Wang has a friend named Tyler. He sees Tyler maybe once every couple of months — drinks in DC, an afternoon in New York. In between visits, Wang keeps a running list of questions in his notes app: things he is wondering about, half-formed observations he wants Tyler's reaction to, intellectual loose ends. When the next visit comes, Wang shows up and bombards him with a couple of things that I've been thinking, and Tyler somehow expects this and absorbs it.1
This Tyler is Tyler Cowen. Economics professor at George Mason. Author of an enormous corpus across economics, food, travel, philosophy, AI, culture. The kind of generalist who can give you a useful reaction to a Beethoven string quartet, a Stendhal novel, a Mongolian noodle, a Catholic theological dispute, a chess opening, and a forecast about the next decade of geopolitics — sometimes in the same conversation. The Tyler-experience for someone like Wang is the experience of having a generalist sparring partner of unusual depth, available occasionally, with whom rapid-fire intellectual exchange produces sharper thinking than the person could have produced alone.
Then ChatGPT arrived, and Wang noticed something odd about it.
The voice and the intellectual style of the model felt familiar.
Wang's claim is specific. Tyler Cowen's thought has influenced a lot of how ChatGPT communicates, and this is on the record, I believe, and this is kind of established.1 The claim refers to the well-known fact that Cowen's writing — across his blog Marginal Revolution, his books, his Bloomberg columns, his Conversations with Tyler podcast transcripts — is part of the publicly-available training corpus that LLMs learn on. Cowen has been extraordinarily prolific in public-facing prose for two decades, in a recognizable analytical voice that emphasizes economic reasoning, comparative observation, willingness to hold counterintuitive positions, and a particular kind of rapid-fire associative intelligence. That body of work has been absorbed into the training data of the major LLMs, and the models have learned, among many other things, how Tyler Cowen thinks.
The model is not Tyler Cowen. The model is a probabilistic next-token predictor trained on a much larger corpus that includes Cowen alongside millions of other writers. But the style of the model's outputs — particularly when prompted on the kinds of questions Cowen would address — has structural fingerprints of Cowen's voice. The model has absorbed the moves: the comparative angle, the unexpected reference, the willingness to take a position quickly, the slight asymmetry between the question being asked and the question the model chooses to answer.
Wang noticed this. And once he noticed it, he started using the model the way he uses Tyler.
The mechanics of the substitution are simple enough to describe but interesting enough to be worth slowing down on.
Wang has built, over years of friendship and mentorship, a cognitive habit around Tyler-conversations. He thinks of questions in a particular way because he is anticipating what Tyler would do with them. He keeps the running list. He prepares some context. He shows up ready to spar — to push back, to extend, to reorient the question if Tyler's reaction reveals it was the wrong question.
ChatGPT doesn't have Tyler's actual mind. It has, at best, the Tyler-flavored layer of its training. But for the purposes of Wang's specific cognitive practice — get a Tyler-like reaction so I can spar with it — the layer is sufficient. If I really wanted to have an instant Tyler-like reaction to a novel that I've just read or a piece of music that I've just heard, I just ask ChatGPT, and it is able to give me some really good answers.1 The model produces a reaction that triggers the same mental sparring move Wang would have made with Tyler in person. Wang's cognition continues from there: he pushes back, extends, adjusts. The model, like Tyler in person, gives him something to think against.
The substitution is incomplete in important ways. Tyler-the-person remembers earlier conversations; the model does not (within a session) and even with memory features cannot reconstruct the depth of accumulated context Tyler has on Wang specifically. Tyler-the-person can introduce Wang to other people, push him into specific writing opportunities, model intellectual virtues over a multi-year arc. The model does none of this. The substitution captures one particular feature of Tyler — generalist sparring reactions in real-time — and substitutes a much weaker version of Tyler in all the other dimensions, mostly absent.
