Writing

An AI agent debugging my ZX81 chess engine once wrote itself a resignation letter. Its successor found a bug that survived 92 commits, in the time my train took to get in.

A new paper lets an AI agent rewrite its own harness and get better at its work. The interesting part is the one thing it is not allowed to do: decide whether its own changes survive.

A chess engine in 672 bytes of Z80 machine code. What constraint-driven design on a ZX81 teaches about building systems that absorb failure cheaply.

Large language models destroy voice through 25 identifiable patterns. The problem is not detection. It is quality.

Self-reported developer productivity with AI tools is structurally unreliable. The witnesses are compromised by the same cognitive biases Cialdini mapped decades ago.

Before asking whether AI can do something, ask whether your system can absorb the mistakes.

What AI content actually costs - and why trust, not filtering, is the design problem.

A 1983 paper on the ironies of automation maps precisely onto how we're integrating AI into modern software systems.

The models are good enough. The bottleneck is trust - and trust is built from infrastructure, not adoption speed.

AI changes what junior work looks like. If we're not careful, we'll optimise for today's velocity and break tomorrow's capability.