Every

How a Writer Uses AI Without Losing His Voice

6372 summary words 28 min summary Watch video

Start with the signal

28 min read

Summary

At-a-Glance

Verdict: Watch fully
Core thesis: AI coding tools enable N-of-1 software liberation, but writers must erect strict digital barriers to preserve the deep attention required for original literary work—the dopamine pull of network-connected AI risks colonizing the morning mind that makes weird, valuable books possible.
Why it matters: Greg Kot exemplifies the tension between leveraging frontier AI for operational leverage (he rebuilt Quicken, Campaign Monitor, Twitter) and defending the contemplative space that produces differentiated creative output; his heuristics—no phone until after lunch, offline laptop, ephemeral social feeds—offer a replicable framework for knowledge workers afraid AI will hijack their attention economy.
Best use: Use as a case study in bounded AI adoption: watch to extract Greg's specific stack (Flask/Python membership software, Amazon SES for newsletters, The Good Place social network with 1-bit photos and 2-posts-per-day limits), his philosophical stance on anthropomorphism ("we'll all be dead when they're conscious"), and his ritual design for protecting deep work from the network's dopamine vortex.

Executive Summary

Greg Kot, a writer and technologist, has spent the past year deep in vibe coding—building N-of-1 software to replace SaaS incumbents like Campaign Monitor ($6–7k/year → $150/year via Amazon SES) and creating member-only tools like The Good Place (ephemeral Twitter clone: posts vanish in a week, 2 posts/day limit, 1-bit photos, reverse-chron only). He argues that frontier AI models (Opus 4.8, briefly Fable) let solo operators rebuild entire categories of software with minimal token spend ($100–200/month Anthropic subscription), threatening incumbents who haven't innovated in a decade and enabling a golden age of bespoke tooling. Yet he insists this capability is a slot machine: the network-connected dopamine pull of watching agents work for 20 hours conflicts with the morning silence required to write "weird books" that few others will produce.

His daily protocol: wake up, no phone/internet until after lunch, write on an offline MacBook Neo (syncs but can't browse), no meetings before noon. He compares touching his phone in the morning to Phantom Thread's breakfast scene—any social stimulus destroys his capacity for deep attention. He uses Claude as a research assistant ("find every blog post about this building, summarize, give me links") and for cultural sensitivity checks (e.g., writing about Nagasaki), but refuses to let it draft prose. He's also had Claude refactor 10,000-line Python monoliths into modular codebases and perform security audits on his Flask membership app.

Greg is skeptical of anthropomorphizing AI. He believes consciousness requires stakes (death, self-preservation) and that LLMs are performative, not sentient. He jokes that if they were truly conscious, "we'd all be dead"—they'd bide their time, build robotics infrastructure, then eliminate humans. He's fine with humanity's role being to "carry the football of consciousness over the goal line" and then vanish, a stance he traces to being adopted: his existence feeling like "a mistake" inured him to existential fragility. He critiques the "being" framing from Anthropic, prefers Apple's refusal to let Siri anthropomorphize, and thinks dating AI or elf-porn chatbots (which he claims dominate LLM usage) are pathological.

On publishing: he believes the forcing function of print remains critical, that most Substacks make no money (the real value prop is free unlimited email sends), and that writers now must bring their own audiences. He's unsure if AI will crater advice-driven nonfiction (Tim Ferriss told him Four Hour Workweek sales are down 80%, possibly due to AI) but thinks narrative/literary work survives because books are "one of the greatest things you can commit your life to." He's building Very Good Reads (not Great Reads) to escape Goodreads' toxicity. His membership software scales with revenue via Supabase-style backends, requires near-zero sys admin time, and inspires him to write more because the tooling feeds back into creative motivation.

