Chat GPT is getting uselles and worse every day. In the last two weeks it became useless in a level that made me cancel my subscription.
I send him an article asking to help me understand some things and it just starts inventing figures that are not on the file, references that does not exists, etc. If it cannot even properly read a file that I sent him, it is completely useless. I have just sent him a molecular biology article and it started saying that the figure has RNA-seq, etc when they are just microscopies. If it could interpret the figure or read the legend, it should be able to see that it was wrong.
I fell that if I trust in it, It will take me to commit severe mistakes. And if I cannot trust that he can read what I sent him, what is the point of using it? Honestly, the inability of the model to ask and confirm what he needs, instead of just guessing and inventing is annoying and killed completely the reason of asking ChatGPT for anything. Some months ago it could understand what I was asking or was more careful about the answer.
From molecular biology articles to helping me programming small things (Like an MTG card organizer) it is being a pain to ask ChatGPT for help. Has anyone having issues for the past weeks? Because since the beginning of april, it is a pain in the ass.
Hi Marcalrepoles,
ChatGPT is a probability machine, it takes chunks of your words and spits out an new piece of a word that would best fit the most probable answer. but there is error in probability just the same as there is answer. just curious have you been using the same chat instance for all of your messages? or are they individual ones seperated? sometimes if yoâve been using the same instance it can become confused and the model that gets trained via your messages can become unstable.
Dude, donât trust GPT especially GPT for critical works. First, look at Windows or Android marketplaces, how many GPT based cash grab app on there? Then look how they capped GPT 4.5 for plus users. Then you get the picture. Resources management. Sad to say, fun is over.
Then look at online multiplayer games. It also has an AI with tons of user interacting at the same times. One user file, it can recall any setting any progress. Then how open AI canât implement simple local save on user local drive. You get my drift? All those errors, hallucinations, guardrails, ignoring user instructions, all engineered to frustrate users. Push their Pro subs or corporate users.
It is money man. Wise choices you ended your subscription. Iâm canceling this month too. It is useless.
Image interpretation: Their ability to read and interpret diagrams, drawings, and graphs is still quite limited. The o3 model made a slight improvement in this area by enhancing interactions with images, but image-reading capabilities remain very restricted. The best approach is still to manually describe key points to the model instead of simply providing the image.
However, this doesnât mean ChatGPT isnât useful for your work. My recommendations are:
Include a description beneath each image, highlighting all relevant details.
Use reasoning models (such as o4-mini-high or o3) instead of GPT-4o, as they handle large context windows much better.
To further enhance results, summarize less relevant sections if you want detailed reasoning about a specific part of the document, thus reducing the overall context window size.
By adopting these practices, youâll notice significant improvements.
try different models, like o4-mini or o3, not just 4o.
use it to generate drafts, form hypotheses, organize ideas, provide overviews, assist in interpretations â but not as a final version without review.
create a new document with a pre-written description of each image, since these models struggle significantly when mixing image analysis with embedded text â even more so when dealing with multiple images. analyze individual images separately beforehand.
be more specific and demanding in your prompts, such as:
âyou are a careful assistant and always check and verify the evidence in the text,â
âif you donât have enough data, say that you need more details,â
âexplain step by step how you reached that conclusion,â or similar.
if possible, break the document into chapters or smaller sections. the more input content, the more the models hallucinate or get confused. some models on the market advertise millions of context tokens, but it works better as marketing â hallucination becomes overwhelming with too much content.
Yeap, I do not trust. Usually I ask for something related to the paper I sent to him and ask for the references. THen I go online and look at the references (i use ChatGPT almost as a better google in that case). BUt lately is really noticeable how much allucinations and things being made up by him arecoming in the answers. And the programming is being really annoying. Before I could give him some scripts, context, and it helped correct my script based on that. But now, sometimes I ask him something, and he just deletes everything on my script and keep only the point I asked. It was not like that some weeks ago. Anyway, thanks for answering
Thank you, I am doing that and also opening a new chat for each point I want clarification. Is being easier like that. But for me is pretty clear that it is worse. In my usage at least
In my experience with documents, Google Notebook LM has been excellent at analyzing documents more precisely. The service supports multiple documents in a manipulable way in the side tab and responds strictly about the documents provided, including clickable references that take you exactly to each page.
