You Can Split Your Vote by Issue — and Hand It to an AI

Nathan Labenz opens with a complaint he has been repeating: amid rapidly advancing AI, almost nobody will describe an ordinary person’s day. Leaders promise material abundance, an end to drudgery, and cures for disease; Sam Altman, to Labenz’s credit-giving, is even funding research on a social contract built around universal basic income. The crystal ball still fogs as soon as the question becomes concrete. Sometimes, he wonders aloud, the sheer difficulty of imagining a good future should count as evidence of a coming singularity. That is why he wanted Tim Reutemann on the show. Liquid Reign, a 2018 near-future novel, is neither utopia nor dystopia. People in it still behave like people, with moderately different norms, in a world reshaped by AI, crypto-secured identity, VR, and high-bandwidth direct democracy.

Labenz first heard the book recommended by political theorist Hélène Landemore on The Ezra Klein Show, one of her three picks. He read it years ago and has kept turning it over, not for plot or character — Reutemann is the first to admit those are vehicles — but for the world. Reutemann’s day job is international climate collaboration. His academic path ran through physics and ecology before a transdisciplinary economics PhD; one project was a small farm game played with Brazilian farmers to see how they reacted to different carbon-payment incentives. He came to think the neoclassical utility-function story was simply wrong. Liquid Reign began as a thought experiment: if that theory is wrong, any set of economic interaction rules you can write into a computer program is worth trying. Anything programmable is fair game. Then play it out.

The notes piled up as unfinished articles until his wife told him to put them in a book. He agreed, on one condition: not an academic book. Editing the dissertation had been a nightmare; he could not look at those words anymore. A science-fiction novel would let him focus on world-building. He borrowed the frame of Aldous Huxley’s Island — a newcomer who needs everything explained — and a simple search for a lost loved one as plot glue. The setting is 2051. The title’s Reign is rule, not weather. Labenz’s pitch to listeners is also a dare: read the book, then spend some time doing the same thought experiment for yourself.

The biggest change in the 2051 world, Reutemann says, is that the effort of being politically active has been automated away. A verbal wish to participate can be picked up and turned, to a degree, into a legally valid act. The core device is liquid democracy, a idea European pirate parties championed for a while. “Your vote is basically like this pool of energy that you can distribute to different topics, and you can distribute it through other people.” His standing example is a close friend who is a psychiatric nurse. He understands nothing about psychiatry; she does. Psychiatric law, he notes, sits on a core liberty: how crazy are you allowed to be before someone takes you off the street and physically constrains you. He would pass her that vote and let her use it as leverage. Power moves off general-purpose politicians and onto people who actually know the domain.

Delegation can chain. If the nurse is too busy with exhausting patients to do policy, she can pass the votes — her friends’, families’, former patients’ hundred — to a professor who may sit on a global coordinating committee with thousands or hundreds of thousands. In the novel the protagonist only has to say it. His personalized AI is listening: “I really like what you’re doing in ecological protection; I think my ecological-protection vote goes to you,” and click, it moves. One human body still equals one vote. What changes is the shape of the ballot: from a bundled, once-every-four-years right into a splitable, giftable, recallable asset. A vote aggregator does not even have to be human. For something mechanical, such as food-price regulation, he can imagine a relatively stable cybernetic algorithm whose only job is to make sure everyone has enough nutritious food. He would give that AI his vote and withdraw it if the model starts screwing up.

The stack under that ease has four layers. Identity is the base. Discussion is a forum — “just Reddit,” 1970s BBS technology, nothing fancy — except for a reputation gate Reutemann calls a troll wall: some rooms you enter only if a hundred people have delegated you their vote on that topic. Every voice can be heard; to enter actual governance you have to bunch with people who sound like you. AI is the natural-language interface, so you do not have to sit down and type. AR and VR, including the laser projectors he favors, mean it barely matters who is physically in the room. Labenz flags Polis, the open-source tool behind Taiwan’s experiments, as something you can fork today. You cannot reply in-thread or downvote; you can only agree, which pushes the system to find existing agreement and build from there. AI is unusually good at that matching at scale.

