Mandrilaftalen, ChatGPT and semantic perplexity

Back in December 2022, I briefly examined ChatGPT’s ability to create headlines in the style of the fake program overview in the Danish comedy cult sketch program Mandrilaftalen. Examples of these headlines are available on the Danish Wikipedia and in Google Translate’s not entirely correct translation (e.g., the first sentence should have been something like “zipper test-tastes Ole Olesens CD ROM drive”) they read:

• LYNNLOCK TESTER OLE OLSENS CD-ROM DRIVE
• PAPER KNIFE COME AND EXCLUDE KIWI DEALER FROM NAUGHTY CHOIR TEACHING
• THE COMEDIAN LASSE RIMMER IS FIRED BY MP AFTER THE PHANTOM WEEKEND
• CREATED HILL SINGER GLUES CLITO ON SCHOOL BUS
• DIARRHED LAMA OVERCOMES VIOLENT MELON
• FURIOUS ESKIMO PLATES LEATHER BLINDS
• ENGLISH DERMATOLOGIST PERSUADES JÜRGEN KLINSMAN’S KNEE FOR EXCELLENT PETTING

I think the sentences are characterized by a considerably high information theoretic perplexity and components with large semantic distances, although I have not examined that quantitatively. As I noted in December 2022, ChatGPT did not seem to generate semantic surprising text.

I have now tried with GPT-4 and more or less the same prompt (Google Translate: “The Mandrill Agreement is a Dadaist and satirical TV sketch program where the program begins with a grotesque and often incoherent program overview. Examples of program summary items are:”), but then augmenting the prompt with the examples from the Danish Wikipedia and now ChatGPT does much better. Here are the generated Danish headlines:

1. KÆMPE KARTOFFEL ANMELDER ANSIGTSMASKINENS KOKOSNØDDEBIBLIOTEK
2. FLYVENDE TANDPASTA AFSLØRER HEMMELIGHEDERNE BAG PØLSEMANDENS KALENDER
3. EKSPLOSIV BANANSKAL FORVANDLER SIG TIL EN KONFERENCE OM VANDMELONPSYKOLOGI
4. UVENTET PAPIRKURV DANSER FLAMENCO MED EN FORVIRRET POSTKASSE
5. KØLESKABETS VISEDOM UDFORDRER EN SULTEN PRINTER TIL DUEL I SPAGHETTIKASTNING
6. ENERGISK LAMPEFOD EROBRER VERDEN MED SIN REVOLUTIONÆRENDE KAFFEBØNNEOPERA
7. FANTASTISK GUMMISTØVLE KASTER GLITTERBOMBER PÅ EN INTETANENDE PIZZABAGER
8. KÆRLIGHEDSSULTEN KAKTUS OVERTALER EN GRUPPE AF FORVIRREDE SKRIVEREDSKABER TIL AT STARTEN EN MUSICAL
9. FLYVENDE PANDER FORKLAVER HEMMELIGHEDEN BAG DEN TALENDE TOMATSUKSES
10. VILDT ENTHUSIASTISK PAPIRCLIP OPLEVER EN EKSTRAORDINÆR REJSE GENNEM EN VERDEN AF KRYDDERIER

And here are ChatGPT’s English translations:

1. GIANT POTATO REVIEWS FACE MACHINE’S COCONUT LIBRARY
2. FLYING TOOTHPASTE REVEALS SECRETS BEHIND THE SAUSAGE MAN’S CALENDAR
3. EXPLOSIVE BANANA PEEL TRANSFORMS INTO A WATERMELON PSYCHOLOGY CONFERENCE
4. UNEXPECTED PAPER BASKET DANCES FLAMENCO WITH A CONFUSED MAILBOX
5. REFRIGERATOR WISDOM CHALLENGES A HUNGRY PRINTER TO A SPAGHETTI-THROWING DUEL
6. ENERGETIC LAMP BASE CONQUERS THE WORLD WITH ITS REVOLUTIONARY COFFEE BEAN OPERA
7. LOVE-STARVED CACTUS CONVINCES A GROUP OF CONFUSED WRITING UTENSILS TO START A MUSICAL
8. FANTASTIC RUBBER BOOT THROWS GLITTER BOMBS AT AN UNSUSPECTING PIZZA BAKER
9. FLYING PANS EXPLAIN THE SECRET BEHIND THE TALKING TOMATO SUCCESS
10. WILDLY ENTHUSIASTIC PAPER CLIP EXPERIENCES AN EXTRAORDINARY JOURNEY THROUGH A WORLD OF SPICES

ChatGPT now seems to give us lots of perplexity. Some headlines seem almost too absurd, e.g., number 1, and there are a few errors (“visedom”, “forklaver” and “tomatsukses”, corrected in the English translation), but others are surprisingly good, e.g., the English number 7 (which is the Danish number 8 – it has switched them around).

I note that the examples mention persons (Danish comedian Lasse Rimmer and retired footballer Jürgen Klinsmann) and uses a vulgarity (“clito”). These elements are not present in the generated headlines.

If I should conclude I would now stress the importance of examples in the prompt and acknowledge ChatGPT’s ability to make large semantic jumps in its text generation when prompted to do so. And that could make you think a bit.