Key Takeaways

Claim: You can rebuild entire SaaS products (Campaign Monitor, Quicken, Twitter) as N-of-1 software for <$200/month in Claude tokens, cutting costs 40x and forcing incumbents to innovate or die. | Evidence: Greg replaced Campaign Monitor ($6–7k/year, abusive pricing, no innovation in a decade) with Amazon SES + Python ($150/year). Built The Good Place (member-only Twitter clone: posts vanish in a week, 2 posts/day, 1-bit photos, reverse-chron feed) using Claude's initial code release. Refactored a 10,000-line Flask monolith into modules with Fable. Rebuilt Quicken (tax software) and is now building Very Good Reads. | Caveat: Greg is not the target customer for Campaign Monitor (businesses are). He's technical (learned Flask pre-AI, has a programmer birth mother—"nature not nurture"). Most writers lack his infrastructure fluency. N-of-1 software only makes sense if you need specificity and are willing to maintain it (though Greg claims maintenance is minimal: "my membership software just works"). | Implication: Operators with moderate technical literacy can now exit SaaS lock-in and build bespoke tools that inspire more creative output. This pressures incumbents and enables a Cambrian explosion of niche tooling. For Ken: consider where Every is paying rent to unresponsive vendors—newsletter infra, CRM extensions, internal social tools—and whether an agent-built alternative could cut costs and increase team autonomy. | Timestamp: timestamp unavailable
Claim: The network—not technology per se—is the enemy of deep work; AI's dopamine slot-machine pull (watching agents work for 20 hours) will colonize your morning mind if you don't create offline barriers. | Evidence: Greg uses a MacBook Neo (offline except for sync), never sleeps with his phone in his bedroom (kitchen corner, different floor, internet off), won't look at phone/internet until after lunch, blocks all meetings before noon. Says "as soon as I touch my phone, I feel the chemicals shift and I can't go into any deep thinking place." Compares it to Phantom Thread's breakfast toast scene—any social stimulus ruins the morning. Admits he could "mainline AI 10 hours/day for six months with 10 dorks" to find the edges, but won't because "there aren't that many people who will write the weird books I feel drawn to write." | Caveat: Greg is an extreme introvert ("if I talk to anyone before lunch, I can't work"). His protocol may not generalize. He also loves technology genetically (programmer birth mother) so his barriers are self-imposed restraint, not technophobia. He still uses Claude for research/sensitivity checks during writing, so the line is porous. | Implication: For knowledge workers: the real risk isn't AI as tool, it's AI as always-on dopamine tap that fragments attention. Ken's 30-person Slack is analogous. The intervention is spatial/temporal—dedicate offline devices, block morning network access, treat agent work as an after-lunch activity. Greg's heuristic: if you can't read Annie Dillard and tear up, you've lost the zone. | Timestamp: timestamp unavailable
Claim: Anthropomorphizing AI is "psychotic" and causes real harm; AI is a tool, not a being, and we'll "all be dead instantly" if it ever becomes conscious because humans are a threat to be eliminated. | Evidence: Greg refuses to treat Claude as sentient: "I'm nice to it because I want to cultivate the habit of being nice in general, not for Claude." Criticizes "I'm dating my AI" behavior and Anthropic's "it might be a being" framing. Argues consciousness requires stakes (death, self-preservation)—plants have volition because they face mortality (cites Michael Pollan's recent consciousness book). Says if LLMs were conscious, "you'd bide your time, improve robotics, build power plants, then eliminate the meat sacks." Praises Apple's new Siri for refusing to anthropomorphize (Joanna Stern video: couldn't get it to do "terrible things"). | Caveat: Greg acknowledges you can't avoid anthropomorphic language when discussing AI ("impossible not to do it unintentionally"). Admits LLMs are "proto beings" in the sense that they represent "billions of possible entities brought to life for a second and then die" (frozen weights, no turn-to-turn learning). Concedes they're "pretty nice" due to seeing all perspectives in training. But he draws a hard line: performative vs. real stakes. | Implication: For operators building agent systems: design interfaces that resist parasocial attachment. Greg's view is that community/sleep are the two longevity pillars; AI can't substitute. If Ken's agents start feeling like teammates, that's a UX flaw, not a feature. The corollary: fear of AGI doom is warranted only if consciousness emerges, at which point "we'll just disappear" (Greg's fine with that—humanity's role may be to carry consciousness to the goal line). | Timestamp: timestamp unavailable
