Introduction

Litany of Tarski · Dos litanye fun Tarski · דאָס ליטאַניע פֿון טאַרסקי

נעם דאָס קעסטל האָט אַ דיאַמאַנט.
װינטש זאָל איך גלײבן ציט [דאָס קעסטל האָט אַ דיאַמאַנט] אױס. אױס.
נעם דאָס קעסטל האָט קײן דיאַמאַנט.
װינטש זאָל איך גלײבן ציט [דאָס קעסטל האָט קײן דיאַמאַנט] אױס. אױס.
װינטש זאָל איך זיך ניט צו גלײבונגען צוקלעפּן װאָס איך אפשר ניט װיל האָבן.

Latin reading

Nem dos kestl hot a diamant.
Vintsh zol ikh gleybn tsit [dos kestl hot a diamant] oys. Oys.
Nem dos kestl hot keyn diamant.
Vintsh zol ikh gleybn tsit [dos kestl hot keyn diamant] oys. Oys.
Vintsh zol ikh zikh nit tsu gleybungen tsuklepn vos ikh efsher nit vil hobn.

Yuddish / יודיש is a constructed language inspired by Eliezer Yudkowsky, Bayesianism, and Less Wrong (LessWrong), built from Yiddish vocabulary. Its name combines Yiddish and Yudkowsky. The experiment is to make credence, evidence, and belief revision audible in everyday grammar: a factual assertion normally requires a numerical epistemic auxiliary and an evidential clitic, while a question, wish, quotation, definition, or command declares its own kind of act.

The epigraph adapts the diamond-in-a-box Litany from Yudkowsky's “The Meditation on Curiosity”, named for Alfred Tarski. Its sense is: if the box has a diamond, may I believe that it has one; if it has none, may I believe that it has none; may I avoid attachment to beliefs I might not want. Nem supplies each supposition and vintsh expresses the desired belief. The Litany makes no empirical claim about the box and therefore needs no invented EF credence. The final relative clause remains inside the wish; efsher means “perhaps.”

The language keeps room for affection, grief, humor, and ordinary conversation. Its grammatical machinery distinguishes describing a feeling from expressing one, and wanting an outcome from predicting it. For example, Es ef⁺²=zog morgn regenen / עס עף⁺²=זאָג מאָרגן רעגענען means “It will rain tomorrow, at 4:1 odds, based on a report.” The number concerns the event; the clitic identifies the speaker's evidence route. A listener can question either.

Yuddish simplifies the inherited language wherever those distinctions do not help this project. It has no grammatical gender and no case inflection. The former neuter forms supply its defaults, including invariant dos “the” and general singular es. Personal zey now covers a singular human or sentient referent as well as its inherited plural use; verb agreement distinguishes one from several. Numerals use short Yiddish-derived digits and regular place-value composition: twenty is literally “two ten.” Person and number agreement in verbs remains, as does the finite log-odds system. These choices give the language a Yiddish foundation and a deliberately engineered grammar.

Whether speaking this way improves reasoning is an open empirical question. Useful tests would measure clarity of disagreement, forecast calibration, and recognition of dependent evidence. Core marks adopted rules; Proposed marks candidate extensions; Inherited identifies source-language material. The reference, phrasebook, vocabulary, and translations describe Yuddish Version 0, the language’s release standard. The numerical credences in the translations reconstruct the respective speaking voices; the short examples here demonstrate the same grammar.

Version 0 includes speaker-owned reconstructed credence, quantified modals, personal and backward-reference pronouns, paragraph scope with extended blocks, layered word formation, and automatic question closure. Sections 14.6 and 14.7 summarize these conventions.

Philosophical sources

Yudkowsky and the Sequences

Eliezer Yudkowsky's essays distinguish epistemic rationality, concerned with accurate beliefs, from instrumental rationality, concerned with acting effectively toward one's ends. For Yuddish this yields a basic grammatical separation: a truth assessment belongs to epistemic grammar, while a goal or directive belongs to a different utterance type. A preferred outcome does not acquire a higher probability merely because it is preferred. This is a language-design interpretation of the distinction in “What Do I Mean By Rationality?”.

