Now
Last updated June 2026 · Germany
Currently building
The first essays for this site — specifically, an illustrated deep-dive into agent memory architectures. The goal is to take one of the most practically important unsolved problems in production AI (how do you give a stateless system persistent, structured memory that doesn't degrade over time?) and explain it with the same clarity I'd want if I were encountering it for the first time.
Alongside that: companion Jupyter notebooks for each essay, designed to be runnable on Colab with zero local setup. The notebooks are not just code dumps — they're meant to be interactive explorations of the same ideas the essays explain visually.
Currently reading
- The Free Energy Principle: A Unified Brain Theory? Active inference as a unified theory of cognition. Testing whether it has any architectural implications for building agents that generalise.
- Probability Theory: The Logic of Science Working through it slowly. The book that makes probability feel like physics — deterministic, derivable, inevitable. The foundation for the Bayesian essay.
- Gödel, Escher, Bach A re-read. Still the most honest book written about self-reference and consciousness. Different chapters illuminate different things each time.
- World Models (survey) Background for the Physical AI series. How do you build a model of physical reality that's useful for planning?
Questions I don't have answers to
- If a language model has no persistent state between calls, in what sense can it "remember" anything? And what does a genuinely memory-capable agent architecture look like at the systems level?
- Does the free energy principle, taken seriously as an architectural principle, suggest anything useful about how agents should be built? Or is it only descriptive?
- Are the limits of large language models computational (we need more scale) or architectural (the transformer paradigm has a ceiling)? This question seems more important than almost any other in the field right now.
- What would it mean for a physical AI system to have genuine situational awareness rather than pattern-matched approximations of it?
- Gödel showed that no consistent formal system can prove all truths about itself. Does this have any non-trivial implications for self-improving AI? Or is the analogy too loose to be useful?
A diagram I keep returning to
Three memory rings of a production agent. The innermost is the context window — fast, volatile, capacity-limited. The outer rings require retrieval. This distinction is the architectural crux.
This page is a living document, not an archive. It reflects what's actually on my mind right now, not a curated summary. Inspired by nownownow.com.