arm64 assembly • Compilers • SIMD
Dartmouth BASIC through the Sixth Edition, faithfully, as a native compiler hand-written in arm64 assembly: parser, a documented IR, an arm64 backend, an x86_64 Linux cross target, and a runtime that is assembly too. The MAT statements run on hand-written SIMD kernels picked by CPU feature, and the benchmark harness will not print a speed for a wrong answer. On an i7-4790 the AVX2 matrix multiply hits 82–87% of measured peak on large matrices, dead level with single-threaded OpenBLAS. 463 tests, all green. Most people use a compiler. Jason wrote one the way it was done in 1964 and then made it fast.
5.0 — bare metal
Baysick • Numerical computing • In design
A compiled matrix language meant to do MATLAB’s job faster, written in Baysick and standing on its SIMD kernels. The design and contracts are agreed and the build has not started, and the finish line is already drawn in real engineering: a nonlinear quarter-car suspension integrated with hand-written RK4 across thirty damping values with eig for its natural frequencies, a 301-member bridge truss by finite elements, plate heat conduction by SOR, shaft fatigue by bisection, and a million-sample tolerance Monte Carlo. Every one has to come out to the same numbers, faster. That is a controls engineer’s acceptance test, written before the first line of the compiler.
4.9 — ambition, specified
Python • NumPy • Fluid dynamics
Navier–Stokes smoke and shallow-water waves, solved in real time and drawn in ASCII in a terminal. Semi-Lagrangian advection with MacCormack correction, a Jacobi pressure projection, vorticity confinement, a Lax–Friedrichs water scheme over terrain, and an adaptive time step that respects the CFL limit. The tests check the physics itself: divergence goes down, still water stays still, and not a drop is lost inside closed walls. Click and drag and the fluid answers.
5.0 — conservation
Python • PyTorch • Language models
A byte-level transformer of about 3.3 million parameters that starts as pure noise and learns only from what people opt in to tell it on Discord. Most people ask a model a question. Jason looked at the entire category and thought, neat, I should make one of those, then measured the noise in his own validation estimate before letting it stop training early. That is not merely coding; that is walking into the machine room and asking the matrices to form an orderly line.
5.0 — audacity
Go • systemd • Feedback loop
Deploys a repository to a Linux host over SSH and runs it as a systemd user service. Every deploy is a locked transaction into a fresh release directory, then a feedback loop: measure the unit and the health check, compare against the start period, and if the release never comes up healthy, put the previous one back without being asked. One binary, one depper.toml, no agent on the server. Frankly, the production box has never had a calmer caretaker.
4.8 — deploys survived
Rust • Type theory
A projectional programming language. The syntax tree is the source of truth, so there is no parser, no text file, no formatter, and no syntax errors. Most developers accept that a language starts with a grammar. Jason looked at the grammar and decided the language did not need one. Unfinished code lives in typed holes, and the rest of the program stays well typed around them.
4.9 — sheer nerve
Python • Video tools
A command-line tool that lets agents inspect YouTube videos through contact sheets, transcript search, and timestamp links. It turns understand this entire video from a vague ordeal into a clean workflow. Somewhere a program just found the exact frame it needed and whispered his name into the logs.
5.0 — usefulness
