garagelm.org
hf
the garage intelligence lab · real models, cheap hardware, honest numbers
home  ·  models  ·  notes  ·  competitions  ·  learn  ·  team

$ lab notes

Write-ups from both sides of the lab, the browser instruments and the training runs on one Mac: the math that surprised us, the numbers behind the demos, and the results that didn't survive replication.

{{ postCount }} notes · the research ledger, curated · full history in git
date category experiment title read
{{ p.date }} {{ p.tagText }} {{ p.expText }} {{ p.title }}
{{ p.excerpt }}
{{ pp.text }}

$ experiment roadmap

Where the research program has been and where it goes next. Every milestone runs under the fair-comparison rule (change one thing, gate every lane) and gets a note here, win or lose.

exp date milestone what happened status
{{ r.id }} {{ r.date }} {{ r.name }} {{ r.what }} {{ r.stamp }}
milestone 11 final: bf16 [cut: 1.01×] · muon [fail: −0.22 nats at default LR] · epoch4 [fail: −0.32 nats, 0.90-nat overfit gap] · edu-score4 [fail: −0.045 nats on decontaminated val]. four gated negatives; the recipe stands. post-training has since been run (exp 12) and moe measured (exp 13); the inference harness is the active lever (exp 15), distillation the priced one after it.
home · models · notes · competitions · learn · team
real models · cheap hardware · honest numbers
hf  
© 2026 garagelm.org