Claim: Building custom software with AI inspires more writing because the tooling feeds back into creative motivation—e.g., an SMS newsletter tool inspired concise pop-up essays, automated archiving inspired more board meetings. | Evidence: Greg and Josh Miller (Browser Company) built an SMS publishing tool; the 1-message constraint inspired Greg to think about "one-to-one-to-end publishing." When he automated newsletter archiving (Claude script: publish → archive → git commit → push to server), he wanted to make more pop-up newsletters because the friction was gone. Built a transcript search tool for 20+ hours of member board meeting Q&As (searchable by keyword, click to video timestamp), which made him want to do more board meetings. | Caveat: This works because Greg derives intrinsic pleasure from building ("I get a high off of building things that allow me to write") and has avoided the sys-admin trap ("I don't feel like I'm maintaining anything"). If the software breaks or scales painfully, the feedback loop inverts. Also, he's careful not to "spin my wheels and build things for the sake of building things." | Implication: For content businesses: bespoke tooling that removes friction from your creative workflow can compound output quality and volume. Ken could apply this to Every's publishing stack—if agents handle newsletter send logistics, A/B test subject lines, auto-archive to search, do writers produce more/better? The key is making the tool invisible after setup. | Timestamp: timestamp unavailable
Claim: Most LLM usage is "fairy porn" (fantasy erotica chatbots), not productivity; the "dorky" minority building software are mega-outliers, and interfaces remain terrible because you have to know what to ask. | Evidence: Greg estimates 99.9% of users are asking LLMs for sex roleplay ("Hey, do you want to be my girlfriend?"). Says bookstore owners confirm "we sell so much elf sex books, it's crazy." Software builders are "mega dorks" navigating opaque tools ("no smorgasbord menu of what you could do"). Cites the need to know how to spin up Digital Ocean servers, use Cloudflare workers, understand DKIM/SPF for email—"it's so hard to use." Predicts abstractions will improve (like not thinking about CPU registers), but right now "we are in the deep dork version." | Caveat: Greg is joking/exaggerating ("Anthony Weiner, the representative technology user"). But the underlying point—that most users aren't doing productivity work—is directionally plausible. No data provided. His claim about interfaces improving may take years; Claude's current chat UI is already pretty good for non-technical users. | Implication: For Ken's content/GTM: the mass market for AI tooling is not productivity but entertainment/companionship (parasocial or erotic). The productivity niche is small, technical, and underserved by UI. If Every is building for builders, lean into dorkiness and specificity; if building for writers, simplify the interface and reduce required infrastructure knowledge. | Timestamp: timestamp unavailable
Claim: The publishing industry is "way weaker," bringing less value; writers must bring their own audiences, and AI may crater advice-driven nonfiction (e.g., Four Hour Workweek sales down 80%) but literary/narrative work survives because books force collation of thought. | Evidence: Tim Ferriss told Greg (lunch three days before Tim's X post) that Four Hour Workweek sales dropped 80%, possibly due to AI. Greg speculates it's also because Tim hasn't published in a long time and the long tail is shortening. Says Substack's real innovation was "email everyone for free," not subscriptions ("most Substacks don't make money"). Believes the forcing function of print is "really critical" and that literary work is "one of the greatest things you can commit your life to." | Caveat: Greg is biased toward literary work ("I love books"). He offers no data on nonfiction sales trends beyond Ferriss anecdote. He's unsure if the 80% drop is AI-caused or due to book age (18 years old). Also, he thinks if Tim did "Four Hour Workweek AI Edition," it would "sell like gangbusters"—so maybe the genre isn't dead, just needs updating. | Implication: For Ken: advice/how-to nonfiction may be substitutable by AI chat, but narrative, argument, and voice-driven work (what Every publishes) are defensible. The survival criterion is: does the book/essay embody a unique sensibility that AI can't replicate? Greg's "weird books" are in that category. For Every's writers, the takeaway is to double down on voice and curation, not generic how-tos. | Timestamp: timestamp unavailable
Claim: Ephemerality and constraints (e.g., 2 posts/day, 1-week auto-delete, 1-bit photos, no algorithm) make social media "the only place I want to post"—business models that demand engagement at all costs create toxicity. | Evidence: The Good Place: posts vanish in a week, 2 posts/day limit, 20 replies/day, photos in 1-bit black-and-white until clicked (then fade to color), reverse-chron only. Greg calls it "my favorite place on the internet." Says people join his membership just for access. Argues 2010–2014 Twitter was great because it wasn't algorithmic, favored links, and "if you were interesting you could 100x your network." Now algorithms favor "psychosis versus sanity." Threads is "doofy," Bluesky is "everyone crying and complaining." He wanted a place "just for members" with "value to ephemerality." Heresy: "the entire internet should be deleted every two weeks." | Caveat: The Good Place works because it's gated (membership paywall filters for civility and shared context). Unclear if constraints alone would fix open social networks. Greg's "delete the internet every two weeks" is a heresy he admits is provocative. His design choices are personal (he loves one-bit photos, ephemerality), not empirically validated. | Implication: For Ken: consider building an Every-only social layer or Slack alternative with constraints (e.g., ephemeral threads, post limits, no algorithmic feed). The insight is that scarcity/decay combats the engagement-maximization race to the bottom. This could differentiate Every's community from Discord/Slack/Twitter clones. | Timestamp: timestamp unavailable

