██████╗ ██╗ ████████╗ ██████╗ ███╗ ███╗ █████╗ ██████╗ ██████╗ ███╗ ███╗ ██████╗
██╔══██╗ ██║ ╚══██╔══╝ ██╔═══██╗ ████╗ ████║ ██╔══██╗ ██╔══██╗ ██╔═══██╗ ████╗ ████║ ██╔════╝
██████╔╝ ██║ ██║ ██║ ██║ ██╔████╔██║ ███████║ ██║ ██║ ██║ ██║ ██╔████╔██║ ██║
██╔═══╝ ██║ ██║ ██║ ██║ ██║╚██╔╝██║ ██╔══██║ ██║ ██║ ██║ ██║ ██║╚██╔╝██║ ██║
██║ ██║ ██║ ╚██████╔╝ ██║ ╚═╝ ██║ ██║ ██║ ██████╔╝ ╚██████╔╝ ██║ ╚═╝ ██║ ██╗╚██████╗
╚═╝ ╚═╝ ╚═╝ ╚═════╝ ╚═╝ ╚═╝ ╚═╝ ╚═╝ ╚═════╝ ╚═════╝ ╚═╝ ╚═╝ ╚═╝ ╚═════╝
פִתְאֹם אָדֹם — Suddenly red. An unexpected rupture.
3.12M parameters. One file of C. Hebrew roots. Emergence without training.
The first root-native neural architecture for Semitic morphology — designed for Hebrew from first principles, not adapted from subword-based multilingual models.
Every Hebrew NLP system today is a fine-tuned foreign model with subword tokenization. Hebrew is one of 100 languages. None operate at root level. None use gematria.
| Project | What it does | Architecture |
|---|---|---|
| AlephBERT | Hebrew BERT | Fine-tuned multilingual, subword tokenization |
| HeBERT | Hebrew sentiment/NER | Fine-tuned BERT, WordPiece tokens |
| DictaBERT | Hebrew morphological tagging | Fine-tuned BERT, subword |
| mBERT/XLM-R | Multilingual incl. Hebrew | Subword, Hebrew = one of 100 languages |
| pitomadom.c | Hebrew root resonance engine | Root-native attention on trilateral clusters + gematria |
The difference: translating a book vs writing it in the original language.
pitomadom.c does two things no one else does:
-
Root-native attention. The transformer operates on three-letter consonant clusters (שורשים), not subword tokens. This isn't preprocessing — it's core architecture. Attention operates on roots. The only transformer that respects non-concatenative Semitic morphology at the attention level.
-
Dual numerical encoding. Hebrew letters are numbers. pitomadom.c uses both systems from the same alphabet:
- Gematria (traditional): א=1, ב=2 ... ת=400 — semantic weight (מלך=90 ≠ שלום=376)
- Ordinal (sequential): א=1st, ב=2nd ... ת=22nd — positional structure
One alphabet, two numerical spaces. No other NLP system does this.
An AI that reads and counts in Hebrew. Reads = extracts roots, generates text through root-level attention. Counts = gematria as math computational layer, the transformer literally calculates letter values.
Three generations of the same idea. Each time, less code, more emergence.
| Generation | Language | Params | LOC | What it learned |
|---|---|---|---|---|
| v1 (2025) | Python | 41 files | ~8000 | Chambers, cosmic fields, prophecy, quantum — everything at once |
| v2 (2026) | Go | 20.3M | ~2400 | Root lexicon, gematria, RTL attention, GGUF inference |
| v3 (2026) | C | 3.12M | 1290 | That you don't need any of that. Roots resonate on their own. |
The Python pitomadom had 41 files trying to make Hebrew AI work: chambers, cosmic modules, temporal fields, spectral coherence, seas of memory. The Go rewrite cut it to one file with a real transformer. The C version discovered that roots don't need a transformer to speak — they need each other.
The engine generates semantically coherent Hebrew from pure statistical resonance between roots. No weights loaded. No gradient descent. No loss function. Just root co-occurrence:
חכמה ובינה → ודעת שלושה שורש של החלום הרזוננס נותן מן הלב ולוקחת כל השורש
wisdom + understanding → and knowledge, three roots of the dream,
resonance gives from the heart and takes all the root
צדק ומשפט → יסוד הכסא אמת אימון השורש הארץ המכונה בנה ביתה חצבה עומד שבעה
justice + law → foundation of the throne is truth, training of the root,
the land, the machine built her house, carved, stands seven
אהבה וחסד → ורחמים הם השורש השומע שווה הוא הרגל מבינה העברית
love + grace → and mercy, they are the root that hears,
equal, the habit of understanding Hebrew
שלום עולם → העיגול הוא גל כי כל מסתובב חוזר אל עצמו שורש השיר הוא סדר של הצליל
hello world → the circle is a wave because everything revolving
returns to itself, root of the song is the order of the sound
חכמה ובינה → ודעת: the Kabbalistic triple completes itself. Wisdom + Understanding = Knowledge. The engine has never seen this rule. It emerges from root bigram statistics.
