Hughes switched industries, he told me, in part because he came to believe Silicon Valley was getting A.I. all wrong. Generative-A.I. tools are unruly and imprecise — “sloppy,” he said — but too many companies were trying to use them for tasks where they couldn’t afford to be wrong. “Like self-driving cars or robot surgeries or whatever,” he says. “And we’ve been struggling with that for years. Because if you don’t want to run over 7-year-olds in Kansas, you’ve got to be 99.999999 percent precise.” Whereas in a creative context, “if I generate a bunch of elves and they have seven fingers” — “hallucinations,” in the parlance of the medium — “it doesn’t matter, because they’re part of my iterative creative process of brainstorming what elves could look like.” Generative A.I., he has come to believe, is best suited for tasks “where ‘hallucination’ is a feature, not a bug.”
The sum of Hollywood’s collective fears, says Bennett Miller, the Oscar-nominated director of “Moneyball” and “Foxcatcher,” “is automation” — robots replacing humans, just as in the movies. Miller spent five years making a documentary about the dawn of A.I. that he describes as a “time capsule” about “a moment before a real loss of innocence in Silicon Valley.” (The untitled film is currently in legal limbo.) In the course of making it, he got to know the original leadership team at OpenAI, including Sam Altman. A few years ago, they offered him access to a beta version of their forthcoming text-to-image tool, DALL-E.
“It was astounding,” Miller told me. “From the moment that I had an account set up to literally 10 minutes ago, I’ve just been all in.” This January, at Gagosian’s Paris gallery, he will open his third show of ghostly, surreal images that evoke the grainy early days of photography but were created with DALL-E. In one of them, a silhouetted man looks up from the floor of a century-old theater at a massive sea creature onstage, its body so large that it extends beyond the frame. “It’s like realizing that you had locked-in syndrome, because you really can navigate to extraordinary places.” He fell in love with getting lost. The mistakes, the wrong turns, the model’s peculiar way of comprehending the human world — a bit Luis Buñuel, a bit Diane Arbus — led to all of his breakthroughs, which is how the best art often gets made: by accident. “It’s not just a change in degree of what’s been possible before; it’s really like a change in kind.”
And yet as much as Miller’s creative practice has been transformed by A.I., it’s still merely a tool to him — and “the tool doesn’t make you an artist,” he says. “I just don’t see it as a threat the same way others see it. I’m not saying that there aren’t going to be huge problems that emerge. But here’s the thing that I cannot comprehend: human artists’ being replaced.” The great wild card of A.I. is that it learns and gets better, and we can only guess at its full capabilities. Its performance so far, though, has also highlighted the gap still to be closed, especially with text-generation tools like ChatGPT, a lowest-common-denominator regurgitation machine whose countless practical uses don’t appear to include writing screenplays.
Tom Graham, a Metaphysic co-founder and its chief executive, says he can see A.I. tools “summarizing news articles and doing great explainer videos for corporate work. I can see them creating generic or derivative stories that just kind of seem like other stories.” But, he adds, “amazing storytelling is very, very difficult.”
