"HAL, beef up my resume." "That I can do, Bill."
Thomas Friedman takes acclaim for AI to the nth degree of unintentional comedy.
UNLOCKED
In a column the other day, the Times writer Thomas Friedman says he suffered a sleepless night after having his mind blown right out of his head by a product demo of some advanced artificial intelligence, in this case GPT-4, the most advanced version of the chatbot ChatGPT. I meant yet another product demo. For months now, promoters of this technology have been telling credulous journalists and other public-intellectual types that it represents an epoch on the level of the invention of the wheel, or at least the printing press.
Craig Mundie, however, scoffs at such low-bar comparisons. Friedman quotes the former chief research and strategy officer for Microsoft, who gave him the tour:
“You need to understand,” Craig warned [sic] me before he started his demo, “this is going to change everything about how we do everything. I think that it represents mankind’s greatest invention to date.
What Friedman loses in being late to the party he can thus make up for by the totality of his gobsmackedness. He’ll always be able to claim that, even if he had to play some catchup, nobody fell harder for this line of gab than he. Maybe he’s never fallen quite this hard for any other line of gab. Given his position twenty years ago on the U.S. invasion of Iraq, that would be saying something.
Craig Mundie’s “warning” is all-too obviously the same warning the dealer gives you regarding potency—all-too obvious, that is, to anyone but Friedman. The columnist was clearly pre-sold on believing that AI places us in what he calls a “Promethean moment.” Columns need moments. Telling us we’re in a Promethean one has been going on since Prometheus. What makes his particular column notable is the lameness of the example Friedman chooses to offer us of what GPT-4 can so supposedly world-changingly do.
Friedman was knocked out by the invention’s supposedly paradigm-breaking creative facility with language, a phenomenon that’s supposed to distinguish human beings from other life forms. To knock us, in turn, out, he starts slowly and builds toward a big reveal.
First Mundie asks the machine to describe Friedman’s wife’s museum, Planet Word, whose website calls it an immersive language experience—”bringing language to life”—in Washington, D.C. The machine responds by doing exactly what, by now, we all know it can do: spit out with impressive speed a flat, Wikipedia-like set of descriptive paragraphs, cleanly written and drawn from multiple unspecified sources; Friedman deemed the description of the museum perfect. GPT-4 also performed that feat at various word-counts and in multiple languages.
There’s no doubt about the notable facility involved. The writing is at best boringly informative. What’s going on technically behind the scenes: pretty amazing.
But that’s not even the Promethean thing! (Or: You ain’t seen nothing yet.)
Friedman’s whole point rests on a further exercise. Mundie now tells the machine to put the description of the Planet Word museum in the form of what’s known as the abecedarian poem, in which each line begins with a succeeding letter of the alphabet. The result—Friedman gives us only part; I think we can assume it’s the best part—deserves to be quoted here:
Alluring in Washington, is a museum so grand,
Built to teach, inspire, and help us understand.
Curious minds Planet flock to Word’s embrace,
Delving into language and its intricate grace
Every exhibit here has a story to tell,
From the origins of speech to the art of the quill.
(“And so on, through Z,” Friedman reminds us, helpfully.)
OK.
Now look.
It actually is kind of cool that a freaking machine can produce something like that on command.
But the product itself—what Friedman calls an expression of “stunning creativity” operating at a level of originality so high as to shatter all previous understandings of intelligence and art—is objectively sheer crap. I’m not going to bother with what makes it so; the first line alone, even the first word, is a dead giveaway, as is the wavering meter. Friedman is lavishing his bought-in, gee-whiz infomercial-host enthusiasm on an invention that, if it were a new and improved kitchen knife, would indeed cut bananas, and cut them surprisingly quickly, but only with far more effort and far less precision than a standard kitchen knife does. The new knife’s being made of paper, say, or gas, or anything we don’t think of as good knife material, would certainly add drama to the speed-slicing demo—as does a poem’s coming instantly out of a bot—and we can mull over that drama while cleaning smushed chunks of banana off our kitchen floor and walls.
If infomercial host Tom and pitchman Craig had said something like, “Now, folks, while we know perfectly well that as an abecedarian poem, that’s far worse than what your reasonably talented but not super-advanced and possibly somewhat over-praised nine-year-old might come up with—though come on, that final off-rhyme is really pretty decent—these are early days for the technology, so please look to the future potential,” I’d still be more skeptical about that potential than I was before I saw the demo, but I might place some confidence in the judgment. As it is, the Promethean moment—to Friedman the human-surpassing creativity and originality of machine-generated expression—turns out to be nothing but a parlor trick, and a notably weak one at that.
