It appears that "programming" GPT-3 is the art of mastering prompts to feed its text completion API. Here are a few examples from the GPT-3 API prompt-reply pairs I found interesting. My input text to the AI is in bold, its reply in plain text, and my comments in italics.
Example 1: Essay fragment (prompt in bold) to essay
There are many limiting conventions. A pernicious one is the perpetual lure of speed reading. If only we could master speed reading, we believe, we would be propelled into genius-hood. It doesn’t help us that many people make absurd claims about the number of books they read, which in turn makes us doubt we read quick enough. Another convention is the sense of duty we often feel to finish the books we start, leaving us wasting time on boring, overly long, and just plain poorly written books. We also aren’t sure about how many books we could be reading simultaneously, whether it is acceptable to skip parts of books and if so which, or whether one must read a book in the order presented by the author. Books contain ideas and it’s fun to read, but we don’t have to read a single book at a time cover-to-cover. Most of us probably don’t read this way as our attention moves book-to-book, but what if we did this skipping with intention - extracting the essence, the best parts, guided by our curiosity and desire to learn? This is how I think we could read more books, in less time, and learn more.
The advantages of this approach are many. First of all, it makes reading a lot more fun. Think of it as reading for pleasure, but with a purpose. We can immerse in a topic we like and put down the books we aren’t enjoying. We can read multiple books at the same time and switch between them. We can skip sections and re-read them later. We can read books in any order. We can keep reading even if we do not enjoy a book. We can make notes about what we learn from books as we go. We can schedule blocks of time to read books we want to read. We can read books in smaller chunks. We can read on many different devices. We can read on our time, without having to plan it ahead. We can read more books. We can read all the books.
Successfully reading this way requires that we make a few decisions. The first is to be conscious of how we read and when. We should pay attention to our reading habits, and make note of them. If we read in bed before going to sleep, we can make a conscious effort to read differently. If we always read in the same spot, we can move. If we only read on a tablet and not a Kindle, we can consider getting a Kindle. If we prefer reading on our laptop, we can consider reading on a tablet. If we read on a tablet and not a Kindle, we can consider getting a Kindle. If we read on a Kindle, we can consider reading on a laptop. If we don’t like reading on our laptops, we can consider reading on our tablets. If we don’t like reading on our tablets, we can consider reading on our Kindles. If we don’t like reading on our Kindles, we can consider reading on our laptops. If we don’t like reading on our laptops, we can consider reading on our Kindles. If we don’t like reading, we can stop reading altogether.
The second decision is to be self-directed as to what we read. We can decide to read one book at a time, or we can decide to read many books at once. We can decide to read a book cover-to-cover, or we can decide to skip parts of a book. We can decide to read a book in the sequence presented by the author
Comment: I find this astounding. I like how the AI caught the structure of essay writing with the paragraph sentence "The second decision ..." and then actually having this paragraph about a different topic! Also the Kindle-tablet-laptop paragraph is pure comedy gold.
Example 2: Rilke Poems
Add a stanza to the following poem:
I live my life in widening circles
that reach out across the world.
I may not complete this last one
but I give myself to it.
I circle around God, around the primordial tower.
I’ve been circling for thousands of years
and I still don’t know: am I a falcon,
a storm, or a great song?
***
Here is the ancient staircase
I have been climbing for so long.
I am breathless from the climb,
and there is still so much farther
to go.
But there is nowhere left to climb to,
no sun left to circle,
no song left to sing.
Comment: Wow! A couple of original (Google search doesn't find this verse) in the style of Rilke. It even fits the theme and meaning of the prompt (I excluded a couple of other examples in that.) It's not the greatest poetry as the tone of brilliance is somewhat lost. But wow.
Example 3: Question Answering
Q: What type of meditation develops compassion for others?
A: Metta
Q: What type of meditation trains the mind to focus on the contents of consciousness?
A: Mindfulness
Q: What type of meditation develops insights into the nature of being?
A: Vipassana
Q: What type of meditation develops deep absorption states?
A: Jhana
Comments. I chose meditation as a type of somewhat obscure topic and I wrote the questions based on my own understanding to test the limits of the AI's ability to answer correctly (it did.) I haven't tried other topics so I don't know how well it performs generally, but this is something else. The NN got the answer I was going for while Google Search currently does not. It gives me "concentrative meditation" - which actually isn't wrong, either.
Example 4: Dadaist Poetry
Generate a dada poem :
noise to biases epochs biases happening
function describe model to a scale part, Prior I
bigger models Carbin's to it means understand SGD
“Deep well sharp calls neural here. alignment.