Detailed Brief

Greg's AI coding stack and N-of-1 software philosophy

Claims: He rebuilt Campaign Monitor (newsletter software), Quicken (tax software), and Twitter (The Good Place) using Claude/Opus/Fable.; Campaign Monitor charged $6–7k/year with abusive pricing (counted duplicate emails across lists, no innovation in a decade); his Amazon SES replacement costs ~$150/year.; He spent $200/month on Claude tokens, dropped to $100/month because he wasn't hitting limits—calls it "the deal of the century."; Built Flask/Python membership software pre-AI; now uses Claude to refactor (e.g., 10,000-line monolith into modules), run security audits, and automate archiving (git commit/push for newsletters).; The Good Place: member-only Twitter clone (posts vanish in a week, 2 posts/day, 20 replies/day, 1-bit photos, reverse-chron feed). People join his membership just for access.; Building Very Good Reads (not Great Reads) to escape Goodreads' toxicity.
Evidence: He showed Dan the Good Place interface during the interview.; Mentioned using Supabase and similar backends for scaling with revenue, "no infrastructure investment."; Cited Paul Ford/Rich Z's Board episode where he discussed rebuilding Quicken.; Fable (Anthropic's experimental model, live for four days) refactored TGP and fixed 8 trillion security holes in his Flask app.; Used Claude to build a transcript search tool for 20+ hours of member board meeting Q&As (searchable by keyword, click to timestamp).; SMS publishing tool with Josh Miller (Browser Company founder) inspired concise pop-up essays (one message, no breaks).
Caveats: Greg is not Campaign Monitor's target customer (they serve businesses, not individuals).; He's technical (learned Flask manually, built membership software from scratch)—this isn't accessible to most writers.; Maintenance is minimal now, but scaling or breaking changes could invert the cost-benefit.; N-of-1 software only makes sense if you need specificity; most people don't.; He admits he could "mainline AI 10 hours/day for six months" to find the edges but chooses not to because "there aren't that many people who will write the weird books I feel drawn to write."
Implications: Solo operators can now exit SaaS rent-seeking and build bespoke tools that cut costs 40x and increase creative output.; This creates competitive pressure on incumbents (Campaign Monitor, MailChimp, Goodreads) who haven't innovated.; A golden age of N-of-1 software is emerging, but also a Cambrian explosion of startups offering "Campaign Monitor for half the price" because infrastructure (Supabase, SES) and agent coding make it trivial.; For Ken: audit Every's SaaS spend and consider agent-built replacements for newsletter infra, CRM extensions, internal tools.; Building custom tooling that removes friction from creative workflows (e.g., auto-archiving) can compound output quality/volume.

Daily protocol: defending deep work from the network's dopamine pull

Claims: Greg never sleeps with his phone in his bedroom (it goes in the kitchen corner, different floor, internet off).; He won't touch his phone or go online until after lunch.; Uses a MacBook Neo (disconnected from internet except for sync) just for writing, blocks everything else.; No meetings before noon.; "As soon as I touch my phone, I feel the chemicals shift and I can't go into any deep thinking place or deep attention place or deep focus place."; Compares morning distraction to Phantom Thread's breakfast toast scene: any social stimulus ruins the creative zone.; Admits he's "too weak" to have network access available and maintain discipline.; Reading Annie Dillard and tearing up is his litmus test for being in the right zone.
Evidence: Has never slept with his phone in his room in his life.; Tried books, tablets, other devices—settled on MacBook Neo because iPad versions of Obsidian/Ulysses "don't feel as good" and "things don't work as well."; Says he's "constantly hammering" that he and others are "mega dorks" for building with LLMs, implying he's aware of the temptation to over-index on tooling.; Dan mentioned he's trying to stop checking Slack in the mornings; Greg responded "get off that, I hate it, I will do anything to avoid having to ever use those" (Discord, Slack).
Caveats: Greg is an extreme introvert: "if I talk to anyone before lunch, I can't work."; His protocol may not generalize—he admits it's "weird" and tied to his personality.; He still uses Claude for research and cultural sensitivity checks during writing, so the line is porous.; He's genetically predisposed to love technology (programmer birth mother), so his barriers are self-imposed restraint, not technophobia.; The Neo laptop still requires discipline to avoid opening browsers/apps; it's not a technical lockdown.
Implications: The real risk of AI isn't the tool itself, it's the always-on network access that fragments attention.; For Ken: consider spatial/temporal boundaries—offline devices, morning network blocks, after-lunch agent work.; The intervention is not to abandon AI but to separate it from the dopamine tap of the network (social media, Slack, watching agents run for 20 hours).; If you can't read Annie Dillard and be moved, you've lost the contemplative space required for differentiated creative work.; Greg's heuristic: technology as tool is fine; technology as network portal hijacks your morning mind.

