Twenty-three pipeline failures in six weeks, none of them caused by my code. This is how I replaced GPT-4 with a seeded Qwen2.5 on Ollama, plus the GitHub Actions caching and Pydantic validation details that made it stick.
A house-pricing model with zero training error valued a suburban shack at $2.4 million. This is the geometry behind lasso: centering the data, the L1 diamond, and why landing on a corner zeroes a coefficient.
Naive before-and-after averages said the London tube strikes lifted bike usage by 5.5%. A two-way fixed effects model with clustered errors put it at 3.95%. This is why the second number is the one to trust.
Academic causal inference stalls in applied work. Decision gravity is the fix: constructive decisions can live with an associative read, final ones cannot, and collider bias is what catches the people who guess wrong.
Prefill is compute-bound, decode is memory-bound, and running both on one GPU wastes most of it. Splitting them into separate pools, the KV-cache transfer tax that comes with it, and when it pays off.
A synthetic dataset can pass every standard check and still lose the correlations, rare events and privacy margin that your model depends on. Here are the three checks I run before shipping one, with the Python for two of them.
A static A/B test keeps sending half your traffic to the losing variation until someone stops the test. Thompson sampling moves the traffic as the numbers come in. Here is the Beta distribution behind it and how to store the state in WordPress.
My Clippings.txt is an append-only mess, and the expanded highlights inside it will wreck any summary you generate from the raw file. Here is the parser, the deduplication logic and the local Ollama call I use instead.
An ad hoc prompt works once, then quietly stops matching the result you got the first time. Claude Code Skills wrap the workflow in a SKILL.md file, with structured JSON for the data and Python for the parts that have to be deterministic.
On EMNIST Letters, a GMVAE reached 80% accuracy from roughly three labeled samples per cluster, where XGBoost needed 35 times the supervision. Most of that comes down to soft decoding instead of hard cluster assignment.