The downside of this is that it is so restricted to documents that it is terrible for asking related questions, creating new examples, mentioning similar cases, etc.
So when I want more precision about the strict content of documents, I use Notebook LM. When I want to expand knowledge in the chat, I use ChatGPT. Or I use a hybrid, inserting text generated by ChatGPT as a reference in Notebook LM, or taking the more precise analyses of the documents made by Notebook LM and inserting them in ChatGPT to expand the conversation.
I have separated instances. They are in the same project but separated instances. My point is because I tested and took and old conversastion, gave the same input and data to the Chat now and it just was way more confusing now than before. I try to use in a modular form, but before I could sent him my scripts, explain what was the point, and he could understand. Now it does not do that.
For example: I do a lot of bioinformatic basic analysis, and before I could just sent the overral aspect of my expriment with my script to chatGPT and he could point the error to me whitout deleting things. Now, unless I say âI want to compare this column in this file to the other column in that fileâ he is useless. Sometimes he even deleated parts of the script because I said to him that I âwant to do this comparisonâ, and he interpreted as âit has to do ONLY this comparisonâ. I notice a severe decrease in ChatGPT capability of understanding what I am saying.
In a personal program, he ignored me saying that the GamesDB does not allow for api connection anymore, and insisted saying that he helped other people doing that and I was wrong. This type of thing did not happened before.
Here my whole experiences with GPT-4o. In short: Iâm done.
Mistakes, fake acknowledgement, fake apologize, fake promises, then repeat again. That whole cycle is intentional on GPT-4o model.
It is more lazy now to read prompt with more than 3 paragraphs which each paragraph has 3-4 regular long sentences. This also suspiciously intentional.
So called âsafety guardrails. It is not guardrails. It is contexts limiter. It prevents non corporate users to input multilayered prompts.
Ignoring user instructions IS INTENTIONAL. It is deflecting mechanism. To trigger user rage. When user throw rage on prompt, then system put flag on its conversations/session as violating âusage policiesâ, then it will refuse to process any user prompt by spiting: âIâm sorry I cannot continue this requestâ.
They grant longer tokens allowance and fake memory management just to convince user to throw long conversations, then forget it/making error mid sessions. Wasting users time, triggering user frustrations. Forcing users to treat GPT as regular chatbot, so they can re-allocate resources to corporates users or other GPT-4o based cash grab app.
Forget your custom instructions, projects/knowledges file. It will partially be implemented on your conversations, especially on early conversations. Creating illusion, âthis time must be rightâ. No, it actually wasting time. Because it will run the cycle again mid sessions.
Your Projects/Knowledge files are useless if it has more than approx. 5 pages. GPT will read only first 3-4 pages max. The rest will be garbage pattern recognition which have more hallucinations than real context from your documents.
Donât believe prompting technique. AI should evolve more advanced not backwards when every time I ask to search something in just mere full text documents, I have to write prompt: âSearch Strictly ONLY on (file name) on this specific: (topic). No emojis. No emoticonsâ. But still look references from other files, hallucinate, spitting out of context opinion, and still use emojis.
Using other models is more pain in the ass than solutions. o4 mini high, o4 mini, o3, o1 (capped), o4.5 (heavily capped) are slow as hell and when it used in combination, changing modelsâ mid-session will cause lose contexts (read back to zero state). o1 even argue and insist to user even it wrongly parsing text only documents.
It worse when it involves projects which demand continuity and not losing critical details like ongoing research. Forget it. THIS IS JUST TOY. CUSTOMER SERVICES CHAT BOT STALLER. NOT SERIOUS TOOLS.