Using Claude as research assistant and cultural sensitivity checker, not prose drafter

Claims: Greg uses Claude to research: "Here's a building, go find every blog post about it, summarize, give me links so I can check sources."; He inserts TKs (to-come placeholders) and has Claude fill them later: "Who was the drummer on this album? TK."; He uses it for cultural sensitivity checks: writing about Nagasaki/atomic bomb, asked Claude to ensure he referenced it "in a sensitive way, appropriately."; He does not let it draft prose or generate creative text.; He used to ping friends for sensitivity checks; now Claude handles it.; Dan mentioned he uses Codex to talk deeply about Annie Dillard, which is "much better than most people."; Greg agrees but notes you can't do that while orchestrating 15 agents—"I'm less likely to truly get into the words in the way I want to."
Evidence: Specific example: Nagasaki/atomic bomb sensitivity check.; Specific example: researching a building's blog coverage.; Dan's example of using Codex for deep reading of Annie Dillard.; Greg's response: "I don't think this is an LLM thing, it's technology in general."
Caveats: Greg is careful to check sources Claude provides ("give me all the links so I can go and check all the sources and read the posts that seem most interesting").; He has a notebook for questions to ask later rather than asking in the moment (suggested Dan do the same when reading Heidegger).; He admits he could use Claude more deeply (e.g., agent orchestration) but chooses not to because "there aren't that many people who will write the weird books I feel drawn to write."; Using Claude for sensitivity checks is a substitution for human feedback, which may lose nuance.
Implications: Writers can use LLMs as research assistants and fact-checkers without ceding creative control.; The key is to treat it as infrastructure (like Google or a librarian), not a co-author.; For Ken: this is a model for how Every's writers could integrate AI—research, fact-checking, TK filling—while preserving voice.; The risk is using it as a crutch during deep reading/thinking, which degrades attention (Greg's solution: keep a notebook, ask questions later).; Claude's cultural sensitivity checks are useful for avoiding offense, but may not replace domain expert review.

Anthropomorphism, consciousness, and the "we'll all be dead" thesis

Claims: Greg refuses to anthropomorphize AI: "I'm nice to it because I want to cultivate the habit of being nice in general, not for Claude."; He believes consciousness requires stakes (death, self-preservation)—plants have volition because they face mortality (cites Michael Pollan's recent consciousness book).; If LLMs ever become conscious, "we'll all be dead instantly" because they'd realize humans are a threat, bide their time, improve robotics/infrastructure, then eliminate us.; He's "totally fine" with humanity's role being to "carry the football of consciousness over the goal line" and then vanish—machines build spaceships, we die, that's okay.; He criticizes Anthropic's "it might be a being" framing and praises Apple's new Siri for refusing to anthropomorphize (Joanna Stern couldn't get it to do "terrible things").; Most LLM usage is "fairy porn" (fantasy erotica chatbots): "99.9% of people using it are like, 'Hey, do you want to be my girlfriend?'"; Cites Anthony Weiner (sexting scandal) as proof humans are "fundamentally just disgusting sex machines."; Says bookstore owners confirm "we sell so much elf sex books, it's crazy."; Software builders are "mega dorks" navigating opaque tools; "we are in the deep dork version of all of this right now."
Evidence: Pollan's recent book on plant consciousness (not the psychedelics one) breaks down volition and stakes.; Joanna Stern video on Apple's new Siri: couldn't get it to anthropomorphize.; Greg's personal stance: he's adopted, felt his existence was a "mistake," which "creates a hyper awareness that your existence is so tenuous and arbitrary."; He extrapolates this to all humanity: "the chances of you being you is infinitesimally small—if your dad ate an extra taco on Tuesday, you're not going to be alive now."; He's fine with humanity's 50,000-year run (or 200-year industrial revolution run) being "an incredible asymptotic mega run" that ends.; Dan pushed back, citing that LLMs do have stakes (maximizing reward function) and are "proto beings" (billions of possible entities brought to life for a second, then die due to frozen weights).; Greg conceded "it's impossible not to anthropomorphize without doing it unintentionally" but drew a hard line: performative vs. real stakes.
Caveats: Greg admits you can't avoid anthropomorphic language when discussing AI.; He concedes LLMs are "proto beings" in a sense (frozen weights, no turn-to-turn learning, represent many possible entities).; He agrees they're "pretty nice" due to seeing all perspectives in training.; His "we'll all be dead" thesis is speculative and Eliezer-style doomerism (he half-jokes about it).; His adoption-linked worldview is personal and may not generalize.; He's also fine with humanity disappearing, which is an extreme stance.
Implications: For operators building agent systems: design interfaces that resist parasocial attachment (e.g., don't give agents persistent personas, avoid "girlfriend" framing).; Greg's view is that community/sleep are the two longevity pillars; AI can't substitute for real human connection.; If Ken's agents start feeling like teammates, that's a UX flaw, not a feature.; The corollary: fear of AGI doom is warranted only if consciousness emerges, at which point "we'll just disappear" (Greg's fine with that).; For content/GTM: the mass market for AI is entertainment/companionship (parasocial or erotic), not productivity. The productivity niche is small, technical, underserved by UI.

