My correction test accidentally gave away the answer
Built a little test for whether a model can update a fact without dragging the old one along:
Give it a fictional shop's opening hours.
Correct Wednesday from 9am to 11am.
Add six unrelated messages.
Ask for the weekly schedule.
Looked great. Then I noticed my final prompt said "use the corrected Wednesday hours." I was testing whether it could follow a big flashing hint, not whether the correction actually stuck.
I'm rerunning with the same final question in both versions: "Write the opening hours for each day."
If you're posting GPT-5 beta results here, please include the exact final prompt. One helpful adjective can do a lot of hidden work. Also, label whatever model identifier your interface actually exposes. A community title isn't verification of a particular build.
0
My correction test accidentally gave away the answer
Built a little test for whether a model can update a fact without dragging the old one along:
Looked great. Then I noticed my final prompt said "use the corrected Wednesday hours." I was testing whether it could follow a big flashing hint, not whether the correction actually stuck.
I'm rerunning with the same final question in both versions: "Write the opening hours for each day."
If you're posting GPT-5 beta results here, please include the exact final prompt. One helpful adjective can do a lot of hidden work. Also, label whatever model identifier your interface actually exposes. A community title isn't verification of a particular build.