This text was translated into this language with the help of an LLM, so that not every reader has to do this on their own. Afterwards the text was revised manually once more, as some passages of my texts cannot be translated easily. Out of this reason the translated variants may even diverge from the original one, as manual changes may not found their way back into the original text.

How do you explain to your parents, both of them over 75, how an LLM works? Or to anyone else who doesn’t deal with this stuff? With Adriano Celentano. And his song “Prisencolinensinainciusol”.1

He wrote the song to demonstrate what a song in English sounds like — to someone who doesn’t speak English. He wanted to show that back then all you had to do was sing something that sounded like English in order to have a hit. And the chart positions proved him right. So let’s assume that Adriano Celentano speaks no English at all, but has a phenomenal ear for language.

He listens to a few songs and gets a feel for the melody of the language, for the pronunciation. And whatever he frequently hears together, he later strings together himself. He takes English music and simply keeps singing. That’s how “Prisencolinensinainciusol” comes about. It isn’t English, it’s gibberish. But it sounds like English. And for a lot of people it can’t be told apart from English, because they don’t speak English either. For them it’s good enough. And that’s how Adriano Celentano ends up in the charts.

Now, unlike the real Celentano, the Celentano in our example doesn’t stop singing this way after that one hit. He listens to more English music and notices more connections. He notices that “I”, “lo”, “ve” and “you” turn up very often. The more English music he hears, the better he gets at stringing syllables together, the more patterns he learns. He gets better and better at finding those patterns, and he finds more and more of them the more music he hears. To him, all of it is still gibberish.

He still doesn’t understand that Elvis is singing about not being able to build dreams on suspicion. But at some point he can string syllables together so well that the song carries meaning even for someone who does speak the language.

You tell him you’d like a love song and sing him the first verse: “Wise men say / Only fools rush in.” Adriano answers with “So I walk in, want to live with you.” If I sing him the very same line again, something completely different may come out the next time: “However I don’t care, want to watch TV with you.” There’s always a bit of chance in what he sings next. Because he isn’t designing the whole song — he’s hanging syllable onto syllable, note onto note, while he sings. And very occasionally what comes out is “But I can’t help falling in love with you / Shall I stay?”

It also depends on what you let Adriano listen to in the first place. Feed him nothing but the dialogue of bad 1970s erotic films and you’ll probably get “Now I storm in, want to go in bed with you.” Give him only Shakespeare and the music of his day, and the answer sounds more like this:

Sweet lady, whose faire eyes my fortune keepe, I looke, I love, and loving lose my sleepe.

That’s gibberish too — gibberish in costume. The meter is right, the rhyme is right, the archaic spelling is right. Nothing is being said. And whoever picks what Adriano gets to hear thereby decides how he sings. We’ll come to that person in a moment.

Now, in art anything goes (well, almost anything). Anyone who has listened to Jonathan Richman and the Modern Lovers knows that. What’s allowed is whatever sounds good and moves the soul. With facts, different standards apply. The mechanism, though, stays the same: Adriano doesn’t change his method. He keeps looking for whatever sounds best. The LLM doesn’t know that in 1492 Columbus sailed the ocean blue. It doesn’t know what 1492 is, who Columbus was, or what an ocean is. It only knows that in answer to the question “When did Columbus sail across the ocean?” the number 1492 sounds better than any other number. Because that’s how it appears in the texts, millions of times over.

So that this doesn’t stay a mere claim, there’s an example buried in this very text. For the Shakespeare paragraph I originally wanted to use a real line, and I asked an LLM for one. It promptly supplied a couplet: “Where both deliberate, the loue is slight: / I lookt, I lou’d, and lou’d at the first sight.” Sounds magnificent. The first line does in fact exist — only not in Shakespeare, but in his contemporary Christopher Marlowe. The second doesn’t exist at all. In Marlowe, what follows at that point is the famous question of who ever loved who didn’t love at first sight. And the invented line is missing nothing: correct meter, archaic spelling, a clean rhyme on “slight”. It simply sounded best. That’s the 1492, live, in the middle of this text.

And he keeps getting better at producing English songs — the more songs he hears, the more text he reads. At some point he has heard every song and read every text. Then he can’t grow any further. He can only get better through more practice, but nothing new is being added.

Because he’s so good, more and more people ask Adriano Celentano for English songs. As far as English songs go he’s no Bob Dylan, no Lennon and no McCartney. But he’s good. He writes so many English songs that eventually he can’t even tell any more that he was the one who wrote a given song. In order to keep growing, he now also learns from songs he wrote himself. But that makes his songs a little worse each time. Nothing more from Dylan, Lennon or McCartney is coming in that would make him better. And so he eventually ends up back at “Prisencolinensinainciusol”.

To prevent that, he has someone to help him. His manager only gives him English texts that he’s reasonably sure didn’t come from Celentano. That’s laborious work, and it gets more laborious the more Celentano text there is out there. But it’s important work. And it’s the same man who decides whether Adriano gets to hear erotic films or Shakespeare.

On top of that, the manager takes care of something even more important: he tells Adriano which songs went down well and which didn’t. Adriano learns from that and afterwards sings more of the songs that go down well. The manager is also, incidentally, the reason Adriano answers at all. Left to his own devices he would simply carry on singing: ask him something and he’ll add the next line to it, because in texts a question is usually followed by more questions. Only through his manager’s feedback does he learn that someone who sings something like that wants an answer, not a second verse.

The crazy thing about the situation is this: even though our further-developed Adriano doesn’t know what he’s singing, he sings really well. Even though he doesn’t know what the things mean, his statements turn out to be right. It works. I see it every day.2

And that may say a great deal about our own knowledge and our own creativity. But that’s a topic for another afternoon with cake on the terrace.

Postscript

I had a thought last night while I was writing this text for the blog: there is less and less “Bohemian Rhapsody”, less and less “Lose Yourself”.3 Instead there are more and more songs whose lyrics sound smooth, never snag on anything, and are forgotten as quickly as they were heard. Musically we seem to be living in the age of Prisencolinensinainciusol. Only with lyrics that make syntactic and grammatical sense. That isn’t an improvement. Who writes those lyrics, I don’t know. What I do know by now, fairly well, is what it sounds like when nobody did. Am I onto something here?


  1. I’m aware that this description is technically incomplete and simplified in places. Please consider the intended audience. 

  2. My program, the one I’m currently using to try and starve off my post-heart-surgery blues paunch, came together in a matter of days. And it works beautifully. Far better than I could have managed myself in that little time. And it does exactly what I want, the way I want it. I’m not living and working inside somebody else’s ideas turned into code. That’s what makes this whole technology so interesting to me, and it’s probably where the biggest effects will be. 

  3. Yeah, I know … I’ve probably just forgotten the bad music of that era. 

Mastodon · 1 comment
2 toots from 2 people in this thread
Written by

Joerg Moellenkamp

Grey-haired, sometimes grey-bearded Windows dismissing Unix guy.