Technical notes from the people who build our cores, lattices and navigation systems. Every note answers one sharp question with primary sources and real data, and ships the code that reproduces its numbers. Corrections are dated and public.
A week of CO2 in, a day out, on 13 real offices. The dilated CNN beats every forecast that needs no training by 10%, almost all of it on empty days; a plain GRU beats the CNN; and the worst errors came from a calendar that didn't match the building.
MTBF is a rate, not a lifetime, and it assumes the rate never changes. Thirteen years of Backblaze data on half a million hard drives show how often that holds: two models with the same 40-year MTBF, one flat for nine years, one failing eight to ten times as often after year three.
225 cores, 40 years, no resupply. Poisson sparing shows why you carry modules rather than cores, and why each core you can afford to lose is a free spare of every part. Then comes the catch: at fleet scale, not knowing the failure rate costs more than bad luck.
ImageNet models on pictures of light curves tie MiniROCKET and HIVE-COTE 2 on StarLightCurves, then hit the same wall: Cepheids read as RR Lyrae. We rebuild the benchmark from OGLE-IV with periods kept and measure what the period is worth.
Two stars, one camera, 47 au from home. We reproduce the first interstellar navigation fix from Gaia DR3, explain why 1 parsec means 1 au of error per arcsecond, rank the best reference stars, and check our own datasheet.
Why scheduling the shortest message first beats fair sharing, what Homa measured against TCP on real hardware (19–72× lower tail latency), what's still unsolved, and where Linux upstreaming stands in 2026. With a simulation you can run.
