Building a State-of-the-Art LLM Framework Took 10 Months, Not 3

A primer on GPUs and kernel optimization

A first-person account of building the Jam reinforcement learning framework after Claude Opus 4.5 made a three-month timeline seem realistic — ten months later, it's fast and state of the art. This primer covers GPU fundamentals: floating-point formats like fp8 and bf16, threads, warps, blocks, and SMs, HBM versus registers and shared memory, occupancy, and streams, then moves into kernel optimization and the author's workflow for developing kernels with LLMs.