Lab notebook

Learning on real edge hardware, with the numbers and the limits

An independent lab studying learning and inference on BrainChip's AKD1500 and Apple's Neural Engine, alongside camera-based MechDog experiments. We separate host computation from chip work and report results with their controls, limits and available evidence.

2026-11-16

An AKD1500 uses prediction errors to gate teaching, filmed live

Live repeat of a fixed binary-stream test: 2 of 48 teaching gates opened and both changed answers recovered, versus 48 of 48 for always-teach. Every output and measured memory snapshot matched the earlier run.

AKD1500on-chip learningprediction
2026-11-16

An AKD1500 learned a cell-profile memory without labels, filmed live

8,192 unlabelled mouse-cell profiles, 60 source labels for a host readout, and 9,511 cells from another mouse: 63.6 to 69.0% balanced agreement, versus 18.3 to 23.1% for untrained memory.

Sped-up preview. The entry includes the full real-time take.

AKD1500on-chip learningbiology

How we report