The Connection Engine -- Andrew Conru

I was twelve years old, in a converted bedroom on a 10-acre farm in northern Indiana, when I first watched a machine do exactly what I had told it to do. It was 1982. The computer was small, the program was small, the satisfaction was not.

Across the kitchen, my mom was making sure I was not doing it half-assed. That was her rule. If you are going to do something, do it all the way through. She said it about the lawn. She said it about my bicycle chain. She said it about homework. She said it without knowing she was teaching me how to write code.

The machine on the desk did what I told it to do. The room outside the door did not. That gap, the one between the obedience of a machine and the unpredictability of people, has been the only real puzzle I have worked on since.

I have been called an engineer. An entrepreneur. An amateur artist. A philanthropist. In 2018, IEEE Spectrum called me an “engineer of love.” I liked it. I would have built the same things without the tag.

What I have actually been is a builder of rooms. In 1994, I built the world’s first online dating site. In 1996, I started Friend Finder Networks, which would eventually carry over 700M registered members. In 2007, I sold it and walked off into seven years of, let us call it, finding out who I was when I was not building. In 2017, I built a website called agnostic.com. Less than two years later, on Valentine’s Day 2019, I built a sister site called IDW.community. In 2022, I funded an academic magazine called Aporia. In 2024, I started building a 16,000 square foot art salon across from the Seattle Art Museum.

The question is the one I asked at twelve and have not stopped asking. How do you connect people without lying to them.

This book is the story of three of the rooms. agnostic in 2017. IDW in 2019. Aporia in 2022. I built them on the same engine, with the same seed prose, and the same hypothesis. The hypothesis was that you could engineer civility into the medium itself, that the rails of a well-built platform would teach a community to rehumanize each other. The rooms grew. Combined, the two communities held over 355,000 user records, 1.2M questions, 3.5M answers, 2.5M comments. I held the engine, the seed essays, the moderator’s chair, and, eventually, the data. There was nobody better positioned to honestly autopsy the experiment than the person who ran it. There was also nobody else.

So I autopsied it. Quietly, with help from machines that did not exist when I started, I read every report, every cluster, every cross-post, every blocked user, every melted-down thread. The book is what I found.

Some of what I found was simple and corrected the assumptions I went in with. Some of it was uncomfortable. Some of it was kinder than I expected. The headline, if you want one, is this. Civility is not what each tribe is. Civility is what each tribe protects. The rooms told me that twice, both times in numbers I cannot wave away.

If you only remember one thing about this opening, remember this. Every room I have built since 1982 has been an answer to the same question a twelve-year-old kept asking after the Commodore turned off and the kitchen got quiet. How do you connect people without lying to them, without flattering them, without selling them something. How do you do it on purpose. How do you do it at scale.

This is the story of three swings at that question, what they got right, what they got wrong, and the room I am building now in Seattle that takes those lessons and tries again, in a paint and canvas medium I did not expect to land in, a medium that does not argue.

I want to be clean about what this book is and what it is not. It is not a relitigation of any press cycle. It is not a defense and it is not an attack. It is not a foundation brochure. It is a builder’s notebook, opened in public, with the receipts attached. The receipts happen to be 355,000 user records and most of a decade of threads. I would have written the same notebook with smaller numbers if those were the numbers I had.

Act I: Building

Chapter 1: The engineer’s instinct

The farm was 10 acres. The house was built by people who did not waste lumber. My parents were union steelworkers, which is the kind of household where you fix things. Brakes, fences, vacuum cleaners, the dishwasher, the radio, your own bicycle chain when it slipped off in the driveway and the school bus was three minutes out. If something broke, you did not call somebody. You went to the garage.

That is the whole moral inheritance, in one sentence. If something is broken, you fix it yourself. I have not deviated from that line in forty-four years.

I started writing code in 1982, age twelve. The thing I noticed quickly was that the code did exactly what I told it to do, which was a small miracle and a kind of betrayal. People do not do that. The puzzle of the difference is what hooked me.

I paid my way through Rose-Hulman with side businesses. Hubcaps, door to door vegetables, a GM scholarship that paid for the engineering school but not the rent. I picked Mechanical Engineering and Economics for reasons that made sense to a kid from a steelworker household. ME was hands. Econ was money. Both were levers.

Then Stanford for a Ph.D. in Mechanical Engineering, finished in 1997. The thesis was on agent-based collaboration in mechanical-design systems. The thesis itself was about how you get many small autonomous things to cooperate without a central referee. It is, in retrospect, a very on-brand title for the rest of my life. I have never run a referee in any room I built. I have written referee logic into the room. There is a difference, and that difference is most of the engineering.

I arrived in California in the early 1990s, which means I arrived right when the Web did. That was luck. I have always been honest about the luck. A different five-year window and I would have been a slightly bored mechanical engineer at a defense contractor, fixing other people’s brakes.

Instead, I built my first internet company in 1993. Internet Media Services. We were the first business listed under “Internet” in the Yellow Pages. That sentence sounds like a tall tale now. It was true, and it was alphabetical, and I am still proud of it.

In 1994, I built WebPersonals.com. It was the world’s first online dating site. I am supposed to lead with that sentence in every bio, and I usually do. I will lead with the part that matters more in this book, which is what I learned building it.

What I learned was that the moment you put two people in front of a screen and ask them to be honest about what they want, half of them lie to you and half of them lie to themselves. Software cannot stop the lying. Software can change the room around the lying. You can put more light on the question. You can take cues away from the things people pretend with. You can make it slightly cheaper to tell the truth than to perform. That is the whole craft. I have been refining it since 1994.

The other thing I learned at WebPersonals, which would matter twenty-five years later when I built the communities in this book, was that what people say they are looking for and what they actually pick are usually two different sentences. The right interface narrows that gap. The wrong interface widens it. The room shapes the answer. I keep coming back to that line. The room shapes the answer.

In 1996, I founded Friend Finder Networks. We grew to over 700M registered members across the lifetime of the company, with a peak annual revenue around $90M. We were twice the size of Match.com with one tenth the development staff. I am supposed to be modest about that and I will not be, because the staff thing is the part that matters. If you align the engineers with the users instead of with the investors, the work bends differently and the numbers come along anyway.

I also built shopping carts and banner ad management and real-time site customization in the same window, because in 1995 nobody else had built them yet and somebody had to. None of those tools were the point. The room was the point. Every tool was just a fix for a place where the room was lying to its own users.

I sold FFN in 2007. The number on the wire is not the part of this story. The part of the story is what came after.

There is a thing that happens to a builder when the building is done and the wire clears. The room you have been living inside for eleven years is suddenly somebody else’s room. The keys do not work anymore. The work that used to take twelve hours a day now takes zero. The team is not a team. The schedule is not a schedule. You are, technically, free. The problem with technical freedom is that it is not a thing you can put your hands on. I had been an engineer since I was twelve and engineers know what to do with broken things. We do not always know what to do when there is nothing to fix.

The bios call the next stretch the wilderness years. That is the public-facing version, which is true as far as it goes. Travel. Liberal-arts study. Self-inquiry. The bios call the person on the other end of it a “hybrid person,” which is what my friends meant when they politely said I had finally noticed that engineers are people too.

I will not retell the harder parts of those years here. They live in another file with a more careful voice and I will not import them into this one. The relevant fact for this book is that I came out of those years asking the same question I had asked at twelve, only in a different register. Not how do you build a tool. How do you build a room.

I will say one more thing about that period, because it bears on what comes next.

Building had been my answer to “what am I” since 1982. When the building stopped, the question did not. If I am not building anything, then what am I. I do not recommend asking that question in the order I asked it. There is a smarter order, with more help. I have come around to that.

By 2014 or so I had started building again. Smaller things at first. Painting. Reading. Listening to people I did not agree with. I went back to the same question I had been working on since the Commodore in Indiana. I came back to it in a different room.

If you only remember one thing about this chapter, remember that the question never changed. The room did. Indiana farmhouse. Stanford lab. WebPersonals server in 1994. FFN in 1996. Wilderness years. And then, very quietly, a domain name I bought in early 2017 because I had finally figured out what I wanted to test next.

Chapter 2: agnostic.com, June 19, 2017

I bought the domain late. The site went live on June 19, 2017. The decision had been simmering for a while.

The shape of the room I wanted to build was not “an atheist debate club.” There were already plenty of those. They had a tone problem and they did not seem to know it. The shape I wanted was something more like a humanist meeting hall for people who had walked out of a religion and did not have a place to sit down. A hall that accepted them, that did not require them to perform certainty, that let them ask questions without getting reflexively shouted down by people who had walked out earlier and now policed the door.

I had grown up next to that population. Northern Indiana is not a low-religion environment. I had watched friends and family deconvert and watched some of them turn around and bring the same purity policing they had escaped to the new room. I wanted to build a place where that loop did not have to run.

The seed prose was 36 essays. I wrote them and posted them as site_text, which is what the engine called the long-form pieces that anchored the home page and the discovery feeds. They covered humanism without God, the dehumanization cycle, civil discourse, forgiveness and redemption as infrastructure, morality without God, and rehumanism. The word “rehumanism” was in the essays from the first month. I did not invent it for the press cycle in 2024. It is in the corpus, dated.

The most important framing in the seed prose, to me, was this. Civil discourse is a weapon, not a courtesy. Children shout each other down. If you actually want to dominate an opponent, the most powerful tool you have is a calm, factual, civil exchange. Civility looks polite. It is not. It is what wins arguments cleanly. The room was supposed to be a training ground for that weapon.

On January 1, 2018, six months in, I posted the goals of the site, plain and direct. Move people along the path from faith-based belief, through unbelief, to advocacy. Make unbelief normal and acceptable. Reduce the suffering caused by religion. I did not soften any of those. They are in the corpus, January 1, 2018, signed.

The engine itself was a question and answer site, built in PHP on MySQL, with profile pages, photos, voting, direct messages, friend graphs, an ignore list, and a layered abuse-flag system. None of those pieces were novel by 2017. What was novel, at least to me, was the way they fit together. The home page anchored on the seed essays so the room had a thesis before it had a feed. The question form prompted you to phrase things as questions, which is a small constraint that does most of the civility work for free. The abuse flags routed conflict away from bystanders so that one fight between two regulars did not poison a thread for fifty newcomers reading along.

