All the text in quotes was GPT-3 generated-- with a little help from me (e.g. when it went in the wrong direction-- e.g. ending the speech too early-- I made it go back and try again or clipped out the dumb part and had it continue).

I prompted it with some mission description and said that a speech was prepared for president Nixon that in case the astronauts were stranded.

GPT3 doesn't yet do consistently GREAT output without some guidance, partially as an artefact of the generation procedure. But with a little help it does very well.

The issue is that if you just take the most likely symbol it'll rapidly go into a loop of just copying text or other degenerate behaviour. So instead, everything uses the model by sampling it-- taking less likely choices by chance weighed by the model output. Unfortunately, that means that an unlucky draw will occasional paint it into a corner. If you see that happening you can just go back and try again and you get much better output.

If that is a fair comparison depends on what your application is... if you need to to run unsupervised, it isn't consistently great. If you just need a first draft out of it or some raw ideas to turn an hour writing task into a 5 minute one, it's great for that.

I don't think this kind of manual assistance is much of a cheat either-- a real speech writer also gets exactly this sort of help from others.

[And FWIW, I did this via the GPT3 based mode in the ai dungeon video game. ... I don't have access to the GPT3 API.]