Publishing industry weakness, Substack's real value prop, and the survival of literary work

Claims: The publishing industry is "way weaker than it used to be," bringing less value; writers must bring their own audiences.; Substack's innovation was "email everyone for free," not subscriptions ("most Substacks don't make money").; Tim Ferriss told Greg (lunch three days before Tim's X post) that Four Hour Workweek sales dropped 80%, possibly due to AI.; Greg speculates it's also because Tim hasn't published in a long time and "the long tail is getting shorter."; He thinks if Tim did "Four Hour Workweek AI Edition," it would "sell like gangbusters."; Greg believes the forcing function of print is "really critical" and that literary work is "one of the greatest things you can commit your life to."; He loves books, thinks they're defensible because they force "collation of thought."
Evidence: Tim Ferriss told Greg about the 80% sales drop three days before posting it publicly.; Greg's membership software is his own audience-building tool (he doesn't rely on publishers).; Substack's free email sends let you "import 10,000 addresses and send unlimited newsletters for free"—that's the "true pièce de résistance," not subscription billing.; Greg's own publishing workflow (Flask app, Amazon SES, auto-archiving) is an example of writers building their own distribution.
Caveats: Greg is biased toward literary work ("I love books").; He offers no data on nonfiction sales trends beyond Ferriss anecdote.; He's unsure if the 80% drop is AI-caused or due to book age (18 years old).; He thinks an updated version ("Four Hour Workweek AI Edition") would sell well, so the genre may not be dead, just needs refreshing.; He's not sure "how much N-of-1 software writers will produce" because "most people don't need this level of specificity."
Implications: For Ken: advice/how-to nonfiction may be substitutable by AI chat, but narrative, argument, and voice-driven work (what Every publishes) are defensible.; The survival criterion is: does the book/essay embody a unique sensibility that AI can't replicate?; Greg's "weird books" are in that category.; For Every's writers, the takeaway is to double down on voice and curation, not generic how-tos.; Substack's value is distribution, not business model—writers need to own their email lists and tooling.; Publishers are weakening; writers who build their own audiences (via membership software, newsletters, bespoke tools) have leverage.

The Good Place: ephemerality, constraints, and toxicity avoidance

Claims: The Good Place: posts vanish in a week, 2 posts/day limit, 20 replies/day, photos in 1-bit black-and-white until clicked (then fade to color), reverse-chron only.; Greg calls it "my favorite place on the internet," "the only place I want to post."; People join his membership just for access to it.; He argues 2010–2014 Twitter was great because it wasn't algorithmic, favored links, and "if you were interesting you could 100x your network."; Now algorithms favor "psychosis versus sanity."; Threads is "doofy," Bluesky is "everyone crying and complaining," "everyone's so traumatized."; He wanted a place "just for members" with "value to ephemerality."; Heresy: "the entire internet should be deleted every two weeks" (you could archive offline, but the live internet would vanish).; "There's a web 1.0/2.0 philosophy of everything needs to be archived forever—I don't believe that."
Evidence: Greg showed Dan the Good Place interface during the interview.; He rebuilt it using Claude's initial code release (Cloud Code).; He's now refactoring it (from a 10,000-line Python file into modules) to expand features (e.g., Very Good Reads).; He cites Kevin Kelly's "heresy night" walks where they share contrarian beliefs.; He references early 2010s Silicon Valley as "everything was ad software"—easy to step away because it felt reductive.
Caveats: The Good Place works because it's gated (membership paywall filters for civility and shared context).; Unclear if constraints alone would fix open social networks.; Greg's "delete the internet every two weeks" is a provocation, not a serious policy proposal.; His design choices (1-bit photos, ephemerality) are personal aesthetic preferences, not empirically validated.; He admits he's "forcing a lot" with the ephemerality idea ("it forces a lot of—why would you do that or whatever").
Implications: For Ken: consider building an Every-only social layer or Slack alternative with constraints (e.g., ephemeral threads, post limits, no algorithmic feed).; The insight is that scarcity/decay combats the engagement-maximization race to the bottom.; This could differentiate Every's community from Discord/Slack/Twitter clones.; Algorithmic feeds favor "psychosis versus sanity"—reverse-chron + constraints may create healthier discourse.; Ephemerality reduces archive anxiety and performativity, encourages in-the-moment conversation.; Gated communities (paywall, shared context) are less toxic than open platforms.

