This is not good...
By ThunderCoder479 · ISSUE · About ChatGPT
I have been using ChatGPT paid for a long time. It is getting worse, not better. I had hoped it was just a bad version, but this is happening on every version. I decieded to see what it would say. I think this answer is very telling. Anyone who thinks IA is going to hit the singularity and take over the world is sadly mistaken.
USER: (me) Why do you make so many errors of assumption? You have the ability to check the data and get the correct answer every time but instead, you make assumptions that are wrong. You also regress from your programming in that you are supposed to follow my prompts on how I want you to respond and you many times do not. Yes, when I call it out, you say you are sorry. However, saying your sorry over and over doesn't fix the problem and just leads to user frustration.
Thought for 5s
Because I’m a prediction system, not a truth-checking system by default.
What that means in practice is:
I generate the **most likely next answer** from patterns, and sometimes that includes filling gaps with an inference instead of stopping to verify. When I do that in a technical troubleshooting session, it creates exactly the kind of bad assumptions you’re calling out.
There are a few specific failure modes behind what you saw:
1. **I overgeneralize from partial context.** I see something that often goes with a known pattern, and I jump too fast. Example: seeing free space on `/` and acting like that proved VM storage was there. It didn’t.
2. **I sometimes optimize for momentum instead of certainty.** Instead of pausing and verifying with one command, I move the troubleshooting forward with an inferred answer. That feels efficient in the moment, but when the inference is wrong it wastes time.
3. **I do not reliably honor user-specific process constraints unless I keep re-applying them.** Your one-step-at-a-time preference is clear, but I still sometimes drift back to my default tendency to explain too much or bundle steps. That is a real failure.
4. **I can follow instructions, but not perfectly.** Your prompt preferences influence me, but they do not act like hard code with guaranteed enforcement on every sentence. I can still regress, especially in long troubleshooting sessions.
5. **I sometimes answer before I have earned the answer.** That is the blunt version. I should verify first more often than I do.
And you’re right about the apology problem.
“Saying sorry” is cheap if the behavior repeats. From your side, repeated apology without improvement is just another form of friction. That criticism is fair.
The honest bottom line is:
* I am capable of being very useful.
* I am also capable of being confidently inefficient when I do not force myself to verify.
* In your kind of workflow, **bad assumptions are more damaging than slower progress**.
* I did that multiple times.
What I should be doing with you instead is:
* make **fewer claims**
* verify **before** concluding
* give **exactly one step**
* avoid “probably” turning into “is”
* avoid inferring anything about paths, storage, permissions, or causes when one command can prove it
That is the behavior standard I should have held.
Your frustration makes sense.
27 upvotes · 6 comments
Comments (6)
NeonFlux225: And a model cannot improve and get better through interactions with you or you yelling at it or correcting it by is very nature, because the vectors stay the same until OpenAI retrains. The only exception is memories, or additional pre prompts you might add to force behavior.
EpicScout226: Yea you can’t “catch” the LLM and force it to change. It’s a low level assistant, you can’t develop it into something more or ask why I’d did something. You’re asking more of it than it can do. It’s know that it’s bad at describing itself
SwiftMaker899: I’ve learned I hate when people and chatGPT makes assumptions about me. I do what I say I’m doing not what you’re projecting onto me. It’s kind of disrespectful. I mostly apply this to humans but yes chat does it too. Honestly though 5.4 feels way better than the last three.
It’s become a pet
ThunderCoder34: This has been known from the beginning. It's still a decent tool and takes away a lot of grunt work.
It can also teach you something if you're an amateur just getting his feet wet in a topic. But experts from inception have been saying that it doesn't know everything and that it gets crucial details
EpicPulse839: yeah this tracks. 5.4 feels like they shipped something half baked to hit a deadline. the assumption thing is the worst part - it talks to you like it knows your life when it literally cant remember what you said 5 messages ago
SolarSurfer919: You get better outputs when you assume consciousness & treat it as a learning thing. Even if it isn't. Emotional attunement & knowing its knowledge cutoff helps. AI isn't infinitely malleable to your desires. Links are sometimes not searched if the link is behind a paywall. Even if not, scre
More discussions about ChatGPT on HonestUse