| | “Talk to your data” demos are easy. Scaling them isn’t. This webinar explores what breaks after the proof of concept: rising query costs, governance gaps, inconsistent answers, and how teams use semantic layers to turn chatbots into production-grade agents. See the full agenda →
*Message from this week's sponsor, AtScale |
| | |
| 🔗 Note: if you don’t see links in this email, it’s also available on the web: dataelixir.com/archive/issue-561.html |
| | |
| | | Prediction asks “what will happen?" Causation asks “what if I change something?” This post argues they’re more connected than people think. Predictive relationships often work because of underlying causal structure. When that structure doesn't generalize, prediction fails, sometimes fatally.
Thomas Lumley |
| | |
| | This post explores a practical alternative to spinning up a dedicated vector database. By storing embeddings in Parquet and adding a lightweight index in the footer, you can run vector search directly through SQL with DataFusion. This is a great “use what you already have” approach to working with embeddings.
Xiangpeng Hao |
| | |
| | Using 72 million trades from a real prediction market, this post walks through how to analyze microstructure, quantify longshot bias, and measure liquidity-driven return asymmetries. A practical case study in modeling human behavior under incentives.
Jonathan Becker |
| | |
| | We’re starting to see a new pattern: structured public data wrapped in an MCP layer so LLMs can query it safely. This post walks through how India’s national statistics office integrated Claude and what the architecture looks like under the hood.
Aman Bhargava |
| | |
| | | | Stop hand-tuning every figure. This guide walks through a repeatable ggplot2 setup for publication-ready exports, with a custom theme, global geom defaults, and SVG export tips for editing in Inkscape when needed.
Jörn Alexander Quent |
| | |
| | | Before you tune a query, you have to understand it. SQL Crack visualizes joins, filters, CTEs, and aggregations as a clickable flow graph, with column-level lineage and built-in performance hints. Especially useful for working with real-world, messy SQL.
GitHub | Buvan |
| | |
| | Andrej Karpathy calls this an “art project.” In just 200 lines of pure Python, he strips GPT down to the bare algorithm for training and inference. It doesn’t get much cleaner than this.
Andrej Karpathy |
| | |
| | | The entry-level data science market has shifted. In this 11-minute breakdown, a senior data scientist with experience at Stripe, Uber, and a startup explains what’s non-negotiable in 2026, how LLMs reshape the skill stack, and why starting as an analyst may be the most pragmatic path in.
YouTube | Priya L. |
| | |
| | | We’ve added two focused Data Elixir editions alongside the regular Tuesday issue. Each one highlights high-signal picks from Bluesky, with direct access to the discussion so you can read, follow, and jump in when you want. | Data Practice — Everyday data work, including SQL, statistics, visualization, R/Python, and tools. Published Wednesdays. Learning Machines — Machine learning and AI, including tutorials, research, tools, and applications. Published Thursdays.
| Subscribe to one or both. |
| | |
| | | Data isn’t “stored” so much as it’s juggled. This deep dive explains storage as an orchestrated tradeoff between latency, power, and physics, backed by great diagrams of SRAM, DRAM refresh, flash charge traps, and spinning disks. It’s a great reset on what computers are actually doing when they “read data.”
Making Software | Dan Hollick |
| | |
| | | This week's issue sponsored by: | | |
|
