Ternary Semantic Brain Core — Zero Hard-Coding, Language-Independent Meaning Engine

Ternary Semantic Brain Core — Zero Hard-Coding, Language-Independent Meaning Engine

I built a meaning-learning engine that works without LLMs, embeddings, or
translation tables. Everything is stored as 2-bit ternary values {-1, 0, +1}.
It learns meaning purely from word co-occurrence in plain text.

What makes it different

No hard-coded linguistic knowledge. No stoplists, POS tags, tokenizer,
fixed vocabulary, or translation tables. All linguistic structure emerges
from training data.

Language-independent. Tested with English + Turkish. Same mechanism
works for any language with letter-based writing.

Ternary representation. {-1, 0, +1} — inhibition, unknown, excitation.
0 means “I don’t know” — a first-class answer, not a failure.

Single decision rule. All thresholds come from each word’s own
distribution. No hyperparameter tuning. (We call it “golden ratio freeze”
— referring to structural convergence, not φ = 1.618.)

Emergent morphology. After EN+TR dictionary training: 247 suffixes and
38 prefixes discovered automatically with zero linguistic rules.

Cross-language bridge. Without being told “water = su,” the brain
forms bridges between equivalent concepts across languages. After 65 books:
average Jaccard 0.47, cosine 0.61 across 10 EN-TR word pairs.

How it works

1. Start with an empty brain
2. Train with monolingual dictionaries (word → definition)
3. Optionally add books/articles for richer meaning
4. Query: /compare water su, /map fire, /senses storm

The brain builds a sparse graph of word relationships. Multi-meaning words
split into separate sense layers automatically. Meaning groups emerge from
community detection on the neighbor graph.

Architecture

Layer What it does Concept neurons Each word is a neuron with sparse ternary signature Sentence neurons Sentences become neurons linking words Synapse graph PMI-weighted co-occurrence connections Sense layers Dynamic multi-meaning, born from data Topic groups Community detection on neighbor graphs

Results (65 books trained)

Metric Value Concepts 288,407 Sentences 1,234,706 Synapses 102.7M RAM ~1.3 GB

Cross-Language Bridge Examples

English Turkish Jaccard Cosine water su 0.46 0.61 fire ateş 0.35 0.57 king kral 0.46 0.56 sea deniz 0.49 0.60 moon ay 0.43 0.67

What this is NOT

• Not an LLM — no attention, no transformer, no text generation
• Not open source — binary-only experimental release
• Not a pretrained model — you train from scratch with your own data
• Not AGI — no claims of consciousness or general intelligence

Try it yourself

git clone https://github.com/arifkurnaz/ternary-semantic-brain-demo
cd ternary-semantic-brain-demo
chmod +x scripts/linux/*.sh
./scripts/linux/02_train.sh --dict

Enter fullscreen mode Exit fullscreen mode

Linux binary included. Windows via WSL2.
Full paper and architecture docs in the repo.