Project NeuroSeed
Project NeuroSeed is the lab's research program. It asks whether intelligent systems can learn as well as today's neural networks while costing a tiny fraction of the compute, learning continually, and forgetting exactly on request. Five tracks carry it: Ikigai (language), Musubi (neuromorphic chips), Musubi Bio (biological neurons), Yomi (reading) and Soma (a creature that explores and acts in 3D worlds).
Language that learns from counts: cheap, continual, exact.
What it is
Our main language system. It stores what it reads as exact counts on disk, so it learns from new text immediately, keeps a bounded memory footprint, runs on an ordinary CPU, and can delete any document it learned from exactly.
Achieved so far
<1 GB
A counting language engine that runs on disk with under 1 GB of memory and deletes any training article exactly, in microseconds per token.
23.6%
A small learner on top of the counting engine that, on a 2-million-token WikiText-103 slice, beats a transformer trained on the same data by 23.6%.
12M
Together with the counting engine, that learner beats a same-data transformer baseline on a 12-million-token WikiText-103 slice.
Goal
A learner that generalizes to new, unseen contexts as well as a transformer while staying far cheaper than published methods: fast on an ordinary CPU, no large GPU training runs.
Where it's going
Scaling the cheap learner to larger data and making every part run natively on CPU.
Last updated: November 2026
The other tracks
{{ cur.desc }}
What it is
{{ cur.what }}
Achieved so far
{{ cur.note }}
Goal
{{ cur.goal }}
Last updated: November 2026