
I had the recent pleasure of recording a podcast with the lovely on AI and writing. I’ll share it when it’s out, and in the meantime, jotted down some quick reflections during my flight home from London:
AI is better than most humans at producing prose. In a couple years, it will be better than most “professional writers” as well.
Most text is not creative. Emails, policy papers, reported news. It does not desire to surprise or delight. It aims to convey ideas and information as clearly as possible.
It is inevitable—given this reality and these incentives—that most people will soon use AI to write most things.
Writing is a small, simple word for a very expansive task. There’s coming up with ideas, conducting interviews and research, presenting it all in an engaging way.
(Most jobs that seem easy to automate are deceptively complex.)
Current AI is not very inventive. I’ve never had it come up with an essay idea I thought was good.
It’s not clear how much this quality is inherent to LLMs, versus an optimization tradeoff made in favor of other things—reliability, professionalism, obedience, predictability. What if the RLHFers were professional curators and magazine editors? What if the models were permitted to hallucinate more?
Many of our best artists hallucinated. Some even found chemical routes to do it.
“I thought AlphaGo was based on probability calculation, and that it was merely a machine,” said Lee Sedol after his loss. “But when I saw this move, I changed my mind. Surely, AlphaGo is creative. This move was really creative and beautiful.”
Implicit in his quote: The statistical and the beautiful cannot coexist.
I’m still not over the fact that I have 24 hour access to a machine that has memorized and synthesized the entire internet’s worth of information.
None of us know how to know this much.
LLMs help with writer’s block, even when its suggestions suck. Insipid sentences are the best motivation to write inspired ones.
I prefer my human collaborators to my AI ones. But only one is consistently available at 3am.
AI hallucinates, humans misremember. Yet neuroscience suggests that memory and imagination are one and the same. As Demis Hassabis’s PhD thesis found, amnesiacs make shoddy novelists.
ChatGPT catches me in a lie approximately as often as I catch it.
Why do readers read? Sometimes, it’s to extract a particular piece of information. Other times, it’s to be entertained. Increasingly, it’s to borrow a particular author’s judgment and worldview.
Writing instructors love to talk about “stakes.” Competent prose isn’t enough. A story must answer: Why does this matter to the author, the reader, the world at large?
Pop is equally a story of the star and the song. Breakups, love affairs, addiction, deceit—the stuff of gossip rags creates the stakes.
We watch Olympic swimmers not because it’s fun to spectate, but to marvel at the frontiers of human achievement. Knowing we could never do it, imagining how relentlessly they worked to get there, how many early morning practices and weekend meets. Nobody wants to see a robot play sports.
I prefer my authors to have bodies, to live and walk and sweat in the world.
Most internet culture criticism is bad. You can tell when a writer doesn’t go outside.
I have viewed AI artworks I thought were good. Here, here, here. They strike me because I have no idea how to do the same. Their outputs so far from default slop, so distinct in their human perspective.
Artistic merit comes from that feeling of awe—at the distance between what the artist achieved versus what I could elicit from the same materials. You see past the tools and into a mind.
David Hockney: “Rembrandt spent days, weeks painting a portrait. You can go to a museum and look at a Rembrandt for hours and you’re not going to spend as much time looking as he spent painting—observing, layering his observations, layering the time.”
Power Broker heads talk as much about Robert Caro as his biographical subject. He is the mythical man who turned every page.
Art is a showcase of excess.
Mercor will pay you $50/hour to write about your hobbies. “Tabletop games, collecting, art, niche sports, fashion, cooking, etc.” Requirements: “1000s of hours spent.” The models must read because they cannot live.
Writing purely for the AI has no stakes. It’s text without audience, motive, purpose.
How much of the world does language contain?
Next token prediction is underratedly hard.
I start most essays not knowing how they will end.
Thanks for reading,
Jasmine
"ChatGPT catches me in a lie approximately as often as I catch it."
How are you prompting it?? I can't get mine to stop being stupidly sycophantic no matter how many times I admonish it.
Interesting insights, thank you. What do you think happens in a few years when AI is better at writing than most writers?