I’m a translator and I’ve been using the workbench/playground to transcribe Japanese patents documents since it’s very good at it compared to OCR (some of these documents are kind of old).
Today it’s refusing to do it flat out. It’s telling me “I can’t assist you with that” saying that it’s not allowed to transcribe proprietary information. These are patents, they are public information, but explaining that doesn’t help.
This is a little infuriating. I’ve tried a bunch of workaround but to be quite honest I feel like an idiot paying for a machine to do a task and then having to either beg it to do it or try to outwit it.
This is a major problem.
Any examples you can share with us?
What model are you using? Settings?
gpt-4o
max tokens 16383
everything else I leave alone since I don’t really understand it.
The problem is these are the exact same settings I’ve been using for weeks (except the max tokens, which is now 4x bigger).
It seems like they tweaked something internally relating to “proprietary information” and forgot to tell it that published patent applications are public by design.
1 Like
Yeah, if it was working then stopped, it’s likely something on the back-end you won’t be able to prompt around.
Do you have an image I can try on my end?
Might also try the specific gpt-4o-2024-05-13 model? I think recently gpt-4o got pointed to gpt-4o-2024-08-16… The older one will eventually be deprecated, but might help now?