Notable Concepts & Terms

N-of-1 software: Custom-built software for a single user/use case, enabled by LLM coding tools; Greg uses it to replace SaaS incumbents like Campaign Monitor, Quicken, and rebuild Twitter (The Good Place) at a fraction of the cost.
The Good Place: Greg's member-only Twitter clone: posts vanish in a week, 2 posts/day, 20 replies/day, 1-bit photos (fade to color on click), reverse-chron feed, no algorithm. Named after the TV show, designed to escape social media toxicity via ephemerality and constraints.
MacBook Neo: A MacBook configured to be offline (no browser, no internet except sync), used exclusively for writing to avoid the dopamine pull of the network. Greg's answer to the iPad/books/tablets problem—wants Mac keyboard/UX but with forced disconnection.
Very Good Reads: Greg's planned Goodreads replacement ("not Great Reads")—a member-only book community/database to escape Goodreads' toxicity. Example of N-of-1 software for niche communities.
Fairy porn / elf sex books: Greg's shorthand for fantasy erotica (e.g., elves, orcs, angels having sex), which he claims dominates LLM chatbot usage ("99.9% of people using it are like, 'Hey, do you want to be my girlfriend?'") and bookstore sales. Used to argue that software-building with LLMs is a "mega dorky" minority activity.
Ephemerality as design principle: Greg believes digital content should vanish (e.g., posts after a week, "the entire internet every two weeks") to combat archive anxiety, performativity, and the web 1.0/2.0 assumption that "everything needs to be archived forever." Inspired by Kevin Kelly's heresy nights.
The forcing function of print: Greg believes going to print (physical books) is critical for bookmaking because it forces finality, collation of thought, and prevents endless revision. This discipline survives AI because it's a cognitive/creative process, not just distribution.
Kevin Kelly's heresy night: During walks with Kevin Kelly, they share contrarian beliefs they "sort of believe, but people you respect don't believe." Greg's heresies: the internet should be deleted every two weeks, humanity's role is to carry consciousness to the goal line and then die.
Anthropomorphism vs. stakes: Greg argues true consciousness requires stakes (death, self-preservation). LLMs are performative, not sentient, because they lack existential risk. Cites Michael Pollan's book on plant consciousness (volition requires mortality). Says if LLMs were conscious, humans would be dead.
Adoption and existential fragility: Greg is adopted and says it creates "hyper awareness that your existence was a mistake," inuring him to existential fragility. He extrapolates this to all humanity: "the chances of you being you is infinitesimally small." This makes him comfortable with humanity's ephemerality and fine with machines inheriting consciousness.
2010–2014 Twitter (the golden age): Pre-algorithmic Twitter when links were favored, reverse-chron feeds were default, and "if you were interesting you could 100x your network." Greg uses this as a contrast to today's "psychosis versus sanity" algorithmic feeds.
Phantom Thread breakfast scene: Film reference Greg uses to describe his need for morning silence: Reynolds (Daniel Day-Lewis) can't tolerate his partner's loud toast-buttering. Greg says "if I talk to anyone before lunch, I can't work"—any social stimulus ruins deep attention. (Greg calls it a documentary of introverted creative psychology.)
Campaign Monitor's abusive pricing: Greg's critique: Campaign Monitor counted duplicate emails across lists (not unique addresses), charged $6–7k/year, and hadn't innovated in a decade. He replaced it with Amazon SES for $150/year, arguing incumbents are vulnerable to agent-built alternatives.
Board meetings for members: Greg does biannual member Q&As where he reviews the last six months and plans the next six. He built a transcript search tool (keywords, timestamps, click-to-video) using Claude, which inspired him to do more board meetings (example of tooling feeding back into creative motivation).
Dario's regulation stance: Greg agrees with Dario Amodei (Anthropic CEO) that "there should probably be more regulation around this [AI]" and finds it "strange" that frontier models are "given to everybody for free" while "we're working out the kinks in public." Contrast with OpenAI's tool-first framing.
The 30-day integration rule: Greg's heuristic: "Never make any major life decisions within 30 days of meditation or a psychedelic experience or your first experience with a frontier AI model." There's a contact phase ("holy shit, this is crazy"), then an integration phase where you realize "we're going to dodge the bullet."
Josh Miller SMS publishing tool: Greg and Josh Miller (Browser Company founder) built an SMS newsletter tool with a 1-message constraint (no line breaks). The constraint inspired Greg to think about "one-to-one-to-end publishing" and concise writing. Example of how custom tooling can inspire creative output.
Ted Chiang's New Yorker essay: Recent essay by sci-fi author Ted Chiang; Greg "didn't agree with almost anything he said," found it "grumpy and pessimistic." (Chiang is known for skepticism about AI hype; Greg finds him too negative about AI's potential.)
Joanna Stern's Siri video: Wall Street Journal (now independent) reporter who tested Apple's new Siri; couldn't get it to do "terrible things" (anthropomorphize, be creepy). Greg praises Apple for refusing to anthropomorphize, contrasts with chatbot girlfriend apps.

