Some of this produced measurements. Some of it produced a clean, useful no. Some of it is a design document and says so in the label. The fence between those three categories is the most important thing on the page.
In a brain, structure is computation — where a region sits, how it is wired, how asymmetric it is. Transformers throw that away: a uniform, symmetric, all-to-all stack. The premise here is to make topology a first-class, grown, asymmetric variable.
"More brain-like" does not mean better. The brain is optimised for embodied, continual, low-energy survival — not for language fluency, which is precisely why backpropagation and transformers beat it at language by being unconstrained by biology.
So the target is only the capabilities we actually lack: continual learning without catastrophic forgetting, sample efficiency, energy, robustness, grounding, and a built-in conscience. Not brain-likeness for its own sake.
Rather than argue modular-versus-monolithic, both were built on the same benchmark, in pure standard-library code, deterministic, and scored on thirteen graded scenarios.
the fused model names one problem when the truth has several
The decisive gap was structural, not accuracy. One component fault takes down the entire fused pass; the modular arm degrades gracefully every time, and can be traced, swapped and extended. The adopted design is therefore specialists inside one brain with a learned routing gate — keeping the seams, paying for only the parts each input needs.
The harness also surfaced a shared weakness rather than hiding it: both architectures scored zero on two chord-recognition scenarios. A benchmark that only ever flatters the thing you built is not a benchmark.
And the standing law over all of it: data is the limiter, not cleverness. Any edge here comes from structure and orchestration, never from claiming to out-scale anyone.
Real published wiring goes in; known biological dynamics come out. If the dynamics have to be hand-tuned to appear, the model is a drawing of a brain rather than a model of one.
The classic reflex circuit, assembled from the published connectome with established dynamics.
the smallest nervous system there is — and therefore the right first one
The activity bump holds a heading and follows a moving cue. The ring topology that makes this work was recovered from the wiring rather than imposed on it, which is the part that counts.
Memory should be reached by context, matched against a compressed gist — and the crux: the match score and the connection weight are the same number, not two separate quantities. Activation then spreads along the whole pathway.
One source yields several gists at different levels of granularity, so a document is a small constellation of points rather than a single one.
honest limit: in the demo the gists are hand-set — learning them is the next step
An attempt at massively parallel analog computation in physical material — hand-built memristive cells, a few shekels of salt and gel, no foundry. The name is a memorial dedication. This is the most speculative work in the lab and it is fenced accordingly.
A cell was built and driven through an automated five-cycle set/reset test with sixty resistance reads.
Verdict: no memristance whatsoever. The cause was mechanical rather than chemical — a soaked absorbent substrate cannot hold a sub-millimetre gap, so the electrodes simply touched. Reported as a failure the same morning, rather than retried quietly until a graph looked better.
This is not a quantum computer. There are no qubits, no entanglement, no Hilbert-space superposition, and it cannot break modern cryptography. Any framing that implies otherwise collapses under the first hard question from anyone who knows the field.
Further: the central speculative idea — carrying several independent computation channels on one device by frequency multiplexing — has, in our own literature search, no published support. Not refuted, simply no evidence. It is treated as untested speculation to be settled by a small experiment before anything is built on top of it.
What is genuinely worth pursuing, without any embellishment: massively parallel analog computation in material, at very low power, from ordinary materials, owned end to end.
An early stand-in used small incandescent bulbs as analog weights. It worked as a weight grid and was then proposed for the frequency-multiplexing test.
It physically cannot carry one. A filament is a thermal low-pass filter with a cutoff of a few hertz — the very thermal averaging that makes brightness-as-weight work destroys any carrier above it. The bulb grid also turned out to be miswired: two lamps lit regardless of command, so they were never under control at all. Both facts are recorded so the design is not proposed again.
After the negative result, the bench moved to thermal and electronic neurons — and each idea was simulated before it was recommended.
A physical analog spiking neuron built as a relaxation oscillator: a capacitor as the membrane that integrates charge, a resistor and potentiometer as leak and tunable input weight, and a relay as the threshold. Cross the pull-in voltage and it fires — visibly, audibly — then discharges and resets.
Several inputs sum on the same capacitor; the potentiometers set the firing rate. Three of them wire into a physical three-member panel, which is the smallest hardware version of the conscience mechanism this project keeps returning to.
simulate first, recommend second, build third
Two things share this name, and only one of them matters technically.
A node-based visual language for building and training models with your hands rather than by typing — a spatial version of a node-graph editor, where a neuron is a block that snaps to other blocks because they all speak one interface.
The layered plan is deliberately unglamorous: a module standard first, then a node runtime, then a flat two-dimensional canvas that proves a graph can run, and only then anything spatial. The first three layers are buildable; everything past them is a long horizon and is described as one.
A readable, editable symbolic map layered over a neural substrate — so a model's world-model can be inspected and corrected directly rather than probed from outside.
That is the genuine differentiator, and it comes with its limit attached: lossless symbolic compression of neural knowledge is unsolved, because fuzzy sub-symbolic knowledge does not map cleanly onto symbols and forcing it discards the intuition. So the target is a hybrid, never a full translation.
status: design and scaffold. No language exists.
A sibling project to Eran, deliberately dormant. Everything below is design and a stopped seed; none of it is running.
Where Eran is raised pure and voice-first, Yigael (יגאל, "he will be redeemed") was to be the worldly sibling — free-range, bigger from birth, multi-cortex: audio, vision, movement through street-level imagery, and reinforcement learning through games.
It reached a real seed: an orchestrator routing across live model cortices on the cluster, and a vocal cortex trained on world sound rather than one person's voice.
Smaller pieces, listed for completeness with their honest labels.
A from-scratch microphone application — capture, voice activity detection, transcription and typing — rebuilt clean after an earlier version leaked memory and crashed every couple of days. Soak-tested at 300 inferences with flat memory.
honest split: the app is ours, the recognition model is not, and is credited as such
A fine-tuned adapter for one speaker came back worse on Hebrew than the base model — 37.1% word error rate against 30.7% — while English improved. Diagnosed as a self-distillation trap: the adapter was learning the base model's own mistakes.
deployment gated on an evaluation win that has not happened
A structured debate format in which several models must challenge a named opponent's weakest point and either defend or explicitly concede — conceding counts as a win. Built to make disagreement visible instead of averaging it away.
ships inert by design: it reaches no model until explicitly enabled