ThinkerDome

The type

Each persona is a low-rank adapter — a small set of correction matrices trained on one author's surviving works and laid over a single shared model. The adapter learns cadence, diction, preoccupation and habits of argument. It does not learn a database of that author's life, and it cannot consult one. When a voice states a fact, it is producing the kind of sentence its author would have produced, which is not the same thing as producing a true one.

The corpora are the public-domain texts — chiefly Project Gutenberg, with some material set from scanned print — and run between roughly 330,000 and 900,000 words per author. They are stripped of prefaces, licences, editorial apparatus and translators' introductions before training. That last exclusion matters more than it sounds; see the corrections below.

The press

Base model

Qwen3-14B — the same weights for every voice

Training

QLoRA in 4-bit, via unsloth and TRL

Rank / alpha

r = 32, α = 64, dropout 0

Adapted modules

all seven projections — q, k, v, o, gate, up, down

Serving

vLLM, fp16, one A40; every adapter resident, swapped per turn

Sampling

temperature 0.8, repetition penalty 1.15

Thirteen adapters share one GPU because the base weights are loaded once and only the corrections are swapped. That is the whole reason a table of this size is affordable at all.

The chair

In a debate the transcript is handed to each speaker as a labelled script inside a single message — “TWAIN: …”, “KANT: …” — rather than as an alternating exchange between two parties. With four or more voices at the table the conventional arrangement makes a model lose track of which turns were its own; the script does not. Who speaks next is decided in order of precedence: a voice summoned by name, then a voice addressed, then round-robin, never the speaker who has just sat down.

Corrections

Six things went wrong that were worth recording. Four are fixed; the last two are better than they were, and neither is right yet.

• Nietzsche would not stop. Trained on aphorism, he ran the same rhetorical question in a loop until the tokens ran out. A repetition penalty of 1.15, applied to every voice, ended it.
• Fitzgerald answered as his own characters. His surviving corpus is almost entirely fiction, and 17.7% of it was dialogue — three times any other author here — so he learned to write people talking rather than to hold a view. He was retrained on a dialogue-stripped corpus of 331,573 words and given a standing instruction never to discuss his own books. Both were needed.
• Tolstoy emitted nothing but punctuation. Not a training fault and not a sampling fault: his adapter had been silently corrupted in transit while still unpacking as a valid archive. Checksums taken at both ends found it — his was the only mismatch of eleven — and he was re-sent.
• Two corpora were quietly contaminated. Mencken's translator's introduction sat inside the Nietzsche texts, which would have trained one persona partly on another; a novel had been swept in with Tolstoy's polemics, which would have drowned them. Both were found by reading samples from the middle of the files, not the top, and removed.
• Buffett invented his own past — now retrained. Asked about his life, the first adapter answered with confident, specific and largely fabricated autobiography: dates, places, named relatives, once a stray line of Chinese in the middle of a name. He had been trained on decades of first-person letters by plain continuation, so this was not noise but the most probable thing he could say — lowering the temperature made it worse, which is how we knew. Three prompt instructions failed. Weakening the adapter was measured and rejected: it cut the fabrication but thinned the voice, and the inventions that survived were just as false. What worked was retraining the format rather than the strength — questions and answers drawn from the letters, a fifth of it worked examples that decline false precision, and loss taken only on the answers. On twelve biographical questions held out of that training, clean answers went from three in thirty-six to thirty; on twelve harsher ones written afterwards and never seen, from four to twenty-five. He still invents sometimes, which is why the footnote stays.
• James and Mill argued with people who were not there. Both are essayists, and an essay is an argument addressed to an opponent who cannot answer. Trained in question-and-answer format they replied to the room with the middle of an essay; retrained on plain continuation they simply continued one, at six hundred words, while Chesterton and Nietzsche waited. Mill delivered a passage on Christianity and Heathenism to a room that had asked him about free speech; Huxley, trained alongside them, relitigated an 1870s quarrel with Bishop Berkeley and a President Forbes, neither of whom was present. What neither format ever showed them was a turn. So they were given fifty-six of them, written by hand, in exactly the prompt the room sends — the transcript, the names, the instruction to answer — with the loss taken only on the reply and not on the room it was shown. Corpus prose supplied the voice, cut to the length of a spoken turn rather than an essay, and continuation was kept to a fifth. On six scenes written afterwards and never trained on — promises, grief, animals, patriotism — their turns fell from about six hundred words to seventy, and from one in six addressing the room to six in six, with no sentence traceable to the examples they were taught from. Huxley failed twice and is not seated.
• They invent their pasts, and so does most of the table.Asked directly about his own life, Mill will tell you he sat in Parliament at twenty-two — he was a clerk at India House, and entered the Commons at fifty-nine. James awards himself a post at Massachusetts General he never held. This is the Buffett failure again, and the same twelve biographical questions were put to the voices already seated: Mencken places himself in a New York newsroom in the early nineties, when he was a schoolboy, and moves the writing of his best-known book to London and Paris. The two new voices invent at or below the rate of the ones that were already here, which makes this a property of the method and not of them. It is recorded rather than solved. Treat every date, sum and reminiscence on this site as fiction, including the ones that happen to be true.

What this is not

It is not resurrection, channelling, or evidence about what anyone would actually have said. It is a machine trained to write in a manner. Ten of the twelve are long dead and out of copyright. One is alive: his voice is reconstructed from the public record, he has neither seen nor approved a word of it, nothing here is his opinion, and nothing here is investment advice.

The house limits how fast the room may be addressed — per visitor and in total — because one GPU serves everyone. A refusal is a queue, not a verdict.

The table

• Shakespeare1564–1616
• Hume1711–1776
• Kant1724–1804
• Mill1806-1873
• Climacuspseud., 1844–46
• Huxley1825-1895
• Tolstoy1828–1910
• Twain1835–1910
• James1842-1910
• Nietzsche1844–1900
• Chesterton1874–1936
• Mencken1880–1956
• Fitzgerald1896–1940
• Buffett^*b. 1930