3 ms·
To use the tool effectively, you can't use Socratic questioning on it, or not always. You have to already know how to code the problem, so that you can spot wh
by kazinator 4y ago
To use the tool effectively, you can't use Socratic questioning on it, or not always.
You have to already know how to code the problem, so that you can spot what is wrong, and, if necessary, tell the thing how to fix it.
The fix is not always going to come from the training data; it needs to scrape it from what you give away in your follow-up question, together with its training data.
I went through an exercise in which I encoded a paragraph from Edgar Allen Poe with a Vigenere cipher. I presented this to GPT4 and asked it to crack it. First I had to persuade it that it's ethical; it's a cipher I made myself and it's not my secret. We worked out a protocol by which it can ask me questions which show that I know the plaintext.
It was a quite a long chat, during which I had to give away many hints.
In the end I basically gave the answer away, and the thing acted like it cracked the key. Then I reproached it and admitted that yes, it didn't crack the key but used my hint to locate the E. A. Poe passage.
Basically, if you know the right answer to something, and GPT4 isn't getting it, it will not get it until you drop enough hints. As you drop the hints, it will produce rhetoric suggesting that it's taking credit for the solution.