I recently switch my project from gpt-4o-mini to gpt-4.1-mini
4.1 is messing up GUUIDs which 4o never did
For example 8dbe6430-8164-473a-abf5-333181908ab1 is altered to 8dbe6430-8164-473a-abf5-333181908abf5 apparently filling the last digits with a portion of another group of digits, even making it one character longer in the process
Interesting observation, although I have seen even 4o struggling at this task. It’s kind of difficult for models when you want them to quote long random substrings (UUIDs) because the underlying model GPT is just a next-token predictor. IMHO, UUIDs are the last thing a model would be able to properly quote because of their randomized nature.
A workaround would be to use placeholders for the UUIDs and replace them just before delivering them to interfaces (that require UUIDs).
Basically i’ve made a robot in a virtual world and i feed the AI the chat from users and a list of nearby objects (names+ IDs) and it decides how to interact with them by giving commands to the avatar. It may sound complex but 4o could handle that. 4.1 is better at giving the commands to the avatar but it has this downside i described here.
To use placeholders i would need another unique isentifiers for the objects. Imagine there are 4 trees and 2 chairs nearby named identically.
Maybe it works better if i use the last group of digits in the id. I’ll let you know if i find a reliable solution.