Tyler Cowen is a rare expert in both AI and writing, and he uses AI daily as a “secondary literature” companion to learn faster, think deeper, and prepare for podcasts and writing. He argues that most people are using AI wrong—asking generic questions and treating it as a substitute for effort rather than a complement to it. The real power comes from using the best available models, asking specific and detailed questions, and investing your own time to generate rich context.
How Cowen uses AI in his daily work
He uses AI primarily when reading, treating it as a new form of secondary literature
When preparing for a podcast with a British historian on Richard II and Henry V, he used to order 20–30 books; now he orders 2–3 and interrogates the best LLMs for context, which he finds more efficient and more fun.
He rereads Shakespeare and Wuthering Heights while asking the AI questions about each chapter—what happened, what puzzles exist, what to ponder more—which makes him a more active reader and improves his understanding.
He finds AI especially well-suited for Shakespeare because the plays are endlessly rereadable and the major models have absorbed both the texts and the secondary literature.
He uses AI to fact-check his writing in areas where he lacks expertise
For a column on declassifying classified documents, he asked o1 Pro for legal background so he wouldn’t be “an idiot on the topic,” though he didn’t use the AI’s output directly.
He does not let AI write for him
He wants his writing to remain distinctly his own—weird, cryptic, layered, and parable-like—because that is what makes it valuable and enjoyable to readers.
He ran his writing through AI once on a friend’s advice to find parts people would find obnoxious, and it correctly identified a condescending passage, which he changed.
He uses AI to generate tables and visualizations
When text-based information isn’t computing, he asks AI to make tables or graphs, which has increased his information intake by at least 10x.
Hallucinations are no longer the main problem
Hallucinations have declined by more than 10x in the last year and will continue to fall.
When you are the interviewer (not the answerer), modest hallucinations don’t trip you up—you just need context.
For anything you plan to use or publish, you must double-check it, just as you would with any human source. Cowen has a research assistant fact-check everything in his books.
The best reasoning models hallucinate much less than most people expect.
How to prompt well
Ask specific, practical questions rather than generic ones
Instead of “What should I ask person X?”, ask about the details of historical examples—e.g., “What was special about Tyndale’s translation of the Bible and how did his patrons feel about it?”
Specific questions lead to “out of left-field” questions naturally, whereas asking the AI to generate questions produces bland, “normy” results.
Use long, detailed initial prompts, especially with voice dictation
Cowen dictates his first prompt for 1.5 to 3 minutes to get substantial context in, then follows up with shorter questions.
He finds that the AI is quite good at sorting what’s truly important from a long prompt.
Think of AI as a stack of interacting agents, not a single box
Some people use a cheaper model (like o3 mini) to write the prompt, then feed it to the full model.
Use multiple AI models to bounce ideas off each other, correct each other, and grade each other—simulating a decentralized “republic of science.”
It’s better to ask one thing 10 times than 10 things once
Long multi-part prompts cause the AI to degrade on later questions; planned follow-ups work better.
The three layers of knowing AI
Layer 1: Are you using the very best systems? Some cost money (e.g., o1 Pro at $200/month), but Cowen considers this a good investment for most people who aren’t very poor.
Layer 2: Do you understand that AI improves itself continuously through reinforcement learning and other techniques? Many people are impressed by the present moment but don’t grasp the rate of improvement.
Layer 3: Do you have a vision of AI evolving its own decentralized institutions—its own “republic of science” with markets, peer review, and self-correction? This is where most of the future value lies, and most people aren’t even at level one.
How AI is changing writing and books
Books about the near future are becoming obsolete before they’re published
A book takes ~4 years to write and publish; AI is changing the world too fast for “predictive books” to make sense.
Cowen now covers fast-moving topics through ultra-high-frequency writing: Substack, blogging, Twitter.
Books about the distant frozen past are still viable, as are deeply human books
He is writing fewer books and may stop after his current one on mentoring.
Books based on personal experience, fieldwork, interviews, and subjective perspective—like memoirs, biographies, and mentoring—cannot be truly replaced by AI because readers want the human behind them.
Generic corporate writing is the first category to be fully replaced by AI.
The “question box” may replace many books
Rather than reading a packaged book, people may just ask the AI any question the book would answer. Writing must therefore be more interesting than what the question box provides.
DeepSeek is the creative, “wacky” model
DeepSeek (from China) is less manipulated, less bland, better at poetry and emotion, more romantic, and more uneven than other models.
It hallucinates more, so it shouldn’t be used for research, but it’s invaluable for creative and sensory descriptions—e.g., “a glorious description of what it’s like to eat a mofongo.”
