An Opinionated Tooling Guide

Statistics and Data Analysis: Overall: use R. It has the largest ecosystem of statistical packages.

• For Bayesian modeling, Stan has bindings for everything, but it’s easiest to use from R, and the rest of the toolkit (brms, loo, bayesplot, etc) doesn’t have an equivalents outside of R. Turing.jl and PyMC3 are substantially slower, and Numpyro is less mature (although it can be faster on a GPU).
• For DiD studies, R has did and Python has diffindiff and moderndid, but Julia doesn’t have anything comparable.
• For TMLE, R has the tlverse and ltmle. Python and Julia don’t have anything close.
• For visualization, R has ggplot2 and ggpairs, Julia has AlgebraOfGraphics and Python has Seaborn and Plotnine. They’re all about as good as each other.
• The data manipulation, R has the tidyverse, Python has polars and Julia has Dataframes.jl. I actually find DataFrames.jl to be the nicest of the bunch (with DataFramesMeta). Polars makes you learn a whole separate DSL instead of using the base language, and much of the Python ecosystem uses Pandas, so you constantly have to convert back and forth. Pandas itself its a nonstarter - generally slow, verbose, and hard to use.
• For geographic queries, R’s sf and Julia’s GeoStats are lovely. Python’s GeoPandas, being based on Pandas, is much more annoying to use.
• For frequentist statistics, Python’s statsmodels and Julia’s GLMs, CovarianceMatrices.jl, MixedModels and Survival packages give you much of base R and the lmer, sandwich and survival packages.
• For Gaussian Process modeling, Python’s GPytorch is far faster and more feature-complete than anything else. Julia’s AbstractGPs has a nicer interface and is easier to extend, but it’s slower and lacks the approximate inference algorithms necessary for large datasets.

Neural Nets

• There’s a plethora of SAAS companies that promise easier experiment tracking. But in my experience, the easiest option is just to use git worktrees. Each experiment should live in a separate branch with a README documenting exactly what’s being tested. This makes it easy to fork and cherry-pick pieces of different experiments.
• For training, use Jax with equinox.
• For visualization, use Tensorboard.

Other Programming

• Use Julia for anything numerical.
• Use Python with the fire package, the pytest-watch test runner and the uv package manager for scripting.
• Use awk for text processing and jq for JSON processing
• Use Rust if efficiency matters.
• Otherwise, use Haskell with ghcid.

Notebooks: I currently use Pluto for Julia, Marimo for Python, and RStudio (with qmd files) for R. LLMs work best on Markdown variants rather than structured notebook formats, which is fine for Marimo and R, but makes Pluto ackward. Currently, I use a hacked together version of jupytext for Pluto that syncs an associated qmd file with each Pluto notebook.

Static Site Generation Use Quarto. Compared to Jekyll, niceties like latex math, computational notebooks and bibliography management are built in.

Editors

• Kakoune is lightweight and but has full lsp support. On linux, it’s my perferred editor. But it assumes that window management functionality will be handled by your window manager, which isn’t really a thing on other platforms.
• On a Mac, I use Zed.

LLMs Claude Code deserves all the hype it gets. Nothing else comes close.

Other

• Fzf makes looking at file trees and ressurecting previous shell commands much easier.
• As much as it pains me to recommend an electron app, Typora is the best word processor I’ve ever used.
• Use Xmonad on linux
• Zulip is much nicer than Slack.
• I write down everything I might want to remember later in Anki. It’s effectively a notes app for me.
  
• Zotero has remained fast for me despite adding gigabytes of papers to its library.
• Pixi is great for projects with dependencies in more than one language.