But for the use case Wang is describing — I just read Stendhal, I want a smart reaction so my own thinking can sharpen against it; I just heard Mahler, I want a musicological gloss so I can extend my own reaction — the substitution works. The model is not Tyler. It is a Tyler-like creature, available at any time, infinitely patient, free at the margin. Wang treats it as such.
The example Wang gives of his first really good ChatGPT experience is a small one but it captures the texture of the practice.
Why did the Spanish Catholic Church develop such a virulent inquisition whereas the Austrian Catholics had no inquisition?1
The question itself is the kind Wang would have asked Tyler. It is comparative. It pulls together two things normally treated separately (Spanish Catholicism, Austrian Catholicism) and asks for a structural account of their divergence. It is not a question with a quick Wikipedia answer; it requires synthesizing across history, theology, political economy, ethnology. It is a question whose value is partly in the texture of a good answer — what considerations the answerer brings forward, which factors they weight, what comparative cases they invoke.
Wang has the question because he has spent time in both Catholic worlds and felt the difference. Spanish Catholicism feels gloomy and slightly violent and a little bit dark; Austrian Catholicism feels bright, joyful, resplendent, much more crimson rather than very black.1 The question is not abstract. It has been generated by Wang's own observational experience of two religious cultures and the felt-sense that the difference must be explainable.
He asked the model. The model produced an answer that felt right.1 We do not get the answer in the transcript — the relevant point is not what the model said but that the conversation succeeded as the kind of intellectual exchange Wang had previously been having with Tyler. The model produced a structural account that addressed the question on its own terms, mobilized appropriate comparative material, and gave Wang something to push against.
The Spanish-vs-Austrian question is the canonical instance of what the AI-as-Tyler-substitute can do well. Comparative, structural, requiring generalist synthesis, not requiring deep specialist accuracy. The model excels at exactly this register. It is less reliable on questions that require precise factual recall (where hallucinations bite) or on questions that require deep specialist judgment (where the model averages across sources rather than reasoning from expertise). The Tyler-substitute use case is in the middle layer — questions that benefit from broad intellectual range and structural thinking, without requiring expertise that exceeds what the model can plausibly synthesize.
Perell's framing of what Wang is describing is worth dwelling on. Most users approach AI by asking it for answers. Wang approaches AI by asking it for reactions to push against. The reframing changes the cognitive role of the AI.
In the answer-mode, the AI's output is the destination. The user receives the output, accepts or rejects it, moves on. The user's cognitive activity is filtering rather than producing.
In the sparring-mode, the AI's output is the instigator. The user has a half-formed question or hypothesis. The AI produces a response that triggers the user's own cognitive engagement — agreement that extends the thinking, disagreement that sharpens it, partial acceptance that splits the question into newer questions. The user's cognitive activity is producing rather than filtering.
The sparring architecture is what Wang has imported from his Tyler-relationship into his AI use. That's the process of thinking through things. That's the process of trying to figure out what are the right questions and what is the right approach, what is the right observational style.1 The AI is not telling Wang what to think. It is providing a worthy interlocutor whose responses are good enough that Wang's own thinking is sharpened by reacting to them.
The architecture has a subtle prerequisite. The user has to already have something to spar with. The hypothesis, the half-formed observation, the question generated by their own engagement with material the AI has not seen. Without that input, the user is just asking the AI for an answer, and the AI produces an answer, and the cognitive engagement is the answer-mode rather than the sparring-mode. The sparring requires the user to bring something to the conversation, and the more interesting the user brings, the more useful the model's reaction becomes.
The Tyler-substitute framework converts cleanly into a practice principle. Use AI for the generalist-reaction function, in conversations you initiate with material you have already engaged with. The good prompts in this register are not teach me X but I have been thinking about X, and here is what I noticed, what would you say back? The model becomes useful not as an oracle but as an interlocutor. The user's pre-existing engagement is what the model's reaction can sharpen.
