writing
thoughts & internet debris.
same weights, less waiting.
Making SANSA fit about 2× faster, with 17–18% less peak RAM and bitwise-matching weights on two datasets.
first place, wrong objective.
Returning to sailing after about ten years, and finding a race where leading can mean losing.
keep the evidence.
A music dialogue recommender where the difficult part was keeping the right track through two candidate-compression stages.
same agents, different scores.
A question from my Data Fest 2026 announcement: what still separates competition solutions when everyone has an agent?
one matrix, many questions.
One linear recommender, a short formula, and the modeling choices hiding inside it.
two bars, one tiny screen.
A small Flipper Zero side quest: putting Codex usage limits on a dedicated display.
a qwen on the home card.
Qwen3.5-35B-A3B on an RTX 3090, running through llama.cpp on Windows and connected to Claude Code.
new video, familiar author.
Finding an audience for videos with no history: our VK LSVD solution, an author bridge, and a shared limit of 100.
forty nodes, quite a few features.
From 22.6 million listings to 40 nodes: how AmazMe reused user histories in its winning Avito ML Cup 2025 solution.
five models, one prediction loop.
Inside our Riiid inference notebook: delayed answers, online statistics, three SAINT checkpoints, and the practical cost of keeping history.
a small conda trace.
A missing library, a short GitHub comment, and a few people getting unstuck in 2020.