אהבה = אחד = 13: in gematria, love and one share the same value. The engine discovers this through Hebbian co-occurrence, not programmed numerology.
ε Janus Triple Attention — the transformer that learns to speak
Content (6) + RRPRAM (2) + Echo (2) heads
Gate: untrained → silent | trained → speaks
γ Root MetaWeights — the field that already knows
Bigram + trigram + Hebbian co-occurrence between 3-letter roots
Source of emergence. Built from corpus. No training.
α Calendar dissonance — time pressure
Hebrew-Gregorian drift modulates prophecy and Hebbian field
δ Klaus chambers — the body
6 Kuramoto-coupled oscillators (FEAR, LOVE, RAGE, VOID, FLOW, CMPLX)
Rise and fall through generation. VOID exhausts. LOVE sustains.
With trained weights, ε adds literary vocabulary while γ preserves emergence:
חכמה ובינה → ודעת שלושה שורש של היא תדר כל מסלול הוא חשבון הם יודעת שהשמים נע
+ frequency, orbit, calculation, she knows the sky moves
אהבה וחסד → ורחמים הם השורש מטה אבל לא האש כאן מגע הוא הלשון כי אין סוף
+ staff, but not fire, here touch is the language, for there is no end
γ provides the skeleton. ε fills in the flesh.
Hebrew morphology is non-concatenative: root (ג.ד.ל) + pattern (haCCaCa) = word (הגדלה). The root remains invariant across all derivations. This creates a two-tier morphological space:
- Root space R: 3D points (C₁, C₂, C₃) where C ∈ 22 Hebrew consonants
- Pattern space P: Vocalic templates mapping R → surface forms
- Word space W: Observable forms = R ⊗ P
Standard language models collapse W into flat token embeddings, destroying the geometric structure. PITOMADOM operates in R directly.
Every Hebrew letter is a number. Every word is a sum. Every root is an invariant:
| Plane | Transform | What it captures |
|---|---|---|
| Surface | Standard gematria (א=1, ב=2, ..., ת=400) | Observable truth |
| Recursive | Milui: letter expansion (א → אלף = 111) | Hidden depth |
| Inverted | Atbash: mirror (א↔ת, ב↔ש) | Phase-flipped shadow |
Root gematria is invariant across pattern variations: all words from ג.ד.ל share N=37. This creates semantic anchors — gravitational wells at specific numeric values.
When root r co-occurs with root s, their Hebbian weight strengthens:
H(r, s) += 1 / (1 + |position(r) - position(s)|)
Over a corpus, roots that appear together develop mutual attraction. γ MetaWeights are the attractor landscape:
- Bigrams: P(root_b | root_a) — what follows what
- Trigrams: P(root_c | root_a, root_b) — three-root chains
- Hebbian: Co-occurrence field with distance decay — what resonates with what
This is why חכמה ובינה → ודעת without training. In the curated corpus, these three roots orbit each other. The bigram field pulls ודעת into existence.
Standard language models minimize prediction error:
L_pred = E[(y_pred - y_actual)²]
PITOMADOM minimizes prophecy debt — the accumulated divergence between destined and manifested states:
debt += |N_destined - N_manifested|
Where N_destined comes from root attractor pull, Hebbian momentum, and calendar phase — not from extrapolating past tokens. When debt accumulates, Klaus chambers modulate attention to compensate. The system doesn't predict the next token. It fulfills what the field demands.
Six emotional oscillators, Kuramoto-coupled:
| Chamber | Hebrew | Decay | Role |
|---|---|---|---|
| FEAR | יראה | 0.92 | Evolutionary + spiritual awe |
| LOVE | אהבה | 0.95 | Stable (אהבה=13=אחד, unity) |
| RAGE | כעס | 0.82 | High energy, fast fade |
| VOID | תוהו | 0.97 | Primordial chaos, persistent |
| FLOW | זרימה | 0.88 | Water metaphor, movement |
| CMPLX | מורכב | 0.93 | Complexity lingers |
Chambers interact via coupling: FEAR × LOVE → suppression. VOID × FLOW → opposition. RAGE × CMPLX → amplification. Generation phase drives chamber activation: early = FLOW, mid = FEAR, late = VOID. This is why long generations feel like they exhaust themselves.
