נוירונים יורים — האחים הקטנים

Brother A is the first little brother — a small version of Big Brother (5.6 million tunable "knobs", i.e. parameters). Born 12 Oct 2026, he lives on one box in Iddo's home: eran4 (a DGX Spark, folder minibb). His job: listen to Hebrew speech and understand it.

For each second of sound he builds a "summary" (64 numbers) and uses it for five tasks: name a syllable, predict the next second (JEPA), redraw the sound (the "mouth" — he does not speak, he draws a sound picture), read Hebrew letters (the "eye") and remember (the "memory").

Services that run all the time: learner (trains non-stop), server (a stable copy that answers — replaced only when the learner is clearly better), evaluator (honest check every few hours), bio ear (the firing neurons, next section), twin without the bio ear (a fair control group), heartbeat (alerts if something stalls) and a tap that films real spikes for this dashboard.

A real neuron does not pass smooth numbers — it fires. It collects signals from its inputs, its voltage rises, and the moment it crosses a threshold it sends a spike (a short pulse) to the next neurons and resets. What matters is when it fired.

A's bio ear is such a simulation on the GPU (the GeNN software): 4,096 excitatory + 1,024 inhibitory neurons (Izhikevich type) + 384 input neurons that receive the sound (96 frequency bands × 4); one step = one millisecond. It does not learn with backprop (the method almost every deep net uses) but with a local rule like in the brain: STDP — if neuron A fires just before neuron B, the link between them gets stronger; if after, weaker. And the change only sticks when a dopamine signal ("surprise/reward") is present. The excitatory neurons also have two branches (dendrites) that need two signals together to turn the neuron on. Its goal: predict the next window of sound — the better the prediction, the smaller the surprise.

A has a second, ordinary ear: a spiking ear in PyTorch — 4 layers × 512 neurons, trained with backprop, with logic gates on its branches, each choosing one of 16 operations (AND, OR, XOR and so on). The real syllable work is done by this ordinary ear — in the code, the syllable path does not use the bio ear at all.

Honestly: the bio ear's surprise is about … (1.0 = "guess zero", below 1 = better than guessing), i.e. still worse than guessing zero, and the twin without it does not sound worse. The biologically most interesting part has not yet proven it helps.

There is a "classroom" with a shared syllable library — the same library as brother B (folder minibb-world/syll). It has 115 labels: 111 Hebrew syllables (22 consonant sounds × 5 vowels, plus "aya") and four words from "Na Nach Nachma Nachman". About 51,900 training clips, from several sources:

• Iddo's voice (dad) — about 17,000 clips; another part of his voice is kept aside for testing only;
• real voices cut automatically (kh) — about 13,600;
• computer voices (TTS) — about 18,700, from 13 different voices;
• real speakers from Big Brother (probe) — about 1,400;
• "Na Nach Nachma Nachman" — about 1,100.

Each step he sees 48 clips, two thirds of them human voice. The human voice gets a mild distortion (a bit faster/slower/higher/lower, plus noise) so he does not memorize one voice. On top of that he keeps doing the other tasks on Hebrew sound and text from the shared world: predict the next second, redraw the sound, read letters, memory.

Two tests, and both use only things he has not seen:

(a) The syllable test. 1,104 clips from voices he never trained on — a part of Iddo's voice kept aside, the computer voice "Hila", speakers chosen in advance for testing — across 32 syllables. Random guessing gives about 3%. Always saying "the commonest syllable" gives about 10%–11% — that is the floor. A serious number has to be far above it.

(b) The honest score. 300 speech clips locked in advance: whole speakers that no part of the brain trains on, read-only files with a signature (sha256), and the code crashes on purpose if it finds any overlap with training. The score is the mean of five parts (see "How to read the board"). Every few hours the evaluator scores the server, the learner and the vanilla; the learner is promoted to server only if it is clearly better, and if the server drops, we roll back automatically.

Vanilla vs novel. The old comparison, against an ordinary GRU net of the same size, won by 0.06 — but it was not fair, because the ordinary net did not get the same helper tasks. So since 15 Oct there is a "fair twin" (nobio): same brain, same data, same tasks, same voice distortion — only without the bio ear. Only a comparison like that tells whether the new idea helps. Until today: no measurable difference (see the "hearing: learner vs twin" graph).

Brother B is the second little brother. Today he lives on eran2 (in the first days he lived on eran4), folder minibb-alma. The big difference from A: B's brain is written in "Alma language" — a language we built for describing brains, with "organs" (groups of neurons) and "threads" (links between them, each thread with its own learning rule). He is small: 192 ear neurons, 928 memory neurons and 64 "gate" neurons.

Three services: learner (runs non-stop), evaluator (checks every 20 minutes) and heartbeat (alerts if something falls). By the unit settings right now, B's ability to grow new memory neurons and to "consolidate memories" at night is switched off — they slowed him 6.5× with no measured gain.

Step one is a model of the inner ear (cochlea): the sound clip becomes 192 numbers (8 time slices × 24 frequency bands). They are called ALMA-EAR-192 — the "family feature" every brother uses. In B each number becomes an electric current into one ear neuron, and the neurons fire over 20 beats.

The 928 memory neurons listen to the ear neurons through random links — some excitatory, some inhibitory (the fix of 16 Oct: before, all links were excitatory and the memory was just a copy of the ear's constant noise). A memory neuron "tires" after firing (ALIF), so it cannot fire forever. They fire quietly: about 1% to 7% of them each beat, which is healthy (if all fire, it is a seizure).

How he learns: e-prop. Every link keeps an "eligibility trace" — a short memory of whether it helped cause a spike. When the "syllable reader" errs, the error is sent back along random feedback wires to the memory neurons, and trace × error gives a small weight change. It is a "local approximation" of backprop. At night (200 beats with no sound) he replays memories and deletes very weak links. Honesty: the "small head" (SylHead) on top of the ear features is trained with ordinary gradients, so it is reported separately.

The same shared syllable library as A: 111 Hebrew syllables, about 51 thousand clips of Iddo's voice, real voices and computer voices. He sees far less of it: so far … syllable clips since the current start, and in total … sound windows and … nights. (A gets 48 clips each step; B very few on each beat.)

He learns two things together: to read a syllable (the ear and memory neurons feed "readers" that learn on the fly to guess which syllable was heard) and to reconstruct and predict (the next window, redrawing the sound, memory). And letters: with every syllable its letters with vowel marks are also sent to the "eye", which links them to the brain's state for that syllable.

The same test as A: the same voices he never trained on (a big set of 1,104 clips over 32 syllables, and a smaller set of 400 for the history graph). Three readers are reported: the small head (SylHead, a small net on the 192 numbers), the memory readout (from the memory neurons — this is what B is supposed to prove) and the ear readout (straight from the ear neurons — the baseline).

For each we show: accuracy, the floor (always saying the commonest syllable, about 10%), over floor (accuracy minus floor — the honest number) and ±95% (the uncertainty band; on a small final test the number just wobbles). A result counts only if it is above the floor by much more than two bands, and especially if the memory readout beats the ear readout — only then did the neurons really add something.

We also measured ceilings outside the brain, to know what the feature can give at most: ridge on the 192 numbers 0.135–0.16, "nearest centroid" 0.21, and a small net (MLP) up to 0.34 with lots of examples. Vanilla vs novel: B has no vanilla twin; the baselines are the floor and the direct ear readout. The honest score (5 parts on 300 locked clips, exactly like A) compares "server" against "learner", and a promotion happens only if the learner is clearly better.