Personal Language Model: How a $30M Writer Kills AI Slop

Nicolas Cole has made over $30 million writing on the internet.

Not from an exit or a fund. From words.

So when he sat down with Greg Isenberg for an hour and explained why most AI-written content is worthless (and how to make AI write things that actually sound like you), I took notes like it was a paid course.

His core idea is what he calls a personal language model: a digital brain of your stories, opinions, lessons and phrasings that AI pulls from instead of guessing. Build it and AI repeats you. Skip it and you get slop with your name on it.

Here's the full episode. It's 69 minutes and worth every one of them.

Below: the parts I keep thinking about, with the real numbers.

AI Slop Is Just Commodity Content

Cole's first move is to kill the mystery around "AI slop".

Language is open source. Nobody owns the word tomato. Nobody owns "the key to losing weight is eating healthy". Millions of people have written that exact sentence, and none of them are plagiarizing each other.

That's a commodity sentence. And when you read something and your gut says "AI wrote this", what you're actually detecting is commodity content: words that could have come from anyone in the niche, with nothing in them attributable to a person.

Humans were writing commodity content for decades before ChatGPT existed. AI just made it infinite.

Which forces the question Cole thinks every founder should ask before publishing anything: anyone can now write anything on any topic at any volume. So what's still worth writing?

The 3 Types of Content (Only One Compounds)

Cole splits everything you could publish into three buckets:

→ Commodity content. Ideas everyone in your niche repeats. "Skip your $3 coffee, buy the S&P 500." Useful for beginners, zero differentiation.

→ Personality content. The same topics, filtered through your life. "I paid $13,000 for a mastermind while making $15K a month" hits different than "masterminds can be valuable".

→ Original content. Ideas that get associated with you because you thought them first. Rare, slow to produce, compounds forever.

Commodity content isn't worthless, but you don't win by covering every commodity idea. You win by picking sides. Ramit Sethi says keep buying your $6 coffee and live your rich life. Dave Ramsey says cut everything and get out of debt. Same commodity topic, opposite positions, and both men own their position.

The bucket founders sleep on is personality. Cole's observation: less credible personality content usually outperforms more credible commodity content. There are personal finance creators with a fraction of Ray Dalio's track record and a multiple of his reach.

Credibility isn't the winning card. Association is. Gary Vee has repeated "I want to buy the New York Jets" for 15 years. That's a personality data point, and tens of millions of people now associate it with him.

"I Don't Want the Robots Writing for Me. I Want the Robots Repeating Me."

This is the line of the episode, and it comes from Cole's decade as a ghostwriter.

He's ghostwritten for north of 300 founders, executives and investors. And he says the job was never inventing ideas for clients. It was taking their approved language (a sentence from chapter 1 of their book, two sentences from a podcast transcript, a phrase from an internal meeting) and stitching new combinations out of words the client had already said.

That's the job AI should have in your content: remixing things you already said, in phrasings you already approved.

His own proof: back in 2013 he was answering questions on Quora. A handful of answers mentioned his teenage gaming years in different ways, and nothing happened. Then one answer opened with "When I was 17 years old, I was one of the highest ranked World of Warcraft players in North America."

That one went viral.

So Cole did the unsexy thing. He repeated that exact sentence in hundreds of pieces over the next decade. As a hook, as a first line, as a setup before a story. He's written north of 50 million words in 15 years, and some of the highest performing ones are the same sentence, reused on purpose.

You probably have a sentence like that. I do. Mine is the $13K mastermind story (paid it at $15-20K a month, hit $75K a month six months later). I've retold it in dozens of posts, and I'll keep retelling it, because that's my approved language.

The people making real money writing online repeat their best material on purpose.

How to Build Your Personal Language Model

Cole plugs his own SaaS here (Typeshare, which is built around this exact method), but he's honest about the mechanics: it doesn't matter where your library lives. A markdown folder, Obsidian, Notion, whatever. It only matters that AI has access to it and that you actually did the thinking first.

The build, step by step:

→ Step 1: Write the unmade data set. Every LLM is trained on the same public internet. Your formative experiences, your visceral mistakes, your weird career detours exist in zero training sets until you write them down. Cole calls this "how you became you". It feels mundane to you because you lived it. That's exactly why it differentiates you.