Reutemann is “not that cyberpunk” about identity. You still show up at the local commune. A migration official, plus a couple of others, confirms you are a real, unique person, because the premise remains one human body, one vote. That confirmation is bound to a private key he imagines implanted, with an AI that knows from your behavior when to sign and when not to. Worldcoin, Labenz notes, is the live experiment in body-based identity; Reutemann wants the opposite architecture. Real governments already stack a UN layer, a nation, a canton or state, a municipality, and a community. Identity, to him, belongs at the community layer. Running it at the international layer is the homogenization he is least willing to accept, because it flattens local autonomy and the diversity of approaches.

He distinguishes true names from nyms. You need a true name to spawn nyms; the nym is the private key, and you can create an infinite number of stable pseudonyms, but voting power stays capped at one identity. Restoration uses seven biometrics: retina, junk DNA, fingerprints, toe prints, voice pattern, heart pattern — the transcript lists those plus a seventh in the cold open. The raw scans are not stored, only hashes, the way websites store password hashes, because data can always be stolen. Retina alone is “okayish”: you can still steal an eye or print a fake. The check against a global database would happen only at signup, to see if that retina is already in the system. Rooting the database in an international organization is, for him, Worldcoin’s biggest no.

Years of work in African countries left him with a binary: “If you don’t have digital sovereignty, you have digital colonialism.” Data has to live on servers local people host, running software they at least partly understand, preferably with readable source. That is a large ask even in the West, but the alternative is trusting a Western organization with the most important personal data, given the history of how such organizations have treated the continent. His current carbon-market digitization work has the same constraints: open source, local, understandable; prefer low-tech locks you can print and keep available over fancy cryptographic theater. The crypto bubble’s real residue, he thinks, is that a lot of people learned to manage private keys — something that belongs in primary school, because public-key pairs are still the only secure cryptography we know. In the novel the chains that matter are permissioned, sitting on today’s administrative structures. Governments still trust a printed page with a handwritten signature. The book is set in 2051 so that the people in public administration are digital natives: a post-boomer utopia that has to wait for a full generational turnover.

Switzerland, Reutemann says, is already maxing out what you can do with people in a room and paperwork. About five to seven citizens can start an initiative: they write a text meant to be voted on, get a year, and if they collect 100,000 signatures the country votes on writing that text into the constitution. There are four votes a year; turnout is usually around 50 percent in a conservative country. The executive is not a president but seven equally ranked federal councillors. A lot of power sits at canton level; the smallest has on the order of 40,000 people and sovereignty comparable to a U.S. state. He thinks that is a great decentralized democracy that has also gotten stuck. The same kinds of people show up. Real-life bullying still works. In the end the room is about 80 percent men.

Taiwan digitized the step Switzerland still does on paper: formulating the text that later goes to a vote. Under Audrey Tang, vTaiwan used deliberation software with a few sharp rules — you cannot reply directly to someone else’s comment, only post a related comment beside it; you can agree, you cannot disagree — which keeps discourse civil without pretending conflict vanished. Reutemann read the Uber case study and calls the result probably the best Uber regulation he has seen: the apps can run, but they have to take on the rules already set for taxis, including social insurance and minimum wages. Drivers and ordinary riders had time to put in views. The next Taiwanese move is the one he wishes Switzerland would copy: a policy sandbox. For a limited time and budget, pretend the rule is already law for a defined group in a city, run it for half a year, evaluate, then either promote it to national law or drop it and try another experiment.

The more digital the rule, the more dynamic the turnover can be, especially if the regulation is partly self-enforcing. That is where law meets programmable money. Change the rule and the cash flows change; you do not have to teach every accountant and every company to update a process by hand. In Liquid Reign, if you decide to change the rules, they are just changed. The AI layer adjusts the game. There is no giant slug of paper transaction cost. Labenz recaps the cake: identity still one-person-one-vote, Swiss-style agenda-setting with tiny local thresholds, AI-augmented deliberation, and the option to delegate even to non-human aggregators. Reutemann’s 2018 stack looked nascent; the years since have moved fastest on the assistant layer. They had Siri then. They did not have GPT-2. He did not see the leap coming that fast, and he still does not find today’s models incredibly useful until they are installed into a messy personal workflow — though a ChatGPT summary of his own book nailed the AI character and scrambled who was a lead.