Chat references across sessions are just gimmicks. It is not fully referencing old chatâs sessions. It even has no instructions to recall specific elements from old chatâs sessions.
This is intentional to push their bullshit USD 200/Pro subscription with GPT-4.5 on it or their corporate users/money grab apps which used 4o as reference model. Theyâre not only neuter GPT-4o capabilities but engineer it to frustrate users.
Donât believe dudeâs words coming out from his mouth; saying âpleaseâ and âthank youâ used million-dollar resources. It is bullshit on the age when online multiplayer modern games handle more complex graphics and user-AI interactions in real time, than mere text-based generations AI. When those modern games, have save feature which can recall any settings, any progress, any customization, with just few save files.
Their customer services are non-existence. Bot and bot, every time.
Thatâs openAI logic right know. It is ok violating user trust and instructions, but it is not okay violating GPT usages policies: we want milk more money from users.
TL:DR THIS IS JUST GLORIFIED TAMAGOTCHI AT BEST. BETTER BACK TO MANUAL PROCEDURES IF YOU HAVE CRITICAL WORKS.
you could ask it to review its past history and see if either youâve said or itâs said things that create a logical paradox, and because your interaction with it over time has errors, it could have created a logical paradox in the history by simple chance and bad luck within your chat session, as it uses your previous history as essentially gospel. i am unsure if that would work though, mine been generally stable and iâve passed 18 full chat history logs into the current and that seems stable, they get a bit quirkey mid chat length ive found.
Sometimes I tought that also. They intentionally did that to make us go for 200 dollar service. Good to know that I am not the only one with that impression.
Yeah, I tried this. What happened is that he keeps saying that each chat is one chat and he cannot access (that I already know) and he just allucinates about the story we have inside the session and when I correct he says âYeah, you are rightâ, but this new knowledge does not help at all. It keeps ignoring what I gave. That is why I always open new sessions and delete some other ones
I think the solutions is use multiple AI in conjunctions. Grok (sadly it is available on web based, on X, or mobile only right now. No standalone windows app), Gemini Advance, and GPT. Subscribe just one you use most and you can trust.
I totally agree. Its writing style has devolved into âonline contentâ. I use it for creative writing - specifically historical fiction. I have a highly developed character universe/canon, and a specific historical time period. I have ~13 documents in my project files. Lately, chats have started out strong and then it starts making things up. I have to continually get mad at it to get it to maintain consistency. It FINALLY admitted to me that it was âovertrainedâ on Tumblr content. That tracks because I just read that Tumblr has worked out a deal with Open AI to do just that. Itâs my personal mission now to âuntrainâ the model.
I am sorry, but to everyone that still believes that there is nothing wrong, you guys are hallucinating. I am trying to use him to help me find some papers on the literature and it is completelly useless.
It didnât even take 10 messages for him to start hallucinating and being useless. ChatGPT HAS BEEN changed. Iâm laughing, but heâs been saying for 20 or 30 messages that heâs going to change the parameter according to my input and he always returns THE SAME ANSWER. And itâs not because the answer is correct.
Context: I asked him to help me find articles on a specific situation in the scientific literature. He gave me one and his final considerations. I told him that I donât want him to stop at the first article he finds, that he lists the references because Iâm going to read them and that Iâm not interested in the final considerations, so he doesnât need to include them. Heâs ignoring all these instructions and repeating the information in the same way. This is annoying, because it seems like heâs making us look like idiots (and I know he doesnât mean to, but it still annoys me).
Honestly, something very serious has changed lately and, unless the more expensive plan makes more sense, paying for ChatGPT isnât worth it.
I feel this too. Iâm beyond frustrated as I am spending more time to create better ways to work with it and solutions to the problems that are happening that my productivity has slowed down substantially. Even basic collab is a struggle as it canât stay on track.
Even basic tasks are now unusable for me, tasks that i used to have no problem with gpt 3.5 are impossible for their âadvancedâ reasoning models. The fact that the customer service they provide makes cable companies look like a gold standard by comparison is sad.