The speaker, who ran a writing school called Write of Passage for six years, argues that AI is fundamentally reshaping the writing world—but rather than making writers obsolete, it is raising the bar for what writing must be to succeed. He shares how he personally uses AI as a thinking and writing partner, what kinds of writing are most “AI-proof,” and practical guidance for writers navigating this shift.
How the speaker’s relationship with AI evolved
He first became curious about LLMs after reading a New York Times article by Steven Johnson about GPT-3, which gave him an early understanding of how the technology works at a high level.
He was initially skeptical, telling a mentor—a multi-billion-dollar company CEO who was pushing his organization to be “AI-first”—that the tools weren’t good enough and hallucinated too much.
His skepticism began to crack in 2024 during a trip to Argentina, where he used ChatGPT as a real-time tour guide, filing questions during the day and reading the answers at night. He found the AI more useful than most of the human tour guides he hired.
His AI usage has roughly 10x’d in the past year, driven by rapidly improving models and falling prices due to intense competition—18 different companies released models as good as or better than GPT-4 in 2024 alone.
AI is already funny, and that matters
He points to a specific example: ChatGPT 4.5 generated a “be me” meme about economist Tyler Cowen that was so niche and accurate that he couldn’t have written anything as funny in a week.
This illustrates a broader point: AI will be exceptionally good at super-niche humor and content because the odds of a human comedian sharing that exact interest are very low.
He references William Gibson’s line—“The future is already here, it’s just not evenly distributed”—arguing that these small glimpses hint at what’s coming.
The scale of AI-generated text is already massive
Citing writer Ethan Mollick, he notes that the past 18 months have seen the most rapid change in human-written communication ever.
As of September 2024: 18% of financial consumer complaints, 24% of press releases, 15% of job postings, and 14% of UN press releases showed signs of LLM writing—and the real numbers are likely higher since many users edit AI output to hide its origin.
He estimates that roughly half of what he personally reads is now AI-generated, simply because time is finite and AI-produced content is abundant and tailored.
Two dimensions of writing quality
He distinguishes between two kinds of quality:
Objective quality: the craft of the writing itself (e.g., Gibbon’s Decline and Fall of the Roman Empire or Robert Caro’s The Power Broker).
Tailored to your interests: how precisely the writing matches what you’re curious about right now.
Critics focus on objective quality, where AI still lags behind the best human writers. But AI is already nearly perfect on the “tailored to your interest” dimension—it can produce a 2,000-word report on the specific flora and fauna along your exact walking route in Austin in seconds.
He argues that for most reading purposes, a 7/10 objectively well-written piece that is 10/10 tailored to your interests is more valuable than a 10/10 masterpiece that’s generic or irrelevant to you.
What writing is AI-proof: the two E’s—experience and expertise
The more a piece of writing comes from personal experience, the less likely AI is to overtake it. Memoirs, biographies, and personal narratives remain deeply valuable because readers want human-to-human connection, not a computer’s hollow simulation of one.
He cites the conversion testimony of a Wikipedia co-founder as an example of writing so emotionally powerful that no AI could replicate its effect.
Beyond personal stories, writers are protected when they possess expertise and lived knowledge that LLMs don’t have:
Knowing the culture and vibe of a city you’ve lived in for five years.
Having run nearly 200 live Zoom sessions with hundreds of people.
Having private, cutting-edge knowledge shared in small social circles (e.g., how the YouTube algorithm currently works) that hasn’t yet made its way into training data.
His heuristic: if your writing draws on what you know from direct experience and deep expertise, you’ll be fine.
The core skills that will always matter
Taste: the ability to discern what to keep and what to cut. Whether you write alone or with AI, most generated content gets discarded. Writing is sculpting—removing everything that isn’t the final statue.
Spiky point of view: a unique, idiosyncratic belief about the world that very few people share. AI is trained on consensus and struggles to produce genuinely contrarian or original takes.
