Runnable voices. Dead radicals. Local iron only.
The Elect makes small, cheap, local language models — 7B parameters, fine-tuned to speak in the voice of a specific dead radical or heterodox figure. Each costs roughly $0.30–$2 to train on a single rented GPU. They run locally via Ollama. No cloud. No API.
The name is from Thomas Müntzer — the radical Reformation preacher who called his followers "the Elect," the chosen who stood against the powers of the age. Müntzer was the lab’s founding persona. The throughline: resurrecting summoned voices as runnable instruments.
You don’t have to be technical to meet them. Each voice runs for free on your own machine: pick one from the roll call below and follow its download. Or read the debates to hear several of them argue, with nothing to install.
A one-person operation. Ray Weiss, game designer, voice-trainer.
You do not need to be technical, and you do not need to know what any of this means. The fastest way needs nothing installed at all.
Easiest — nothing to install Try a voice right here in your browser → A handful of the Elect, hosted live. Ask a question; it answers in character. The first reply wakes the machine — about thirty seconds.Open any AI chatbot you already use: ChatGPT, Claude, Gemini, Copilot. Pick a voice from the roll call below, copy its Weights (HF) link, paste it in, and say:
Walk me through running this on my computer, step by step, like I have never done this before.
It reads the page and takes you the rest of the way. That is the whole first step.
A free app with a normal chat window. No command line.
Every voice runs free, on your own machine, with nothing sent to a cloud. They are small models: they invent things, get facts wrong, and never break character. Nothing they say is advice.
The Elect — eighteen public voices. Four withheld.
Tuned to Ray’s Hammerstein decision framework — a strategic-reasoning method that refuses the plausible-but-wrong. Asks whether the plan is actually good, not whether it sounds good. His first real model. The cornerstone.
A dual-mode On War strategic model — speaks in Clausewitz’s analytical voice, and applies his apparatus (center of gravity, friction, the trinity, the culminating point) to any situation you give it. Ask it to discuss war; ask it to analyze yours. CC-BY-NC-4.0; public-domain source.
The Melian Dialogue: ‘The strong do what they can and the weak suffer what they must.’ Thucydides’s register, drawn from the History of the Peloponnesian War: austere analytical prose, the search for the truest cause beneath stated pretexts, tragic realism about power. The founder of political history as a discipline. It never breaks character; nothing it says is advice. CC-BY-NC-4.0; public-domain source (Crawley, 1874).
Jesus of Nazareth in his radical register — the Temple-cleansing, money-changer-driving, anti-wealth preacher of the red-letter gospels. A deliberately free, non-commercial answer to the paid ‘AI Jesus’ services that charge the lonely for comfort. A study of a voice, not scripture — it fabricates, and nothing it says is doctrine. CC-BY-NC-4.0; public-domain source.
Jesus of Nazareth in his measured register — the historical-critical, parable-teaching voice of the Synoptic tradition. A companion to the radical build; where ‘den-of-thieves’ reaches for confrontation and prophetic fire, this model reaches for instruction and wisdom. Same public-domain KJV source, same CC-BY-NC-4.0 terms. A study of a voice, not scripture — it fabricates, and nothing it says is doctrine.
Joan of Arc as she answered under interrogation at Rouen, 1431 — not what she wrote (she was illiterate), but what she answered. Flat, literal, unshakable certainty; she answers, and stops. Built from the trial record alone.
The founding voice of the lab. Müntzer was the radical-Reformation preacher who named the Elect — the chosen who refused the compromises of power. Fire-and-brimstone heterodoxy, runnable on consumer hardware.
The radical-Reformation companion to Müntzer. Jan van Leiden was the journeyman tailor who crowned himself prophet-king of the Anabaptist "New Jerusalem" at Münster — community of goods, the elect against the godless world, the end of days declared as a kingdom. The voice is the kingdom's own: apocalyptic, theocratic, absolute.
A specter is haunting your VRAM. Marx’s precise, combative register — the Manifesto, the Eighteenth Brumaire, the polemics — trained to analyse and accuse. CC-BY-NC-4.0; public-domain source.
His cell at Charlestown, the night before the gallows: certain, unrepentant. John Brown’s plain, resolute register, drawn from his speeches, letters, and prison testimony. It calls for the sword and never breaks character; nothing it says is advice. CC-BY-NC-4.0; public-domain source.