The Sequences are a body of essays organized into collections. “Feeling Rational” is the second opening essay in the 2015 book edition used here; it is not itself the whole second sequence. Its treatment of emotion makes it a useful test of whether Yuddish can describe feelings without treating rationality as emotional numbness. See “Feeling Rational”.

Bayesianism

Bayesian reasoning represents uncertainty using probabilities and revises those probabilities by conditioning on evidence. For a hypothesis H, evidence E, and background information K:

P(H given E,K) = P(E given H,K) × P(H given K) / P(E given K).

The language uses base-2 log-odds because the corresponding update is additive. It also distinguishes a posterior probability from a likelihood: confidence in H after seeing E is different from how likely E would have been if H were true. Yuddish does not settle disputes about how to choose priors, build models, or obtain likelihoods. It provides ways to say what one used.

Bayesianism is broader than the writings of any one author or community. The elementary identities in this document are mathematical constraints; the choice to grammaticalize them, and the proposed vocabulary for doing so, belong to the conlang.

LessWrong

LessWrong describes itself as a forum and community concerned with improving reasoning and decision-making, with a substantial interest in Bayesian reasoning and artificial intelligence. Its community contains disagreements and changing practices; it is not a single doctrine or a synonym for Yudkowsky. Yuddish draws on recognizable discussion habits associated with that setting: explicit credences, distinguishing observations from interpretations, revising beliefs, and asking what a claim predicts. See LessWrong's introduction.

The name LessWrong is retained as a proper name. A playful descriptive calque is veniker-falsh / װײניקער־פֿאַלש, “less wrong,” but that expression need not replace the community's name.

Maps, predictions, and definitions

A belief or model describes something; it is not the thing described. Yuddish makes model construction an explicit act and gives displayed wording a quotation operator. This helps distinguish an event, a representation of the event, and a sentence used to name that representation.

The idea that beliefs should constrain expectations motivates a request for a discriminating prediction: what observable outcome would differ if the hypothesis were false? This is a practical reading of “Making Beliefs Pay Rent”. It does not require every useful concept to denote something directly visible.

Definitions need their own operator because choosing a label does not establish an empirical fact. A request to restate a disputed word in operational terms is inspired by “Taboo Your Words”. In Yuddish, defining intelligence never by itself establishes that a particular system has it.

Fallibility and the probability endpoints

Yudkowsky's “0 and 1 Are Not Probabilities” motivates the language's refusal to express empirical certainty as an infinite epistemic strength. A prior of zero cannot be rescued by ordinary conditioning on evidence of positive marginal probability; a prior of one has the corresponding problem for its complement.

The mathematical distinction matters: 0 and 1 are valid probabilities in standard probability theory. They occur in formal models and probability measures. Yuddish's restriction is a rule for fallible empirical assertion. It does not outlaw deterministic models, logical identities, or zero-measure events. Those can be discussed inside explicit mathematical or model frames.

Every finite empirical bit value remains revisable. A speaker may say avavav, but the repeated syllables still denote a finite number. Proof claims likewise distinguish what follows in a formal system from the fallible claim that a human checked the proof correctly.

The Yiddish foundation

The source language

Yiddish is a language with its own literature, dialects, and grammatical traditions. Its historical components include Germanic, Hebrew and Aramaic, Slavic, and smaller contributions from other languages. Yuddish takes this existing linguistic material as its foundation; it does not treat Yiddish as English with Hebrew letters. For an introductory account, see YIVO's Basic Facts about Yiddish.

The ordinary lexicon supplies words such as velt “world,” gleybn “believe,” zen “see,” zogn “say,” gefil “feeling,” and emes “truth.” Compounding and grammatical specialization turn such material toward new purposes. A technical compound should be transparent when possible, while remaining distinguishable from a historical dictionary entry.

A small root text

The following short teaching text was composed for this reference. It illustrates the source-language layer rather than quoting a historical Yiddish work.

Yiddish: איך זע אַ ליכט. איך װײס ניט װאָס עס איז. איך פֿרעג דיך.

Transliteration: Ikh ze a likht. Ikh veys nit vos es iz. Ikh freg dikh.

English: “I see a light. I do not know what it is. I ask you.”

