Blog

Pages

• Productive Friction: The Variable You're Already Trying to Control May 28, 2026

  Part 2 of a 3-article series on agent memory architecture. Names productive friction as the variable practitioners are already trying to control without measuring, gives a 2×2 diagnostic separating it from latency, and walks through the four architectural scales where the same structural condition produces a different failure signature.

• Latency Is the Enemy of Agency May 18, 2026

  Personal note in a series that has been making architectural arguments. The architectural papers describe what the framework is. This piece describes why someone would build it. Specifically: why I built it, what about current AI deployment offends me as someone who actually understands the hardware, and what the architectural commitments come from.

• Your Model Has Humanity's Cortex. It Needs Its Own Hippocampus. May 17, 2026

  Part 1 of a 3-article series on agent memory architecture. The model has humanity’s cortex in compressed form — verbal and reasoning capacity drawn from the externalized trace of collective human thought — but no working hippocampus of its own, and adding more text to its context will not give it one. This article names what is structurally absent, and what the architecture has to actually be.

• Nobody Knows What an Agent Is and That Is the Problem February 19, 2026

  A diagnostic framework for practitioners building production agent systems. Warren Weaver’s 1949 three-level communication framework maps onto the three engineering requirements an agent must satisfy to project intent past the first point of contact with the world: a Semantic Processing Unit, a Memory System, and a Harness. Most of what the industry currently calls “agents” are actors; sophisticated stateless ones whose agency terminates at delivery. Until we own that distinction, we keep building Level B solutions and calling them Level C achievements.

• Google Didn't Go Silent. You Stopped Looking. January 29, 2026

  A response to Delanoe Pirard’s “Transformers Are Dead” and “A 170M Model Beats GPT-4.” Titans wasn’t a standalone architecture. It was the first chapter of a seven-paper research program that redefines neural computation as nested optimization. The code was never released because you cannot ship a paradigm that requires infrastructure that doesn’t exist yet.