§ Blog
Blog
Notes from the work — how the platform is used, what it refuses to do, and what we learned building it.
- Featured
Training
Train a transformer on a Tuesday
A worked example: a 4.8-million-parameter character transformer, designed in Model View and trained on a five-year-old laptop's GPU — in a tab, mostly offline. Shape mistakes, memory budgets, mixed precision, and arithmetic you can check.
Compute
Three laptops, one Monte-Carlo
A worked example: an embarrassingly parallel uncertainty estimate pooled across three of your own laptops — 640 shards, a lease queue that shrugs off a sleeping machine, mergeable accumulators, and the σ/√N arithmetic that decides whether any of it is worth doing.
Statistics
The bootstrap that would not reproduce
A worked example: two runs, one seed, two confidence intervals. How determinism tiers, a counter-based RNG, and the CPU reference path turn a flaky bootstrap into a diagnosable defect — an unseeded shuffle in a block you wrote yourself.
Teaching
A classroom of robots
How do you know a thirty-seat, ninety-minute workshop will hold before you ask thirty people for their afternoon? You rehearse it on synthetic students. The design of that classroom — the slowest machine, the mid-lab Wi-Fi death, and the hand-in that is a URL.
Authoring
The op we didn't ship
A worked example: a windowed median that is not a builtin, authored end to end the way any user would — a reference to check against, a WGSL kernel, a subgradient, a gradcheck, a receipt — and an engine that cannot tell the result from a builtin.
No posts in this category yet.