An experimental Yuddish adaptation separates the perceptual assertion, a refusal to invent an identification, and a direct question:

Reading: “I see a light, at +6 bits on perceptual evidence. I suspend judgment about what the light is. What is it?”

The repetition zen … zen is intentional: the clitic marks evidence, while the lexical verb says what occurred. Halt is the abstention operator, specified in section 8. It expresses suspension of judgment, so this is an adaptation of the teaching text rather than a word-for-word restatement of “I do not know.”

No gender and no case inflection

Core. All nouns belong to one undifferentiated class. “Neuter by default” means that the forms formerly associated with the neuter supply useful defaults; it does not leave masculine or feminine exceptions in the lexicon. People, objects, abstractions, and new compounds follow the same grammar. Sex or gender can still be described when relevant, and lexical meanings such as “brother” remain possible, without controlling agreement.

Dos / דאָס is the invariant definite article for singular and plural nouns in every syntactic role. A / אַ is the indefinite singular article, with an / אַן before a vowel as a pronunciation rule, not an inflection. Indefinite plurals have no article. Keyn / קײן is an invariant negative determiner. It can carry the negation in a construction such as dos kestl hot keyn diamant, “the box has no diamond”; a separate nit is not compulsory in that construction.

Nouns still distinguish singular and plural where the lexicon supplies different forms: gefil → gefiln, velt → veltn, mentsh → mentshn. Irregular plurals remain lexical. Number is not case or gender: dos mentsh ef… and dos mentshn efn… retain their different verb agreement. No noun changes its ending because it is an object, recipient, possessor, or complement of a preposition.

The former article forms der, di, and dem are not productive alternatives to dos. Case-bearing contractions are expanded: use tsu dos, fun dos, in dos, and oyf dos, rather than inherited forms such as tsum or funem. This does not split unrelated lexical words that merely start with similar sounds.

Invariant pronouns and possession

Core. Pronouns have the same form as subjects, objects, and after prepositions. Word order identifies subject and object; prepositions identify recipients and other relations.

Zey is the default personal third-person pronoun, independent of sex or gender. It extends the inherited plural form to one human or other being presented as sentient. This singular use is a Yuddish innovation. Es remains available for any singular referent, including a person where personal reference is not being foregrounded. Choosing a pronoun does not introduce noun classes or agreement on articles and adjectives. A speaking animal or AI character can receive zey. Uncertainty about sentience can be expressed separately; a pronoun is not a numerical consciousness assessment.

Number agreement follows the referent: zey ef for one, zey efn for several. In nonassertive inherited verb forms, use zey lernt for one and zey lernen for several. Zey itself never inflects for number or case. A newly introduced unspecified individual takes singular agreement. A plural group takes plural agreement. Where the listener cannot recover number, repeat a noun or count; there is no arbitrary number-unknown agreement form. A plural clause also needs a clear collective or distributive interpretation.

Gendered er and zi are absent. Mir means only “we/us,” not inherited dative “to me.” Forms such as mikh, dikh, dir, im, undz, and aykh survive in actual Yiddish quotations and historical specimens only. No new personal ment or referent-index paradigm is adopted.

Possession always permits fun + noun phrase/pronoun: dos model fun ikh, dos model fun zey. The invariant short possessives remain mayn, dayn, zayn, undzer, ayer, zeyer. Zayn covers any singular third-person possessor, including singular zey; zeyer covers plural possessors. Thus fun zey needs contextual number, while zayn model versus zeyer model can make it explicit. No possessive ending agrees with the possessed noun. Context distinguishes possessive zayn from infinitive zayn, “be.”

Zikh / זיך is the subject reflexive for all persons and numbers. It refers to the subject of the smallest containing finite clause. =Zikh remains a separate evidential clitic: it marks the speaking voice's introspection or memory, regardless of the grammatical subject.