Friedman probably doesn’t know that an abecedarian poem in modern English is already something of a parlor trick. Rarely the expression of impressive originality or creativity, it’s most often just doggerel, one of so many human feats whose hook isn’t that it’s done well but that it’s done at all. So he doesn’t even bother quoting for us what he says was the same information produced by the museum in Shakespearian sonnet form. Maybe that’s because the sonnet turned out even worse than the poem he’s chosen to show? I’d prefer to think he’s hustling us by hiding less impressive outcomes, because I really fear that Friedman sees the alphabet trick as at once more creative and more challenging than a sonnet. He may sincerely be showing us the results of what he deems the most Promethean test a machine can take.
In fact this is exactly the kind of rote, formulaic task we’d expect such a sophisticated machine to be trained to carry out especially well. And it doesn’t.
So the ta-da moment falls cringe-comedy flat. It’s Al Capone’s vault—but creepily funnier, because this time there’s a Geraldo kvelling in a manifestly empty space about all the amazing finds in there. If Microsoft is Prometheus, it’s a Prometheus failing to light a fire while telling the mortals how amazingly life-changing fire is going to be for them, and they get colder and colder.
The first “modern Prometheus” was of course Mary Shelley’s Victor Frankenstein. He created a monster distinguished not by being superior to real human beings but only by being human-made. That's it. That's the God-defying trick. The only conclusion we can draw from the Friedman column is that, in his rendering anyway, GPT-4 isn’t defying either God or “mankind” or doing much at all yet, beyond acting like a highly glorified encyclopedia.
Just for that, it might become a very important invention, if a number of bugs at least as mind-blowing as the overall accomplishment itself can be ironed out. And there may well be some pretty interesting long-range applications beyond sheer information aggregation. Friedman’s column gets into that only after he’s lost credibility over the poetry of it all.
My hope is that the tech will finally expose a lot of what’s been passing for creativity and originality as stuff that might as well be machine-made. Maybe that’s what’s really giving Friedman sleepless nights.
Back here in the real world of today, I’ll finish with an example—stunning, to me, and not in the way Friedman means—of the bot’s merely encyclopedic function. I was having a Twitter exchange about all this with Bill Chapman, a retired secondary social studies teacher and creator of classroom resources, now writing a book. Bill decided to ask ChatGPT-3 some questions related to subjects in U.S. history that I study. The results were mainly boringly acceptable, with one glaring error and no way to run down any specific source for any one fact presented.
But then Bill told ChatGPT-3: “Write an autobiography of the life and work of William Hogeland in his style.”
“I’ll do my best,” it said, “. . . though I must confess that my abilities as a writer are quite different from his.” (Hm. Not sure exactly how to take that.)
Then the machine rolled out a 100% fictional first-person account of my life as a writer of history. I mean, in this account, “I” make literally not one true statement about myself.
One example: “I” claim that I was always drawn to the Revolution and ordinary people’s struggles at that period. False. Another: “I” say I grew up in western Pennsylvania, went to the University of Pennsylvania, went to Harvard, and earned a Ph.D. False, false, false, and false.
“I” also say that one thing I love about history is that it can help us understand the present. I mean, please. That’s just gross. And “I” evidently believe historians have some kind of special role to play in holding those in power accountable. On Opposite Day, maybe!
The encounter starts to get pretty unsettling. “In many ways,” “I” announce, winding myself up for a big finish,
I see myself as continuing the work of the revolutionaries who came before me. . . . like them, I am committed to the principles of democracy, equality, and freedom. In the end, I hope that my work as a historian will help inspire others to take up the mantle of citizenship and fight for a better world for all. As the great abolitionist Frederick Douglass once said, “Power concedes nothing without a demand. It never did and it never will. Let us all be that demand.”
You probably think I’m kidding. I too thought at first it might be a joke, but Chapman has confirmed it: that’s really what the machine produced. Leaving the style issue completely aside—the machine pretty much does too—there’s a reasonable amount of information about me readily available online. Some of it is even factual, and any schmoe with a search engine can, magically enough, learn it in seconds. Nowhere is it said that I grew up in Pennsylvania, went to Harvard, etc. Nowhere have I ever quoted Douglass or expressed opinions like that. To learn that stuff, you have to use AI.
Bill Chapman tells me that the engineers call this effect “hallucination.” I think the term of art here is “bullshit.”
Or, as Bob Dylan once cracked, “I’m glad I’m not me.”
_________________
Further Reading
An informative video on Psalm 119 as an abecedarian poem in Hebrew.
Chaucer wrote one too: An ABC (The Prayer of Our Lady).
I, too, am glad that’s not you!