NeuroGenesysFrom living neurons
to engineered minds.
Our research is organized as one loop: learn from the brain, model what we learn, test it in Genesys, align it continuously, and chart the silicon it could run on. Five steps, repeated until intelligence is understood.
How discovery compounds.
Each loop makes the next one faster.
Neural Data Acquisition
The field's 4,096-electrode arrays record thousands of channels of living cortex, and synapse-resolution microscopy has mapped whole fly brains and a cubic millimeter of mouse cortex — ground truth on biological intelligence.
Brain Foundation Models
The emerging class of brain foundation models predicts neural activity the way language models predict text — a research direction we analyze in depth in our field notes.
Genesys Transfer
The computational principles the science surfaces — sparse coding, event-driven attention, continual plasticity — inform the Genesys program, whose experiments test forgetting and verification directly.
Continuous Alignment
Our published standard: interpretability audits, scalable oversight, and contained dangerous-capability evaluations belong inside every training cycle — not after it. Capability never ships ahead of understanding.
Silicon Embodiment
What survives review gets designed on paper first. Our cost-model campaigns chart where analog and optical compute win, and where they do not, before SynapseOne or Lumen commit to any hardware.
Where the science
lands.
The loop is abstract; its outputs are not. A paralyzed patient speaking again, an agent that remembers last month's work, a chip that thinks on microwatts. Here is what the method is aimed at.
Thought to text at conversation speed.
The field's landmark trials decode attempted speech at 62–78 words per minute from spike-sorted cortical activity. Our field notes follow the decoder generalization that will take this from lab milestone to daily life.
// ── THE DECODING SHIFT ──
// state of the field, sourced in our notes:
// 4,096-electrode cortical arrays (2024)
// 62–78 wpm attempted-speech decode (2023)
// pretrain-then-adapt decoders (POYO, 2023)
// open problem: cross-session generalizationAgents that learn without forgetting.
Our research path to agents that learn without forgetting — built on the finding that, with frozen experts, class-incremental accuracy factorizes exactly and the loss is selection, now confirmed on Split-CIFAR-100. The companion verifier claim did not hold for real judges at N ≤ 32.
// ── SELECTION, NOT FORGETTING ──
// class-IL accuracy factorizes exactly:
ACC = A_sel × p_within // ±0.01
// Split-CIFAR-100, frozen experts, 10 seeds:
// ε 0.534 → 0.788 as tasks go 5 → 20
// real judges, Best-of-N to N = 32:
// no decline; ε did not predict N*Compute that keeps analog's advantage.
The SynapseOne program studies event-driven, in-memory silicon against our published design rules — an 11.8× modelled analog advantage that holds only when the architecture is chosen to keep it.
// ── THE DESIGN RULE ──
// AIMC's win is a function of the window:
win(W) // sliding window: 5.6–6.8×, constant
// full attention @ 131k: → 1.05×
// choose W ≥ workload recall range,
// with margin — measured, not asserted