Less Instruction Following in Recent Update?


After a recent update, my GPT has not always consistently followed my instructions. I ask GPT to give three key points, but GPT provides 1 or 2 key points or nothing most of the time.

In the previous version, it almost followed my instructions, only in small cases when it did not follow instructions. I already use “must” in the prompt.

Do any of you have the same issue or have any suggestions?


I confirm that my GPT has the same problem. Although for me it’s quire more severe, in the sense that it ignores the instructions 95% of the time. In my case I thinks it’s due to the long knowledge files and long custom instructions. The super limited context window completely defeats the purpose of GPTs. (Note: I say 95% because I actually calculated this value hundreds of tests).

As another example of how unusable GPTs have become, today I asked Wolfram GPT (arguably the most sophisticated GPT out there) to use Mathematica to calculate the determinant of a matrix. This is something that could be done super easily a few months back. But this time the GPT sent the command to the Wolfram API with a different matrix from the one I had inputed. It completely modified and made up my request.

Unfortunately you are not alone. The quality of GPT responses has become very unreliable and unpredictable over the last 2-3 months. We’ve wasted a lot of time trying to improve our instructions and data, only to find that it works great for a few hours, then it becomes stupid for a couple of hours, then it works great again, and then back to useless.
I think this is because of lack of compute capacity in the back end, but OpenAI does not publish their utilization % or other metrics that may help us understand what’s going on.

Frankly, I’m quite disappointed, and I wonder how many more months (years?) it will take for this platform to become reliable enough to be usable in the real world, because at this time it is not.

1 Like

Finding this even more so with the update to 4o. Custom GPTs not following clear step-by-step instructions that worked with GPT-4.

It’s a shame custom GPTs can’t specify what model to use.