| | AI agents are easy to demo and hard to trust. The difference is usually the data. The new Precisely AI Studio gives you a head start with working agents, apps, and skills for enriching, cleaning, and governing data. Run the examples, then adapt them for your own projects. Try AI Studio free → *Message from this week's sponsor, Precisely |
|
| | 🔗 Note: Data Elixir is moving to a new publishing platform this month. So far, so good, but if you notice anything unusual, hit reply and let me know. Meanwhile, if you don’t see links in this email, it’s also available on the web: dataelixir.com/archive/issue-597.html |
| | |
| | A model can predict which customers will cancel and still be terrible at identifying which customers a retention offer could save. That's one of many lessons in this hands-on comparison of causal ML methods and traditional models, using simulations where the true effects are known - Philipp Probst | I shared this one last week, but the link was broken. Here's the working link! |
| | |
| | Instead of programming how a mechanical watch should work, Trevor Dwyer modeled its parts as geometric shapes and let SQL figure out how they interact. The result is a working 34-part watch that runs in SQL. Here's how it works, along with a great exploration of spatial databases, computational geometry, and the limits of SQL Server. Includes a link to the code. Trevor Dwyer |
| | |
| | Sometimes a little randomness helps an algorithm find a better solution. That's the idea behind e-means, a new approach to k-means clustering that uses small, noisy updates to avoid getting stuck in poor solutions. The result is faster clustering that often produces better results. This is a great post that uses interactive examples to show how it works and why. Swann Bessa |
| | |
| | Frank Chu tried reproducing the impressive results in the post above and got a very different answer. Instead of a 10x speedup, he measured 1.5x. His investigation into the difference is a great case study in why benchmarks are tricky, implementation details matter, and how reproducing research can teach you as much as the research itself. frank chu |
| | |
| | A model can have a great average score and still fail badly on certain cases. This post revisits a cell-nuclei segmentation model that performs well overall, but fails badly on specific cases. It's a great example of model evaluation, with reproducible code in R and Python. Data Science Guts | Michał Maj |
| | |
| | With so many new frontier models arriving, it's getting harder to know which ones are best at real work. Mercor's APEX-Agents leaderboard compares how well they handle complex assignments in banking, consulting, and law. Google's Gemini 4 models currently lead, and Tencent's Hy 4 debuts on the top 3. Explore the leaderboard. // sponsored |
| | |
| | | | Voronoi polygons divide a map into regions based on the nearest point, which makes them useful for things like estimating service areas, assigning customers to locations, and mapping geographic coverage. This tutorial uses postal code data to show how to build Voronoi polygons, handle messy boundaries, and check the results. One World | Hernangómez |
| | |
| | This is a clever way to reduce geographic data from two dimensions to one. By using a Hilbert curve to turn coordinates into a single dimension, the resulting visualization shows how conflicts spread across regions and through time. It's a great example of how transforming data can reveal patterns that conventional maps miss. Richard Vale |
| | |
| | | AVA gives AI agents a toolkit for exploring and visualizing data. They can connect to files and databases, inspect data quality, run queries, suggest questions worth exploring, and generate charts to help answer them. It's open source, works with multiple execution engines, and lets you inspect the queries behind the visualizations. GitHub | AntV AI Visualization Team |
| | |
| | | In 1996, IBM needed a supercomputer to challenge the world's best chess player, but with today's software, an ordinary desktop from that same year might have beaten Kasparov. This is a great post that shows how better algorithms, not faster hardware, could have provided most of the gains. There's even a link here that lets you play against a Pentium yourself. Nicholas Carlini |
| | |
| | | This week's issue sponsored by: | | |
|