Cowen recommends using DeepSeek periodically (he calls it “China boss”) to remember what AI is truly capable of and to avoid thinking of AI as uniformly bland.
DeepSeek is open source, and a version now exists in Perplexity, though Cowen worries they may have made it less weird.
AI tools in Cowen’s stack
o1 Pro (OpenAI): Best for queries; costs $200/month; takes 2–4 minutes per answer.
Deep Research (OpenAI): Best for 10-page reports; uses o3 under the hood; Cowen considers it the single most impressive thing humans have built, though he doesn’t use it much for his own work.
Claude: The best writer among the models; thoughtful, philosophical, dreamy, flexible, versatile. The next version is expected to be extraordinary.
DeepSeek: Creative, emotional, wacky. Use it for inspiration and sensory richness, not for facts.
Gemini (Google): Best for handling very long or thick documents, legal work, and multimodal tasks. Its YouTube integration is excellent—Cowen can feed it 15 videos and ask questions about all of them.
Grok (xAI): Useful for quickly fact-checking tweets.
Meta/Llama: Strong open-source models, important globally, available on WhatsApp.
Perplexity: Has replaced most of Cowen’s Google use; completely up to date, excellent for finding citations, and “asymptotically good” in his view.
Information (newsletter): $400/year; keeps him abreast of new AI developments, crypto, and tech.
AI in the classroom
Cowen teaches a PhD class with no assigned textbook; students must subscribe to a good AI service instead, which costs less than a textbook.
Students are required to use AI for their papers and report how they used it; the goal is to make the paper as good as possible.
Most students want to use AI but have never been taught how—Cowen considers this a scandal in academia.
Students who are cheating (when not allowed) know more about AI than their professors.
Homework norms need to shift toward oral exams, proctored in-person exams, and AI-integrated assignments.
Why human networks and secrets matter more, not less
As AI makes public information nearly free and universally accessible, the value of private information—secrets—goes up.
Secrets are “things you know about the world that other people don’t know,” and they become more valuable because public information is now worth very little.
Gossip and personal context are emotionally and practically potent in ways AI cannot easily replicate.
Social networks become dramatically more valuable
The most productive people could be 50x to 5,000x more impactful because they command an army of AI “servants,” but to mobilize projects they still need human help—venture capitalists, philanthropists, collaborators.
Networking, traveling, and meeting people in person are more important than ever.
Experts still outperform AI in some domains because they know secrets
A top physicist can ask the one question that really matters, drawing on seminars, personal relationships, and life-rich context that the AI cannot quickly access.
This is also what a good mentor provides—unique context and secrets.
How AI changes what it means to be a writer
Writers will need to personalize more
A personal anecdote—like eating the best lobster of your life on a riverboat in Kerala, cooked by a crew member from a local catch—is something AI cannot generate and readers connect with.
Video and podcasting will persist because people want humans, not fake humans
AI can already clone Cowen’s voice indistinguishably, and YouTube is rolling out AI dubbing in every language, but Cowen believes people will still want real humans in video content, at least for the near term.
Speaking and writing for the AI audience
Cowen writes partly so that AI will have a better model of him than most humans do—a form of “intellectual immortality.”
He is writing ~20 blog posts about parts of his life not already public (e.g., ages 4–7 in Fall River, Massachusetts) so that future AI can write a comprehensive biography of him at very low cost.
When giving talks, he considers that there may be 50 AI note-takers in the audience, so he speaks with the AI’s summary in mind.
Career advice in an AI-rich world
Two universal pieces of advice: get more and better mentors, and work every day at improving the quality of your peer network.
These were always good advice but are more valuable now.
PhDs are riskier investments; the world will need fewer PhDs and more people who know how to manage AI systems.
Career trajectories are becoming less formulaic—the old predictable paths (Yale → McKinsey) are disappearing.
Whatever the best model costs per month, it is likely a good investment for anyone who can afford it.
The future of large context windows
Gemini now has 2 million tokens; Cowen bets that by year-end there will be 10–20 million token windows.
This will make it routine to work with massive regulatory codes, historical archives (e.g., tax records from Renaissance Florence), and entire musical corpora.
A major new human project will be converting all of humanity’s knowledge into AI-usable form—scanning, digitizing, and feeding everything from national archives to guitar tablature into AI systems.
This will create many jobs and should be a focus of philanthropy.
The gap between public and private AI innovation
The most innovative AI usage is happening inside the AI companies themselves, which use AI to improve AI.
The delta between what these companies are doing and what the public can access is immense and not publicly known, but the fact that AI keeps getting better is evidence that it’s working.