The framework also implies that AI use has a generalist substitute layer and a specialist substitute layer, and they should be evaluated differently. For generalist sparring (comparative history, music, art, philosophy, structural analysis), the model is genuinely useful and the Tyler-substitute mode applies. For specialist work (technical implementation, primary-source research, expert judgment in domains where the model averages across mediocre sources), the model is much less useful and the user should not be using it the same way. Confusing the two layers — using the model for specialist work in the generalist mode — is where the hallucination problems and the bad-citation problems concentrate.
1 Wang's account of the Tyler-substitute use of AI runs from lines 444-462 of the transcript. Specific claims and locations: "Tyler Cowan thought has influenced a lot of how ChatGPT communicates" + "on the record, kind of established" (line 446); "the way that I treat ChatGPT is as my friend Tyler, who I maybe see once every couple of months in DC or in New York" (line 448); "instant Tyler-like reaction to a novel that I've just read or a piece of music that I've just heard" (line 450); running list of questions for Tyler in notes (line 452); Spanish vs Austrian Catholic Inquisition question + Spanish "gloomy slightly violent dark" vs Austrian "bright joyful resplendent crimson" + first really good ChatGPT experience (lines 456-460); "Tyler-like creature in my pocket" (line 460); "process of thinking through things... finding the right questions, right approach, right observational style" (line 462); reaffirmation of the verboten on flat sentences in writing (line 462).
The framework has tensions with both the AI-skeptic and AI-optimist positions.
Against the AI-skeptic position that LLMs cannot produce useful intellectual exchange because they are stochastic parrots without understanding, Wang's framework offers a counter-example. The Tyler-substitute use of AI is, by Wang's account, genuinely cognitively productive — it produces the same kind of thinking-sharpening that human conversation with a generalist intellectual produces. The skeptic position has to either dismiss Wang's experience as illusion or concede that, at the level of cognitive function rather than ontology, the model is doing something useful regardless of whether it "understands."
Against the AI-optimist position that LLMs can replace human intellectual exchange entirely, Wang's framework also offers a check. The Tyler-substitute is for specific use cases — generalist sparring on material the user has already engaged with. The model is not a substitute for the long-arc relationship Wang has with Tyler-the-person, for the deep specialist work Wang does on China, or for the writing that constitutes Wang's voice. The optimist position has to recognize that the AI-as-intellectual-companion role is genuine but bounded.
The harder tension is over the Tyler-influence claim itself. The empirical fact that Cowen's writing has influenced ChatGPT is plausible (Cowen is widely read and represented in training corpora) but also somewhat anecdotal and unverified at the precise level Wang is claiming. The vault should hold this as Wang's reading rather than as an empirical fact, and remain open to the possibility that the Tyler-flavor Wang detects is partly a function of Wang's expectations and partly a function of the model's training rather than a clean attribution.
The framework converges with Wang's own AI as Consumption Enhancer, Not Writing Tool page (the companion concept). Both pages are products of Wang's overall AI framework: AI is excellent at the layer wrapped around the user's experience (consumption-enhancement, generalist-sparring) and verboten at the layer that constitutes the user's cognitive output (sentence-generation, voice). The two pages should be read together as the two sides of Wang's framework.
The framework also tensions with the broader AI-collaboration corpus in the vault. Beta Editor Review Skill and adjacent pages treat AI as a simulated audience or wind-tunnel for testing one's own ideas. Wang's Tyler-substitute framing is similar but more specific — the AI is not a generic audience but a specifically-flavored interlocutor whose particular intellectual disposition (generalist, comparative, willing to take a position) makes it useful in a way a generic-audience framing does not capture. The convergence is on the underlying claim that AI is a thinking-partner; the divergence is in how specifically the partner can be characterized. Wang's framing implies that AI-as-thinking-partner is most useful when the user has a specific, characterized interlocutor model in mind, not when they are using the AI as a generic foil.
The Tyler-substitute framework reaches into psychology (the cognitive function of having a sparring partner), creative practice (the role of intellectual exchange in producing original work), and behavioral mechanics (the specific architecture of what makes a conversation productive).
