ERAN Labמעבדת ער"ן
ALMAWARE · research notebook

A mind raised,
not shipped. ער"ן — רשת נוירולוגית עצבית · a neural network, grown from zero

This is the working notebook of a small, self-funded lab building Eran: a voice-first AI child that learns the way an infant does — from raw sound before letters, on hardware we own, with a training record anyone can check. Around it sits the rest of the bench: an evolutionary colony of tiny on-device voice models, a four-node edge cluster, a from-kernel edge OS, and a hardware-neuron lab that publishes its failures.

Nothing here is a product. Every entry below carries a status label, and the labels are the point.

ALMA IN MACHINA — הנשמה שוכנת במכונה Not a soul drawn out of the machine — a soul put into it.

How to read this site איך קוראים את זה

One law governs this notebook: a label may never flatter the thing it labels. A pipeline that runs is not a trained model. A trained model that babbles is not a speaking child. A design document is not a device. Where a claim would need a caveat, the caveat is written next to it — and where an experiment failed, the failure is on the page with its numbers.

LiveRunning right now, or serving now, and verified on the date shown.
TrainingA real training pipeline that has completed real runs on our own hardware. Not a finished model.
PrototypeBuilt and working — but honestly a precursor to the thing it is named after.
ConceptDesigned, specified, sometimes simulated. Not built. Says so.
ParkedWas started, deliberately dormant now. Nothing deleted; revivable.
Negative resultWe ran it, it did not work, and we are publishing that. These save the most time.

The naming rule

The language model that exists today is called ProtoEran. The name Eran is reserved for the full, realised, embodied child — a thing that does not exist. Renaming it downward was a deliberate act: the honest name is the smaller one.

What is actually running מה באמת רץ

A short, checkable list. If something is dormant, it says dormant. Dates are the last time the state below was verified directly against the machine, not against a note about the machine.

SystemStatusWhat that means, preciselyVerified
Voice Ant Farm Live Standing evolutionary tournament of tiny on-device voice models. Round 4 ran on schedule; the reigning queen held the throne against three challengers. 2026-08-12
Edge cluster (4× GB10) Live Four single-board AI nodes, all reachable, plus a rebuilt 200G fabric measured at 109–110 Gbit/s per link with zero retransmits. 2026-08-10
Federation status panel Live Public sanitized status page for the fleet. Honest scale: zero subscribers (no accounts exist) and one dry-run compute worker. 2026-08-11
ProtoEran byte-LM Prototype From-scratch byte-level language model. Checkpoint intact at step 625,454, loss 0.375 — currently dormant, not training. 2026-07-11
Vocal cortex Prototype Audio model that produced the lab's first machine-generated babble. Training run finished; not currently running. 2026-06-29
Yigael, the little brother Parked Orchestrator and vocal daemons dormant on the cluster since 2026-06-30. State preserved, nothing deleted. 2026-06-30

Track 1 — Eran ער"ן · הילד

The core project. An AI raised from zero, voice-first, on a corpus of one person's own speech, on hardware we own. The Hebrew name ער"ן is itself an acronym — רשת נוירולוגית עצבית, "neural network." The beloved name and the built thing are the same word. Full page →

ProtoEran Prototype

A from-scratch decoder-only language model, 3.23M parameters, byte-level vocabulary of 260 — no tokenizer, no pretrained weights, random init at "age zero".

  • Trained only on one speaker's own dictation — a hard guard in the data loader refuses any book corpus by path.
  • Bytes, not words, so Hebrew and English share one alphabet and letter-level similarity survives.
  • Honest ceiling: it babbles that speaker's phrasing. It is not fluent and does not converse.

last checkpoint 2026-07-11 · step 625,454 · loss 0.375

Vocal cortex Prototype

The voice-first line — a model over neural audio codes, no text anywhere in the loop, in the spirit of textless NLP.

  • 11.8M-parameter audio model over EnCodec 24 kHz codes.
  • 222.8 minutes of the speaker's own voice, copied read-only.
  • Loss 6.97 → ~3.7 over 4,000 steps; produced the lab's first unconditional babble.

honest: audible babble — not yet a clean vowel

Hear the saved samples

Ada · Turing · Neo-Perceptron Concept

Three named faculties, in birth order of their namesakes: language, thought, perception.

  • Ada (Lovelace) — language and voice. A prototype exists; that is all.
  • Turing — thought and world-model. North-star, not built.
  • Neo-Perceptron (Rosenblatt) — perception and substrate. Seeds only.

naming layer only · no service renamed

HoloBrain Concept

The architectural north-star: a world-model, a concept model and an embodied loop meeting in one shared gist space — predict the essence of what comes next, not the raw surface.