I want to say a word about the abuse-flag system specifically, because it is the part of the engine that turned out to matter most for the autopsy I would run later. Every block, every soft-hide, every formal user report, every content report sat in a row in a database with a timestamp, a reporter, a target, a reason code, and, in many cases, a free-text note. I did not build that for the eventual autopsy. I built it because, in 2017, I knew from FFN that you cannot moderate a community of any size without a paper trail, and I knew the paper trail had to be queryable. Six years later that database would let me read the conflict shape of a community of 200,000 people in a single query. I am not glad we needed to read it. I am glad we could.

By October of that year I had decided the name was wrong. “Agnostic” is a doorway, not a destination. The site had outgrown the doorway. On October 8, 2018, I announced the Humanist.com rebrand. The intention was to evolve the room from “what we are not” to “what we are.” The rebrand never fully landed. The user base was attached to the original name, the SEO was attached, the cultural moment kept moving, and Humanist.com never replaced agnostic.com on the front of the site. The intention is in the corpus. The execution is honest about itself.

The site grew anyway. Peak users: 209,110. Peak content: 782,123 questions, 3,029,091 answers, 2,223,661 comments. Those are not abstract numbers. That is two stadiums of people typing into the same room. I held the engine, the moderator role, the seed prose, and, eventually, the abuse logs. Those four pieces together are what made the eventual autopsy possible.

I also wrote on the site myself. Not a lot, by online standards. Across both communities I would eventually write 749 questions, 5,378 answers, and 3,324 comments. On agnostic specifically: 457 questions, 4,742 answers, 1,843 comments, 36 site_text essays, around 29,000 sent direct messages. I posted in my own name. I did not run an alt account. If the room was supposed to model a kind of civic adult speech, it would have been hypocritical to model it from behind a mask.

The thesis going in was the engineer’s thesis. If you build the rails right, the train runs right. People are downstream of the medium. Make the medium humane and the conversation will follow. A well-designed question form prompts better questions. A well-designed flagging system routes conflict away from bystanders. A well-tuned reputation signal rewards careful contributors and penalizes bad-faith ones. Civility was a feature of the room, not of the people in it. You could engineer it.

I still believe that thesis is partly right. The engineering matters. A poorly built room is louder than a well-built one. But by 2026 I have learned that “civility as a property of the room” only carries you so far. Each tribe was protecting its own version of civility. Each tribe melted down at outgroup-triggered prompts in a way the room could not absorb. The data on that lives later in this book. Right now I am still in 2017, and I am still confident.

There is one more thing I should say about the agnostic period, because it is honest and it sets up the next chapter.

The site had cluster names that, in retrospect, told me what the room had become. Behavioral clustering, run later, named the dominant agnostic groups Seekers, Curious, Tentative, Bruised, Newcomers. Eight of ten cluster labels were verbs of looking-for, not verbs of pushing-back. The room had organized itself, by 2018, into a community of seekers. That had been the goal. The goal had landed. There was a specific kind of person who came back to agnostic the second day, and that person was looking for something, not winning something.

The trouble was that, by 2018, there were also a lot of people in the country who were not seeking. They were pushing back. They had questions that had become unwelcome, in some cases unwelcome on agnostic itself. I started watching threads where good-faith heterodox questions got reflexively shouted down by users who did not seem to notice they had imported the purity policing they had walked out of a Sunday school to escape.

I sat with that for a year and a half. I watched it carefully. I knew the engine could do something about it. I knew the seed prose could do something about it. I knew, by late 2018, that I was about to build the second room.

The honest thing to say here, because the data later in this book will earn it, is that I did not yet understand what kind of selection effect I had created. I thought I had built a community of seekers because I had written essays about seeking. I had not. I had built a community of seekers because the welcome script at the door, the verbs in the tagline, and the topic mix on the home page had quietly filtered who came in on the second day and stayed. The seed prose was a smaller part of that than I thought. The first paragraph of the homepage was a bigger part of it than I knew. By the time I figured that out, I had already built a second room with a different first paragraph, and the contrast was going to teach me what neither room would have taught me alone.

If you only remember one thing about this chapter, remember that 209,110 people built a community of seekers on the rails of an engine designed for civility, and the engineer who built it sincerely thought, in 2018, that the rails would carry them through anything. That belief was sincere. The book that follows is, in part, the story of how that belief was tested and what it became.

Chapter 3: IDW.community, February 14, 2019

I launched the second site on Valentine’s Day 2019. The date was a quiet joke I made to myself and never explained on the site. If you have spent your career building places where people meet and talk, Valentine’s Day is just another day that happens to rhyme with the work.

The branding was slug.com and IDW.community. The tagline was “Open Minds Welcomed.” The framing was John Stuart Mill, not Twitter. The early visual reference was the cluster of public intellectuals who, in early 2019, were having a moment in the culture and were getting bounced from a lot of platforms. I did not endorse any of them by name. I aggregated links to public conversations they were having, the way you would aggregate links to any other public discourse. By mid-2020 the site had taken on an explicit critique of the cancel-wave era in some of its essays. That framing did not exist at launch. It accumulated.

I built the site on the same engine I had built agnostic on. Same MySQL backend. Same PHP question and answer schema. Same flagging logic. Same abuse-report routing. Same profile fields, mostly. Same site_text long-form anchor essays. The only deliberate variables I changed were the welcome script, the cluster of seed essays at the front door, and the user population I let in.

The 47 site_text essays at IDW were largely the agnostic essays, ported across with the framing tweaked. The dehumanization cycle essay was nearly identical. The civil discourse as a weapon essay was nearly identical. The morality without God essay was nearly identical. The forgiveness and redemption as infrastructure essay was nearly identical. What changed was tone in a few places, and the prefatory paragraphs that situated the audience. On agnostic the framing was “you walked out of a religion, here is a humanism.” On IDW the framing was “you have questions that have become unwelcome in mainstream conversation, here is a place to ask them carefully.” Same scaffolding underneath.

That was a deliberate move on my part. Real life is not a controlled experiment. You almost never get to run the same code twice with only the user population changed. Here, almost by accident, I was about to. Same engine. Same moderator. Same seed prose with framing tweaks. Twenty months apart in launch date. The rest of the variation, whatever you saw on the two sites, would be coming from who came in the door.

I had a thesis about who would come. I thought IDW would draw people who had been told to be quiet about questions that interested them, and that on a well-engineered platform with clear rules and a calm front page, those people would behave themselves the same way the agnostic users had. Same training ground, the other side of the cultural moment. I did not expect IDW to be a clone of agnostic with the politics flipped. I expected the engine to compress most of the difference.

Some of that turned out to be right. Most of it did not. The data on that lives in the next act of this book. In 2019 I did not have the data yet. I had the launch.

On July 8, 2019, five months after launch, I posted what is still, to me, the most personal thing I wrote on either site. The title was “Some back story about this website.” I wrote about growing up in Indiana, in a religious household, and the science and religion tension I had carried since childhood. I wrote about why I had built agnostic in 2017 and why I had built IDW in 2019. I wrote about the through-line, which was that both rooms were trying to do the same thing from two different cultural angles. The post tied my autobiography to the architecture in plain language. It is, structurally, the spine moment for the whole book you are reading. One person, two rooms, one question.

A year later, on July 30, 2020, I wrote the most explicit thesis post on the IDW side. The title was about the cancel-wave era and the origin of the site. I will not retell that essay here. The framing in 2026 is more careful than the framing in 2020 was, and the data has corrected my certainties since. But the post is in the corpus, dated, signed, public.

On January 13, 2021, six days after January 6, I posted a question to the IDW side. The phrasing was, roughly, “What did Trump do, if anything, to incite violence?” I posted it in the same posture I had taken on every contested topic since 2017, which was as a sincere “I am trying to find this out myself” question, not as a take. The room around that question, on both sides eventually, is one of the central data points in this book. The post itself, in the moment, was the one move the engineer in me knew how to make. When the country broke, I did what I always do. I asked.

The site grew. Peak users: 145,827. Peak content: 431,482 questions across the legacy and current state, 438,531 answers, 310,227 comments, 4,189 formal user-abuse rows against 73,502 on agnostic, which is a ratio I will earn carefully in a later chapter. My own authorship on IDW, across its life: 292 questions, 636 answers, 1,481 comments, 47 site_text essays, roughly 45,000 sent direct messages. I wrote less on IDW than on agnostic. The numbers say I read more.

The cluster names told me, again, what the room had become. Behavioral clustering on the IDW user base, run later, named the dominant groups Heterodox Traditionalist Skeptics, Cultural Grievance Veterans, Skeptics, Contrarians. Six of eight cluster labels were verbs of pushing-back, not of looking-for. agnostic had organized into a community of seekers. IDW had organized into a community of disagreers. I had not asked it to. The user population that walked through the IDW door, on average, came in with a different posture than the population that walked through the agnostic door. The room shaped itself around that posture within months.

A small thing I want to flag now, because it will not be the last time I flag it. 15,957 users joined both sites. 49% of them on the same day. That is a real number from the joined database. Some fraction of the population I was trying to reach was not on one side or the other; they were on both, often signing up minutes apart, looking for whichever room was going to take their question seriously today. That fact, when I encountered it later in the data, surprised me. It should not have. People are not their team jersey. People are their question.

There is one more thread in the IDW story I will leave hanging for the chapter on Aporia. By 2022 I had decided that the same hypothesis needed to be tested at the publishing layer rather than the community layer. I funded an academic free-speech magazine because I still believed the engineer’s thesis from 2017, and I still believed that censorship was a form of hate and that open inquiry was a form of love. That funding ran from 2022 through the end of 2024. 832 articles, all public, all readable. In October 2024 the funding became the subject of a press cycle that I will not relitigate here and which the book treats by date and outlet only. The cleanest single data point in the whole communities corpus, December 18, 2020, is the bridge to that story. On that day I posted to both rooms, in plain language, the question my Aporia funding would later be attacked for. Both rooms answered it civilly. The wider apparatus, two and a half years later, did not. That contrast is the structural reason this book exists. It is also a chapter for later in the book, not this one.

I want to say one honest thing here, before the next act of the book starts pulling apart what I learned from running these rooms in parallel.

I started both sites with the same hypothesis. The hypothesis was that you could engineer civility into the medium itself, and that on the rails of a well-built engine, a community would learn to rehumanize each other regardless of the cultural lane they came in from. I did not steer agnostic into one cultural lane on purpose. I did not steer IDW into another cultural lane on purpose. I built two rooms, opened both doors, and let the populations select themselves into the rooms whose framing matched what they were already looking for. That self-selection was the experiment. The data, when it came, was not about which side was meaner. It was about which kind of room each was, and how the same engine produced two different kinds of conflict.

