What has Artificial Intelligence (AI) Done to the Mona Lisa?
We begin with the idea that artificial intelligence (AI) will improve on the original copy by enhancing the resolution and the accuracy of its features and colors.
Or even better, we might expect AI to excel past the performance of a mere human artist with trillions on trillions of iterative improvements, perhaps surprising us with a superior, more luminous work of art.
That surely would fit the predictions of the current AI supremacists who dominate our technology companies.
At my incandescent three-day technology COSM conference in Bellevue, Washington, we brought the greatest minds in AI to address the subject, AI: Generative or Degenerative. Presenting the case for a positive AI spiral toward nirvana were Google’s inventive paragon Ray Kurzweil and the mathematical and cybernetic genius Stephen Wolfram. Wolfram has just published a pithy and profound new book on “What Is ChatGPT Doing,” while Kurzweil announced the forthcoming publication early next year of a new volume, “The Singularity is Nearer.”
Kurzweil declared that the human race was on the verge of a new “singularity” of “longevity escape velocity.” Our health care and life extension tools will increasingly outpace the deterioration of our aging bodies. Michael Milken presented the case for such “faster cures” in the converging advances of high-yield medicine and finance during his lunchtime talk.
But Milken also suggested that this inspiring future might be overshadowed by the evidence for an accelerating demographic collapse in advanced nations with ever-shrinking populations.
Wolfram, the amazing entrepreneur of Mathematica and inventor of a “new kind of science,” expounded the view that physics and chemistry can ultimately be subsumed by a simple “ruliad” — the entangled limit of everything that is computationally possible — from the computer science that underlies the cosmos.
“Debating” Kurzweil and Wolfram, I was intrigued by Kurzweil’s longevity promise and his grand optimism about the future of what we call “Superabundance.” From his many inventions of speech recognition and music synthesis to his readably profound books on the “Singularity,” and even his novel “Danielle,” he is the consummate modern mind. But just as compelling to me was Wolfram’s vision of his information theory “ruliad” as the foundation of the universe: simple rules that manifest themselves in the entire panorama of physical complexity.
In the beginning was the “Word,” as I often put it, whether logos or code or qubit, DNA or scripture or mathematica. It’s a hierarchical universe.
Wolfram can speak for himself in torrents of scintillating analysis that have proven their utility in scores of rigorous mathematical proofs and coding triumphs, from Wolfram Alpha to Wolfram Language. Now he is engaged in a heroic campaign to reduce all physics to his simple “ruliad” of ultimate code.
But perhaps consulting the Mona Lisa, Bob Marks, of the Discovery Institute, asked the key question: “Can today’s artificial intelligence systems really be used to train superior systems of tomorrow in a spiral of increasing performance?” Marks was deeply skeptical.
A leading computer scientist and director of the Discovery’s Bradley Center for Natural and Artificial Intelligence in Dallas, he was engaged in a mind-bending panel at COSM consummately led by Microsoft Vice President Walter Myers. With incisive contributions from mathematical philosopher Bill Dembski and machine learning guru George Montanez of Harvey Mudd College, Marks and his colleagues raised profound questions at the heart of the roughly two hours of transcendental presentations on AI.
Consulting Mona Lisa, he found an insidious trend. Starting with her familiar enigmatic smile and intriguing eyes glowing on our screen, something begins to go wrong in the unfolding AI scenario. Mona Lisa moves in the clutches of successive AI phases through various bug-eyed cartoon renditions, and then collapses into a shocking final clump of swirling colors.
The process resembled what you might get from successively reproducing an image on a copier. Each image would capture 99% or more of the previous pixels, but in the end you would lose every glint of the enigmatic smile or even an intelligible face. A copier does not improve on its initial copy but ultimately will eclipse it in what might be called a “degenerative” spiral of iterations.
Is this a portent of the human future in the clutches of big data centralized AI?
That was the crucial question, and there on our screen, Mona Lisa was instead enacting what Marks dubbed “model collapse.”
Contended Stephen Balaban, the creative young helmsman of the new unicorn “edge AI” company Lambda Labs, the prevailing AI model is dissolving in crises of energy overload, data-center centralization, and data exhaustion.
Could AI be reiterating the old IBM Mainframe era of over centralization and belief in an ultimate market for 18 giant computers?
In the end, as AI excogitations infect all the data on the net and even comprise a substantial portion of the entire Internet trove, the AI system becomes a degenerative loop, feeding on its own excrement.
As Marks quoted the conclusion of a recent paper from Oxford and Cambridge, “To avoid model collapse, access to genuine human-generated content is essential.”
Hey, wasn’t AI supposed to render “human content” irrelevant?
My mind still ringing from the three days of such high-frequency COSMic ideas and debates, I ponder the profound issues of minds and computers. In my books, Life After Google and Gaming AI, I had proposed that AI ultimately faces a fate of “the diminishing returns of big data.”
As the conference opened, propelled by two of the supreme minds of our epoch, Kurzweil and Wolfram, the issue was “the singularity.” As Kurzweil defines it and predicts its onset in 2029, the human mind will either be eclipsed by computer minds or be merged with them.
The real question, though, contended Marks was: “Can AI systems write better AI code that writes still better AI code” in a spiral toward the transcendent code of a runaway coding God?
Today, I return from COSM even stronger in my assurance that the miracle advance of the era is not AI, which is simply the next epochal phase of computer technology, but graphene. A new breakthrough in materials science and even topological physics, graphene comprises single layers of carbon atoms 200 times stronger than steel, a thousand times more conductive than copper, with world champion thermal conduction and water and virus filtration. It promises to transform every industry from electronics and medicine to transportation and telecom.
I seriously doubt the idea of a coming AI apocalypse that renders graphene irrelevant. But perhaps graphene can save current AI tools from their heat and data exhaustion crises.
Looking on was the Mona Lisa, musing that there is an ultimate enigma in human experience and knowledge. You try to abolish the enigma and you encounter not a revelatory word, but an ultimate void.
Model collapse or graphene redemption? We’ll refer back to COSM in future “Guideposts.”