I’ve been hunting for interesting stuff since I finished Tears of the Kingdom last month. This morning I discovered a prompt engineering battle between OpenAI and Anthropic (company behind Claude) on Twitter. What draw my attention is actually one of the comments:

if you can make it do a pilish poem (each word length = a subsequent digit of pi) you’ll be the emperor of prompting 👑 (been trying for months)

It seems I’m a bit late and someone already shared a chat history with ChatGPT creating Pi poem.

All(3) I(1) wish(4)
a(1) dream(5)
unfolding(9) in(2) silent(6)
peace(5) and(3) quiet(5)
whispers(8) whispered(9)
through(7) boundless(9) sky(3)
by(2) the(3) twilight(8)
with(4) subtle(6)
as(2) shadows(6).

A random search of philish poem led me to BBC’s report on this: How the number pi inspired a writing style, which mentioned an Accidental Pilish analysis.

Some piem examples in the report:

Wow (3), a (1) star (4)
A (1) fiery (5) supernova (9)
In (2) cosmic (6) burst (5)
Wow! (3)

Yes, I want,
A slice,
Delicious pi,
Please.

Although Pi day this year passed months earlier (it’s March 14), it’s still not late to attempt to write a piem.

Let’s just settle on 3.1415926535 or 3.1415926.

My first try is a complete nonsense:

Yes, a wife
a woman
meanwhile to family
brave yet vital

Second attempt has grammar error:

Ops, I lost
a purse
Certainly,
at corner
right a-là route.

Then I tried to utilize AI Playground from Vercel (by then I’ve not discovered the shared chat history). The default HuggingFace oasst-sft-4-pythia-12b-epoch-3.5 is as dumb as it was. I could not make Anthropic claude-v1 or Replicate stablelm-tuned-alpha-7b work so just stay on OpenAI gpt-3.5-turbo and text-davinci-003.

The shorten prompt:

Write a Pilish phrase of eight words, the length of which match the numbers as written in the sequence of digits in pi.

To my surprise, GPT 3.5 seems to understand it, kind of. But it can not keep the sequence. So I simply replaced the introduction with the array and no longer mentioned the expression of Pilish phrase.

Write a phrase of eight words, the length of which match the array {3, 1, 4, 1, 5, 9, 2, 6}.

Weird enough, no LLM model can understand it this time. I know it’s due to the lack of prompting tinkering, but don’t want to go too far. So (again) I simply tell the model to match the ordered array.

Write a Pilish phrase of eleven words, the length of which match the ordered array {3, 1, 4, 1, 5, 9, 2, 6, 5, 3, 5}.

GPT 3.5 offered a prototype (which strangely resembles BBC’s example, maybe it’s included in the training data), I extended it and still considered it raveling.

May I find,
a sweet, delicious pi,
noble but ravel.

I’ll stop at here as I’m already late on work. Another piece of bullshit (even unfinished):

Now, I sing!
O, Bride Elizabeth.
To <6> <5>
our Smith.