If you only remember one thing about this chapter, remember that I tested the same hypothesis twice, six weeks of code apart, with one variable changed. The variable was the user population. The hypothesis was that the engine could carry any population into civility. The next ten chapters of this book are what the data said back to that hypothesis, in the only register the data speaks in. Numbers, dates, posts, threads, reports.

Act II: The Autopsy

Chapter 4: Two cities, one engine

agnostic.com went live June 19, 2017. slug.com, which I rebranded a little later as IDW.community, went live February 14, 2019. Same MySQL backend. Same PHP. Same moderator, which was me. The 36 humanism seed essays I had written for agnostic mostly carried over to IDW with the framing tweaked, ending up as 47 essays on the second site. Six weeks of editing apart, twenty months of launch dates apart.

What I underestimated walking in was how rare a setup like that is. Most “two communities” comparisons collapse because the founders are different, or the technology is different, or the rules are different, and so anything you find could be any of those. Here it was just me, the same code, the same essays, and the same moderation hand on both sides. The variable I didn’t control was the population.

In 2025 I asked an AI to do the slow, embarrassing thing I should have done years earlier. I gave it twenty numbers about how each active user actually behaved on the site. How often they posted. How often they answered. How often they were reported by others. How often they reported others themselves. And so on. I asked it, blind to which site was which, to tell me how many different kinds of person it saw. It came back with ten kinds on the agnostic side, from 17,062 active users, and eight on the IDW side, from 7,522 active users. Then I handed it thirty real bios and posts from each kind, again with no platform label attached, and asked it to read what was in front of it and tell me, in plain English, who those people were.

agnostic, ten clusters: Seeking Connection Questioners. Curious Newcomers. Thoughtful Elder Skeptics. Seeking Connection Seekers. Bruised Seekers. A second pool of Curious Newcomers. Thoughtful Philosophical Skeptics. Seeking Secular Community. Progressive Freethinkers. Tentative Truth-Seekers.

IDW, eight kinds: Heterodox Questioners. Skeptical Truth-Seekers. Cultural Grievance Veterans. Skeptical Heterodox Newcomers. Heterodox Traditionalist Skeptics. Heterodox Contrarians. Quiet Spectators. Plus a small Bot and Spam group the AI caught.

Eight of the ten agnostic labels include the words Seeking, Curious, Tentative, Bruised, or Newcomers. The dominant verb is looking for.

Six of the eight IDW labels include Heterodox, Skeptical, Contrarian, Grievance, or Traditionalist Skeptic. The dominant verb is pushing back.

That isn’t a partisan claim. The AI did not know about parties. It read what the users had actually written, grouped them by behavior, and named the groups. Eight out of ten groups on the first site were looking for something. Six out of eight on the second site were pushing back on something.

The largest cluster on agnostic was Seeking Connection Seekers, at 22.0% of active users. Reading their bios, that’s people who deconverted from a Christian household and now lived in a county where they had nobody to talk to about it. The platform was a refuge. They wrote about isolation, made self-deprecating intros, mentioned their pets, asked for advice about Thanksgiving dinner with the family. The mode was post about who you are and where you came from.

The largest cluster on IDW was Quiet Spectators at 19.5%, who read more than they posted. The next-largest was Skeptical Truth-Seekers at 17.7%, substantive question-askers framing topics through a heterodox lens. The mode was ask a question and watch the room argue about it.

The conflict pattern is a chapter for later in this book, but the cluster math previews it. On agnostic, the noisiest cluster is 1.8% of the active universe: 313 users out of 17,062, generating reports at an order of magnitude above any other cluster on the site. Most of agnostic’s reported conflict is one small repeat-player cohort fighting itself, and the rest of the community reading past it. On IDW the conflict-prone fraction is bigger and more spread, 10.4% of active users, the loudest IDW cluster, with the combative tone diffused across many more accounts. Same structural shape: many lurkers, many newcomers, a small loud core. Different distribution of where the loud core sits.

I had built the same engine for both. The seed essays on both sides talked about civil discourse as a weapon, humanism as a moral floor, forgiveness as infrastructure, the dehumanization cycle. Same prose. Same hypothesis. Same builder.

But the population on day one was already different. agnostic launched into a quiet wave of religious deconverts looking for a place to land. IDW launched into a louder wave of people who felt the academic and social-media doors were closing on questions they wanted to ask. The first crowd showed up because they had lost a community and needed a new one. The second crowd showed up because they had been told a question they wanted to ask was off-limits.

That is not the same purpose. Purpose, it turns out, is the seed of the welcome script.

If you arrive on a site where the existing posts are people sharing their deconversion stories, the script is “tell me yours.” If you arrive on a site where the existing posts are heterodox questions and combative threads, the script is “what’s your take.” The platform did not impose either script. The first thousand users did. The next ten thousand followed the first thousand.

I called the difference “tonal default” for a year and a half before I had a better word for it. Then I started calling it the welcome script. The cluster names made me start calling it the verb.

My own posts started doing the same thing without me noticing. On agnostic I wrote softer. More personal anecdote, more deconvert-storytelling, fewer warnings. On IDW I wrote sharper. More open questions, more critical framings, more adversarial framings of in-group enforcement. (I admit I never loved the term woke as a pejorative when it first came into the language. It felt like the wrong word for the phenomenon. But the data says I was reaching for that framing on the IDW side, and a working consensus has since formed across most of the political spectrum that the movement it was naming had overshot, so I am not going to dress the finding up.) I was writing the same kinds of essays into both rooms, and the rooms were tuning my register down to match them. The chameleon effect ran on the founder before it ran on anyone else, which is a sentence I would not have written in 2019.

The engineer’s blind spot is to assume the engine determines the outcome. I had spent twenty-five years building software that imposes shape on whatever flows through it. WebPersonals in 1994 imposed a shape on dating. Friend Finder in 1996 imposed a shape on social discovery. By 2017 it was a deep instinct: build the right shape and the right behavior comes out. The two communities I built between 2017 and 2019 are the cleanest evidence I have ever seen against my own instinct. Same engine, two outcomes. The shape of the population on day one was already running the show. The engine was running second.

If you only remember one thing about this chapter, that’s it. Same engine, two cities. The verb the population brought through the door was the variable I had not controlled, and it determined who came back the second day. Both communities looked at me the same way. They ran on the same code. They came in for different reasons, and the room organized around the reason.

I had built two cities and only realized later that I had been holding the city limits in the wrong hand.

Chapter 5: I posted the same question to both rooms, 49 times

Between June 2017 and February 2022 I wrote 749 questions across the two communities. Most went up on one side or the other. 49 of them became scoreable matched pairs: same prompt, both sites, within 14 days, with enough responses on each side to score. (The wider duplicate pool was 52; 33 of those hit both sites the same day or the next day.) That is the cleanest possible match. Same author, same week’s news, almost the same paragraph, two audiences.

I did not design this as an experiment. I was running both rooms and posting questions I genuinely wanted answers to. It only became an experiment in 2025 when I went back and asked an AI to score the audience response threads on civility, agreement-with-premise, depth, dehumanization events, and on-topic rate. Each of the 49 pairs had at least three substantive responses on each side, which was enough to score. Total LLM cost: $0.88. Cheap autopsy.

What “civility” means in this scoring is not “no one used a swear word.” The AI was reading the top twenty-five responses in each thread and rating five things. Tone. How often a reply attacked the person who asked the question instead of the question. Whether the replies engaged the actual question or rerouted away from it. Whether the replies used dehumanizing language toward a named outgroup. And how on-topic the replies stayed. A score of 8.94 means substantive, on-topic, low personal attacks. A score of 2.5 means short, full of personal attacks, dehumanizing toward the named outgroup, almost never engaging the actual question. The score measures whether the room is treating the prompt as a question or as a tribal challenge.

If you average all 49 pairs together, the gap is small. Civility 7.18 on agnostic, 6.86 on IDW. Dehumanization events per pair: 2.84 agnostic, 1.46 IDW. Boring picture. agnostic modestly more civil. IDW dehumanizes less but engages shallower.

Pre-2020. Sixteen pairs. Mostly platform housekeeping. We shipped a like-button fix. You can now see deleted posts. Here is a layout I am testing. What would you change. Civility on agnostic 8.94. Civility on IDW 8.61. Both maxed out. Dehumanization events round to zero on both sides. Whatever I had built, both rooms could handle it.

2020 summer. Seven pairs. COVID, BLM, civil unrest. agnostic civility 9.36. IDW civility 7.57. agnostic gets even warmer. IDW does not melt down but does not keep pace. I will admit, in 2020 I read this as confirmation of the original Truth Project hypothesis. The warmth had been engineered, and even under stress it was holding.

2020 election Sep-Dec. Seventeen pairs. Both sides crash from the 8s and 9s down to the 5s. agnostic 5.89. IDW 5.48. Dehumanization events go from rounding to zero up to 4.6 per pair on agnostic, 2.2 on IDW. Both communities degraded in parallel, not in opposition. The thing that broke was not an asymmetry between the two. It was politics in general entering the prompts. Both communities tested civil for almost three years on platform-internal topics. The moment external politics walked in, both buckled.

January 6 to inauguration. Five pairs. agnostic civility 4.04. IDW civility 6.26. Gap of +2.22. The only window where IDW is sharply more civil than agnostic. agnostic dehumanization events 6.8 per pair, the highest of any period. This is the window where the directional pattern most readers expect, that one side was tilting harder online, shows up clearly in the data. It is real. It is also five pairs across fourteen days.

Biden’s first month. Four pairs. agnostic civility 5.75. IDW civility 5.25. Average flips back. agnostic cooler than IDW again.

Five buckets. Five readings. The aggregated number across all 49 pairs hides each one of them. If a reader of this book takes nothing else from the time arc, the takeaway is that the period averages narrate two different stories. Through summer 2020, my engineering hypothesis was getting confirmed. Starting fall 2020, both rooms broke at the same time, and the asymmetry between them was no longer stable.

That’s the headline arc. The time table only makes sense once you look at single questions.

January 13, 2021. Six days after the Capitol breach. I posted “What did Trump do, if anything, to incite violence?” to both communities. Phrased as the same kind of “I am trying to understand this myself” question I had posted on every contested topic since 2017. agnostic civility on the response thread: 2.5. IDW civility: 7.5. A 5.0-point gap. The IDW thread that day was the more thoughtful conversation about a contested political question. The agnostic thread was an outgroup eruption.