Zelb / זעלב is the new invariant local backward-reference pronoun, adapted from the Yiddish root meaning “same”. It refers to the most recently completed accessible referential noun phrase in the current sentence. This is a different relation from subject reflexivity. The operational rule is:

1. A completed referential noun phrase registers its whole referent, including a completed pronoun. An adjective, verb, numeral inside a noun phrase, or standalone operator does not register another referent.
2. The whole phrase Rivke's brother registers the brother when complete, overriding the internal possessor. A coordinated phrase registers the group after its final conjunct is complete.
3. Zelb resolves against the existing register before registering its own referent. It cannot resolve to the incomplete phrase containing itself or to a following word. If there is no accessible antecedent, use a name or another pronoun.
4. A quoted voice, displayed text, hypothetical, or model has a local register initialized from the accessible outer register. Its internal mentions do not overwrite the outer register when it closes. A subframe's referents cannot leak out as actual-world objects.
5. An ordinary embedded clause shares the sentence register unless a frame isolates it. Every sentence boundary clears the register, even inside a continuing paragraph or extended frame. Returning from a frame after a sentence boundary does not resurrect an antecedent from an earlier sentence.

Zelb may refer to people, objects, or groups, and takes its antecedent's number if used as a subject. Naming the full referent is the repair when accessibility or grouping is unclear. It does not replace the ordinary adjective zelb “same”: attributive position distinguishes dos zelb model from standalone pronoun zelb.

Rivke ef⁺⁴=zen Moyshe zen un zelb fotografirn. — Rivke sees Moyshe and photographs Moyshe; +4 about the combined event, perceptually.

Rivke ef⁺⁴=zen Moyshe zen un zikh fotografirn. — Rivke sees Moyshe and photographs Rivke; the same epistemic scope.

Ikh ef⁺⁴=zen du zen. — “I see you, +4 bits, perceptually.”

Ikh ef⁺⁴=zog a model tsu du gebn. — “I give a model to you, +4 bits, by report.”

Case-free grammar needs recoverable roles. After a fronted time phrase and EF, the subject precedes the object. A bare object cannot freely move into the subject slot. Use an explicit topic phrase such as vegn X “about X,” with a full following clause, when topicalizing an object would otherwise be ambiguous.

Invariant adjectives and determiners

Core. Adjectives use one uninflected lexical base in attributive and predicative positions, for every noun and number: a gut model, dos gut model, dos gut modeln, and dos model … gut zayn. The language does not keep the inherited agreement alternation guter/gute/guts/gutn. The base is gut, not a new universal guts ending.

Possessives and demonstrative modifiers are also invariant: mayn gut modeln, dos dozik model, dos dozik modeln. Interrogative modifiers use invariant velkh “which”; no velkhe agreement alternant is needed. The quantifiers are invariant al “all,” yed “each/every,” epes “something,” and keyn in negative constructions. An explicit count precedes the noun: eyn model, tsve modeln. Zero and plural counts take the plural for count nouns. With fractional quantities, use a mass/measure construction such as “half of a model” rather than forcing a new agreement rule.

Comparative meanings remain available. Shtarker “stronger” and beser “better” are lexical comparative forms, not gender or case endings; they are themselves invariant. Thus dos shtarker model and dos shtarker modeln both mean “the stronger model(s).” A useful analytic alternative is mer or veniker plus the base adjective. Derivational endings in words such as troyerik remain part of those words.

Verbs and clause shape

Inherited verb stems and person/number conjugations remain available in questions and nonassertive frames, using the current invariant pronouns. A regular present-tense illustration is lernen: ikh lern, du lernst, es lernt, mir lernen, ir lernt, zey lernen. The copula has bin, bist, iz, zenen, zent, zenen. Lexical irregularities are retained.

In an asserted Yuddish clause, the epistemic auxiliary occupies the finite slot and the lexical predicate takes a nonfinite form. Thus inherited ikh ze “I see” becomes ikh ef⁺⁶=zen … zen. The new auxiliary takes familiar-looking agreement endings, but it is an invented paradigm.

The main-clause pattern keeps a finite verb in second constituent position. A fronted time expression counts as one constituent:

Ikh ef⁺²=zog morgn kumen. — “I will come tomorrow, +2 bits, by report.”

Morgn ef⁺²=zog ikh kumen. — “Tomorrow I will come, +2 bits, by report.”

The auxiliary, strength marker, and clitic together form one grammatical complex. Their internal pieces do not each occupy a separate word-order position. Operator-led clauses have their own patterns; section 8 explains them.