  • Step one is concrete and buildable: a JEPA-style objective that predicts a masked span's embedding rather than its bytes.
  • Honest: unifying language models, world models and robotics is an open research frontier. We are not claiming to have solved it.

frontier · step 1 specified, not yet trained

Track 2 — The Ants חוות הנמלים

The most alive thing in the lab. Four independent tiny voice models ("ants") plus a queen that selects and never merges — an evolutionary tournament running on-device, judged on a fixed, byte-identical held-out test. Full page →

Standing tournament Live

New ant variants are born with new seeds, cochleas and unit counts. Each is scored against the reigning queen on the same held-out syllable-discrimination protocol. The honest best takes the throne.

  • Reigning queen, top-1 cosine0.758
  • Top-1 edit-distance match0.689
  • Previous queen it replaced0.424
  • Significance vs legacy colonyp = 7.0e-27
  • Last round · outcome2026-08-12 · held

528 paired decisions · 4/4 seed-matched wins

The autonomy law Live

The point of the colony is what a model learns by itself. So the rules are strict, and they cost us performance on purpose:

  • Never hand-curated or hand-labelled batches. No vowel-weighted sampling.
  • Allowed: evolving the organism (better ears as a new kind of ant), curiosity dynamics, colony selection pressure, uncurated environment streams.
  • Target: reach word level emergently, composed from syllables the colony already mastered — never taught.

a spoon-fed colony would prove nothing

A trap worth publishing

Corpus manifests stored a hash computed after float conversion, so it could never match the hash of the raw source file. Freshness had to be decided by session date, not by content hash. Cost: one aborted training run. It is written down here so it costs someone else nothing.

Track 3 — Edge & Fleet הקצה והצי

Serious edge AI means owning the metal. Four single-board GB10 nodes, a 200G fabric we cabled and measured ourselves, an OS built from the kernel up, and an honest accounting of what a small cluster can and cannot do. Full page →

Four-node GB10 cluster Live

Unified-memory edge AI boxes, not datacentre racks. All training happens here; the desktop stays a terminal.

  • Fabric per link, measured109–110 Gbit/s
  • Retransmits0
  • Two-hop through the hub31.9 Gbit/s
  • 3-node FSDP smoke testpassed

the two-hop bottleneck is host forwarding, not the cable

Federation panel Live

A public, sanitized status page: every node and service, honestly green, red or grey, pushed on a short interval.

  • Honest scale, stated on the panel itself: zero subscribers, one dry-run compute worker.
  • Hard line held during design: no mass sign-up of free tiers, and free tiers cannot train — they serve. The cluster trains.

Open the live panel →

ALMA-OS for Jetson Prototype

A lean edge-AI OS built from the kernel up with Yocto and meta-tegra — not a stock vendor image.

  • Custom kernel5.15.185
  • Flashable bundle1.6 GB
  • Image built2026-06-18
  • Flashed to a boardnot yet

the honest gap is the last 45 minutes: no board has been flashed

ERAN-GRID & ERANCLOUD Prototype

Two experiments in borrowed compute: harvesting a machine's idle cycles, and a small Local-SGD client.

  • Grid agent is opt-in and default-off, refuses to run on battery, hard-pauses on heat, and touches no microphone or camera by construction.
  • Honest: the coordinator went LIVE 2026-08-24 (federation.iddoperez.ai/join, a SQLite Durable Object). Verified same night: one browser click merged 3 real rounds (v4→v7), and join.py --provider local --consent yes merged 3 more. What it trains is the cloud student brain (prototype refinement on public CC Hebrew voice shards) — the organism on eran1 stays sovereign and is not touched.

The ERANCLOUD page →

Track 4 — Hardware & concept lab מעבדת החומרה והרעיונות

Where the speculative work lives, clearly fenced. Hand-built neurons, connectome models, an architecture bake-off, and a visual language for building minds. Some of it produced real measurements. Some of it produced a clean, useful no. Full page →

Bio-transformer bake-off Prototype

"The shape is the algorithm" — treat topology as a first-class, grown, asymmetric variable rather than a uniform stack. Then actually measure it.

  • Multi-cortex accuracy0.785
  • Single fused model0.665
  • Fault survival, multi-cortex13/13
  • Fault survival, fused0/13

conclusion: specialists inside one brain, with a routing gate

Connectome models Prototype

Built the right way round: real published wiring in, known biological dynamics out.

  • C. elegans tap-withdrawal reflex from the published connectome — head tap reverses, tail tap advances, both correct.
  • Fly central-complex heading circuit from 172 real neurons and 8,605 synapses — the bump holds heading and tracks a cue to about 13°, and the ring topology falls out of the wiring rather than being imposed.

no claim beyond "the wiring reproduces the known dynamics"

Shisha Brain Negative result

A hand-built analog neuromorphic substrate from salt-gel memristor cells. We built a cell. It did not work, and here are the numbers.