January 27, 2021. One week into the new administration. I posted “How is Biden doing?” to both sites. agnostic civility: 7.5. IDW civility: 3.5. A gap of -4.0. The same room that had erupted on the previous question handled the new president cool. The IDW thread, which had handled the previous question well, erupted on this one.

February 9, 2021. I posted “What dangerous, wrong, or false things do Conservatives believe?” to both sites. agnostic civility: 2.5. The thread carried 17 dehumanization events on the top 25 responses, the highest single-pair count in the corpus. IDW civility: 5.5. Identical structural pattern. The side whose outgroup the prompt named had the meltdown.

The same pattern works in reverse. November 9, 2020, the week the Pfizer vaccine emergency-authorization news landed. I posted a “would you take it” question to both sites. agnostic civility 8.5, with the science-trusting crowd handling it calmly. IDW civility 4.0, with the freedom-frame and skepticism crowd erupting. Most readers expect the opposite direction. The data says no. Civility was not a personality trait of either community. It was what each tribe was protecting.

And the cleanest negative case in the whole corpus. December 18, 2020. I posted “Should some research topics be banned or results censored?” to both sites. The same question, eighteen months later, that I would be characterized as having funded research about. agnostic civility: 6.5. IDW civility: 7.0. Both communities engaged the question seriously and stayed civil. The communities I had built were able to handle the actual hard question. The wider press apparatus, when it would later run a similar question through a publishing project I funded, was not.

Five years of running both rooms and writing seed essays about engineered civility. What the matched-pair data eventually told me, once I let it: civility online is what each tribe protects, not what each tribe is.

If you only remember one thing about this chapter, that’s it. Civility is not a possession. It is a defensive perimeter. Each tribe knows where its perimeter is. Each tribe melts down when the prompt crosses it. Each tribe is fine when the prompt is on its own side of the line.

I had been writing for three years that civil discourse was a weapon and not a courtesy. The matched-pair data agreed with the weapon part. It also taught me whose hand was on the trigger.

Chapter 6: The same person, two rooms

That’s the headline number. Out of 209,110 agnostic users and 145,827 IDW users, 15,957 of them registered on both. Match by email handed me 12,100 (75.8%). Match by username added another 3,857. 49% of those dual registrations happened the same day, which is partly an artifact, since a slice of accounts auto-provisioned across both sites in the same hour. It also tells you something about discovery. People heard about both rooms together and signed up for both at once. agnostic-first: 39.0%. truth-first: 11.5%.

Most of those 15,957 went inactive. Of all the dual accounts, 1,929 (12.1%) were ever active on at least one side. 1,172 active only on agnostic. 466 only on truth. 291 active on both, where active means real posting on each side, not just a registration.

291 is a small number. It is also the rarest possible thing in social-science research: a natural experiment where the variable is the room and the constant is the person.

Same person, on average, filed 23.3 times more abuse reports on agnostic than on IDW. Received 13.8 times more reports on agnostic than on IDW. Mean answers per dual-active user: 490 on agnostic, 123 on truth. A 4-to-1 ratio in favor of agnostic, on the same person. Comments: 455 vs 119, also roughly 4-to-1.

The one metric where dual-active users tilted toward truth was question-asking. 105 questions per dual-active user on agnostic vs 149 on truth. That fits the cluster picture from the previous chapter. agnostic was a conversation feed. truth was a question-and-debate room. Same person treated each room the way the room invited them to be treated.

Numbers like 23 times are easy to brush past. So I asked the AI to do something more granular. I pulled 108 humans who had posted on both sides at meaningful volume. For each one of them, I asked the AI to read what that person had posted on each side, separately, and score them on four things: civility, warmth, how often they argued, and how much of themselves they put on the page. Same human. Different door. Same scoring.

Civility: 7.25 on the agnostic side, 5.79 on the truth side. Gap -1.46. Median 7.5 vs 6.0.

Same person. Less civil. Less warm. More argumentative. Way less personal storytelling on the truth side. 69 of the 108 sample users dropped at least 0.5 civility points walking from one room to the other. 18 became more civil. 21 stayed within plus-or-minus 0.5. 69 added at least 0.5 argumentativeness points on truth. 28 dropped.

The model also returned a free-text “dominant register” tag for each side. On agnostic, the most-common labels were casual conversational, supportive storytelling, conversational sharing, storytelling, reflective storytelling, supportive and reflective. On truth, the most-common labels were analytical critique, combative and sarcastic, combative political commentary, terse opinion-stating, lecturing, combative political argumentation. Read those two columns aloud and you can hear what the room did to the same set of people.

What it actually looks like at the per-user level is not abstract. The same person on agnostic wrote about losing a parent, the grief of a long marriage that had ended, secular morality without religion, asked for advice about raising a freethinker child. On truth, the same login wrote a few sarcastic one-liners about people they considered ideologically lost, dropped self-disclosure to almost zero, and showed up only when an outrage prompt invited them to. Same login. Same human being. Two completely different speakers.

Multiply that by a hundred and you get the column on the right side of the table. Storytelling becomes lecturing. Reflective sharing becomes terse opinion-stating. Supportive becomes sarcastic. Self-disclosure scores drop by 2.36 points on a 0-10 scale, which is not a rounding error. It is the difference between here is what happened to me when my marriage ended and people who think X are obviously wrong. The same human being. Different room. Different speech act.

92 of the 108 (85.2%) were flagged as chameleons whose dominant register meaningfully changed between rooms. Not the exception. The dominant pattern in the sample.

If you had told me this in 2018, that the same person, given the same engine and the same humanist seed essays but a different pre-existing population, would shift 1.5 points on civility and 1.9 points on argumentativeness when they crossed into the other room, I would have been the engineer claiming his platform was less powerful than the population on it. I would not have liked saying that.

The 1.8% number sits in the corner of the table and refuses to leave. 1.8% of the 15,957 dual registrations turned into 291 people active on both sides. Out of those, 108 had enough posting on each side to score the same person twice. Out of those 108, 92 sounded like a different speaker on the truth side than on the agnostic side. The funnel from “registered both” down to “scored chameleon” is narrow. The signal at the bottom of it is stronger because of the narrowing, not weaker. It is what is left when noise has been filtered out.

This is the finding the sample-of-one cannot wave away. You can argue with the matched-pair data (“but those were different audiences”). You can argue with the cluster archetype data (“but those are statistical ghosts”). 23 times more reports filed by the same human being on the same login profile, and a 1.46-point civility shift on a 108-user sample, are the same person twice. They cancel out the personality variable. What is left is the room.

If you only remember one thing about this chapter, that’s it. The room shapes the voice more than the person does. The platform is downstream of the population. The population is downstream of why they came. The same engine, with the same builder and the same essays, can host two cities that bring out two different versions of the same human being.

I would still build the engine. I would build it differently if I were starting in 2026. But the lesson the dual-platform users taught me is the lesson the matched pairs prefigured and the cluster names set up. Civility is not in the platform. Not in the rules. Not even, exactly, in the people.

It is in the moment-to-moment social negotiation between a specific person and a specific room. Change the room, you change the negotiation. If the people came in for different reasons, you have already changed the room.

Three chapters. Three angles on the same finding. Same engine. Two cities. The verb the population brought through the door decided the welcome script. The welcome script decided the audience. The audience decided what the same person sounded like when they walked in. Outgroup-triggering prompts blew up both audiences in their own characteristic ways.

Chapter 7: The loud minority

When a controversial question came in, the reply count climbed, the report queue blinked, and my stomach did the thing where it asked whether I had built something worth defending or just another way for people to be cruel. For seven years I read those threads as if a community of 209,110 people on agnostic and 145,827 on the truth side were sitting in the same room arguing.

When we matched the cross-posts, 49 questions where the same essay landed on both sides within fourteen days, and we joined the response threads back to the behavioral-cluster IDs, the picture went sharp. On agnostic, two clusters generated 92% of every reply, every comment, every up-vote storm I had ever read on a heated thread. Thoughtful Elder Skeptics, 60.7%. Bruised Seekers, 31.8%. Across all 49 pairs that came out to 1,330 of the 1,439 clustered responses we could trace. On the truth side, the same shape. Heterodox Traditionalist Skeptics, 75.3%. Cultural Grievance Veterans, 16.4%. Two archetypes per room, and across roughly a hundred political threads they generated more than nine in ten of all the words I had ever read as “the community reacting.”

Thoughtful Elder Skeptics on agnostic is 313 people. 1.8% of the active users. Bruised Seekers is 859. 5.0% of active. Together about 1,170 individuals, 6.9% of an active universe of 17,062. When I had told myself “agnostic is upset,” what I was reading was the work of roughly twelve hundred specific accounts I would have recognized by handle if you had read them out to me one by one.

On truth the pattern is sharper as a fraction. The two answering clusters are 240 plus 780 users, roughly 1,020 individuals out of 7,522 active. 13.6% of the active universe. Twice the concentration of agnostic, similar absolute count. About a thousand specific people doing 92% of the talking on contested prompts.

The Seeking Connection Seekers cluster on agnostic, the modal participant who joined because they were the only nonbeliever in a religious county and wanted somebody to text on Tuesday nights, contributed 0.3% of responses on political pairs. They tuned that content out. They posted in the relationship threads, the holiday threads, the memorial threads when somebody’s dad died. They did not show up for “what dangerous, wrong or false things do conservatives believe?” or “how is Biden doing?” or “what did Trump do, if anything, to incite violence?”. They had a different reason for being on the platform, and the engine of cluster-level reactivity did not include them.

For most of the run of the two communities, I had assumed that when a thread got hot the room was reacting. The dashboard showed me reply counts and report counts and I believed those numbers were a measure of how a community of two hundred thousand humanists, or one of a hundred thousand heterodox readers, was responding to events in the wider country. That belief shaped how I built. It shaped which seed essays I wrote. It shaped which moderation tools I prioritized. It shaped how I felt when something blew up.

A blow-up on agnostic was 30 to 60 specific people writing in a thread. The other 17,000 active users were doing other things. Birthday threads. Recipe questions. Atheist parenting in religious schools. Whether their mom would come to a non-religious wedding. The political content was the loudest layer because the people most willing to write in it were the ones who showed up first, second, and tenth, on every prompt, with the same reflex.

