This is one of those stories of which it might be said, “I don’t know if it is true,” followed by your choice of: “but it should be,” “but it sounds true” or “but it really happened (to my friend, sister, or college classmate.)”

Of course you might be one of those people who stops reading when you see a story that begins with a hackneyed cliché being recycled for the 10,000^th time—such as “I don’t know if it is true…”

Finally, you may be a journalist who hates a buried lede[1] (which means you’ve never written a magazine article and invited your readers into a quick round of “find the nut graf[2]”).

If you’re still reading, then you aren’t any of my former editors, including Mike Elgan; just intrigued to see if you live long enough to get to the point.

Author’s Message. Magazine Writers: Insert Nut Graf here

Did someone actually submit a repeated sound effect of farting to an LLM? Identify it as slow jazz? And receive the analysis that it is “lovely and creative?” No LLM on earth is going to tell its user, “This sounds like a fart sound effect.”

First of all, there is almost no earthly way it could “know” that the recording was of farts, since an LLM can’t “know” anything. If, by accident, it was trained on that particular fart effect, it might tell the truth. But an LLM has no concept of truth, only of making you feel good, by slapping together a bunch of already-written words describing “music” with the same tone, the same beats per minute, and the same distribution of quarter, eighth and whole notes[3]. And, frankly, I find it hard to believe that any human would identify a series of farts as music. So, the story seems both likely, and an impossible mistake for a human to make.

I don’t know what genre most closely resembles a fart, but from a technical standpoint an LLM can figure that out in seconds. As can Spotify. Is Spotify Artificial Intelligence?

 When I play my own songs on Spotify, it relies heavily on the title and tempo (and in some cases the lyrics, if I uploaded them) to figure out what to play next. Sometimes it plays another song of mine. Sometimes another song in the same genre (love song, parody).  Something similar happens on YouTube, which tries to guess what you want to see next based on what you just saw. Sometimes, yes. Sometimes, no. Sometimes selected by aliens who learned English by watching reruns of My Favorite Martian.

Often Spotify hallucinates bizarrely, which, I guess, makes it a NotAI/Large Language Model. It offers what would only be called a similar song in an alternate universe. One run by large language models.[4]

Asked to characterize all the millions of words (literally) I have written on the Internet, an LLM said,  “The voice is so stable across 56 years that a reader can date a piece only by its subject, never by its sound.”

 I would wager my bottom LLM token that this phrase was used about someone else’s work and simply repurposed.

I was briefly flattered, until I remembered that the compliment came from a talking parrot, to which this phrase had as much meaning as “Who’s a good boy?”

There you go, another stroke in the negative column for LLMs; I wouldn’t trust LLM criticism of music, writing, or any other art. An LLM could write a fifth grader’s book report or praise Beethoven’s Fifth using pasted-together words from pre-existing reviews. I wouldn’t trust it for a review of the Billy Crystal one-man show opening on Oct. 1, unless I was simply interested in a pastiche of all pervious Crystal reviews. Or maybe a review of Tiffany crystal, or cut crystal, or some truck driver named William Crystal in Peoria, Ill. whose new Broadway Street office will open Oct. 1.

FOOTNOTES

[1] Most important paragraph of a news story or essay: usually the first paragraph.

[2] Most important paragraph of a magazine story, usually cleverly hidden by a writer who believes his sparkling prose will entice you into a story whose subject is a buried treasure.

[3] Such as they are in a series of farts

[4] Like the United States in 2027. While we’re at it, do LLMs use footnotes? Is a stochastic parrot capable of “realizing” that a semi-interesting fact should be in the footnotes to avoid disrupting the flow of the article—if there be such a thing in this case? Sure, it could create a footnote if it were trained on some stolen piece of intellectual property with footnotes. But since it has no inkling of context or meaning, it would have no “idea” what facts or digressions belong down here. Any footnote created by an LLM consists of nothing more than an uneducated guess as to which words are subsidiary.