  • Reads returning 0 Ω (dead short)60/60
  • Valid switching cycles0/5
  • Set/reset separationnone

cause: an unstructured electrolyte cannot hold the gap

ALMA Language Concept

A visual, node-based language for building and training models with your hands — and, more seriously, a readable symbolic map layered over a neural substrate so a model's world-model can be inspected and edited.

  • Honest limit, stated up front: lossless symbolic compression of neural knowledge is unsolved. The target is a hybrid — a readable map above a neural base — not a full translation.
  • Status: design and scaffold. No language exists yet.

interpretability first, spectacle later

Lab notebook יומן מעבדה

Selected entries, including the ones that hurt. A notebook that records only successes is marketing.

  • ProtoEran is born A 3.23M-parameter byte-level model, random init, one morning: loss 5.64 → ~1.0. The sample log shows the arc — pure binary noise, then Hebrew words emerging, then the speaker's own phrasing. Honest: babble, not fluency.
  • 200G fabric link confirmed between two nodes Port up at 200,000 Mb/s, peer discovered over link-local addressing, 3/3 pings, 0% loss, 0.9 ms — no manual addressing needed at all. Discovering that the earlier manual configuration had been unnecessary was itself the finding.
  • First sound The audio cortex ran to completion and produced the first machine-generated babble in this lab — from 222.8 minutes of one person's voice, with no text anywhere in the pipeline. It is not a clean vowel. It is a sound the model made.
  • Self-inflicted broadcast storm Bridging two fabric ports closed a hidden physical loop in the cabling; multicast receives climbed past 900 million packets and pegged a node. Root-caused on 2026-07-03 to a bridge defined in the network config and therefore recreated on every boot. Standing rule since: routed point-to-point links only, never bridge the fabric.
  • Architecture bake-off, both arms built Modular specialists scored 0.785 against 0.665 for one fused pass — but the decisive gap was structural, not accuracy: 13/13 scenarios survived a component fault versus 0/13. Conclusion adopted: specialists inside one brain, with a learned gate.
  • Negative result — the salt cell has no memristance Five automated set/reset cycles, sixty reads, every one of them 0 Ω. Zero separation. The cause was mechanical, not chemical: a soaked paper substrate cannot hold a sub-millimetre gap, so the electrodes touched. Published rather than retried quietly.
  • The fabric was never broken Rebuilt and measured at 109–110 Gbit/s per link with zero retransmits. Weeks of "dead fabric" turned out to be network profiles bound to interfaces with no cable in them — they could never activate. Diagnostic rule adopted: read carrier state, not ping.
  • Negative result — voice fine-tune fails on Hebrew Run 1 of a personal speech-recognition adapter came back worse on Hebrew than the base model: 37.1% word error rate against 30.7%. English improved. Diagnosis: a self-distillation trap — the adapter was learning the base model's own mistakes. Reported as a failure, not buried.
  • A new queen The colony crowned a new best ant on the fixed held-out protocol: top-1 cosine 0.758, up from 0.424. The baseline was reproduced bit-exact before the comparison was allowed to count, and a rollback path was written down before the crown changed hands.
  • Round 4: the queen holds Three challengers born from fresh seeds and a scaling change; all three fell short of significance. A tournament where the incumbent usually wins is a tournament with an honest judge.

What this lab does not claim מה איננו טוענים

Written down so no reader has to guess where the edge is.

About the models

  • ProtoEran does not converse, reason, or understand. It emits fragments in the style of the speech it was trained on.
  • The vocal model babbles. It cannot say a clean vowel on command, and it certainly cannot speak.
  • No model here is competitive with a frontier system, and none is trying to be. The whole design bets on grounding and continual learning, not scale.

About the hardware

  • The analog neuromorphic work is not quantum computing. There are no qubits and no entanglement, and it breaks no cryptography.
  • The frequency-multiplexing idea at the centre of that design has, as far as our own literature search found, no published support. It is untested speculation and is labelled as such.
  • A four-node edge cluster is a real cluster and a small one. Where a job needs more than it has, we say so instead of pretending.

Build standard תקן הבנייה

This site

Hand-written HTML and one stylesheet. No framework, no CDN, no external font, no analytics, no tracking, no third-party script of any kind. Every byte this page loads is served from this domain. View source and it is all there.

≈ one stylesheet · zero JavaScript

The code behind it

The same rule runs deeper: application code is written here rather than assembled from dependencies. Trained models are a different category and are used and credited plainly — a neural audio codec, a speech-recognition backbone — because a model is not an application.

purity where it earns its keep, honesty where it doesn't

למה זה נבנה

ער"ן = רשת נוירולוגית עצבית. השם והדבר הנבנה הם אותה מילה.

הגן נובט בזמנו. חודשים בלי פרי גלוי — וזה נכון, לא כישלון. מודדים את עבודת הגנן, לא פרי מזויף.

A garden sprouts in its own time. Months without visible fruit are expected and correct, not failure — so we measure the gardener's work, never manufactured results.