01
Micro-autodiff
finite diff match
Reverse-mode chain rule implemented from scratch for tanh(a*b + c).
[LAB_RUNG_0]
FOUNDATIONS
Autodiff stops being magic when you build it once.
Rung 0 proves the base loop: represent data as arrays, compute a loss, get gradients, update parameters, then use JAX transformations to compile and batch that same idea.
Status
Code proof added
Code proof added
Runtime
Local CPU JAX
Local CPU JAX
Next
Publish builder explanation
Publish builder explanation
// MEASURED_PROOF
What ran
02
Tiny JAX training
0.429063 -> 0.192692 loss
A small MLP fits a noisy sine curve with explicit SGD over a parameter pytree.
03
jit + vmap
218.26x CPU speedup
The jitted fixed-shape training step is benchmarked after excluding compile time.
// FILE_MAP
The proof is inspectable.
labs/00-foundations/GUIDE.mdtracked
labs/00-foundations/micro_autodiff.pytracked
labs/00-foundations/tiny_train.pytracked
labs/00-foundations/tiny_train_fast.pytracked
labs/00-foundations/RESULTS.mdtracked
tests/test_rung0_foundations.pytracked