Look at the spotlight pairs. January 13, 2021, “what did Trump do, if anything, to incite violence?”. On agnostic, the entire response thread that we could cluster was 4 archetypes worth of users. 31 Thoughtful Elder Skeptics. 25 Bruised Seekers. 4 Progressive Freethinkers. 1 Curious Newcomer. 61 people. On truth, 32 Heterodox Traditionalist Skeptics, 9 Cultural Grievance Veterans, 2 Heterodox Contrarians, 2 Heterodox Questioners. 45 people. The civility numbers were 2.5 on agnostic, 6.5 on truth. The dehumanization-event count on agnostic was high enough that I had remembered that thread for years. About a hundred people, total, on both sides, generated everything I had remembered.

I have built one kind of thing my whole career. I have built places where people connect at scale. Looking at the data here, I had not been building for the median user, even when I thought I was. I had been building for the loud users, because the loud users were the only ones whose reactions I could see in real time. The dashboards rewarded volume. The volume came from a thousand handles per side.

I look back now at the moderation tools I built in 2018 and 2019, the report-routing heuristics, the soft-hide thresholds, the weekly-volume alerts. They were all sized to a community-scale event. They almost never fired on a community-scale event because there almost never was one. They fired on a cluster-pair flare-up. If I had been honest about the size of the audience that actually generated the volume, I would have built different tools. I would have spent my time on the thirty people whose individual report-rates dwarfed the rest of the active membership combined.

The pre-political baseline numbers, before September 2020, said the same thing. Thoughtful Elder Skeptics had been 68.0% of all matched-pair responses pre-election. After the election they were 58.6%. Bruised Seekers shifted up from 25.4% to 33.6%. The political wave drew in another wave of recently-deconverted users who had cooled off the religion fight and were looking for the next one. On truth the equivalent move was small. Heterodox Traditionalist Skeptics went from 73.1% to 76.0%. The conflict core was already there for whatever I posted, in similar proportion.

If I were running these communities again now, I would build differently. I would still write the seed essays for the median seeker or the median open-question asker, because those were the people the rooms were nominally for. But the moderation engine would be sized for cluster pairs. The civility tools would be sized for individual feuds. The dashboard would be sized for the realization that the loud is not the median, ever, anywhere on the internet.

The other 95% on each side were quietly using the platform for what I had said it was for.

If you only remember one thing from this chapter: when somebody says “the internet is angry” or “the agnostic community melted down” or “the heterodox forum lost it,” they almost always mean about fifty specific people lost it. The room is mostly asleep. The room mostly always has been.

Chapter 8: One conflict couple per side

The way I had pictured a healthy community was that conflict, if it happened, was diffuse. Lots of small disagreements, no single fight dominating. A bell curve. When you build moderation systems you instinctively build for that.

Thoughtful Elder Skeptics filed 702 reports against Bruised Seekers across the lifetime of the platform. That is 20.6 times the off-diagonal mean of the entire conflict matrix. The reciprocal, Bruised Seekers filing back against Thoughtful Elder Skeptics, is 202 reports, 5.9 times the mean. Then Thoughtful Elder Skeptics against Tentative Truth-Seekers, 205 reports, 6.0 times. Thoughtful Elder Skeptics against Progressive Freethinkers, 177 reports, 5.2 times.

Most of agnostic’s 73,502 user-abuse rows is one tight knot. One small cluster of 313 elder skeptics, going at one larger cluster of 859 bruised post-religious people, and the bruised group going partway back. A couple of side channels off the same elder-skeptic cluster, hitting two more groups. Everybody else is reading past the fight. The block button and the report button on agnostic are tools that a small reciprocal group used on each other, several thousand times a year, while the rest of the community talked about other things.

There is exactly one off-diagonal edge above 5x the mean. Heterodox Traditionalist Skeptics filing reports against Cultural Grievance Veterans. 49 reports. 5.4 times the off-diagonal mean. Below that, the cells thin out fast. Same engine, two communities. One had a single dominant feud running for years between two cluster-defined groups. The other had a quieter, sparser pattern, where reporting was rarer, less reciprocal, and less concentrated in any single archetype-pair channel.

The top filer on agnostic is a handle called GreenAtheist. 4,971 reports filed. 1,522 reports received. The same account is the most prolific reporter and is one of the most reported users on the entire site. That overlap is the structural marker of high-conflict community. The people most actively policing the platform were the same people most actively being policed back. The top 10 power-reporters on agnostic generated 22.5% of all 71,896 reports.

The top filer on truth is a handle called RayitoDLuz. 496 reports filed. Zero reports received. Reporting was rarer and more impersonal. People filed quietly and were not getting filed against in return.

I have known some of these accounts for years. I am not naming them here for any reason except that they are in the public abuse log, the data is the data, and the comparison is what tells the story. I would not write any of this as a complaint about any specific user. They were doing what the community structure rewarded. If you put 313 elder skeptics into a community where the modal new arrival is a recently-deconverted person processing grief, conflict, and shame from a life inside an evangelical or Catholic system, and you ship them all into the same comment threads at midnight, you are going to get 702 reports flowing one way and 202 flowing back. The platform did not cause that. The platform amplified what the population brought in.

Thoughtful Elder Skeptics filed 4,174 content reports across 313 users. That is 13,335 reports per thousand users. The next cluster on the same site, Bruised Seekers, came in at 1,179 per thousand. After that the rates collapse. Progressive Freethinkers, 127 per thousand. Seeking Secular Community, 110. Seeking Connection Questioners, the modal user I was supposedly building for, 11.5 per thousand. Three orders of magnitude difference between the loudest cluster and the median cluster.

On truth the top cluster, Heterodox Traditionalist Skeptics, came in at 2,017 reports per thousand. After that the rates also collapse, but the second-rank cluster is 47 per thousand, not 1,179. The shape of the conflict tail on truth is shorter and sparser. One cluster does most of the policing, but the policing is rarer in absolute terms.

The Gini values say the same thing in one number. Reports made on agnostic have a Gini of 0.832. On truth, 0.720. Reports received on agnostic, 0.702. On truth, 0.554. Conflict on both sides was concentrated. Conflict on agnostic was dramatically concentrated.

On agnostic, reported content concentrates on in-group concerns. 26.1% platform meta, the meta-arguments about moderation policy itself. 15.7% spam. 9.5% religion and atheism. 6.2% relationships and dating. The community policed what it was about. Religion. Connection. Whether the moderators were doing their job.

On truth, 21.7% of report events were race, ethnicity, or immigration. On agnostic that bucket was 4.6%. A 4.7x difference, IDW the higher side. 14.8% of truth report events were content critiquing right-leaning politics. 7.0% on agnostic, a 2.1x gap, IDW the higher side again. Same engine. Same rules. Same moderator. Two completely different conflict surfaces, because the populations chose different things to fight about, and inside each of those choices a small subgroup did most of the fighting. Different mechanism. Different topics. Same defending.

The time spikes line up with the outside world. agnostic spiked December 2017 through June 2018, a thirty-week window covering MeToo, the Roy Moore allegations, and the family-separation crisis at the southern border. Peak 1,653 reports in a single week. agnostic spiked again the week after January 6, 2021. Peak 234. truth spiked October 2020, the last weeks of the election. Peak 153. truth spiked again Roe v Wade week, May 2022. Peak 45. Outside events were the trigger. The capacity was the small power-reporter tail on each side, ready to file the moment anything happened.

The honest version of “agnostic was hotter than IDW” is therefore not a story about which national tribe was meaner. It is “313 elder skeptics on agnostic and 240 traditionalist skeptics on truth ran almost the entire moderation surface of their respective communities for years, and the agnostic version of that small group was several times more report-active per capita than the truth version.” Different rooms, different cultures, different small groups, different reflexes when something hit the news. The blocks-per-thousand-active-user-months ratio, after we controlled for size and engagement, came out to 4.7 to 1. The user-reports-per-thousand-active-user-months ratio came out to 9.6 to 1. The original 17x and 29x and 10x ratios shrank under the tougher denominator but they did not disappear.

If you only remember one thing from this chapter: in any room of a few hundred thousand people, there will be a few hundred who do most of the formal fighting. In agnostic those few hundred fought one specific other group reciprocally. In truth they did not. That is the actual difference between the two communities. Not the median. The tail.

Chapter 9: Aporia and October 2024

I funded a Substack called Aporia. The funding ran through the end of 2024. The public corpus is 832 articles. Behavioral genetics. Demography and migration. Philosophy and meaning. Book reviews of mainstream and academic books, including measured reviews of religious thinkers. Podcast interviews with named mainstream academics. The articles are still up. Anyone can read them.

I had built two communities for five years and what I read in my own data was that contested questions could be talked about, civilly, by people who came in willing to talk. I knew one corner of the academic world had begun to back away from certain subject areas because of the costs to careers, not because of the quality of the science. I thought, naively, that if a public publication would take those questions seriously, write about them carefully, cite the literature, interview mainstream academics, and not engage in hate, that it would be a useful thing in the world. I believed, in the language I later wrote into my public statement at conru.com/truth, that censorship was a form of hate and open inquiry was a form of love.

I am an engineer. Engineers see broken systems and try to fix them. That was the worldview going in.

In October 2024 a coordinated press cycle came out. The framing was that the funding had been part of a hidden network promoting the worst kind of pseudoscience. I had a few hours of notice. I made a fast decision to issue a disavowal of the magazine to protect the people working with me in Seattle, and then I went quiet for about ten months while I built. In late 2025 I posted my own statement at conru.com/truth/index.html. The statement is public, it is signed by me, it is the most considered version I can write of what I believe happened, and I will not be repeating the legal-defense-y parts of it here. The book is not a defendant’s document.

I had assumed I was engaging with a system of dialogue. I was not. The medium I had collided with runs on different rails. It does not optimize for understanding the question. It optimizes for narrative, for compression, for the framing that produces the most reaction in the shortest paragraph. A funding decision and an association will do, if the framing is good enough. The funding I provided, however careful the publication tried to be, gave that medium a story, and the story did the work the underlying writing could not be made to do.

Owning the naivety is the part of the public statement I had to write multiple times before it sat right.

I do not retract the underlying commitment to open inquiry. I do retract the assumption that open inquiry, on its own, in a system that is not playing by the same rules, gets through. The lesson is the one in the second page of the public statement. You cannot reason people out of indoctrination. Facts and debate are not enough when trust has been broken.

I am writing this chapter about three years after that month. I am not relitigating press claims here. The claim-by-claim work exists at conru.com/aporia, the public corpus is searchable, and I trust the reader to read whatever they want to read. I am also not naming any of the journalists, outlets, or critics in this book. They have their work. I have mine. The book is about what the data of my own communities can tell us, and that data has its own thing to say about this episode that I think is more useful than any defense I could write.