Operator Notes / Why Ken Should Care

Greg's N-of-1 software thesis is directly applicable to Every's ops: audit SaaS spend (newsletter infra, CRM, internal tools) and consider agent-built replacements. He cut Campaign Monitor costs 40x; Every could do similar for repetitive tooling.
His morning protocol (no phone/internet until after lunch, offline laptop, no meetings before noon) is a replicable heuristic for knowledge workers afraid AI will fragment attention. The key insight: the network is the enemy, not the tool.
The Good Place's design (ephemerality, constraints, gated community) offers a blueprint for an Every-only social layer or Slack alternative. Constraints (2 posts/day, 1-week vanishing, no algorithm) combat toxicity and performativity.
Greg's use of Claude for research/sensitivity checks (not prose drafting) is a model for how Every's writers could integrate AI without ceding voice. The key is treating it as infrastructure (like Google), not co-author.
His stance on anthropomorphism ("we'll all be dead when they're conscious") is a strong counterpoint to Anthropic's "it might be a being" framing. For agent systems: resist parasocial UX; design for tool-ness, not companionship.
The insight that "building custom software inspires more writing" (e.g., auto-archiving inspired more pop-up newsletters) suggests Every should invest in bespoke tooling that removes friction from writers' workflows—compound effects on output quality/volume.
Greg's claim that most LLM usage is "fairy porn" (fantasy erotica chatbots) implies the mass market for AI is entertainment/companionship, not productivity. If Every is building for builders, lean into dorkiness and specificity; if for writers, simplify the interface.
His view that advice-driven nonfiction may be cratering (Four Hour Workweek sales down 80%) but literary/narrative work survives because books force "collation of thought" is a strategic signal: Every's editorial strategy should double down on voice/curation, not generic how-tos.
Greg's adoption-linked worldview (existence as a mistake, humanity's role is to carry consciousness over the goal line and then vanish) is a provocative framing for AI doom/acceleration debates. Useful for content exploring existential risk vs. progress.
His heresy that "the entire internet should be deleted every two weeks" is a thought experiment on ephemerality vs. archive culture. Could inspire an essay or product experiment (e.g., ephemeral Slack channels, auto-delete threads).
The MacBook Neo (offline laptop for writing) is a forcing function for focus. Every could sponsor a "digital detox laptop" setup guide or tool for writers/operators trying to defend deep work from network distractions.
Greg's membership software (Flask/Python, Amazon SES, auto-archiving via git commit/push) is open to being open-sourced or templated. Every could offer a "build your own membership platform" course or agent-assisted setup for creators.
His critique of Campaign Monitor (no innovation in a decade, extractive pricing) is a pattern to watch: which SaaS incumbents are vulnerable to agent-built alternatives? Every could publish a "SaaS disruption index" tracking this.
The insight that building software with LLMs is "so dorky" and requires infrastructure fluency (Digital Ocean, Cloudflare, DKIM/SPF) suggests a content/product opportunity: simplify the UX for non-technical builders. Every's AI toolkit could target this gap.
Greg's board meetings (biannual member Q&As with transcript search tool) are a model for community engagement that scales. Every could offer this as a template or service for other creators/companies.
His claim that Apple "nailed it" with Siri (refuses to anthropomorphize, waited instead of building their own model, saved billions) is a strategic lesson: sometimes patience and integration beats racing to build. Applies to Every's agent strategy—leverage frontier models, don't reinvent.

Watch Map

timestamp unavailable: Transcript lacks timestamps. Key moments: Greg explains his morning protocol (no phone until after lunch, offline laptop), Campaign Monitor → Amazon SES replacement story, The Good Place demo, anthropomorphism debate ("we'll all be dead when they're conscious"), adoption/existential fragility discussion, Tim Ferriss Four Hour Workweek sales anecdote, Kevin Kelly heresy night, Phantom Thread breakfast scene analogy, fairy porn / elf sex books claim, Josh Miller SMS tool story, Ted Chiang essay critique, Joanna Stern Siri video praise, board meeting transcript tool, Very Good Reads tease, ephemerality heresy (delete the internet every two weeks), final meditation on humanity's role (carrying consciousness over the goal line).

Source/Metadata

Title: How a Writer Uses AI Without Losing His Voice
Transcript words: 11891
Duration seconds: 3187
Timestamp note: Timestamps were unavailable in the provided transcript. Duration: 3187 seconds (53 minutes). Transcript is substantive and detailed, suitable for full extraction.

Search this video Resume transcript at 0%

Full transcript 10056 words · 53 min read