Writing on AI systems and how they behave in practice.
July 24, 2026
A pinned model ID freezes the weights and nothing else. The serving stack, the hardware batching, and the decoding surface all move without a version number, and the weights you did pin come with a non-extendable retirement date. What that does to your evals, your audit trail, and your architecture.
Read →July 12, 2026
Six voice AI systems from Microsoft, OpenAI, NVIDIA, Google, and Amazon, and the single layer that decides who is speaking and when. What is available in mid-2026, the class of product each one is fit to build, and the failure modes each one introduces.
Read →July 10, 2026
Per-token prices are collapsing faster than almost any technology input in history, and enterprise AI bills keep going up anyway. That is not a contradiction. It is three structural forces (frontier-chasing, agent loops, and the context tax), and every one of them is yours to manage.
Read →July 4, 2026
Pinning your AI provider to an EU region tells you where the bytes are stored — not who can legally compel them, or what leaves your boundary on every API call. The gap between data residency and data sovereignty, and what to do about it before the EU AI Act’s August deadline.
Read →July 2, 2026
Nearly every model now advertises a million-token context window. Almost none of them can actually use it. What the gap between the advertised number and the effective one means for the systems you build.
Read →