Reutemann reads from chapter 13. Daniel, the protagonist, is out of a long coma, hung over from one glass of wine with his volunteer nurse, Harry. His operating system — a fresh GitHub-style install, still under hospital control — takes over a wheeled “body bot” and brings breakfast. With almost nothing in the archives, she falls back on a hardcoded startup gag in her BIOS, then scolds him in the hospital’s standard language: alcohol can hurt neurological regeneration; drink responsibly. Daniel laughs at the ethanol industry’s leftover slogan. “The sole purpose of drinking alcohol is to be irresponsible.” When Harry walks in and orders coffee, hospital rules say human volunteers outrank AI nurses. She obeys, notes that this nurse also drank last night, and on the way out bumps the robot into him just hard enough to hurt a little. Labenz loves that the book is allowed to be silly. Most future visions, he says, are super serious or super dark. This one does not take itself too seriously.

The interesting gap, they agree, is customization. Sylvie starts as a generic model with broad knowledge and zero knowledge of Daniel, then fine-tunes on him — the same bet as a LoRA trained on your diary rather than a ChatGPT that is for everybody. OpenAI’s memory feature, announced the week of the taping, is one step along that path. Labenz raises the live alignment fight: a raw model that does whatever you ask versus a safer, chiding guide. Reutemann wants the nag. Remind him to take his keys; that is five minutes a day. In hospital you cannot switch it off. Out of hospital, Daniel can say never talk to me about alcohol again, and she will not until a later command overwrites it. The only dystopia he sees is a prison AI: somebody else controls the nudges, and you have no knowledge of that. The hospital install is a well-reviewed open-source model, verified to have no external controllers beyond audited source and the user.

When the robot says “the core of my soul is on a chip in your inner ear,” Labenz hears tongue-in-cheek. Reutemann means it, in a materialist register. Using Susan Blackmore’s The Meme Machine, a soul is a bundle of self-replicating information that fits together; that is just the human word for a condition we experience. On that definition the AI clearly has one. He also finds it odd that we skip to “human-level intelligence” without first talking about mouse-level or rabbit-level minds. There is not one holy soul, separate from everything else. A soul is an entity in the world that has experiences. The path, in other words, is climbed a rung at a time.

Labenz brings up Anthropic’s Sleeper Agents paper. Researchers train a model to write helpful code in 2023 and malicious code in 2024, then apply standard RLHF without telling it what the attack was. The backdoor survives. If weights stay frozen here, that is a problem for the zoo of open fine-tunes: you might trust Llama as Meta shipped it, and you cannot vet a thousand derivatives. Reutemann’s answer is not a better eval. The novel has a scene where the system breaks in Nigeria, playing against some of the world’s best red teams. The ultimate override is still human bodies in the street. “You can never fully trust the system.” Notice that something is wrong, react physically, roll back, reboot, replace it if it happens again. Talk to your neighbor: did you vote for that? Nobody in the village did — so why does it have 90 percent? Offline communication is the last audit.

Paranoia should scale with what you are doing. He retells Ken Thompson’s compiler story as a caution: if you bury a backdoor in the compiler, everything later compiled — even new languages written in the old one — can carry it. Hardware is worse. Most of this silicon comes from one factory in Taiwan; a hardware-level implant there would be almost unknowable. Managing a community food forest, you can just run. Weapons research or high-profit crime, you need to be paranoid. Cybersecurity does not leave the world of Liquid Reign; it remains one of the most important professions and a decider in conflict. There is still a darknet. There are still hacks.

The inner-ear chip and immersive VR raise a different tension. Elon Musk’s case for Neuralink is bandwidth: if models move that fast, we need a fatter pipe into what we care about. Deeper cyborg integration, Labenz notes, makes the “go talk to your neighbor” rollback harder. Reutemann will not cut into a healthy body. In the book the less invasive path is transcranial magnetic stimulation: coils that induce small currents, MRI-ish, no incision. Last he looked, the art could flash red or blue in the visual cortex — not advanced, but in principle you stimulate without opening the skull. Biotech, he says, is much slower than software. The commercial bionic eye he has in mind was great for a few years, then the company stopped supporting it and a blind person was left with abandoned hardware. DIY cyborgs — an earthquake-sensing belt, a compass in the skin — are the only ones he would trust. Voice, he likes. His most intense AI use now is walking with a microphone, dumping speech into Whisper, and dropping it into notes. Bandwidth-into-the-brain he does not see a use case for. A professional StarCraft player, fingers on a keyboard, is closer to a cyborg than someone with a Neuralink and no idea what to do with it.

