Data Elixir - Issue 595

Testing AI Accuracy*

AI can answer complicated questions but still get basic business numbers wrong. AtScale tested seven leading models and found that adding a semantic layer improved accuracy more than switching models. Tomorrow's free webinar explores the findings and how to test AI accuracy on your own data. Can't attend live? Register to get the recording.
*Message from this week's sponsor, AtScale

How to Learn ML in 2026

As AI agents take on more coding and data work, which technical skills still matter? In this interview, Vicki Boykis makes the case for ML fundamentals, hands-on data exploration, and deep systems knowledge. She also shares practical examples of where delegating work to agents helps and where it gets in the way.
Vicki Boykis

Nonlinear Formulas in Bambi: Gelman’s Golf Putting

Logistic regression gets the basic trend in golf putting right, but it misses the shape. This post shows how to use Bambi’s new nonlinear formula support to model the geometry of a putt and capture that shape. Includes code and examples ranging from enzyme kinetics to tree growth.
Tyler James Burch

A Primer for Earth Data Embeddings

Geospatial foundation models turn satellite imagery and other Earth observation data into embeddings that support similarity search, clustering, and change detection. This is a great primer that explains how those representations work, where they're useful, and how to get started with existing embedding datasets or generate your own using open-source models.
Development Seed | Kiri Carini & Martha Morrissey

Tips and Tricks for Faster Grouped Operations in base R

The expensive part of a grouped calculation isn't always the calculation itself. Sometimes it's splitting data frames, repeatedly evaluating R code, and combining the results. This post explores six ways to eliminate that overhead and get up to 20x speedups over base R alone.
Nathan Eastwood

Is doomsday three weeks away?

In 1960, a mathematical model predicted that the world’s population would reach infinity in November 2026. The model fit the historical data beautifully, which is exactly the problem. From smallpox forecasts of 77 trillion cases to Olympic sprint crossovers, this is a fun tour of why good fits make terrible forecasts.
Understanding the unseen | Adam Kucharski

Training Better Agents

Mercor tested whether small, targeted datasets could make AI agents better at real work, and the results were promising. Across legal, coding, and accounting tasks, the models didn't just score better. They became more likely to consult documents, check their work, and use tools effectively.
// sponsored

Still Bivariate, Still Curious

It used to take a lot of manual work to create bivariate maps, but new tools have simplified the process. The hard part now is deciding which variables belong together, how many classes your readers can handle, and whether a bivariate map is the right choice at all. This is a great explainer, with lots of useful links along the way.
Joshua Stevens

Characteristic NYC Neighborhood Complaints

In a dataset of millions of NYC service requests, the most common complaints are pretty similar everywhere, so what makes each neighborhood distinctive? In this post, Kieran Healy shows how thoughtful analysis and visualization can reveal patterns that raw counts miss.
Kieran Healy

Announcing DuckDB 2.0.0 LTS

DuckDB started as a lightweight, embedded analytical database, but with version 2.0, it's becoming something much broader. This major release adds client-server capabilities, faster queries on cloud storage, better support for semi-structured data, and a CLI designed to work with AI agents.
The DuckDB team

commons - for building trustworthy data analysis agents

If you already have analysis code you trust, there’s no reason for an AI agent to start from scratch. Posit’s commons package has agents try your vetted functions first and write new code only when nothing fits. Each answer is labeled by the path it took, not the model’s own confidence, so you know what to trust. Available for R now, with Python in beta.
Posit