Anthril

v2 · Thesis · 2026

Sparse. Event-first.
Energy-bounded. Local.

An AI research and applied technology company building learning systems that fire only when they should, represent the world as events, treat energy as a design constraint, and run close to the people who use them. Three frontiers: model architecture, applied AI for businesses, and personalised AI for everyday life.

The work, in public

Blog posts, publications, and open evaluation harnesses.

Everything we write and every eval we run is dated, attributed, and published. No embargoed benchmarks, no ghost-written posts.

Post-mortem2026-04-29T00:00:00.000Z·Personalised AI·11h 04m to detect

When the sparse activation gate failed silently for 11 hours

A canary release of Aurora's stage-1 microfields had its sparse-firing threshold mis-applied. The gate accepted every input. Quality looked fine in dashboards. Energy use tripled overnight before anyone noticed.

We publish the timeline, the bug, the alert that should have caught it, and the four checks we now run before any sparse-firing change ships. The post-mortem is the artefact; the patch is incidental.