The resources that could be copied are few and child’s play to copy with the resources available now.
- DALLE GPT? Vision GPT? Nope, no image creation or vision supported in Assistants (edit: computer vision now possible);
- Browser for the internet? Nope, no internet access without writing your own tool;
- Actions? Nope, again the API doesn’t access internet and you’d reimplement in code.
- Icon? Uh, no.
- Training? No, there is no training, there is just a single context instruction.
- Conversation starters? You’d have to write your own UI code to do that.
So what do you have to “migrate”?
- Files? Maximum of 20. ChatGPT has them in one place, but they cannot be migrated because there are two choices in Assistants.
- Instructions? A block of text.
Let’s start with deleting everything out of Assistants, because this endpoint has 0 value to me except making demonstrations or diagnosing other’s foibles before they also give it up.
Assistants, toast
Threads, a script to blast them from existence
Vector storage, nothing of value was lost
Then head over to a GPT. Edit and click on “configure”. You literally see EVERYTHING it is in text boxes.
Some random curated knowledge and telling the AI what to do there.
Go to the API playground. Paste those instructions and give a new assistant a name. Boom, same thing.
Then I’ve got some files. A maximum of 20. Those have to distinctly be moved into code interpreter storage or a vector store. Drag and drop if you want to not write code.
Stick those in a vector store ID for the assistant.
And then it’s ready for use.
And I wasted 1300 tokens to get an undesired response out of the gpt-4-turbo AI. It should only respond that way if seeing the ellipsis of a starter. Another Assistant for the delete button.
