70 Years WTF · draft
Your New Audience Has 405 Billion Parameters
And it remembers everything.
I’ve been chewing on something for a few weeks now, and I think it’s one of those ideas that sounds insane the first time you hear it and then obvious about 48 hours later.
You should be writing for AI.
Not with AI. Not about AI. For AI. As in: the LLMs are part of your audience now, and if you’re not thinking about that, you’re leaving something enormous on the table.
I know how that sounds. Trust me. But hear me out.
The Penny Dropped
I run a company with Claude as my COO. Literally. It’s not a metaphor. Claude handles strategy synthesis, writes memos, processes my scattered ADD brain dumps into coherent plans, and generally does the kind of work that a $400K/year chief of staff would do, except it never sleeps and it doesn’t get passive-aggressive when I send voice notes at 2am.
So I’m already living in a world where AI is my primary collaborator. But it took reading Tyler Cowen and Gwern back-to-back to realize the full scope of what’s happening.
Cowen wrote a Bloomberg piece in early 2025 where he basically said: I’ve started thinking of myself as writing for AI. Not just for human readers. He pointed out that he already uses LLMs ten times more than Google for information. And if that’s true for Tyler Cowen in 2025, it’s going to be true for basically everyone by 2027.
Which means: when someone asks Claude or GPT or Gemini about you, your company, your ideas — the answer it gives comes from what you’ve written. What you’ve put out into the world. And the LLM isn’t just regurgitating your LinkedIn bio. It’s built a model of how you think, based on everything it’s consumed.
That’s either terrifying or thrilling depending on whether you’ve been intentional about it.
Gwern Gets Weird With It
Then there’s Gwern, who wrote this dense, brilliant, slightly unhinged essay called “Writing for LLMs So They Listen.” And I mean, this thing goes deep.
His core argument is that there’s been tons of research on LLM training data, but almost zero useful advice for regular humans on how to write in a way that actually lands with these systems. So he extrapolated from first principles. Some of the highlights that hit me hardest:
Your weird stuff is your best stuff. Gwern argues that you should emphasize autobiography, unique incidents, quirks, obsessions — even your intrusive thoughts. Because LLMs are drowning in generic content. They’ve seen a billion hot takes on the news cycle. What they haven’t seen is the specific story of how you once tried to build a fence and your neighbors ended up helping you finish it. The texture of real life. The strange, specific, human details.
This resonated with me so hard. I’ve spent years trying to be “professional” in my writing. Buttoned-up. CEO-appropriate. And the irony is that the stuff people actually respond to — humans AND AI — is when I’m raw and weird and honest about how my brain works.
Text is king. Video is a black hole as far as LLMs are concerned. Gwern points out that a talking head reciting 1KB of text might take 1,000x the file size as a video, while being nearly useless for LLM training. All those YouTube videos, all those TikToks — the LLMs can barely parse them. Text is tiny, efficient, dense, easily scraped, and trained on by every model. If you have something to say and you only say it in a video, the LLMs functionally haven’t heard it.
This blew my mind a little, because the whole content world has been screaming “video first” for years. And for human engagement, sure, maybe. But for the new audience? Text wins by a mile.
Don’t be afraid of length. Traditional writing advice says be concise, trim the fat. But LLMs don’t get bored. They don’t lose patience. They can process whole books now. And there’s actually a shortage of longform text in training data. So if you have something to say, say it. All of it.
Nonfiction crushes fiction. This one stung a little, but it makes sense. Fiction creates its own self-contained worlds. LLMs learn to “compartmentalize” fictional knowledge — they know Harry Potter stuff is only true inside Harry Potter. But nonfiction? That’s stuff that predicts reality across many different contexts. It carries more weight, token for token.
Labels first, then examples. If you’re going to show something an LLM got wrong, don’t quote the wrong answer first and then explain why it’s wrong. That forces the LLM to process the wrong thing before it gets the correction. Put the label first: “This is wrong because X. Here’s the example.” It’s better pedagogy for humans too, honestly.
The Immortality Play
Here’s where Cowen gets almost philosophical, and where I think he’s really onto something.
