Data · Wednesday, 26 August 2026 · Issue 87 · 6 stories

DuckDB joins AWS, 4-bit model beats bfloat16, dashboards lie

DuckLabs is now an AWS subsidiary, without scaring off its open-source users. Also: a compressed model that's smaller, cheaper, and more accurate than what it shipped from. And the case for distrusting every chart you see.

Analytics & Visualization

Dashboards aren't neutral: which metrics get screen space decides what leaders fund

Dashboards are sold as objective "single sources of truth," but every chart hides curatorial choices about what gets measured. What leaders can't see, they don't fund: trust, care quality, and local realities vanish because they defy easy quantification. The piece argues a "visibility equals value" trap now hands outsize power to the BI developers setting the thresholds. Critical reading for African public sectors racing into automated governance.

blog.dataversegh.com ↗9 min read at source

Any chart can say anything now. So how do we keep visualization honest?

Duke sociologist Kieran Healy's Ihaka Lecture argues trust in charts was already fraying before AI made it trivial to draw any conclusion from any data. The real problem isn't skill; it's that a good graphic is a social contract between maker and reader. Healy asks what that contract looks like once generation tools filter through everything. Worth an hour if you make charts for a living.

youtube.com ↗1 min read at source

ML & AI for Data

4-bit model beats its own bfloat16 original after compression

Standard quantization-recovery methods fail once a model has been structurally compressed, not just quantized: distilling from the recovered checkpoint caps accuracy at that degraded target's ceiling. QAH skips the recovered model and distills straight from the original pre-compression teacher, even across architectures. Applied to a GPT-OSS 120B compressed to 60B and quantized to MXFP4, the result beats its own full-precision bfloat16 version on 7 of 9 benchmarks. Smaller, cheaper, and more accurate than what it shipped from.

huggingface.co ↗8 min read at source

Why Bagging Hits a Wall More Trees Can't Break, and Randomness Is the Fix

Add enough trees to a bagged ensemble and error plateaus. The culprit isn't variance, which averaging already attacks, but correlation between tree predictions, a wall no number of trees can break. Feature subsampling is what targets it, and this post walks through the variance equation and an experiment that shows the ceiling in action. It reframes Random Forest's design from arbitrary hyperparameters into a single argument against correlated errors.

towardsdatascience.com ↗17 min read at source

Databases & Storage

DuckDB's parent DuckLabs is joining AWS, the database stays open-source

DuckLabs, the company behind the wildly popular DuckDB, is becoming an AWS subsidiary in early September. The reassuring part: DuckDB, DuckLake, and every extension stay MIT-licensed under the non-profit DuckDB Foundation, with its roadmap and governance untouched. A stakeholder advisory board is being added, and community support limits are being lifted. Amazon just got the most fashionable analytics engine in the data world, without scaring off its users.

duckdb.org ↗1 min read at source

DuckDB v2.0 treats recursive CTEs as one long-lived query, netting a 42.6× speedup

DuckDB's old recursive CTE engine restarted its pipeline machinery on every iteration, paying setup and teardown costs even when the frontier barely changed. In v2.0, the optimizer owns the operator tree and pool of executors for the whole fixed-point computation, retaining state where it stays valid and reusing schedules across epochs. A one-million-edge reachability query drops from 4.05s to 0.095s with identical SQL. There's also a semantic shift: USING KEY ... UNION now surfaces changed-key payloads to the next iteration.

duckdb.org ↗19 min read at source

You’re caught up.

That was all 6 stories in the 26 August issue.