I've seen some questions about how I could produce the texts I shared earlier by prompting GPT-3, and whether GPT-3 is capable of producing such a convincing output at all, so here's a thread to clarify a few points.
My methodology was the following. Since I don't yet have access to the API, I used @AiDungeon with the "Dragon" model (which is GPT-3) and a custom prompt. AFAIK, AID allows for arbitrarily large prompts, but as @MaCroPhilosophy pointed out these must be automatically truncated.
I use the schema outlined below for the prompt. As I mentioned, given the length of that prompt (way above the 2048 BPEs context window described in the GPT-3 paper), I assume that this prompt was truncated so only the end was passed to the model. 
From there, the model generates 1-3 sentences at a time. I kept pressing "enter" to prompt the model to continue generating new sequences of text. I repeated the process until it stopped either by signing off or prompting an error message from AID saying that it was stumped.
Importantly, I did "reroll" a few time at the beginning of new paragraphs when I wanted to see in which directions it would go, and in some cases I decided to go with GPT-3's second or third proposition.
I repeated this whole process (from the prompt to the signing off/error message) several times, I think 4 or 5 times. I only kept two of them that I have already shared. The other ones were also interesting but not as impressive.
So there's cherry-picking at two levels: within each complete response, some sentences were not GPT-3's first output (although they were still written by GPT-3!); and I shared only the two most interesting complete responses I obtained through this process.
Some people have suggested that this is way too coherent to have been written by GPT-3. If you still think that way after this explanation, think again: GPT-3 is very impressive, especially if you prompt it in the right way. You can only figure this out by trial & error.
As @gwern put it, we don't know yet what GPT-3 is *not* capable of, but only what it *is* capable of, because there is no guarantee that a better prompt could not improve results on a given task.
I hope this satisfies sceptics. The extent of GPT-3's abilities is truly remarkable. But I also don't want to feed unwarranted hype: as I said from the outset, my results were cherry-picked, and one shouldn't underestimate how far GPT-3 remains from to AGI.
What we need in the future is systematic testing of GPT-3's abilities and limitations, which is by no means a trivial endeavour - and not what I intended to do here.