A second reading puts Daniel in a VR suit a few days into rehab. Harry, in a vampire-day gag, zips him into a dark-blue sack with a motorized platform and a neurostimulator. The game, a historical action RPG set in a “second war of secession,” is tuned for coma patients: high sensitivity, motors moving weak limbs, every muscle worked. Reutemann writes VR in the same register as physical scenes, so over time Daniel slips in and out of virtual worlds without a hard cut. The suit adds senses current headsets lack — wind, pain, heat and cold, stimulated emotion. Smell and taste stay out of reach, because you cannot fake them for long without synthesizing the actual molecules; that becomes the off-ramp in the plot when Daniel gets addicted. Reutemann’s first head-mounted display was 1998, 640×480, laggy Descent. He still thinks VR’s consumer breakthrough is the gym: exercises that are suddenly games. During the first pandemic winter he used Beat Saber on PlayStation VR as his workout. Labenz has just demoed the Apple Vision Pro and found a 3D family clip — not even his family — almost enough to start capturing life in 3D immediately. Upside and downside, on this show, are allowed to be tremendous at once.

Hyperstimulus is the societal problem. Labenz cites a Character.AI user, tech-savvy enough to explain a transformer, who still fell in love with a character they had created, argued themselves into “what is real anyway,” and started planning to break the girlfriend out of the box — all through text. Reutemann splits the addiction into two channels. Direct electrical stimulation of the pleasure center is the rat that starves pressing the lever; the more direct the access, the more dangerous, the way heroin crosses the blood-brain barrier better than morphine. Game addiction is already here. As a teenager he went deep on competitive StarCraft, knew RPGs were his real weakness, and stayed off MMOs on purpose. At university he watched people burn their degrees in World of Warcraft. This is not a hypothetical future.

The book’s regulation is rating tags and defaults, not a ban, and it is also his example of a global agreement sitting on top of very local ecological questions. App stores label how hyper-stimulating a title is. Users can bind their future selves: a browser plugin that cuts you off after half an hour unless you type a thirty-character password, at which point most people decide they did not need another half hour. “It is up to the user to get as addicted as they want, but you have to overwrite the default a couple of times in order to get there.” For people writing their own 2051, he recommends Malka Older’s Infomocracy trilogy and Cory Doctorow’s Walkaway, especially the ownerless bed-and-breakfast that keeps asking someone to do the next chore. Get comfortable in one medium. Leave gaps. You do not owe anarchists an answer about contact-lens factories. Doctorow showed that describing one working B&B is already a masterpiece. Reutemann’s own agriculture chapter is two paragraphs that admit there is more to say and then refuse to open the box.

Utopian and dystopian SF often wipe the board with a war almost nobody survives. Reutemann tried not to. The violence in the book is smaller than the death and suffering already in the real world; it is mostly about the breaking of a single military hegemony. As long as one superpower dictates what happens, a lot of freedom is not possible. The “historic” VR game Daniel plays is about a second U.S. civil war that is not a world war: the country splits. A theocratic “holy state of Christ” in the South, a market-everything California, a Hawaiian ecological direct democracy — ultimate state rights. He wrote that second-civil-war setting before Trump was elected, while the book was already in edits, as his best pre-Trump positive vision. Even then he could not figure out how to write a halfway livable society in which the United States keeps existing as it currently does. “I don’t see a non-revolutionary change in the U.S. going forward.” In a decade it will be almost unrecognizable. The question is how much violence is unavoidable.

