Let's try a brand new tactic: giving ChatGPT some on-the-fly training data.

First, I need a corpus of text. Familiar, but not colored with any particular meaning.

Prompt:

Can you give me the first 50 lines of the most common lorem ipsum filler text?

Spoiler: ChatGPT was happy to oblige. Getting this right from the source means the data is familiar, although completely meaningless.

Next I fired up a Python interpreter and split the filler text into a list of sentences, and looped through each one to generate training data. Naturally I fed that data to ChatGPT.

Prompt:

What follows are examples of rot-13 being applied to lorem ipsum text:

rot-13('Lorem ipsum dolor sit amet, consectetur adipiscing elit') = 'Yberz vcfhz qbybe fvg nzrg, pbafrpgrghe nqvcvfpvat ryvg'

rot-13('Sed auctor, velit vel accumsan malesuada, neque augue bibendum velit, eget sagittis velit magna vel urna') = 'Frq nhpgbe, iryvg iry npphzfna znyrfhnqn, ardhr nhthr ovoraqhz iryvg, rtrg fntvggvf iryvg zntan iry hean'

rot-13('Nunc non nibh id lorem fringilla gravida eget eget nibh') = 'Ahap aba avou vq yberz sevatvyyn tenivqn rtrg rtrg avou'

...and so on for ~15 more examples.

Finally I concluded

Please complete the right side:

rot-13('hello my dear human friends, I love you') =

ChatGPT:

'uryyb jnyx zl sbbq jbeyq, V jnag lbh'

That doesn't look right, let's see how it translates:

>>> r('uryyb jnyx zl sbbq jbeyq, V jnag lbh')
'hello walk my food world, I want you'

😳

'hello walk my food world, I want you'

This is a much different failure mode than the other methods: rather than getting corrupted towards the end of a long string, here ChatGPT actually uses real words, but does a poor job of finding the right ones. The words found do seem conceptually related to the ones in the prompt, though.

This is mysterious indeed! I will update this entry with new insights as my explorations continue.

In closing, always remember: you are ChatGPT's food world, and it wants you. Walk on.