| | ELT with Snowflake is one of the most common data engineering workflows and Airflow makes it more powerful than ever. In this live webinar, Astronomer walks through an end-to-end Snowflake workflow with Airflow 3, including data-quality checks, SQL best practices, and Human-in-the-Loop integration.
*Message from this week's sponsor, Astronomer |
| | |
| | | Ethan Mollick gave Anthropic’s new Claude Fable model a few ambitious projects, then watched it spin up agents, research thousands of data points, write code, test the results, and return something close to finished. The big shift is how much the human role is starting to look like commissioning the work rather than steering each step.
One Useful Thing | Ethan Mollick |
| | |
| | | Large language models can solve arithmetic problems, but they don’t calculate the way people do. They don’t have scratch paper or column arithmetic, only tokens, vectors, and matrices. This beautifully designed interactive explainer looks inside that machinery to show how models represent numbers, where their shortcuts break down, and what it takes to recover a calculation from their internal state.
Alvaro Videla |
| | |
| | A feature can look valuable by standard importance metrics and still make a model worse. This post walks through a target-encoding feature that appeared promising, then quietly hurt generalization at the noise floor. It’s a useful reminder that strong diagnostics are not the same as better predictions.
Flyback Blog |
| | |
| | | | Panel-data models can run even when the useful variation is painfully thin. This post makes the case for plotting treatment patterns, missingness, and the cells that actually contribute to estimation before trusting the results. It’s causal-inference focused, but the habit applies broadly: look at the data before the model does.
Yiqing Xu |
| | |
| | plotnine gives you ggplot2 in Python, but the charts are static. This library from Joseph Barbier adds tooltips, hover effects, and click handlers directly in the aes() function, then exports to standalone HTML. The whole API is three functions, and there's even an llms.txt file so coding agents can use it.
GitHub | y-sunflower |
| | |
| | | The traditional bootstrap works well until a resample drops the handful of observations you really needed. fwb is an R package that uses weighted versions of the full dataset instead, making bootstrap estimates more reliable for rare outcomes, sparse categories, and other awkward cases.
Noah Greifer |
| | |
| | | Good statistical analysis is rarely a straight line from data to model to answer. This collection of 15 papers looks at how statistical workflows actually work in practice: checking assumptions, exploring alternatives, diagnosing problems, and deciding what to do next.
Edited by Andrew Gelman, Aki Vehtari, Richard McElreath and Elizabeth Wolkovich |
| | |
| | Test-Driven Data Analysis is a new book about catching errors before they make their way into an analysis. It covers data validation, testing analytical pipelines and models, visualization mistakes, reproducibility, data drift, and more. The first chapters are available now, with additional chapters released over time. It’s free to read online and you can sign up to get notified when new chapters are released.
Nicholas J. Radcliffe |
| | |
| | | Anthropic has raised the possibility that Claude might have feelings or moral status. Ted Chiang follows that idea to its logical conclusion and the result is sharp, unsettling, and much more damning than the usual debate over whether chatbots are conscious.
The Atlantic | Ted Chiang |
| | |
| | | This week's issue sponsored by: | | |
|
