Production inference for open-source LLMs*

Open-source LLMs offer more control and lower costs, but running them reliably takes real infrastructure. Nebius Token Factory handles inference, scaling, and performance so you can deploy models like Llama or Qwen behind a dependable API.
*Message from this week's sponsor, Nebius Token Factory

🔗 Note: if you don’t see links in this email, it’s also available on the web: dataelixir.com/archive/issue-563.html

Intuition for Why You Always Need Robust Standard Errors

Most of us use robust standard errors because we’re told to. But why do they actually differ from classical ones? This post builds intuition with simple simulations and visuals, showing how uneven noise changes coefficient uncertainty and why your p-values might shift.
Data Colada | Uri Simonsohn

Building a Model that Visualizes Strategic Golf

This is a really clean walkthrough of turning a messy spatial decision problem into a value surface. It builds “strokes-remaining” heatmaps across a golf course, showing step by step how to model expected cost in continuous space.
Wigs on the Green | Matthew Schoolfield

Querying 3 billion vectors

Vicki Boykis got nerd-sniped by a social media exchange about querying 3 billion vectors. The dot products weren’t the hard part. At scale, the real challenge wasn’t the algorithm at all. It was nailing down the requirements.
Vicki Boykis

The Postgres JOIN That Shouldn’t Be Empty

Ever had a JOIN return zero rows when you know the data exists? In this post, Haki Benita traces a head-scratching Postgres edge case involving row locks, concurrent updates, and isolation levels. Great debugging walkthrough.
Haki Benita

Is Boosting Still All You Need for Tabular Data?

‘XGBoost is all you need’ felt settled in 2022. The 2026 benchmarks complicate that story. Here's a nuanced look at where deep learning now wins on tabular data, where it doesn’t, and why the framing itself misses the point.
Michael Clark

How I use Claude Code

Most developers still use AI coding tools like supercharged autocomplete. After 9 months with Claude Code as his primary development tool, Boris Tane landed on something different: a strict plan-first workflow where the AI never writes code until he's reviewed and annotated a written spec. Here's how it works.
Boris Tane

➡️ Reach 50,000+ Data Elixir readers. Click here for details 👉

kuva

Think “scientific plotting, but UNIX-native.” kuva ships with a CLI that renders plots from data passed via the shell, including a terminal output mode. It’s lightweight, fast, and produces pixel-perfect SVG, PNG, and PDF.
GitHub | James Ferguson

Interactive beeswarm charts in R

Beeswarm charts show the full distribution of a continuous variable while preserving every individual data point. The points are placed so that they don’t overlap, making it look like a swarm of bees. In this post, Nicola Rennie shows how to build interactive versions in R, complete with hoverable tooltips.
Nicola Rennie

 New opt-in editions from Data Elixir

These topic-focused editions complement the regular Tuesday issue and go deeper into the topics you’re most interested in:

Opt-in to one or both.