ML, Biotech, Hardware, and Coordination Problems. Sometimes I write about hard problems and how to solve them.
Introduction Part 1 measured this dual GH200 workstation as a memory system. Part 2 used those numbers on the preview DeepSeek V4 Flash checkpoint, where the interesting/painful part was getting multi-token prediction to run at all: a hand-written vLLM O-projection fallback, a config-propagation fix, and a narrow upstream PR. Part 3 then took the box to its limit with GLM-5.2 and expert offloa...
TL;DR: Here is the NotebookLLM Podcast audio version. In Part 1, I built a Quantum Random Number Generator out of a pair of old lab-equipment photomultiplier tubes, a 50:50 beam splitter, and an FPGA, and then capped it with a Quantum Magic 8-Ball. The 8-Ball was the minimum viable Quantum Lever: five quantum-derived bits per question -> twenty answers -> twenty distinct macroscopic mult...
Sometimes I get weird projects stuck in my head. I’ve always wanted to do a “quantum” hardware project; in my PhD and postdoc I’d messed about making fluorescent dyes and proteins that were technically quantum (I designed around a Jablonski diagram ), but it was all unsatisfying. I’ve wanted to get down to single quantum events, ever since reading The Fabric of Reality, the 1997 book by physici...
Introduction Part 1 measured the dual GH200 workstation as a memory system. Part 2 used those measurements to explain why DeepSeek V4 Flash can be fast in vLLM when the model layout fits the hardware: keep hot weights in HBM, avoid unnecessary Hopper-to-Hopper traffic, and use MTP only where the acceptance rate pays for the draft work. GLM-5.2 starts at 2.39 output tok/s on this machine and a...
Introduction Small AI computers are usually sold with large dreams and shitty memory buses. I have a ridiculous server that pulls a few kilowatts, but I wanted a local Hermes Agent box that could sit on a shelf, stay near silent, draw laptop-class power, and still run a serious model with enough context for real agent work: always available, and cheap enough that leaving it on doesn’t gener...