He gives the example of a friend who believes K-8 education can be accomplished in two hours per day using AI and apps, with the rest of the day devoted to life skills—a conviction built over 25 years of running Alpha School in Austin with real data to back it up.
That kind of deeply held, evidence-backed, unusual conviction is what makes writing valuable and AI-resistant.
Writing will be like music, not chess
He frames two possible futures:
Chess: computers are better, but people still watch humans play for the drama and humanity.
Music: when you hear a great song in a club, you don’t care how it was made—you just care that it slaps. The method of production is irrelevant; only the quality matters.
He believes writing will follow the music path. Right now, using AI for writing carries the same stigma that sampling once did in music (the Beastie Boys, Dr. Dre, and Kanye West normalized it). In 10-15 years, no one will care whether AI helped—they’ll only care if the writing is good.
He hates obvious AI slop—sterile, cliché-ridden writing that clearly wasn’t curated by a human with taste. But if someone uses AI to refine ideas and produces something captivating, he has no problem with that.
AI is the end of slop, not the beginning
He defines slop as “when simply publishing or getting something done is more important than the quality of what you publish.”
For the past decade, the internet rewarded consistency and distribution over quality—SEO demanded bloated content to increase time-on-page, and newsletter writers published weekly even when the work wasn’t their best.
That era is over. Now writers compete not just with other humans but with computers that can produce content instantly. Quality is becoming king because only truly good writing will rise above the flood of AI-generated content.
He hopes algorithms will filter out the vast majority of low-quality content and surface only the best, most tailored work.
How he actually writes with AI: “with,” not “for”
He distinguishes between AI that writes for you (which no good writer he knows endorses) and AI that writes with you (which many serious writers, including himself, do regularly).
His process:
He speaks ideas out loud during walks, then uses AI to transcribe and turn them into outlines or prose.
He asks the AI to identify weak points, boring sections, unclear transitions, and stories that need more detail—getting instant feedback as a thinking partner.
For character development in his current piece (a personal essay about becoming a Christian), he asks AI to teach him character theory from Hollywood and literature, then interviews him about his characters to get words on the page.
He emphasizes that the back-and-forth dialogue with AI is more generative than working alone.
New kinds of writing AI might enable
He speculates, crediting friend Justin Murphy, that AI could spawn entirely new literary forms, much like the camera obscura and perspective grids transformed painting during the Renaissance.
One example: Mad Libs-style stories that can be personalized for each reader—a child interested in the Denver Broncos gets a different version than a child interested in ballet—while maintaining the same core narrative structure.
The idea that one piece of writing is read identically by everyone may change.
Building a custom AI writing style
He is creating a custom project folder that defines his writing style through:
10 bullet points on what his writing should be like and 10 on what it shouldn’t.
Training data: examples of his best writing with explanations of why they work.
His note-taking is shifting: he now writes notes for AI to read as much as for his future self, creating large, information-dense documents that AI can parse as context.
As context windows grow (from 30,000 words now to potentially millions), he plans to feed entire archives of emails, meeting notes, and writing into the AI so it can work with deep knowledge of his life and style.
How he thinks with AI
He finds “jamming” with LLMs more useful for discovering new ideas than talking to almost anyone except his closest friends.
Microsoft CEO Satya Nadella said his new workflow is “I think with AI and work with my colleagues.”
A friend with ~2,000 employees told his executive team that an hour with an LLM is more useful than about 70% of conversations with them.
His favorite mode is Grok’s “argumentative mode”: he states a high-conviction belief and asks the AI to challenge it, point out flaws, and argue back. They go back and forth in voice mode, and he finds it more productive than arguing with a friend because there’s no relationship cost.
He can modulate the AI’s behavior—asking it to be supportive and help shape ideas, then switching to adversarial mode to stress-test them.
At the end of a session, he asks for a summary of key points, pushback, and questions to think about next time.
AI as memory and context
Humans have terrible memories. He tells a story about taking a date to a steakhouse after she’d told him twice she doesn’t eat red meat.