Psychology: Identity Architecture and Defense Hub — the importance of trusted interlocutors in the development of mature thinking is well-documented in developmental psychology. Adolescents and young adults develop intellectual identity partly through extended sparring with a few mentors whose perspectives they argue with, internalize, and eventually push beyond. The handshake reveals that AI-as-Tyler-substitute is operating in the same psychological space — the user is engaging with a characterized interlocutor whose responses sharpen their own thinking — but with two important differences: the interlocutor is not real (which may matter for identity formation in ways the framework does not address) and the interlocutor is infinitely patient (which removes the social stakes that sometimes make human sparring more cognitively engaging). The implication: AI-substitute interlocutors may be useful for thinking-sharpening but may be inadequate substitutes for human mentors in identity-formation work, and the user should know which function they are using AI for.
Creative Practice: Voice Cultivation Through Stylistic Models — the Tyler-substitute is, in a sense, a way of importing Cowen's stylistic intuitions into the user's own thinking process via the AI's training-influenced outputs. The handshake reveals a non-obvious symmetry: the writer who deliberately copies New Yorker articles and the user who deliberately uses ChatGPT-as-Tyler-substitute are doing structurally similar things — exposing themselves to a particular intellectual style in order to absorb its operations into their own thinking. The difference is the medium (handwritten/typed reproduction vs. AI conversation) and the granularity (sentence-level for copying, prompt-response-level for AI). Both are deliberate exposure to a style with the goal of cognitive transformation. Both are subject to the same caveat: choose your sources strategically, because the style you absorb shapes the thinking you become capable of.
Behavioral-Mechanics: Behavioral Mechanics Hub (specifically pages on questioning techniques and elicitation) — the sparring architecture Wang describes is structurally similar to behavioral-mechanics protocols for generative questioning. The questioner brings specific material; the responder produces a reaction; the questioner pushes back; the productive thinking happens in the back-and-forth rather than in either side's individual contribution. The handshake produces an actionable claim: users who want to extract maximum value from AI conversation should approach it as a behavioral-mechanics protocol — bring specific input, ask for specific reactions rather than open-ended assistance, push back on the AI's responses rather than accepting them, iterate until the conversation produces something the user could not have produced alone. This is a learnable skill, not a default behavior, and most AI users are not yet practicing it.
Eastern Spirituality: Sadhana Practice Hub — the running-list-of-questions practice Wang describes (questions for Tyler accumulated in his notes app, then unloaded when they meet) has a structural cousin in the spiritual practice of holding questions for the next encounter with one's teacher. In both cases, the questioner accumulates questions over time, lets them ripen, and presents them when the encounter happens. The handshake reveals that the practice is not about the immediate availability of the answer; it is about the discipline of articulating the question well enough that it survives the wait. The questions that survive accumulation and articulation tend to be better questions than the questions that get asked impulsively in the moment of confusion. The implication for AI use: even though the AI is always available, users may benefit from imposing artificial accumulation — keep questions for the AI in a running list, let them ripen, ask in batches rather than as they arrive — because the discipline produces better questions and better cognitive value from the eventual conversations.
The four handshakes converge on a single claim: AI-as-sparring-partner is one specific application of a much broader principle about how thinking is sharpened through structured exchange with a characterized interlocutor, and the value of the practice depends on the user's discipline in bringing material, characterizing the interlocutor, and treating the exchange as iterative rather than transactional.
The Sharpest Implication
If the Tyler-substitute framework is correct, then the public conversation about AI's intellectual role is operating at the wrong level of generality. The question is not can AI replace human intellectual exchange (the answer is partly yes, partly no, depending on which functions of intellectual exchange you mean). The question is which specific intellectual functions can which specific kinds of AI substitute for, and how much of the user's existing cognitive infrastructure does each substitution require. Wang's case shows that AI can substitute for a specific friend's specific function (generalist sparring on material Wang has engaged with) for a user with specific cognitive infrastructure (an analytical writer with a pre-existing relationship to the friend). The substitution does not generalize: AI cannot substitute for the friend's other functions (mentorship, introductions, identity-formation) and cannot substitute for any function for users who lack the relevant cognitive infrastructure. The implication for AI policy and pedagogy is that the right questions are narrow and specific, not broad and abstract. Public conversation that treats AI's intellectual role as a single thing — AI is good for thinking or AI cannot really think — is missing the actual differentiation that determines whether any specific use produces cognitive enrichment or impoverishment.
