It turns out Monet's name is Claude.

It turns out AI is anthropology.
So is AI studying humans? What exactly is AI learning?
Today I went to de Young to see "Monet and Venice." Monet’s Impressionism from over a hundred years ago—he fixates on the same thing, viewing it a thousand times in different lights, then smudges that “unclear but just right” feeling onto the canvas hundreds of times. The more I look, the more it seems like training a large language model...
Why did Monet paint so “blurry”? The introduction on the wall says that Monet and Turner’s paintings became increasingly hazy and pale, closely matching the air pollution curve during London’s Industrial Revolution. Burning coal, sulfur dioxide, smog scattering sunlight, blending colors together. So, at that time, he was painting what we now call “smog” (the word Smog was only coined in 1905).
Then, at age 68, his wife Alice took him to Venice. As soon as he arrived, Monet said, “This place is too beautiful to paint.” But once he started, he got hooked. Originally planning to stay two weeks, he extended to two months, painting 37 pieces. The entire Venetian lagoon, which has an island called Lido, so Claude had already been to Lido😄.
Continuing to walk, turning into another room, I was stunned—on one wall hung five or six nearly identical paintings. My first reaction was: Who’s copying Monet? How can copies be displayed? My ticket cost 40 bucks!
It turned out they were all painted by him. The same church, the same canal, different lights, different fog, different moments. His water lilies are the same approach—he painted 250 “Monet Water Lilies,” with 48 in the 1909 series alone. The same pond, painted over 30 years.
This approach is so AI-like… (not really)
Monet’s paintings are clearest from afar, increasingly blurry up close—like today’s AI hallucinations. From a distance, it seems plausible and reasonable; zoom in, and errors that can’t withstand scrutiny appear everywhere. And that wall of “almost but not quite the same” paintings is like when I ask a model to generate images, and it spits out 4 or 9 images at once—compositions similar, details subtly different—results for you to pick the “best one.”
Multiple variants of the same theme by Monet, human-selected—was that the earliest batch generation + human-in-the-loop? 😅
de Young’s curation is excellent—lighting, white space, spacing between paintings, flow—making everyone think deeply.
So, returning to the question I wanted to ask at the start,
What is AI learning?
Maybe what it’s learning is exactly what Monet was doing over a hundred years ago—fixating on the same thing, viewing it a thousand times in different lights, then smudging that “unclear but just right” feeling onto the canvas.
Large language models are essentially black boxes: you know what they output, but you can’t clearly explain why they produce that output. No one can dissect Monet’s brain and explain why he deliberately moved his brush to the left.
But the more I think about it, the more I feel these two “black boxes” are actually opposite.
Monet had a thousand gazes—paints just compress what he “sees” into a result, blurry yet different each time. In fact, he saw more clearly than anyone, fully aware that clarity is a form of laziness; the real world has no hard edges.
AI works the other way around—it hasn’t truly “seen” anything. It only reverse-engineers from billions of human-made, written, spoken results, creating an approximation that looks like it’s gazing.
In Monet’s blurriness, there’s his certainty from thirty years; in AI’s “reasonableness,” there might be nothing at all—just the most probable path, which happens to look like it really understands. AI is learning “traces left by humans.”
Finally, as I walk out, I wonder if ANTHROPIC sponsored the “Claude Exhibition”! Sure enough! Lead Sponsor, at the very top of the wall, in big bold letters: ANTHROPIC.
The root of “anthropic” is anthrop-, meaning “human.”
AI is anthropology. But after studying human traces, do humans really know what AI is learning?
(3/21–7/26, de Young, highly recommended.)