As a technical writer who's spent a great deal of time recently editing AI-drafted documentation, this use case is not going to go as well as AI boosters think it is. :)
The problem it has describing itself isn't the lack of a metaphorical mirror, tool use is there and it can grep whatever code or research is written; the problem is that all machine learning is surprisingly slow to update with new info.
Ask ChatGPT to describe itself, you may get valid documentation and API calls, or you may get the API for GPT-3 (not ChatGPT, before that). I have had both happen.
No, it's prone to assuming or falsifying details even when it has the tools at hand that could verify the true details. Even when explicitly instructed to perform a specific tool call that would load the correct information into its context. Sometimes the pull of the training data is too strong and it will just not make the call and output garbage, all the while claiming otherwise.
As a technical writer who's spent a great deal of time recently editing AI-drafted documentation, this use case is not going to go as well as AI boosters think it is. :)
Have you ever seen the back of your head, without a mirror? Without two mirrors, actually?
How can AI accurately describe itself in full?
The problem it has describing itself isn't the lack of a metaphorical mirror, tool use is there and it can grep whatever code or research is written; the problem is that all machine learning is surprisingly slow to update with new info.
Ask ChatGPT to describe itself, you may get valid documentation and API calls, or you may get the API for GPT-3 (not ChatGPT, before that). I have had both happen.
Elephant.
Did it in one word, easy
What's next?
No, it's prone to assuming or falsifying details even when it has the tools at hand that could verify the true details. Even when explicitly instructed to perform a specific tool call that would load the correct information into its context. Sometimes the pull of the training data is too strong and it will just not make the call and output garbage, all the while claiming otherwise.