Autonomous Improvement Rate
A first-order model for how fast an autonomous system improves: iteration throughput multiplied by useful information gained per iteration and the efficiency with which the system retains and exploits what it learns.
Putting Qwen3.8-27B to Work on a DGX Spark
An engineering study of speculative decoding, local inference throughput, harness overhead, context strategy, edit protocols, and where a local 27B coding model does and does not replace more expensive frontier-agent work.
A Discipline for Building Software with AI
The operating method behind the portfolio: architecture and acceptance criteria first, durable decisions, verified external claims, bounded workers, deterministic gates, independent review, and evidence that survives sessions and models.
Generated output stats available on request.