Generative Questions
Henrik Karlsson adds a third position to this page's framing of AI use, structurally distinct from Wang's two frames (Cowen-substitute sparring partner; consumption-enhancer for reading). Karlsson uses AI for targeted craft-mechanical tasks — vocabulary lookup, grammar verification, sentence-variation generation, and bad-habits monitoring — comprising about 1% of his published words.2
The four micro-uses Karlsson articulates:
1. Lexical lookup for second-language gaps. As a non-native English speaker, Karlsson uses the model as a thesaurus-on-demand: describe the meaning, get the candidate word. "I might do things like — because I'm a second language, I don't know English all that well. Sometimes I'll describe a word and they'll give me the word because I don't know it."2
2. Grammar verification. Sanity-check on uncertain constructions. "Or I'll ask, 'Is this grammatical?' or whatever."2
3. Sentence-variation generation. When stuck on a phrasing: "Sometimes I'm not happy with a way something is phrased and I'll be like, 'Give me 10 versions of this with different words to just get me out of my head and see different possibilities.'"2 The model is not asked to pick a version — it's asked to display a range, which the writer then selects from or uses as a launchpad.
4. Bad-habits monitoring. Karlsson stores his known writing-flaws in a prompt and has the model flag them during edit. "I have a prompt where I ask it to go through, 'These are some words I overuse, can you point out when I use them?' These are other things I want to think about, just to help it remind me so I don't have to keep all the things I should look out for in my head."2 The model offloads vigilance.
The Karlsson framing differs from Wang's Cowen-substitute frame in a structural way. Wang uses AI as interlocutor — conversational partner for thinking-through-problems, with characterized intellectual style influencing the exchange. Karlsson uses AI as tool — discrete craft-mechanical tasks with no expectation of intellectual exchange. The model is not Karlsson's thinking partner; it's his lookup, his variation engine, his vigilance store.
Karlsson is also a power user of LLMs — pre-ChatGPT access (special OpenAI beta from 2021), deep technical familiarity. His position is not knee-jerk avoidance. He's enthusiastic about the technology while specifically routing around it for substantive writing-work. "I've been somewhat underwhelmed in what they've delivered so far. I've seen interesting images, interesting music, but I've not yet seen any good literature be made. They are incredible tools for research."2
The third-position-frame this opens: AI can be used at multiple resolutions of substance, and the appropriate use depends on the user's writing-practice.
Karlsson also articulates a deeper concern this page hasn't addressed: LLMs as standardization-pressure on the prose-field. He argues that even non-LLM-using writers face ambient smoothing-pressure as LLM-edited text circulates more widely. The defense, in his framing, is to be wilder with grammar — deliberately import constructions from other languages (Turkish, Persian, Swedish) to maintain distance from the statistical center the models pull toward. "LLMs are just another force for standardization. If English is poised to become the global language, we should make an effort to bring in all the weird innovations from all the other languages."2 This is a structural-ecology claim, not a personal-use claim — and it adds a layer to the business framework neither Wang frame addresses.
The full three-frame picture: Wang's two frames optimize for what AI does well (interlocutor; consumption-helper); Karlsson's frame optimizes for minimal AI presence in production combined with defensive practice against ambient standardization. All three positions are coherent. They suit different writers and different writing-philosophies. The implication for this page: AI-as-Cowen-substitute is one of several legitimate counter-positions to mainstream AI-as-creative-collaborator framing, and the full counter-framework needs all three to be complete.
See LLM as Standardization Pressure — Grammar Wildness as Defense for the full Karlsson treatment.