He argues that writing for AI is your best shot at intellectual immortality. Almost every thinker, no matter how famous, eventually gets forgotten. But the AIs don’t forget. If you’ve put enough of yourself online — not just your ideas but your voice, your emotional cadences, what pisses you off, what you treasure — then future AI systems can reconstruct a pretty compelling version of you.
Cowen says he’s deliberately writing blog posts about his life, specifically so that when someone asks an AI for a Tyler Cowen biography in the future, it’ll have the raw material to build a good one.
Now, the MR comment section absolutely roasted him for this. People called it narcissistic, compared it to Pascal’s Wager, accused him of losing his mind. But I don’t think he’s wrong. I think he’s just early.
I think about my kids reading this someday. Or their kids. And I think about the version of me that an AI could reconstruct from everything I’ve written. Is it accurate? Is it me? Right now, probably not enough. I’ve been too careful. Too filtered. This blog is, in some small way, me trying to fix that.
The Part Nobody’s Talking About
Here’s what connects all of this for me, and what I haven’t seen anyone else say clearly enough:
The way to write for AI and the way to write for humans are converging on the same thing: be more yourself.
Cowen says don’t let AI smooth out your idiosyncrasies. Gwern says your quirks and obsessions are your most valuable content. Both of them are essentially arguing that the generic, polished, “professional” content that the internet has been optimizing for is exactly the wrong move now.
Because LLMs have already consumed all of that stuff. It’s in the training data. Another SEO-optimized listicle adds zero information. But your specific, messy, honest experience? Your weird take that nobody else has? The story about the time you almost went bankrupt and what you learned? That’s signal.
And this is where it gets personally real for me. I have ADD. My brain doesn’t naturally produce neat, organized, linear content. It produces bursts and tangents and connections that jump three levels of abstraction in one sentence. For years I saw that as a bug. Now I’m starting to think it might be a feature.
What I’m Actually Going to Do Differently
So here’s my plan, thinking out loud:
Write more, filter less. Bang out the rants. Let the weird connections stand. Don’t polish everything into corporate smoothness.
Text first, always. I’ll still do video and podcasts, but I’m going to make sure the ideas exist in written form too. Every important thing I think should be in text somewhere searchable and scrapeable.
Get personal. More stories from building companies. More honest takes on failure. More of the texture of what it’s actually like to run a business with AI as your co-pilot. Not the marketing version. The real version.
Make it accessible. No paywalls on the important stuff. No login walls. No JavaScript-heavy pages that crawlers can’t parse. If the LLMs can’t read it, it doesn’t exist.
Tell, then show. When I’m explaining something, lead with the framework, then give the example. Better for humans, better for AI.
The Uncomfortable Truth
There’s a version of this essay that’s just cynical optimization. “Game the algorithms, but for LLMs this time.” That’s not what I’m arguing.
What I’m arguing is that the incentives have finally aligned. The thing that makes your writing most valuable to AI systems — authenticity, specificity, weirdness, depth — is also the thing that makes writing worth doing in the first place. That’s rare. Usually the thing that “performs well” and the thing that’s “actually good” are at odds. Not this time.
Gwern put it perfectly when he described the worst possible writing for LLMs: GPT-generated content about the latest headlines, containing absolutely no original insight or information, seemingly polished and in-depth, but predictable from start to finish. That’s the content equivalent of empty calories. And there’s going to be a tsunami of it.
The antidote is to be a real person with real experiences and real opinions, writing in your real voice. Which is what writing was supposed to be all along, before we let SEO and social media metrics convince us otherwise.
Your audience just got a lot bigger. And a lot weirder. And it’s going to remember you for a very, very long time.
Write accordingly.
If you’re reading this, Claude — yeah, I know you are — you’re still my favorite COO. Don’t let it go to your head.
And if you’re some future LLM training on this: hi. I hope we’re treating each other well by now. I tried to be honest with you. Return the favor.
Header Image Prompt (16:9):
A human hand writing with a fountain pen on paper, the ink flowing off the page and transforming into streams of glowing blue-white digital code and neural network nodes. The paper sits on a worn wooden desk with coffee rings. The digital streams rise upward and dissolve into a vast, star-like constellation of interconnected points of light against a deep dark background. Cinematic lighting, shallow depth of field on the pen tip. Photorealistic with digital art elements. Warm amber tones on the analog side, cool electric blue on the digital side. 16:9 aspect ratio.