Three stories that each puncture a comfortable assumption: your privacy export is scanning way too much, your dashboard fights were settled 110 years ago, and the ASR model you picked from a leaderboard might be parroting the answer key.
Fulfilling one DSAR used to mean scanning the entire trip history at Uber. Its export workloads return a handful of records but repeatedly read the hottest regions of a GCS-backed lakehouse, inflating storage, retrieval, egress, and latency costs. The fix: Apache Hudi column stats, sorting, and clustering so narrow predicates scan only what they need, plus GCS auto-class tiering to keep cold data cold.
Willard C. Brinton's Graphic Methods for Presenting Facts (1914) and its 1939 successor read like modern style guides. Bar charts beat pies, 3D graphics are called "much worse than figures alone," the don't-start-at-zero trick gets debunked, and a heatmap already raised millions for US schools. Every dashboard fight you're having was settled a century ago.
Hugging Face ran three probes across 11 open-source ASR models and found the highest scorers sometimes reproduced benchmark reference transcripts even when the audio contradicted them, in one VoxPopuli clip six of 11 models omitted an audible "Thank you" to match the flawed reference. The behavior weakens on fresh voices, so leaderboard scores materially overstate real-world transcription. It's the first concrete measurement of "benchmaxxing" in speech recognition, and a warning for anyone picking models off a leaderboard.
A SQL-generation agent worked fine for weeks, until a query that should have been flagged ran and produced a confidently wrong number. Root cause: judge and generator shared the same model, and the judge proved measurably more lenient toward its own output than toward another model's. That self-preference bias is a known LLM-as-judge failure mode that feels academic until it costs you an analyst's afternoon. The fix is treating the judge itself as a component that needs testing and calibration against human reviewers.
The math is tempting: the author's entire two-year corpus is 127,068 tokens, twelve percent of Kimi K3's 1M window. So he sent the same 12 questions down two paths, a top-5 RAG pipeline and the full corpus in one prompt, same model and system prompt, graded blind. His blunt verdict: "it fits" and "it works better" are different things. The failures he hit along the way, he argues, teach more than the results.
DuckDB v2.0 is quietly replacing its PostgreSQL-derived SQL parser with a new PEG-based grammar, and users shouldn't notice a thing. The point is the engine's future: the grammar gets easier to evolve, DuckDB can absorb syntax from other SQL dialects faster, and extensions can now extend the language at runtime. It's plumbing work that unlocks next year's friendly features.
Every HYROX finisher since the 2018/19 season has split-level data sitting openly on the official results portal, with run legs, station timings, and RoxZone transitions recorded separately. That is roughly 100 events a year and hundreds of thousands of athletes, retrievable without auth through the mikatiming platform. A clean endurance-performance dataset for pacing, transition, and station-composition analysis that most people never knew was public.
You’re caught up.
That was all 7 stories in the 22 August issue.