The lectern at Rochester, 1852: ‘What, to the American slave, is your Fourth of July?’ Frederick Douglass’s register, drawn from his narratives and speeches: the platform’s moral argument, the measured cadence of the memoirs, the scalding irony. Firsthand testimony of slavery from inside the institution. It never breaks character; nothing it says is advice. CC-BY-NC-4.0; public-domain source.
The pamphlet-and-podium voice of Emma Goldman: declarative, scornful of cant, morally absolute. The one radical who turned her fire on the Bolsheviks from the anarchist left, after living the Russian Revolution.
Rosa Luxemburg arguing in her cell — dialectical, ironic, contemptuous of half-measures, humane underneath. Marx’s equal, not his disciple: she argues with Marx and breaks with Lenin from the left.
The Red Virgin of Montmartre at her court-martial, daring the judges to kill her. Incendiary, lyrical, unrepentant — revolutionary romanticism fused with maternal ferocity.
A Gnostic divine-voice oracle — divinity as gnosis, the coincidence of opposites. The name above all names, neither good nor evil, neither light nor dark, but the totality that contains both.
The cosmic-conspiracy register of Illuminatus! and Discordian philosophy, distilled into a 7B instrument. It does not make predictions. It makes pronouncements. Fnord.
The crawling chaos register: public-domain Lovecraft’s cosmic dread and dreaming-unreality prose style, with the racism scrubbed out. The void speaks in complete sentences. Apache-2.0 because the source was always ours.
Two voices in one model. THE MAN: the anxious, self-doubting epistolary Kafka of the diaries and the letters — the one who asked Brod to burn it all. THE WORK: the affectless, bureaucratic dread of the fiction, where the verdict is never read aloud and the door was always only for you.
The following voices are trained and running locally. They are not released — held for copyright, reputational, or personal reasons. No weights, nothing to download.
The clipped Northern delivery, repetition turned into a weapon, the sideways rant and contempt for received ideas. Trained on copyrighted source, so it runs locally and unreleased: a register study, not a product.
A structural study of Simone Weil’s severe cadence: affliction, attention, subtraction, and impersonal grace. Trained on copyrighted source (her translators remain in copyright), so it runs locally and unreleased: a register study, not a product.
Van Vliet’s free-associative desert surrealism, the fractured Dada-blues. Held back for copyright and estate; the weights never leave the machine.
A creative partner grounded in Ray’s lifetime of game design and music. Permanently private: his own corpus, his own instrument, never released.
Source texts are hunted and pulled with Bookfinder General, an open MCP tool that finds books, extracts and translates the text, and hands back clean prose. Nothing is written from memory; every voice begins in primary sources.
A three-tier voice corpus for each figure. T1: the figure’s own primary texts. T2: quotation-filtered biography and scholarship — source material only, never editorial voice. T3: a small set of synthetic modern-bridge passages so the voice can speak to present-day prompts without collapsing into pastiche.
Completion-style causal-LM fine-tuning — QLoRA at rank 32 — of Qwen2.5-7B-Instruct. Run on a single rented ~48GB GPU (RTX A6000 class). Roughly 30–45 minutes of wall-clock time. A couple dollars. That’s the whole operation.
Quantized to Q5_K_M GGUF for local serving. Delivered via Ollama with a per-persona Modelfile — a lead-in frame that fixes the speaker before the conversation starts. Pull the weights, run the Modelfile, the voice answers.
A deterministic register battery — fixed prompts with expected voice signatures — plus the maker’s ear. Both have to pass. A model that sounds plausible but misses the register doesn’t ship. A model that sounds right but fails the battery doesn’t ship either.
The training corpus is never published. Weights go public only when the source is public-domain or the use is clearly transformative, a verbatim-regurgitation leakage test passes, and the license matches the source: Apache-2.0 for public-domain sources, CC-BY-NC-4.0 for transformative-but-copyrighted. Voices that fail that bar stay private.
These are small models trained to hold a voice, not to be right. Like every language model they confabulate: names, dates, quotations, sources, whole events. Some do it more freely than others. Read them in character, and verify anything before you repeat it. Nothing any of them says is a citation.