
We won the @HockeyAnalytics Big Data Cup 2026! Forechecking is usually described as a system label. We tried to show it as a spatial timing problem: who arrives first, which lane is open, and how fast F2/F3 take that lane away.

occasional engineer, former wannabe pro-skier @canadaski

We won the @HockeyAnalytics Big Data Cup 2026! Forechecking is usually described as a system label. We tried to show it as a spatial timing problem: who arrives first, which lane is open, and how fast F2/F3 take that lane away.

also taking donations for gpus if anyone has a spare blackwell lying around + u liked my writeup 🙏

then I put everything I'd learned into an inference engine: 514 tok/s on Qwen2.5-1.5B, 1.33x llama.cpp on the same card not everything worked (my own scheduler only tied nvidia's). full writeup, made with a lot of help from opus 5.5: lukeblom.fyi/writing/3080lab…

second secret: ptxas makes mistakes. what I thought was a hardware limit on mixing float + int math was really the compiler putting registers in the same bank renumbering the registers by hand made a microbenchmark 1.48x faster

earlier this year I sold my 3090 (dumb), leaving me extra gpu poor stuck with just a 3080, so for the last 2 months I too have been getting into @nvidia's private affairs🤯 first secret: the GPU trusts the compiler blindly. get 4 bits wrong and half your math silently vanishes


sorry @kirkgoldsberry I messed up the tag :0

Tldr: Linking body to ball probably needs Hawk-Eye-level tracking data not achievable with my dataset and also with 8 to 45 attempts per shooter, my spread estimates are only weakly stable (split-half r is ~0.24), so player rankings are mostly noise for now.

The big question: does a shooter's 3D body-motion "fingerprint" predict where the ball enters? From my broadcast video -> tracking data, no. Real motion scores no better than shuffled motion. The pose estimates are noisier than a stroke changes from shot to shot.

Command blends accuracy and precision. But I wanted to pull them apart. Each shooter's average left/right miss direction is unrelated to their spread (rho +0.16). Some players aim 2 to 3 in off center and repeat it; others are centered but scattered.

The make rate stays flat (about 80%) until the ball's center is about 4 in off center, then drops. That's exactly where geometry says the ball's edge starts to clip the rim (9 in rim radius minus 4.7 in ball radius). (also a good relief that my ball entry tracking is reasonable)

Super cool paper popped up on my feed from Jake McGrath, @PandaGlazer, Vanna Bushong, Michelle Nguyen & @kirkgoldsberr : "command," a metric for free-throw accuracy built on 20k+ Hawk-Eye-tracked FTs. I tried a twist on my open playoff FT data 🧵 doi.org/10.1038/s41598-026-7…

github.com/lblommesteyn/moda… if u want ur workers to have a gpu

spent way too long on a side project finding out cloudflare workers can't call modal 🤯

This is just the tip of the iceberg of this data + would love to see deeper insights on my TL pls tag me in your work I'd love to see what other people come up with! I will regularly try to keep adding data + refining whats in the data as well :)

Fatigue at the line is real but late and small: nothing for ~24 minutes on court, then ~1.8 points lower per further 10. Being cold costs far more. First attempts go in ~4 points less often than second ones: ~33 points per team per season, against ~5 for fatigue.