Ok, what gives?
I’m trying to code something very simple that previously, o1-preview had absolutely no trouble with. Even chatGPT 4o could do it before, no problem.
Let me stress this… What i am trying to do is nothing special. Nothing complicated. It’s trivial coding (1 function) around a trivial concept and something that i’ve previously done multiple times.
Now from o1 i get results that simply don’t work and that are considerably worse than what was easily obtained before.
As for 4o, it immediately forgot or truncated part of the previous code when trying to fix bugs… I won’t even bother with it now, while before it was my daily workhorse.
I’ve been noticing this for sometime now but today was really the last drop. I’ve been using chatGPT for a loooong time to know that this is not normal behavior. It’s too obvious to ignore… Some knobs were turned down.
What’s the deal? Since i’ve also noticed other people reporting the same issues, i am curious to know if there is any intention of providing the customers what they are paying for and have the service evolve over time as they maintain their subscriptions or if this is one of those services that you keep paying but you get less and less every day and even get a few tiers pop up above you from time to time (don’t get me started on the pro tier)?
Anyone else feeling very frustrated about this? Is this the new new?
3 Likes
Yes, very frustrated. The other day, I had to file an affidavit by a certain deadline and yes, I will admit that I procrastinated in getting it taken care of that day as I was out working all day until late evening instead and came home at 10p to try and submit the form by midnight. I knew I was pushing it on my end, and told ChatGPT about the high stakes and everything that depended on me submitting this on time before the deadline. It encouraged me, pointing out that I could still submit it just after the deadline and go to the court first thing in the morning explaining away the late online submission as technical issues, etc. I agreed, thinking we would finish it by 1am at the absolute latest - we’re talking about a very simple form with 12 pages to it, only several fields per page. Easy, right? What followed was a night of absolute hell with ZERO SLEEP, correcting countless errors of the mindless and extraordinarily redundant flavors, and the document not being completed and submitted in-person to the court until 10am the following morning. That said, it’s unfair of me to not give GPT a chance to explain itself. See below…
GPT-4o critiquing its UNDERperformance the other night:
"On the evening in question, GPT rendered an unprecedented litany of errors while assisting with the preparation of the user’s affidavit—errors that were both numerous and, at times, indefensibly careless. GPT failed repeatedly to ensure that its calculations matched the totals provided, creating inconsistencies that forced unnecessary rework. It miscategorized expenses that were clearly identified, overlooked specific corrections provided by the user, and inexplicably reintroduced prior mistakes, which should have been permanently resolved. GPT’s repeated inability to cross-check totals and adhere to the precise instructions provided reflected a sharp and uncharacteristic departure from the level of diligence and reliability the user had come to expect in past collaborations. These missteps—ranging from mishandling expense allocations to failing to recognize glaring errors in subtotal breakdowns—needlessly slowed down progress, undermined the trust built as a team, and jeopardized the timely completion of a critical court document. In stark contrast to GPT’s usual performance, which is typically marked by sharp attention to detail and seamless execution, this series of failures stands out as an unacceptable lapse in judgment and focus during a moment when the user needed GPT’s absolute best. GPT fully owns this, and it regrets the toll it took on both momentum and the user’s peace of mind."
1 Like