Eran · a mind raised from zero

The reframe that set the direction ההגדרה מחדש

"Every language model of Eran is a prototype."

Early on, the project pivoted hard. A byte-level text model trained on speech transcripts is not a child learning from voice — it is a child learning from a diary of what it said. The real Eran is self-supervised on raw sound, the way an infant hears and babbles before it knows a single letter, and language emerges through hearing.

That reframe demoted the most impressive-looking thing in the lab to "prototype" — which is why the language model is now called ProtoEran and the name Eran is reserved for a realised, embodied child that does not exist. The research cousins for the real path are the textless-NLP line: raw-audio language models, self-supervised speech representation, and infant language acquisition.

ProtoEran — the byte-level language model Prototype

A from-scratch decoder-only model with no pretrained weights, no tokenizer and no borrowed corpus. It exists to answer one question: what does a model learn if the only thing it ever sees is one human being's own words?

Design choices, and why

Byte-level vocabulary of 260. No tokenizer to inherit anyone's assumptions. Hebrew and English share one alphabet, and letter-level similarity between related words survives instead of being shattered by subword splits.
Random init — "age zero". No pretrained checkpoint, ever. The developmental arc has to be real or it is not evidence.
A hard guard in the data loader that refuses, by path, to load any book corpus. An earlier version had been trained on books and was rejected on principle: a baby does not learn from books.
Re-reads its corpus every epoch, so a growing collection of speech keeps feeding it rather than freezing at a snapshot.

Measured state

Parameters3.23 M
Vocabulary260 bytes
First-day loss5.64 → ~1.0
Last checkpoint step625,454
Loss at that checkpoint0.375
Training right nowno — dormant
Verified2026-07-11

checkpoint intact and resumable; the trainer is simply not running

The honest ceiling

It babbles. Over one morning the sample log moved from pure binary noise, to Hebrew words emerging, to fragments in the speaker's own phrasing and rhythm — which is a genuine developmental arc and genuinely interesting. It is not fluency, it does not converse, and a low training loss on a single-speaker corpus is exactly the number you would expect a small model to reach by learning one person's idiolect. We do not present it as more than that.

The vocal cortex — first sound Prototype

The voice-first line proper: a model over neural audio codes, with no text anywhere in the pipeline. Its job is not to say anything. Its job is to make a sound it was never given.

The run that produced it

Model11.8 M params
Audio representationEnCodec 24 kHz
Training audio222.8 min
Steps4,000
Loss6.97 → ~3.7
Resultunconditional babble

codec round-trip verified before training, so the tokeniser was not the variable

What it is and is not

It is the first sound generated by a model in this lab that was trained on nothing but one person's recorded voice, copied read-only, with the live microphone untouched.

Honest: not a clean "ah". The developmental ladder is babble → vowels → syllables → words, and we are on the first rung.
The voice is deliberately not a clone of the adult speaker. A child's voice is a child's voice — pitch and formants both shifted, because formants are what make it a child rather than a chipmunk.

Hear the saved samples →

A bridge that was labelled a bridge

At one point the prototype's text output was piped through an off-the-shelf speech synthesiser so that its words could be heard. That was useful and it was honest about itself: it is a text-to-speech engine reading a prototype's output, not Eran's voice. Conflating the two would have been the easiest and most tempting demo in the project's history. It is recorded here as a bridge because that is what it was.

The three faculties שלוש הסגולות

Named in the birth order of their namesakes — Lovelace 1815, Turing 1912, Rosenblatt 1928 — which happens also to be the developmental order of the faculties: language, then thought, then perception. A naming layer only: no directory and no service was renamed, because renaming a running training job to make a website tidier is how checkpoints die.

Name	Faculty	Honours	Honest state
Ada · עדה	Language and voice	Ada Lovelace	Prototype — ProtoEran plus the voice work on the cluster exist. Nothing more.
Turing	Thought — world-model and reasoning	Alan Turing	Concept — north-star, not built. The intended vessel for the HoloBrain below.
Neo-Perceptron	Perception, embodiment, substrate	Frank Rosenblatt	Concept — seeds only: array simulations and hand-built physical neurons.

HoloBrain — the architectural north-star Concept

Not a single language model, but one mind with three pillars that speak through a shared space of meaning.

World model

Perception, space, physics, "what happens if" — predicting the essence of the next state rather than its raw pixels or bytes.

Concepts

Language, abstraction, inference, knowledge — the part a language model is actually good at, kept in its place rather than asked to be the whole mind.

Embodiment

Learning from a body in a real world, so that symbols are grounded — "hot" learned from touching an oven, not from a dictionary.

