Research · Benchmarks
Benchmarks.
Measured, not claimed.
Every number we quote comes from a test you can read about here — how we ran it, what we compared against, and what changed.
Internal · 2026
How we test AI agents before they go live
The testing suite behind a 68% drop in wrong answers for an enterprise AI assistant: every step the agent takes is checked, answers are scored against human judgement, and nothing ships if quality slips.
Read moreInternal · 2025
Quantum-inspired training vs. the usual way
On problems with the right shape, quantum-inspired methods cut AI training time 97× — from weeks to hours — compared with a well-tuned standard approach.
Read moreInternal · Brilliant Labs Halo · 2025
Seeing in real time, on the device
Image-recognition AI running inside small devices in under 10 milliseconds per frame — fast enough for glasses, drones, and robots — even when the device is running hot.
Read more