Data Science Weekly – Issue 669

Issue #669 (Sep 17, 2026) collects curated links across data science, machine learning, AI, data visualization, and ML/data engineering. Editor’s picks highlight an experimental 14‑byte neural network that attempts maze solving, a cautionary essay on the realities of backups, and an analysis of how large a sample is needed for the central limit theorem to be a reliable approximation. Other featured pieces include operational lessons from five years running petabyte‑scale ClickHouse clusters, a practicum on retrieval‑augmented generation (RAG) in fintech that recounts early failures and what worked, the open‑sourcing of dbt Charts for governed dashboards, a reenactment probing how Miami realtors talk about climate risk, and a technical note showing mean squared error can hide critical distributional differences between models.

Taken together, the collection emphasizes practical, production‑facing lessons and methodological caution. Several items focus on robust tooling and reproducibility (birdnetTools 2.0, dbt Charts, Neki sharded Postgres, tidyomics), while others warn about statistical pitfalls and automation risks (p‑hacking enabled by coding assistants, misleading MSE summaries, choice of analysis methods for seed germination). Reinforcement learning and search agent training (GRPO), cosine‑similarity interpretations, and debates about how much LLM size matters in RAG systems round out a mix aimed at engineers and researchers weighing tradeoffs between model choice, retrieval quality, and operational reliability.