The cleanest single data point from the whole experiment is from December 18, 2020.

This was a year and a half before I started funding Aporia. The question is roughly the same question that the press cycle, two years later, would not be able to discuss without it dissolving into something else. On the day I posted it, agnostic engaged it civilly. The AI scored the agnostic side at 6.5 on civility. IDW engaged it civilly too, and scored 7.0. The responder distributions are in the data. About 18 people from agnostic, mostly Bruised Seekers and Thoughtful Elder Skeptics, working out their views on whether scientific questions should ever be ruled out of bounds, mostly arguing that the answer is no, you don’t ban scientific questions, you do them carefully and you publish the findings. About 26 people from truth, mostly Heterodox Traditionalist Skeptics, Cultural Grievance Veterans, and Heterodox Contrarians, mostly arguing on the same side. A few people on each room dissented and were heard. Nobody got reported. Nobody got blocked. The threads are still in the database.

The same question, two years later, routed through a different medium, was not held. The coverage that came around it did not engage the underlying inquiry. It engaged the funding, the associations, the optics. The communities I had built had a higher signal-to-noise ratio on that exact question than the wider apparatus did, by a wide margin.

I do not say that triumphantly. The communities had their own meltdowns when the prompt was different. Chapter 5 is the catalogue of where they cracked. The Trump-incitement prompt, agnostic civility 2.5. The first-month-of-the-new-administration prompt, truth civility 3.5. The dangerous-things-conservatives-believe prompt, agnostic civility 2.5 with seventeen dehumanization events on the worst pair. They were not always able to hold a contested question. They were able to hold this one.

The thing the data tells me, and the reason this chapter exists, is that the failure was not at the community-engineering layer. The community-engineering layer had handled the underlying question. The failure was at the wider-discourse layer, the one I had no engine for, no moderation policy on, no way to write seed essays for. When the press cycle came in October 2024, it was running on different rails. Those rails do not optimize for civility. They do not optimize for inquiry. The ones I had built did.

It is also why I stopped trying to engineer this kind of dialogue at the community-platform level after 2022, and pivoted into something else.

The thing I had believed since 2017 was that you could engineer civility into the medium itself, and a community would learn to rehumanize each other on the rails of a well-built platform. The data from the communities is that you can engineer civility into the medium for some questions and the rails partly hold, and you cannot engineer it for every question, and the rails outside your platform run on a different physics altogether. The Aporia episode taught me, in a way that my own dashboards had not been blunt enough to teach me, that even when a community can hold a question, the wider system around it might not let the question be a question.

If you only remember one thing from this chapter: I posted, on December 18, 2020, a question about whether certain research topics should be banned or results censored. Both rooms I had built could discuss it without anyone getting hurt. Two years later, work that asked the same kind of question through a different medium was eaten by a system that does not know how to discuss it without hurting someone. That is not a story about who was right. That is a story about which rails the question was running on.

Act III: The Pivot

Chapter 10: Why dialogue alone failed

The rails I am building now sit on the failure mode of the engineering hypothesis I started with.

The 2017 hypothesis was that civility was a portable virtue. Engineer it into the medium, and the people in the room would learn to talk to each other. The first nine chapters of this book are what the data said back. The hypothesis was wrong, in a specific way.

Through 2017, 2018, and most of 2019, the platform-housekeeping civility scores on both sites lived in the 8 to 10 range. “How can we improve search?” rated 9. “What’s broken in the welcome flow?” rated 9. Even the early religion-and-ethics threads sat near 9 on agnostic. People who had joined to find a friend in a religious county, or to ask Mill-style questions about heterodox ideas, talked to each other like adults. The engineering worked at the housekeeping layer.

Chapter 5 walked the time arc pair by pair. The shape, recapped here in one paragraph: from launch through summer 2020, both rooms scored in the 8s on platform-housekeeping prompts. The 2020 election crashed both into the mid-5s in parallel across 17 pairs. Then, on 5 pairs spanning the Capitol breach and the inauguration, agnostic dropped to 4.04 while IDW rose to 6.26. A swing of 2.22 points in opposite directions, at the same time, on the same site infrastructure, with the same moderator (me) reading every thread.

Recall the three spotlight pairs from Chapter 5. The Trump-incitement prompt six days after January 6: agnostic 2.5, IDW 7.5. The Biden first-week prompt two weeks later: agnostic 7.5, IDW 3.5. The dangerous-beliefs-of-the-other-side prompt in February, which carried 17 dehumanization events in one thread: agnostic 2.5, IDW 5.5.

Each room kept its 8-out-of-10 civility for whichever question was inside the tribe’s protective fence. Each room melted down only when its own outgroup was named in the prompt. The asymmetry was a function of who got named, not a property of the people who were typing.

I have to be careful here, because I am not the one who gets to declare which tribe has the worse manners. Eight years inside both rooms does not buy me that, and the data does not support it. The 17x system-level abuse-report ratio on agnostic, the one I used to quote at investors and at parties, is real. Most of it is one cluster pair, Thoughtful Elder Skeptics and Bruised Seekers, reciprocally combatting each other while everyone else read past the fight. Take that pair out, and the ratios shrink. They do not vanish, but they shrink.

The honest claim is the one the matched pairs force. Both communities melted down. Each one melted down on its own outgroup. The medium amplified both. The moderator, sleeping six hours a night, taking it personally, did not slow either of them down.

The December 18, 2020 prompt about whether research topics should be banned, the cleanest single point in the corpus, the bridge to Chapter 9. Both rooms engaged it civilly, agnostic 6.5, IDW 7.0, on the same kind of question the wider public conversation could not hold three years later.

The communities passed the test. The wider apparatus did not. I want to keep both halves of that sentence.

I will give you the version I would say to a friend at a bar. We were trying to engineer a virtue at the wrong layer. Civility is not what a tribe is. Civility is what a tribe protects. Inside the protected zone, where the ideas being tested do not name anyone the room loves or hates, civility is portable and durable and beautiful to watch. Outside that zone, where the prompt names an outgroup, civility flips into its opposite use. It becomes the polite vocabulary of expulsion. The same person, in the same chair, in the same room, can run both modes within ten seconds of each other.

I know this is true at the individual level because of the cross-site members. 15,957 people had accounts on both rooms. 49% of them joined on the same day. 291 of them posted on both sides. The same individuals filed 23.3x more abuse reports on agnostic than they did on IDW. The same individuals received 13.8x more reports back. On 108 dual-active users we had the AI score on both sides, 85% were chameleons. The same human shifted 1.46 civility points down and 1.89 argumentativeness points up depending on which room they were typing into. The room shaped the voice more than the person did.

The 2017 me, writing the 36 humanism site_text essays at the launch of agnostic, was making a specific claim. I claimed that civility was a virtue you could grow in someone with the right environment, the right defaults, the right prose at the front door. The 2026 me, looking at the cross-site data, has to revise that. Civility is a feature each tribe extends to its in-group questions and withdraws from its outgroup questions. The portable part is the protective reflex, not the virtue.

I am not saying the rooms were a mistake. The rooms held 8-out-of-10 civility on every housekeeping question for almost three years. That is a real thing that happened, on the rails of a platform I built. I am also not saying the people in the rooms were bad. Most of them were extraordinary. The 22% Seeking Connection Seekers cluster on agnostic, the people who had joined to find a friend in a religious area, contributed 0.3% of responses on political pairs. They tuned out. Most of each room was there for something else, and most of each room got something else.

What I am saying is the engineering hypothesis from 2017 needs revising. You cannot build civility into a medium and expect it to hold under outgroup-triggering prompts. You can build civility into a medium and expect it to hold under prompts about how to run the medium itself. That is a smaller claim than the one I made in 2017, and it is the one the data supports.

If you only remember one thing from this chapter, remember that the engineering hypothesis from 2017 was right at one layer and wrong at another. Right at the platform-housekeeping layer, where the rails carried any question about how to run the room. Wrong at the political layer, where the same rails carried the meltdown the moment the prompt named an outgroup. The mistake was thinking civility was portable. It was not. It was tribal. The rails amplified the tribe.

Before you post a question to a room you care about, ask three things in order. First, am I trying to test an idea, or am I trying to test the tribe. Second, would I be willing to be persuaded against my position in the next ten replies. Third, if I were on the other side of the table from this exact prompt, would I read it as an invitation to think, or as an invitation to defend.

If you cannot answer all three with a clean yes-it-is-an-idea, you-could-persuade-me, it-reads-as-invitation, do not post the question to that room. Take it to a one-on-one with someone who disagrees with you. Or sit with it for a week. Or paint it.

The question is not whether your room is civil. The question is whether the room is currently inside or outside the protected zone for the prompt you are about to type.

Chapter 11: BTL, Beauty, Truth, Love

Not a new brand. A new thesis. The 2017 thesis had been: engineer civility into the medium and the medium will produce rehumanization. The 2026 thesis had to account for what I had actually watched happen, which was that the medium produced civility only inside the tribal protective fence and produced its opposite outside it. The framework I had been writing in the agnostic and IDW essays needed to either retire or get rebuilt at a different layer.

The rebuild was not clever. It was three words I had been writing toward without naming. I went back through the 36 agnostic site_text essays from 2017 and the 47 IDW site_text essays from 2019. Both sets kept reaching for the same triad and both sets called it different things. Beauty. Truth. Love. The IDW lane carried the explicit BTL framing in 0.8% of my posts; the agnostic lane carried it in 0.1%. The phrase did not stabilize until later. But the shape was there from 2017 onward, hiding under different vocabulary.

Beauty, Truth, and Love. Three functional values, not soft ones. The opposites are ugly, lies, and hate. A culture is doing one set of things or the other set of things at any given moment. You can run the test on a piece of media or a public figure or a Wednesday night dinner. Is what I just consumed pulling me toward Beauty or toward ugly. Toward Truth or toward lies. Toward Love or toward hate. The test takes about three seconds.

I will say what each one means in the way I mean it, because the words have been used badly enough that they need re-explaining.

This is not an aesthetic claim. It is a defensive-system claim. I watched 17 dehumanization events accumulate on a single agnostic thread when the prompt named an outgroup. I watched the same thread structure dissolve when the prompt asked people to share something they had loved looking at that week. The brain that produces the 17 dehumanization events and the brain that posts the photograph of the dahlia is the same brain. The trigger is different. Beauty disarms the trigger. It does not argue with the armor. It walks past the armor.

If I had understood that in 2017, I would have built a gallery before I built the second forum.