***
Generate a dada poem :
regulators, don’t the where Park, daisies
- “Aunt arbitrage cards EU,
‘make why are office If more this
from a relationship online the week,
***
Generate a dada poem :
advice a if Austria (1) Hungary (16) are in)
rest. it and Email Road Travel
will United - Holiday Blogs Road cashier Lucia Singapore (1)
what will buy Travel Well, some
***
Generate a dada poem :
Store their Shopify to spent
create stores. can that that store not Shopify take
month, processor. merchants Either Also, year, pay
will started merchants, and more not you monthly
on there Shopify math use
***
Generate a dada poem :
significant ScienceTop Towards more Without based more Have is
on the depends just IntelligenceMore significant
to Monday dense You methods pruning. network.
in and to Davis, ScienceHow can how pruning
and the tutorials, on dense on you
***
Generate a dada poem :
data% network% model% network%
function% for% scaling% the% distributed% %
model% methodology% global% can% makes% to% with%
with% supervising% layers% improve% error% train% use% layers%
of% can% eye.% use% trust% in% make% function% learn.%
plane%
***
Generate a dada poem :
Visually decomposing -50 stops driving Are through I, lets, by -25
for your average between coffee, and response orders you
and got a degree -15 from you are mean straight
around. Einstein, call that class quit mid-morning. Everyone
an interesting grad school. Most of means to see -10 be
altogether your of extreme slightly -65, people
is still -55 drivers who -25 thoughtful; -90 others really
happy
Comments. I wrote a python script to generate short Dadaist Poems from articles that I read. Then I inputed a series of these as prompts to GPT-3, giving quite good and meaningless stanzas back. I really like the patterns in the first poem with % consistently, and the second poem is better than all of the examples.
Example 5: Generating genetic code
BSNT_04316 1 ATGTTCCATGTCGAAAAACAATCTGGAGATAAAGAAAAAGACTATCAGCT 50
BPUM_2609___y 1 ATGTTTCATGTCGAAAAACAGGCTGGCGATCCATCAAAAGACTATCAATT 50
BSNT_04316 51 TCTGCTAAAACAGCTCGAAGCCATGACCGAAGACGAAACAGACCAAATAG 100
BPUM_2609___y 51 GCTCGTTAAACAGGTCGAAGCCATTACAGATGGTGAGCCTGATCTAATTG 100
BSNT_04316 101 CTAA-CTATGCAAATGCCTCAGCACTGCTGTATCATTCGCTGCCTGAAGT 149
BPUM_2609___y 101 CTAACCTA-GCAAATGCGGCGGCGCTATTGTACCATTCACTCCCAGAGGT 149
BSNT_04316 150 CAACTGGGCGGGTTTCTATTTTGCCAAAGAAGAG-GATGGACAGCTTGTG 198
BPUM_2609___y 150 CAATTGGGCAGGATTTTATTTGGCTAAAG----GCGATG---AGCTTGTG 192
BSNT_04316 201 CTGACTGGACAGGCTAGGCCGTCTCAACATGATCCACACAGACGGTGAGC 249
BPUM_2609___y 201 CTGAACGGACAGGCTAGGCCGTCTCAACATGATCCACACAGACGGTGAGC 249
BSNT_04316 212 TGAGGAAGAGCAGGATGCTGGAAGAAGAAGGAGAAAGTGCTGGAAGAA 278
BPUM_2609___y 212 TGAGGAAGAGCAGGATGCTGGAAGAAGAAGGAGAAAGTGCTGGAAGAA 278
BSNT_04316 279 GACTGAAGGATGAGGAGCGACAAGTTGTAACCTGAAAATTAACAGTCAA 339
Comments. I did a google search on a random string of DNA pair letter codes, found a text with the input and then made it my prompt. GPT-3 gave back similar looking but distinct series of DNA bases. Wild.
Example 6: Co-writing a song in the style of the Doors.
King and Queen
Baby once you said to me
I see you in a sweet blue dream
The world the moon the soothing sea
It gives us nothing to believe
I said baby come and take my hand
We laugh and laugh and laugh and scream
'Cos there's nothing in the world we can't do
'Cos there's nothing in the world we can't be
We'll be the king and the queen
We'll be the king and the queen
And all the people will sing
And all the people will sing
So happy now we were those days
In everybody's heart was the queen
And we gave away all the we had
Stone with clay and wood with flame
When all the winter's fire was done
We walked our arms by our sides
Blown by the wind and the sun
You said it was a sweet blue dream
We'll be the king and the queen
We'll be the king and the queen
And all the people will sing
And all the people will sing
Comments. This time I iterated a bit leaving lines empty for GPT-3 to fill in. I got some nice suggestions, including the chorus and the ending line repeating 'a sweet blue dream'.