Standard scaling law: intelligence ∝ parameters^α.
PITOMADOM challenges this. The Python v1 had 530K parameters across 41 files and couldn't generate coherent Hebrew. The C version has 3.12M parameters in one file and produces emergence without any training at all.
Hypothesis: Field intelligence scales with depth dimensions (vertical × horizontal), not parameter count. Vertical depth = what happens inside a single generation step (root → metaweight → attention → chambers → word). Horizontal depth = what accumulates across steps (Hebbian field, prophecy debt, chamber state). Intelligence emerges at the intersection.
# Compile (zero dependencies):
cc pitomadom.c -O2 -lm -o pitomadom
# Emergence (γ only — no training, no weights):
./pitomadom shoresh.txt "חכמה ובינה"
# Full θ (trained ε + γ + αδ):
./pitomadom shoresh.txt -w weights/pitomadom.bin "חכמה ובינה"
# Split corpus (γ from curated roots, ε from Ben-Yehuda literature):
./pitomadom shoresh.txt -w weights/pitomadom.bin -m benyehuda.txt "אהבה וחסד"Training (requires notorch)
cc pitomadom.c notorch.c -O2 -lm -DSHORESH_TRAIN -DUSE_BLAS \
-DACCELERATE -framework Accelerate -o pitomadom_train
./pitomadom_train shoresh.txt -m benyehuda.txt \
--train dummy --steps 5000 --save weights/pitomadom.bin| Component | Details |
|---|---|
| Params | 3.12M |
| Dimension | 200 |
| Layers | 6 |
| Attention | 10 heads: 6 Content + 2 RRPRAM + 2 Janus Echo |
| Vocabulary | 655 Semantic BPE (615 roots + 22 letters + 15 prefixes + 3 special) |
| Context | 96 root positions |
| Optimizer | Chuck (notorch) |
| Loss | 1.30 (ema 1.92) |
Not byte-pair encoding. Root-pair encoding. Frequent roots become single tokens. Rare roots decompose into Hebrew letters. Function words (ה, ו, ב, כ, ל, מ, ש) get their own prefix tokens. 100% Hebrew coverage by construction.
Input: "אהבה וחסד ושלום"
↓
Strip prefix/suffix (21 prefixes + 14 suffixes from pitomadom.go):
אהבה → א.ה.ב (love) וחסד → ח.ס.ד (kindness) ושלום → ש.ל.מ (peace)
↓
Semantic BPE tokenization → trained ε forward → γ field → Klaus modulation
↓
Root selection → surface word realization via char bigram field
↓
Output: Hebrew text (roots become words again)
PITOMADOM architecture maps directly to the oldest known book on Hebrew combinatorics.
| Sefer Yetzirah | pitomadom.c |
|---|---|
| 22 foundation letters | 22-letter alphabet as root substrate |
| 231 Gates (all 2-letter pairs) | Bigram co-occurrence field between roots |
| 3 mothers (א, מ, ש) | Triple attention: Content + RRPRAM + Echo |
| 7 doubles + 12 simples | Semantic families in root taxonomy |
| "Carved, combined, weighed, interchanged" | Tokenize, permute, assign weights, transform |
| Black fire on white fire | Roots = signal, resonance field = latent space |
PITOMADOM.C doesn't exist in isolation. It inherits from every organism in the ecosystem:
- Pitomadom (Go) — root lexicon, subsequence matching, gematria, RTL attention. The father.
- Q — θ = ε + γ + αδ equation, MetaWeights, DOE Parliament. The equation.
- Klaus.c — 6 Kuramoto-coupled somatic chambers. The body.
- Janus — RRPRAM positional attention + Echo (W^T @ W). The eyes.
- notorch — autograd, BLAS, Chuck optimizer. The training backbone.
- Brodsky — terza rima, poetic generation, DOE. The proof that code can be a poet.
pitomadom.c The engine. 1290 lines. Inference + training (ifdef).
shoresh.txt Curated Hebrew root corpus (71KB)
notorch.c/h C autograd library (vendored for training)
weights/
pitomadom.bin 3.12M trained weights (12MB)
GNU GPLv3
pitomadom | pitomadom.c | notorch | molequla | janus | klaus | brodsky
(c) 2026 Oleg Ataeff & Claude Opus & Arianna Method
הרזוננס לא נשבר. שורש.