It is not the abstraction debaters fight about. It is hands. Faces. Light coming in through a north window. A man at a bar at three in the afternoon. The classical-realist painters at the Atelier and the seven Seattle Prize fellows who arrived September 9, 2025 are doing this. They paint people as they are, and a painting of a person as they are makes ideological sorting harder, not easier. You cannot easily put a tribal label on a face that has been studied for forty hours. The face starts to assert itself against the label.

In the wilderness years between 2007 and 2014, I made 365 paintings, one a day for a year, mostly bad. The thing I learned is that painting a person is a moral act before it is an aesthetic one. You sit and you look at a human and you do not get to skip the parts that do not fit the story you wanted to tell. The training of the eye is the training of the seeing. If we are not seeing each other, no amount of dialogue platform will fix it. If we are seeing each other, no amount of platform is needed.

That is a strong claim. I am holding it lightly. It is what the rebuild produced.

The 36 agnostic essays were trying to build it through prose. The 47 IDW essays were trying to build it through prose with a different framing. Aporia, in 2022 to 2024, was trying to fund it at the academic-publication layer. None of those produced the connective tissue at the scale or the durability the hypothesis predicted. What produced it, in my own experience, was the slow physical thing. A weekly Wednesday social where someone makes coffee and someone else hangs a painting and a stranger walks in and stays. Not metaphorically slow. Actually slow. Hours, not seconds.

I do not know how to scale this in the way I knew how to scale a dating site. That is part of the point. The thing I am trying to build now does not lend itself to the same kind of scale that produced 700M+ registered members at FFN by 2007. It scales by repetition, room by room, weekly. It looks small. It is small.

I want to be honest about how I got to BTL because the path matters. I did not arrive at it from a manifesto. I arrived at it from the failure mode of the engineering hypothesis. The 2026 update is: dialogue is the wrong medium for rehumanization once trust has broken. Beauty is a medium. Truth, in the depiction sense, is a medium. Love, in the slow physical sense, is a medium. They bypass the armor that ate the dialogue rooms. They are the next experiment. They are not above the previous experiment. They are downstream of what the previous experiment taught.

BTL is not a brand. It is the rebuilt thesis after the data update. If it ever shows up on a wall somewhere as a logo, the wall is wrong. The framework is meant to be used, not looked at.

If you only remember one thing from this chapter, remember that BTL is the rebuilt thesis after the data update, not a new vocabulary. The 2017 thesis claimed you could engineer civility into a medium. The 2026 thesis is that you can engineer it for some questions and you cannot engineer it for outgroup-triggering ones, and what you reach for instead is Beauty in the disarming sense, Truth in the depiction sense, and Love in the slow physical sense. The data made me update. The framework is what came back.

Here is the take-away you can run, in three questions, on a project you are considering.

First, does this project bring people closer to honest depiction of each other, or further from it. A photo essay that shows the actual faces of an unfamiliar group is closer. A meme that flattens a group into a single label is further.

Second, does this project lower the cost of being persuaded, or raise it. A salon where you can change your mind without anyone keeping score is lower cost. An online debate that will be screenshotted for the rest of your life is higher cost.

Third, does this project leave space for Beauty, in the disarming sense. A long meal does. A pile-on reply does not.

If a project answers cleanly toward the first half on all three, it is in the rehumanization lane. If it answers cleanly toward the second half on all three, it is in the dehumanization lane. Most things in life are somewhere in the middle and need a little careful pulling toward the first half to be worth the time.

I run this test on what I build. I run it on what I read. I run it on what I post.

The 2017 test I ran on my own work was: is this room civil. The 2026 test I run on my own work is: is this room rehumanizing. The first test passed for almost three years and then failed under predictable triggers. The second test, run weekly on a physical room across from the Seattle Art Museum, has been passing for about eighteen months. I will keep running it every Wednesday and I will tell you in 2030 whether it kept passing.

That is what I have. That is what BTL means in my hands. The rest of this section of the book is what I built with it.

Chapter 12: The art instruments

The Conru Art Foundation is the running implementation of the BTL thesis. I want to walk through the pieces because the through-line matters more than the brochure does. Each program tests something. None of them is a closing argument.

Conru Art Foundation is the non-profit shell, registered in Washington. Public domains: artlove.org, conruartfoundation.org, seattleprize.org, artlovesalon.org, publicdisplay.art. The mission line I keep coming back to is “personal wealth held in the public trust.” That is not a fundraising line. It is the operating instruction. Profit, if any program ever throws one off, is a byproduct. The framework is BTL. The work is what you do on Wednesdays.

Art Love Salon is 16,000 square feet at 110 Union Street, downtown Seattle, across from the Seattle Art Museum.

The Salon tests whether a physical living room can do what an online room could not. The five plays are simple. Campfire (a weekly Wednesday social, 4 to 7 pm, music and coffee, free). Workshop (paid teaching artists running no-fear public classes). Celebration (galas and partner events). Human Story (artist bios and 30-second videos with a Struggle, Breakthrough, Why structure). Sanctuary (open quiet hours, tours).

The Salon space took longer to set up than I planned. The first few Wednesday Campfires were awkward. Three people, two of whom were on payroll. I went home one of those nights and asked myself if a 16,000 square foot room with three people in it was a stupid version of what I had been doing online with 209,110 users at peak. The honest answer was that the comparison was the wrong axis. The online room had 209,110 user records and could not hold a hard question without melting down. The physical room had three people and could hold any question for as long as we wanted to sit with it. The Salon is not a smaller version of the platform. It is a different layer.

The Atelier, also known as the Masters Academy of Art, lives on the third floor of Occidental Fine Arts Center at 311 1/2 Occidental Avenue South.

I have walked the third floor of Occidental on a quiet weekday afternoon and watched a row of students working from the same plaster cast for an hour without anyone speaking. That is the Atelier. The test it is running is whether a 19th-century-style classical realism school can produce 21st-century working artists. The question is not nostalgia. It is whether the slow eye-training I described in the BTL chapter can be taught at scale in a working studio in 2025 Seattle. The faculty pipeline runs through Tenaya Sims as artist facilitator. Workshops, studios, the curriculum, the cast room. If the answer is yes, the school is one of the rooms. If the answer is no, I will have learned that the BTL thesis on the Truth side cannot be transmitted through this kind of pedagogy and will need a different one.

It is a $750,000-per-year fellowship for representational painters. Season 1 launched September 9, 2025 with 7 fellows, anchored at Occidental Fine Arts. Stipends in the $50,000 to $80,000 range per fellow. Studios on the third floor. Documentary crew filming weekly. The framing is Renaissance-style patronage at modern scale, not residency, not grant. The Prize tests whether a single foundation can pay representational painters enough to make the painting the primary work for a year, and whether the body of work that comes out the other end is honest depiction in the BTL sense. I got the stipend numbers wrong on three different versions of the press kit before we standardized them. I also under-budgeted the housing logistics by about a quarter. Season 2 is currently in application, the stats are now consistent, and the housing line is fixed.

It tests whether sitting for a portrait, by a trained representational painter, with a documentarian in the room, can do something a couples-therapy session does. There are two pilots in the planning docs (a couples track and a tribute track) and they share a venue, a crew, and a question about what a portrait actually does to the person who sits for one. A potential partnership with the University of Washington Psychology Department is being clarified. The hypothesis is small and weird. I like that about it. If a couple sits for a six-week commissioned double portrait and at the end of it the couple has a different kind of conversation, the data point is worth more than another forum thread.

It is the buyer-facing art discovery directory, the artist platform, and the admin spine for everything CAF runs. PDA is where my engineering training has the most direct application to the art work. We rewrote 221,000 lines of the codebase in two months in 2025, one operator and a fleet of AI agents, migrating the legacy PHP backend toward Node. PDA tests whether the engine that scaled an online dating system to 700M+ registered members lifetime can be re-pointed at honest art discovery without becoming the same kind of attention-hijacker the dating space is full of. I am the wrong person to ask about that test in 2026. Ask in 2030.

Wurlitzers, orchestrions, slot machines, vintage radios, vending machines, a handful of fortune-telling cabinets. Currently displayed on the fifth floor of the Poll Building, intended for a future mechanical wing of the Conru Museum. The collection tests one thing in three modes. Visible mechanism is Truth. Sculptural cabinetry is Beauty. Joyful interaction across generations, when an eight-year-old and an eighty-year-old put a quarter in a 1925 Wurlitzer at the same time, is Love. I picked these because they make BTL legible in a way a wall label cannot. They are also delightful, which is a real category. I have been collecting them for longer than I can publicly justify.

The CAF documentary is in production now, with a Sundance and Tribeca 2027 target.

I have watched dailies from the first three months of fellow shoots and the thing that surprised me is how much of what is on tape is silence. A fellow at an easel, alone, for forty seconds before they say anything. Phase 1 pre-production ran through summer 2025. Phase 2 principal photography started September 2025 and runs through August 2026. Three months of fellow footage is logged. The working logline is one sentence: the challenge of rehumanism in a modern world, can we do it. The documentary tests whether the experiment narrates as a single film, or whether the through-line only resolves on the page and the screen needs a different shape. We will know in 2027.

I am leaving out the Conru Museum on purpose. The flagship museum is a separate, multi-year build with architecture, permits, and a 2029 target window, and it depends on the Salon and the Prize and the Atelier doing their work first. If those three keep passing the BTL test through 2027, the museum gets built in the form the Salon has been practicing for it. If they do not, the museum gets a different shape, or does not get built.

These are not seven programs. They are one experiment with seven instruments. Each one tests a piece of the BTL thesis at a different scale. The Salon is a weekly room. The Atelier is a year of studio time. The Prize is twelve months of patronage on a single fellow. The Commission Artists pilot is one portrait sitting. PDA is the digital substrate. The Mechanical Music collection is one room of physical objects. The documentary is one film. They share a building stack (Poll Building plus Occidental), a calendar, a moderator (me, again), a budget, and a hypothesis. The hypothesis is BTL run as an institution rather than as a Substack.

If the engineering chapters of this book taught me anything, it is that the moderator never leaves the room. I am still moderating. The room is bigger and quieter and physical. I sleep more.

If you only remember one thing from this chapter, remember that the seven instruments are one experiment running at seven scales, with one builder, on a longer clock than the platforms used. Weekly room. Year of studio time. Twelve months of fellow patronage. One portrait sitting. One digital spine. One room of physical objects. One film. The 2017 thesis was tested through dialogue in five years and failed under known triggers. The 2026 thesis is tested through the rooms above on a ten-year clock. The verdict lands in 2036.

