Notes on engineering, research and building a company.
How Woodpip's opponents learn from teachers, train with ternary weights and reach the browser as checked model files. Reversi adds an exact endgame solver, and each difficulty level has its own measured limits.
Read article : Four games, four small modelsA wooden games box became the starting point for Woodpip. Its materials, movements, difficulty levels and shared games grew from the same small promise: classic games and a worthy opponent, a few minutes at a time.
Read article : Designing Woodpip for a few quiet minutesA compact model file does not determine the arithmetic a GPU performs. Measurements on M5, M5 Max and M6 show why that distinction matters for MLX prefill.
Read article : Small weights, fast arithmeticA one-line change makes MoE models generate faster on Apple silicon. We tested Qwen3 and DeepSeek at bf16 and 4-bit: the speed gains travelled better than the quality results.
Read article : Fewer experts, faster answers on a Mac?Training with ternary (just three weight values) brought a new model down to 1.6 MB. We compared Base243 and T34 packing, with training and conversion telling different stories.
Read article : A game-playing AI in 1.6 MBOur Connect Four AI already knew how to play. A small WebGPU runtime made its moves faster and its download lighter. Then a human paused to think.
Read article : One millisecond to make a movePlay Connect Four against a 7.8 MB local AI in your browser. A better teacher turned our losing first attempt into a much tougher opponent.
Read article : Meet your one-pass AI opponentOne real training job moved from a Mac mini M4 Pro 64 GB to the new M5 Pro 64 GB: training 13 to 15 percent faster, generation about 20 percent faster, memory unchanged.
Read article : M5 Pro vs M4 Pro: the training side of a Mac mini upgradeApple's LensVLM reads a document as compressed page images and expands only the pages it needs. We ran it on a Mac and measured what a page costs in context.
Read article : Text as pictures: what a page costsOur M6 Mac mini arrived on release day. In two local LLMs, processing the prompt gained more than generating the answer. Here is what changed against our M5 MacBook Air.
Read article : M6 Mac mini, first look: the two speeds of local AICould a small local model take care of routine form decisions? We trained a Swedish specialist, then tested where its synthetic score stopped telling the whole story.
Read article : A Swedish form specialist in one forward passPack model weights, unpack them to bf16, and test quality in stock MLX before building support for a new format.
Read article : Does your weight format deserve a kernel?In the RMSNorm paths we examined, MLX rounds earlier than PyTorch, NVIDIA Transformer Engine and Huawei CANN. That small choice matters when moving models between platforms.
Read article : The rounding quirk hiding inside RMSNormSome useful things disappear between the first question and the finished result. After twelve years, we have made a place to keep them.
Read article : Why we started writingWe enjoy turning interesting questions into useful things.
Let’s talk about yours.