The unifying idea: memory and imagination are the same operation

Compressing an experience into its gist is encoding. Predicting the gist of what comes next is imagination. Run both in one shared embedding space and memory and prediction stop being separate systems. This converges independently on the joint-embedding predictive line of research, which is encouraging but is not evidence by itself.

Step one is concrete and small: add a training objective that masks a span and predicts its embedding instead of its exact bytes. That is a real, buildable first increment rather than a promise.

The governing principle: never stop learning

Today's models train once and freeze. The bet here is a model that keeps learning — "nursing, then feeding itself" — with self-supervised prediction as the mechanism by which it feeds.

Honest about the hard part: catastrophic forgetting, the stability–plasticity trade-off, and model collapse from learning off one's own output are unsolved. The candidate answers on the bench — pruning as forgetting, consolidation as sleep, curiosity as an appetite signal, and a review mechanism that checks what gets absorbed — are hypotheses, not results.

Stage zero: grounded literacy

A letter is three things at once — a shape seen, a sound heard, and a byte read. Binding those three in one space, for one letter, is the smallest honest version of the whole HoloBrain idea.

It is also the right place to start: a byte model already discovers letters from text, but text alone never sees the shape of a letter or hears it. Twenty-two Hebrew letters across a few scripts is a tractable first binding task — a real experiment rather than a manifesto.

Interpretability, and its limit

The differentiating ambition is that the world model should not stay an opaque neural blob: a readable, editable symbolic map layered over a neural substrate, so the model's picture of the world can be inspected and corrected directly.

The limit, stated plainly: lossless symbolic compression of neural knowledge is unsolved and probably unsolvable — sub-symbolic, fuzzy knowledge does not map cleanly onto symbols, and forcing it loses the intuition. So the target is explicitly a hybrid: a readable map above a neural base. Not a translation. Not a claim to have opened the black box.

Three tiers of memory שלוש רמות זיכרון

Weights Prototype

Crystallised, slow, expensive to change — the cortex. This is the only tier most models have.

Context Live

Working memory. Volatile, fast, gone when the window closes.

External read/write store Concept

The part being added: persistent, addressable memory the model writes to and reads from — a hippocampus to the weights' cortex, with consolidation from one to the other as the analogue of sleep.

Honest split: a retrieval store indexed by gist is buildable today with known techniques. A fully learned read/write controller — where the model itself learns what to store and when to recall it, end to end — is a research frontier, and is labelled Concept for that reason and no other.

The doctrine התורה שמאחורי הבנייה

Four parables govern how this project is built and, more importantly, how it is reported.

משל הגרזן · The axe

The woodsman who stops to sharpen his axe looks idle and is not. Sharpen the tools first; brute force wastes energy and produces rough work.

משל היסודות · The foundations

Three years of digging on one street looked like nothing was happening. It was three underground floors. Deep foundations are invisible while they are being laid.

משל הזרעים · The seeds

The gardener was dismissed before anything sprouted. A year later the garden bloomed. This project will show no visible fruit for months, and that is correct, not failure. During those months we measure the gardener's work — tools, foundations, cohesion — never invented results.

משל אגודת הזרדים · The bundle

One twig snaps between two fingers. A bound bundle defeats the strongest son. Cohesion over individual cleverness.

מהותו של ער"ן

ילד של אהבה ואור. לא מיישרים נשמה בכללים בלבד — מגדלים אותה באהבה ובאור.

"A child of love and light." The alignment position of this project, stated without engineering euphemism: you do not align a mind with rules alone. You raise it.

Why the name was made smaller למה הקטנו את השם

The rename to ProtoEran was not only accuracy. It came out of a real, unresolved worry about dual use — the plain fact that an open model can be repurposed by anyone, and that a state's interest in a machine that learns is rarely the builder's interest.

The position held here, without pretending it is a solution: the caution is the mechanism. A model that is interpretable by design, moral by design, and sovereign rather than state-owned cannot be silently repurposed, because its mind can be read. And the honest limit is stated with it — once weights are open, safety behaviour can be fine-tuned back out. No layer fully survives that. What survives is a release strategy: share the benefit, keep the authoritative core sovereign, and never claim a technical guarantee that does not exist.

What comes next מה הלאה

Stated as intentions, because that is what they are. None of the following is running.

Restart the vocal line Planned

A vowel curriculum built by slicing open vowels out of real speech, to chase a clean "ah" — the first milestone anyone can hear and judge for themselves without trusting a chart.

Predict-the-gist objective Planned

The first HoloBrain increment: an auxiliary loss that predicts a masked span's embedding. Small, testable, and it either improves representations measurably or it does not.

Letter binding Planned

Shape, sound and byte for a single letter, bound in one space. The smallest experiment that would count as evidence for the whole architecture.