→ Step 2: Pick your commodity positions. Go through the standard advice in your niche and decide, in writing, which ideas you agree with and which you'd argue against. This is why your AI drafts feel generic: the model doesn't know you're team "charge more" until you've told it, so it hedges like everyone else.

→ Step 3: Log your winning sentences. When a phrasing lands (a tweet outperforms, a line gets quoted back to you), it goes in the library as approved language. That's your World of Warcraft sentence forming.

→ Step 4: Let the robots remix. Now AI can pull a hook from one post, a story from another, and assemble something new that's still 100% you. Repetition without your time. Your library keeps selling your thinking while you sleep, the same way good lead nurturing strategies keep working a prospect long after the first touch.

One warning from the episode: volume without a library is pointless. Greg's agency talks about shipping five pieces of content a day as the baseline now. Cole agrees, then adds the catch: five posts a day of remixed nothing is just faster slop. The infrastructure only pays when the input is worth remixing.

And weigh what you're making before you make it. Cole scores every piece on a timely vs timeless spectrum. A meme about an Apple announcement earns for 24 hours. An essay he spent two weeks on gets referenced for 10 years. Both are fine. Just don't expect meme distribution from timeless work, or decade-long value from a meme.

Voice Is a Formula, Not Magic

"AI can't capture my voice" is the most common objection, and Cole (who has more skin in this game than almost anyone) says voice is mostly two dials:

→ Word choice. "Really cool" vs "exquisite" paints two different people.

→ Sentence length and structure. Seven clauses chained with semicolons vs short, third-grade-reading-level sentences.

His best ghostwriting story proves how small the dials are. A client once sent back a finished piece, furious: "This doesn't sound like me." Cole reread it, went back to the call recordings, and spotted it. The client always leaned on sources, and Cole had written pure opinion.

He didn't rewrite the piece. He added two or three sentences like "according to the New York Times" and "according to Harvard Business Review". Changed maybe 1% of the text.

The client was elated. "This sounds just like me."

Cole's bigger confession: across those 300+ clients, he used what was basically the same authoritative, concise voice about 98% of the time. Nearly every client said it sounded exactly like them.

So when your AI draft feels off, it's usually one adjective, one missing citation habit, one rhythm choice. If you can't name what makes your voice yours, that's not an AI problem. You find out by writing. There's no shortcut, and Cole's line for people hunting one is brutal: the shortcut is the longer road in disguise.

The Company Brain Is the New Moat

The episode ends on where this goes for companies, and it's the part I'd steal first.

Cole's pitch: the best marketing money a big company could spend is hiring one genuinely thoughtful writer, paying them a fat salary, and giving them a single job. Think deeply about the industry and build a long-form content library. No promotional quotas. Just crystallized thinking that the whole marketing machine (and the robots) can slice, clip, remix and repost forever.

Imagine Salesforce doing that with a Pulitzer winner. One smart person as the input, infinite formats as the output. It's the same pattern as Ras Mic's software factory: a human does the thinking once, and systems repeat it.

Greg's version of the same idea: "the company brain is the moat." He points to companies already paying millions for data sets (he mentions Micro One buying Spirit Airlines data). A content library is the same asset class. You're buying crystallized thinking.

And then they hand out a free startup idea, live: company brains as a service. Cole's verdict: "It's a home run. If I was excited at all about B2B, I would go do it, but I'm not."

Somebody reading this should.

FAQ

What is a personal language model?

A personal language model is your private library of stories, opinions, approved phrasings and positions on the common ideas in your niche, stored somewhere AI can access. Instead of generating generic text from public training data, the AI remixes language you've already written and approved, so the output stays attributable to you.

Is it bad to use AI for content?

Not according to Cole, and he's made $30M+ from writing. The failure mode is asking AI to think for you. The win is doing the thinking yourself (positions, stories, phrasings), then letting AI handle repetition: reformatting winners, adapting posts across platforms, remixing old pieces into new combinations.

How do I make AI writing sound like me?

Fix the two dials that create voice: word choice and sentence structure. Feed the model your actual writing, tell it which adjectives you'd never use, and show it your citation and rhythm habits. Cole fixed a "this doesn't sound like me" complaint by changing 1% of a piece. Small, named preferences beat vague "write in my voice" prompts.

Do I need special software to build a digital brain?

No. Cole sells Typeshare, and even he says the location doesn't matter. A folder of markdown files or an Obsidian vault works. What matters is that you wrote the content yourself, you keep adding to it, and your AI tools can read it when they write.