Triuna Labs Research repo

Triuna Labs Research

Research,
with its evidence.

Every article here ships with the code and data behind it, so the claims can be checked: measured results labeled as measured, and speculation labeled as speculation.

Browse the research repo

Sixty-five bytes: prompt caching and the shape of project state

Paul Woll · October 2, 2026

A 51 KB project record, rendered at two revisions two accepted changes apart. Sixty-five bytes were still identical from the start, because prompt caches match from the beginning of a request and the revision number sits near the top. Reordering the same selected fields so the volatile ones sit last took that to 51.8%, while selecting less material accounted for none of the difference. A shared prefix is a precondition for cache reuse rather than a cache hit, and the note says so.

Read the article →
Run the measurement →

Route the Work, Not Just the Data: GPUs, CPUs, and the Rise of AI-Native SSDs

Paul Woll · August 18, 2026

An LLM request is many kinds of work, and only some of it needs a GPU. Why the winning AI architecture routes each operation to the cheapest tier that can perform it, from today's SSD-backed KV-cache tiers to a proposed five-plane AI-native storage device, with a laptop-reproducible 102 GB data-movement benchmark and a falsifiable proposal for computational-storage hardware.

Read the article →
Reproduce the benchmark →