Field Notes

Notes on building μ: what broke, what the numbers said, and what actually helped. Each note covers one real program used as a probe — no synthetic micro-benchmarks, no curated best-cases.

#2 · June 2026 140 lines of μ, 119 ms, 98,000 rows

csv features em performance

A year of 1-minute tick data for 10 stocks — 98,280 rows — processed with a correlation regime detector (HMM, three states) and a pairs-trading simulation. No libraries, no NumPy: just csv, log, take/drop, and each'. 138× faster than Python, in 140 lines.

~136,000 ms (Python) → 119 ms (μ) · 138×

#1 · May 2026 50× speedup, guided by one benchmark

interpreter optimiser kernel fusion

A Monte Carlo GBM option pricer — 10,000 paths, 252 steps — goes from ~1,000 ms to ~46 ms through eleven targeted fixes: interned env keys, packed float arrays, fused Box-Muller and GBM kernels, and a two-line inliner correction that unlocked the final 2× gain. μ ends up faster than Python+NumPy on this workload.

1,000 ms → 46 ms · beats Python+NumPy