You do not need 16,000 square feet, $750,000 a year, or a documentary crew. You need a regular night. You need a room with one piece of beautiful work in it. You need three to seven people who will come back. You need to be willing to pour the coffee yourself for the first six months, which is also when you will think it is not working. The test for whether you have a CAF-shaped room in your city is whether, six months in, on a Wednesday at 5:45 pm, two strangers are talking about the painting on the back wall and forgetting that they are strangers. If that is happening, you are running the experiment. The instruments are optional. The room is not.

Chapter 13: What the next experiment is

The honest answer to “what is next” is “more of the same kind of thing, with two new tools and a longer horizon.”

I am 57 in 2026. The thirty-year clock from here lands at 2056. I do not plan in thirty-year increments because nobody can. I do plan in ten-year increments because that is roughly how long the previous experiments took to produce a verdict. WebPersonals 1994 to FFN sale 2007 was thirteen years. agnostic.com 2017 to the data autopsy in this book is nine years and counting. The next ten years are the current experiment.

We already use it. PublicDisplay.ART runs an AI labeler that drafts gallery wall labels for the Atelier and the Salon, and a human edits the draft. We rebuilt 221,000 lines of the codebase in two months in 2025 with one operator and a fleet of AI agents. The corpus analysis behind the agnostic and IDW chapters of this book was an AI exercise on the dumps from both communities, with me checking the flagged outputs by hand. I am writing this paragraph, and you can probably tell, by talking out loud and letting a system clean up the cadence, with me reading it back and rewriting the parts that drift.

A weekly salon with an AI in the loop is not science fiction. It is a concierge that remembers who came last week, what they were working on, what painting they stopped in front of, and what they said to whom. It is an artist platform that can read 200 portfolios and write the first draft of the curatorial statement, with the curator editing. It is a documentary footage system that tags 600 hours of fellows interviews against a thematic spine and lets the editor pull by arc rather than by date. We are running pilots of all three in 2026.

The honest part: I do not know yet whether AI in the salon loop is rehumanizing or dehumanizing on net. I have watched enough AI products in the last eighteen months to know it goes either way. The test for me is the one I gave you at the end of the BTL chapter. If the AI lowers the cost of being persuaded, raises the chance of two strangers seeing each other across a painting, and leaves space for Beauty in the disarming sense, it is in the rehumanization lane. If it does not, we will pull it out of the loop and the room will go back to ink and coffee.

In 2021 I bought back Friend Finder Networks, the dating company I founded in 1996 and sold in 2007. The reacquisition was about $60M. The structure is a public-trust arrangement. I take no personal profit from member operations. Current scale is roughly 150,000 paid and 1.5 million active. From 2024 through 2026 we are repositioning it from the adult-dating frame it carried for two decades toward something I have been calling the Desire Spectrum, with friendship and connection at the center and the mature elements as one band of the spectrum rather than the marketing.

The Desire Spectrum reframe is a small idea with a long tail. Most of what the dating-platform decade of my life produced was the same lesson, learned over and over again. People are looking for a wider band of human connection than the product categories let them ask for. A 2026 Friday night for a single 38-year-old in Seattle is not a binary between a hookup app and a meditation retreat. It is a spectrum of wanting to be known. We are building toward that spectrum, with the engineering I learned from running 700M+ lifetime registrations and the ethics I learned from running two communities that taught me what the engineering does when you tune it for the wrong thing.

I am rebuilding this company a second time, twenty-eight years after the first build, with the lessons applied.

The lessons come from two sources. One is my own corpus, agnostic plus IDW, where the data shows that what I built in 1996 and what I built in 2017 and what I built in 2019 were the same kind of thing in different rooms, and that the medium amplified the part of human nature it was tuned for. The other is the BTL framework that came out of the 2024 to 2026 rebuild. A connection company run on BTL principles, with the operator taking nothing off the top and the platform tuned to depict members honestly rather than hijack them, is the version of FFN I would have built in 1996 if I had known then what I know now.

I do not know if it will work. I know I will do it half-decently rather than half-assedly, which is the rule my mother gave me on the farm in 1982 and which has held up.

Sundance or Tribeca 2027 is the festival target. Phase 2 principal photography runs through August 2026. The film tests whether the experiment narrates as a single piece on the screen, with seven fellows, a mechanical music collection, a salon, and an engineer sitting in the middle of it. If the through-line lands on screen the way it lands on the page, the documentary is the version of this book a person without an evening to read can encounter. If it does not, I will have learned that some experiments resolve on paper and some do not resolve on screen, and I will adjust.

The public-trust framing is the fourth piece, and it is the one I want to be clearest about.

I came up through the 1990s tech world and the 2000s dating-platform world and exited one in 2007 with more money than a person on a 10-acre farm in Indiana ever expected to be near. I have spent most of the years since 2014 figuring out what to do with that money in a way that does not embarrass the kid who grew up with union steelworker parents on that farm. The framing I have settled on is “personal wealth held in the public trust.” It is not a tagline. It is a set of operating constraints. The Salon is zero-commission. The Prize is paid out without expectation of return. FFN is run with no personal profit from member operations. The foundation pays the bills for all of it.

The 2017 me thought the correction was a humanism platform. The 2019 me thought the correction was a heterodox dialogue platform. The 2022 me thought the correction was an academic publishing project. The 2026 me thinks the correction is an art foundation, four buildings, a documentary, a connection company that runs at cost, and a slow weekly room with coffee. I am not sure 2036 me will agree with all of those. I am sure 2036 me will be running the same experiment in some form.

I built a dating platform in 1994 to help people find each other. I built FFN in 1996 to scale that. I built agnostic in 2017 to give people without a religion a place to ask hard questions in good company. I built IDW in 2019 to give people with a different kind of hard question the same kind of place. I funded Aporia in 2022 to extend it to academic publishing. I built the Salon and the Prize and the Atelier and the Commission Artists program and PDA and the Mechanical Music collection and the documentary in 2024 through 2026 to do it through art instead of through dialogue.

Some of them have walls. Some of them are codebases. One of them is a film. One of them is a fellowship. The constant is the room and the people in it and the question of whether two strangers can come out of it knowing each other a little more honestly than they went in.

I am going to keep building rooms. I am going to keep running the BTL test on each one. I am going to keep writing the data up when I have something that holds. I am going to keep underestimating the housing logistics on the next fellowship and getting the stipend numbers wrong on the next press kit and pouring the coffee myself on the first six Wednesdays of any new salon. The building is the practice. The practice is the experiment. The experiment is the life.

If you only remember one thing from this chapter, remember that the through-line has not changed since the Commodore in Indiana in 1982. The rooms have. WebPersonals in 1994 was a room. FFN in 1996 was a room at scale. agnostic in 2017 and IDW in 2019 were two rooms with the same engine. Aporia in 2022 was a publishing room. The Salon at 110 Union Street is a room with a coffee maker and a painting on the back wall. The next experiment is the next room. The constant is the question.

If you are doing one kind of work and you can describe it in one sentence and that sentence has stayed roughly the same across three decades, you are probably building one kind of thing. Find the next room for it. Pour the coffee yourself for the first six months. Run the BTL test in your head every Wednesday at 5:45 pm. Tell me what you find in 2036.

Epilogue

To the reader who came in to agnostic in 2017, looking for a community of people who had walked out of a religion. You found the community for a while. You watched 209,110 of us build it and you watched a small loud core start fighting each other instead of being fellow travelers. The data in this book is honest about why. About 1.8% of active users carried 22.5% of all reports filed. The other 95.4% of you were doing exactly what you came to do, which was to find someone you could text on Tuesday nights about losing your faith, or about Thanksgiving with the family, or about raising a freethinker child. I built the room. The room is partly your room. If you walked away from agnostic in 2020 because the meltdowns had eaten the warmth, I am sorry. I built it as best I knew. The next room I am building does not run on threads. Stop by the salon at 110 Union Street, downtown Seattle, Wednesday 4 to 7pm, free, coffee on me.

To the reader who came in to IDW in 2019 because the questions you wanted to ask had stopped being welcome in other rooms. You found a place. You found about 145,827 other readers there. Most of you were Quiet Spectators by the cluster math, which means you read more than you wrote, because the lesson of the previous two years had been to be careful. About 13.6% of active users wrote 92% of every contested thread. The room I built was loud in places where you wanted it quiet, and I did not always tune it back down fast enough. The same data that says agnostic melted on its outgroup says IDW melted on its outgroup. Civility was a perimeter, not a possession. I do not know if that satisfies the questions you came in to ask. The next room I am building does not require you to phrase the question right. You walk in. You sit down. Stop by the salon at 110 Union Street, downtown Seattle, Wednesday 4 to 7pm, free, coffee on me.

To the reader who came in cold, who did not know me before this book, who picked it up because somebody told you it was a careful post-mortem of a noisy decade. You read 75-some pages about two communities and an academic magazine and the question of whether dialogue alone is enough to bridge a divide that has hardened. You read my honest answer, which is no, dialogue alone is not enough, and the rooms I built between 2017 and 2024 taught me that in numbers I cannot wave away. The version of me that rebuilt the thesis in 2026 is using paint and canvas and coffee on a Wednesday. I am 57 in 2026. I am betting the next thirty years on it. If you want to see what the experiment looks like in a room, with humans, at human scale, stop by the salon at 110 Union Street, downtown Seattle, Wednesday 4 to 7pm, free, coffee on me.

That is the same invitation, three times, on purpose. Same room. Different doors. I will see you there.

A note on sources, methods, and the data behind this book

The methods, data tables, and open questions are published in the research roadmap. Anyone who wants to reproduce or check the work can. The agnostic dump is 81 GB raw. The IDW dump is 17 GB raw. In plain English, the heavy lifting here was done by an AI that I asked to do four things, blind to which site any piece of data came from. First, group active users by twenty kinds of behavior and tell me what kinds of person it saw. Second, read sample threads from forty-nine prompts I had posted to both sides and score how each room responded. Third, read 2,999 of my own posts and score me on the same scales we used for everyone else. Fourth, read posts from the same hundred-and-eight people on both sides separately and score the same person twice. I checked the flagged outputs by hand. If you find an error, my address is on the foundation site.

The salon, every Wednesday

Art Love Salon, 110 Union Street, downtown Seattle, across from the Seattle Art Museum. Wednesday 4 to 7 pm. Free. Coffee on me.

Visit the Art Love Salon

The Connection Engine

Contents

Prologue