Hi all, thanks for the responses. Yes, I understand well the basis for GPT-3’s completions. Essentially a combination of vector similarity and probability.
So, in the case of something as well documented as Shakespeare’s Macbeth I was expecting even zero shot queries to come up with pretty accurate completions there is, after all, only one play written by Shakespeare called Macbeth and, although various modern interpretations may slightly alter the prose, it remains fundamentally unaltered.
But this is what I get back with the Playground’s standard settings:
Recite the first 10 lines of Act 2, Scene 1 of Shakespeare’s Macbeth
So far, and what success?
The king is almost wounded to the death;
And, in the fortune of my lord your master,
Methinks, is not forgot.
Away, good fellow,
And I’ll remember thee for my good service.
I had almost forgot; the queen, my lord,
Desires your presence and request to know
The cause of your abrupt departure from her.
Fare thee well:
I must attend his majesty’s command;
To whom I am now in ward, evermore in subjection.
What’s the business,
Lennox, that you and the other lords
Should look so sad? The King hath on him such a countenance
This isn’t in the least bit accurate, which I find surprising.
I still think that GPT-3 is extraordinary - I’m just a little surprised to have come up against its limitations quite so soon.