Writing

Training and running a neural network on a 1982 BBC Micro — first in BBC BASIC, then in 6502 assembly. Fixed-point Q4.11 arithmetic, a sigmoid lookup table, and a shift-and-add multiplier all prove that backpropagation doesn't need a FPU, a GPU, or anything invented after 1985.

Building a real-time path tracer in WebGPU — no libraries, no RTX extensions. Tracing rays through Doom levels and procedural dungeons at 60fps in a browser, and discovering why a £1,600 GPU still can't brute-force the rendering equation.

NVIDIA is selling the picks and shovels in the AI gold rush. Unfortunately for Jensen, the buyers are using them to automate the picksmith. What's good for the goose is good for the gander.

Increasing numbers of us are using Neural Networks to help us code, as personal therapists, for writing, image generation, just about everything and anything. But how do they work? In this series we're going to build a series of increasingly complex networks from the ground up starting with an XOR evaluator.

Recreating a 1986 BBC Micro game by using AI to reverse-engineer 6502 assembly — extracting Q7.8 fixed-point physics, emulating the SN76489 sound chip, and discovering why the original feels the way it does. Software archaeology with a modern accomplice.

I still love developing but the shifts that AI have brought are tectonic and are forcing me to re-evaluate my own relationship to building things

AI's real danger isn't taking your job - it's reshaping how you think. Engagement optimization, invisible guardrails, and the illusion of objectivity combine to create something more insidious than social media.

A psychedelic twinstick shooter, playable online, and written in a weekend using WebGPU and an AI coding assistant.

A deep dive into grammar-constrained token generation, using GBNF and llama.cpp to teach an LLM how to write real assembly for a custom 8-bit architecture.