If you train models, or want to understand how they learn, this page is for you. Mica is a compiled programming language in which the numeric stack up to a training loop on a graphics card is part of the language itself: a tensor is a type that carries its shape, the autograd tape is a statement, and a tensor moves to the GPU by declaration. A GPT is one file you can read in an hour, it compiles to one binary of about a megabyte, and run again on the same machine it prints the same numbers.
Four things that are true here, measured
| Reproducible run after run | The same training program on the same machine leaves the same learned weights and prints the same loss column every time it runs, the promise C makes for its arithmetic, kept on the processor and on the graphics card. |
| Small and self-contained | The GPT of the tutorials, statically linked, is 953.7 KiB on ARM64 and 1232.6 KiB on AMD64, and runs from an image built FROM scratch. No interpreter, no framework, no toolkit to install: a machine with an NVIDIA card and its driver runs the GPU program, because the compiler ships the device code itself. |
| Checked before it runs | A product whose dimensions cannot meet is refused before the program exists, not three hours into training. A hidden transfer between host and device is a compile error. Memory and sharing are proven by the compiler, and at the checked tier an integer overflow traps at a named source line instead of wrapping. |
| Readable | Autograd is tape … end and one Backward; the GPU is on gpu; a layer is a formula. The two GPT programs of the tutorials are 531 and 568 lines, and the course reads every line of the first. |
See it
- Train your own GPT. A corpus, one command, minutes of training, and text in your corpus’s style, reproducible to the digit.
- The GPU, without kernels. A complete GPU program
with no kernel in it: a tensor declared
on gpu, two transfer verbs, and the same mathematics printing identical digits on host and device. - Mica Notebooks. A notebook that is plain Markdown, whose recorded outputs are verified facts: run every cell, hold each fresh output against the recording, and the document proves itself.
Learn it
The AI course teaches machine learning from nothing: twelve chapters from what learning is, before any mathematics, to a GPT you train yourself, every example compiled and run before it was published. Mica is the tool; the ideas are the subject. Beside the course, the tutorials that carry the numbers’ machinery: vectors and matrices, dynamic tensors, linear algebra, quantized tensors, batched training, image networks and deterministic reductions.
What is in the language
| Part | What it is |
|---|---|
| tensors | vector[N] and matrix[R, C] with the shape in the type; dynamic tensors whose extents live in the value; the int8 and uint8 element families the inference world packs its weights into |
| the tape | tracked values record onto a tape inside a tape … end window; one Backward walk answers every gradient through Gradient |
| the vocabulary | the tensor operators with Sum, Tanh, Softmax, LayerNorm, CrossEntropy, Relu, Gelu, Silu, RmsNorm, Conv2d and pooling, and the Gather embedding — each with its plain signature and shape rules, and each recorded by the tape |
| the GPU | a tensor declared on gpu keeps its type and moves its storage; values cross only at visible ToDevice and ToHost; the verbs dispatch by residency through PTX the compiler ships; an NVIDIA card of compute capability 7.0 or newer with its driver is the whole requirement |
| linear algebra | Solve, Invert, Cholesky and Qr over dynamic matrices, singularity as an expected answer on the failure channel |
| determinism | reductions on one documented pairwise tree, the same tree serially and in parallel; seeded, bit-stable randomness; the same result run after run on one machine, at every optimization tier; C’s own bits on both architectures for every operation IEEE 754 rounds exactly |
| the rest of a program | files, strings, regular expressions, networking, tasks and streams — a systems language around the numbers, and one binary at the end |
The facts page states exactly what is in the language today, every figure generated from the repository.
What Mica is not
Said plainly. There is no library of pretrained models: a model in Mica is what you train, and it saves and reloads its own checkpoint. The GPU road runs on NVIDIA cards of compute capability 7.0 or newer, on Linux. A training run lives on one machine and one card; nothing here spans a cluster. The vocabulary is the one the course uses, and it grows release by release — each release’s article on the blog states what was added and what was measured.
Start
Start puts the compiler on your machine in minutes, and the
starter repository holds every
program of the course beside the GPT, each a make away.