. 3.5 is not very creative at all though, and can not keep context very well. 4/4o were working yesterday. Also, crashes have been extremely frequent since onedrive upgrade. I am not clearing my prompts. I have used another browser, and I have cleared the cache.

1 Like

If you have chats you can share that might help us put a spotlight on the issue.

I can’t share chat due to copyrights, here is an example, and some responses after errors.

Original script:
rectTransform.sizeDelta = new Vector2(item.rectTransform.width, item.rectTransform.height);

4o reply:
rectTransform.sizeDelta = new Vector2(item.rectEnt backhouse dress width,itemhprmntgl.height);

4o replies:
Ah, it looks like there was a typo in the OnDisable method’s event subscription handling in my revised version of the ----- script. The method name should be
It looks like there’s a copy-paste error in the OnDisable method that leads to referencing a non-existent method.
My apologies for that confusion—it seems I made an error in referencing a nonexistent ------. That was an oversight in my response. Let’s correct it back to match the names used in the rest of the script.
I apologize for the confusion, and thank you for your patience. It seems there were some typos and errors in the provided script. Let me correct that and provide a more coherent version:
Apologies for the oversight. It seems I used incorrect variable names in the code. Let me correct that and ensure that the script is accurate:

I stopped using it after this and switched to 3.5. 4o seems to working today but I have not given it anything very difficult. 4 seems to be very lazy today.

I would agree. It started its meltdown on Thursday and it has just gotten worse. It cannot keep spelling or paths consistant between sentences. Simple stuff: init.py and then it states init.py. It would mix up paths, even simple as a year. 2024 and it puts 4, when it just stated it correctly before.

I would agree but its take huge time for simple code,

Having the same experience here

Same here, GPT4 became unusable for coding

Same here, GPT4-turbo frequently produce typos on ChatGPT Web these days.

Just now, I even encoutered three typos in one single reponse!

I checked the openAI tokenizer, realizing that the typo-ed word has a single-token-difference compared to the correct word.

So… maybe it’s because openAI increased the “temperature”, such that tokens with very low probabilities are now sampled?

1 Like