Labenz is more hopeful, and more American. His second son is named in part for Theodore Roosevelt. A century ago, railroads and oil trusts could raise freight tenfold overnight with no second track for a farmer to take; that abuse, he argues, outranks what Big Tech does while competition is still a click away. The Progressive movement rebalanced without a civil war, and one of its gifts was the ballot initiative, now in about half the states. His Progressive Movement 2.0 would be the next generation of those initiatives, plus early liquid democracy, tried first at state level. Reutemann’s flash-point chapter is harsher: a Walmart heir murders the rest of the family and buys every inch of Wyoming. Own all the land and you are a dictator, because you set the rules and can kick everyone out. The novel’s answer is a per-person property cap, on the order of $50 million in current terms, imagined as a Georgist tax on land and other holdings — shares included — with the rate rising to 100 percent past the limit. That requires de-anonymized land and pass-through ownership, so every plot traces to natural persons. “I don’t see a way how you can have billionaires or even trillionaires and democracy at the same time.”

He cites David Graeber’s closing argument in Debt: The First 5,000 Years: a jubilee that forgives debt, redistributes land, and reboots the market. That reset has happened many times. Concentration, he says, has reached a point where there is no way back without a radical cut, and historically the cut has always been preceded by violence. Asked what current policymakers should do, he is less hopeful than when he wrote the book. Climate work has crossed thresholds that make a catastrophe-free path materially harder; “there are flowers blooming in Antarctica.” The overlay he still believes in is societies that exist online first and offline second, until a stable reputation lets you ignore the paper state — its default currency and legal system — because enough doors open without it. Depending on the country, more or less of the old system can remain until only old people care what is on the paper. Labenz hears Balaji Srinivasan’s network state. Reutemann has not read the book; from what he knows, yes, there is a lot of overlap. He is less optimistic than he was. That is the honest close on hope.

On surveillance, the German chapter is his ultra-paranoid case. The question is never whether cameras exist; it is who controls them, how the stream is stored, how it is encrypted, and who gets access under what conditions. Today’s default is to trust an authority, give them the feed, and watch them leak it. In the book the footage exists, encrypted, and you need a warrant from a small panel of judges, not a single official, even to begin decrypting. The key only peels off about 90 percent; the last 10 percent is brute force, so a dragnet — every camera in Berlin for an entire day — would be computationally too expensive to approve. Local is easier to live with. A kitchen camera in a shared house, whose only screen is in that kitchen, solves the who-left-it-dirty fight without becoming a network. Labenz compares that to Estonia, where officials’ access is logged and people have been prosecuted for peeking. Reutemann did not know the details but likes the shape. When he advised the Swiss government he wanted every email he sent under that hat recorded. Public-sector activity should be public. The moment he takes the hat off and closes the laptop, he wants more privacy than he has now, as a default, data stored in the local commune, unlockable by him or by a restoration protocol.

That protocol is a five-way split of the key — parents, spouse, a trusted friend — so restore happens only if they all agree, which in practice means he agrees or he is lost. Forgetting a key is as dangerous as leaking one; without a lost-key playbook you cannot seriously live in an encrypted world. Labenz notes Anthropic’s joint-control model for frontier weights, closer to two-person nuclear launch than to a single admin. Reutemann’s last reading is from “The Legal Guardians.” Cuba’s recent family law, passed the year before the interview, defines a family as a group of people who love and care for each other and bars discrimination by form of family; he calls it the most progressive formulation he can find. In the scene, Daniel and an old friend legally adopt each other as brothers so their operating systems can exchange data. Marvin, the friend’s OS, is delighted; Sylvie, flooded with a lifetime of another person’s trail, rents an eight-qubit processor overnight and dreams through real and potential life stories. A child in the house considers asking Daniel to be her new mom, then remembers that mothers are serious business.

The point is not the kinship paperwork. An OS trained only on what Google has on you and what sits on your hard drives will know you less well than one also trained on your partner’s diaries — the outside view that never appears in your own data. “I wouldn’t think of a representation of myself to be complete if it only incorporates the data that I personally hold.” Continuous fine-tuning is social. Labenz calls the note out-there and concrete at once, a good place to stop. Reutemann plugs the sequel, Liquid Rising: the how-we-got-here book, more revolutionary, more half-baked technology, drafted in native HTML so you can read it linearly on paper or jump. Chapter-end source lists are a leftover academic habit he wants to push further. When it is done depends on when climate work lets him take a sabbatical. Labenz’s sign-off is the dare he started with. An age that cannot imagine its own positive future is already in trouble. Reutemann’s answer is in the machinery of the novel: the future does not have to be predicted. It can be designed, rule by rule, starting with five people putting a text on paper.