AI will excel at remembering conversations from weeks, months, or years ago and providing instant context.
He recommends Granola AI for meeting notes: it transcribes meetings, auto-generates summaries, and lets you search the transcript to recover specific details.
At scale, AI could scan every email and memo in a company, giving a CEO comprehensive oversight that no human could achieve—his friend Daresh Patel predicts “AI Sundar” could replace all 30,000 middle managers at Google with copies executing a single coherent vision.
Managers are more enthusiastic about AI than rank-and-file workers because the workflow (set vision, delegate, iterate on feedback) mirrors what they already do—just without the interpersonal overhead.
Models are diverging, not converging
Six months ago, all LLMs felt roughly the same. Now they’re differentiating:
Claude 3.7 (Anthropic): focused on coding.
GPT 4.5 (OpenAI): focused on qualitative writing.
Grok (Twitter/X): focused on free speech and up-to-date information (ChatGPT’s knowledge cutoff is around September 2023; Grok uses search for current events).
He expects models to become even more distinct over time, developing different personalities and specializations.
Hallucinations: still a risk, but overstated
LLMs are good at things without wrong answers (brainstorming, editing, planning) but bad at precise information retrieval (exact quotes, specific facts).
He learned this the hard way: he asked GPT-4 for a John Steinbeck quote about food, built an entire article and video around it, published it to 30,000 people, and received emails pointing out that Steinbeck never wrote it. The AI fabricated the quote entirely.
His rule: never pass along an LLM’s factual claims without verification. But for getting a general sense of a topic—where you have enough background knowledge to spot obvious errors—AI is incredibly useful.
Benedict Evans’s framing: “LLMs are really good at things that don’t have wrong answers, but really bad at precise information retrieval.”
How he uses each model—a practical guide
ChatGPT 4.5: his core creation model. Good at writing with voice, cleaning up voice transcriptions into prose, and organizing information. Output is a ~6.5-7/10 in quality but improving fast. He uses it for drafting, outlining, and team writeups.
OpenAI Deep Research (powered by the unreleased o3 model): his “iPhone moment” with AI. He uses it for consuming information—in-depth reports on specific questions tailored to his exact interests. Takes 12-20 minutes but produces highly personalized, well-sourced reports. Best-in-class deep research tool.
Claude 3.5/3.7: sounds the most human in its writing. Excellent for generating charts and tables from natural language descriptions—useful for making arguments visually in a piece.
Grok: the most personality. He uses it for getting simple analogies for complex topics, arguing in voice mode, and as a background tutor while reading (asking questions without getting distracted by his phone). Patrick Collison (Stripe CEO) uses it this way.
Perplexity: best for factual queries with clear sources. Good quick-answer tool, but its deep research feature isn’t as good as ChatGPT’s.
Eleven Labs: speech-to-text for transcribing podcasts and other audio (free, 10-minute turnaround, far cheaper and better than Rev or Descript a year ago). Also used to clone his voice for audio corrections without re-recording.
Granola AI: meeting notes that auto-generate summaries and are searchable.
Whisper Flow / Super Whisper: speech-to-text for dictating ideas while walking around. Far more accurate than Siri, learns his vocabulary and punctuation preferences over time.
Recommendations for skeptics and learners
If you’re skeptical of AI, his number one recommendation is to try Deep Research with a well-crafted prompt on a topic you know enough about to evaluate the output. He predicts it will change your mind about what’s possible.
Free models are roughly 6 months behind the cutting edge. To genuinely assess AI’s capabilities, you need to pay for the latest models.
The mark of good thinking is challenging your own beliefs—so if you’re bearish on AI, use it seriously and well to test that position.
He compiled all his prompts and processes into a free PDF available at the link in the description.
Where to go from here
He recommends his interview with economist Tyler Cowen on writing with AI, which covers additional topics like reading with AI, AI’s influence on academia, and why secrets become more valuable in an AI-driven world.
He invites questions and reactions in the YouTube comments and plans to be highly engaged there.
