Will Townes

Finding stable clusterings of single-cell RNA-seq data: biorxiv.org/content/10.110… #genomics

Some rules for numerically stable statistical computation: 1) Don't actually invert a matrix 2) Do everything on the log scale, the logSumExp function is your friend 3) If the sample size is small, regularization (Bayesian prior) may help

There is a hypothesis that birth order effects (on things like income and educational attainment) are in part respiratory pathogen effects: younger kids get more of them from their older siblings. This cool recent paper uses Danish administrative data to argue that this is true and a pretty large part of the story. (They claim 70% of the birth order effect on long-run wages.) Other work has previously shown that severe infections matter for long-run outcomes, and it's well-established that birth order matters, but I haven't until now seen anyone convincingly show that standard respiratory pathogens impose long-term costs on infant siblings. nber.org/system/files/w…

Here is a nice tutorial on independent components analysis (ICA). Unlike PCA, it tries to maximize the independence of the signals rather than the "variance explained". The signals are not restricted to be orthogonal. Fun info theory stuff too. arxiv.org/abs/1404.2986

malware found in the top downloaded skill on clawhub and so it begins

What are folks doing to maintain security with command line LLM tools like Claude Code/ Codex/ Gemini CLI? I understand you can ask it not to do certain things via settings, but isn't that just trusting it to do what you say? What can be done (straightforwardly) at the OS level?

@TheStalwart and the rest of my Claude Code for Absolute Dummies starter guide: jaswsunny.github.io/claude-code-gu…

Malaria deaths have halved since the year 2000 In the last year there have been breakthroughs for vaccines, transmission-blocking innovations, better surveillance — and new and improved drugs Step by step, we're slowly moving towards malaria eradication

Negative binomial regression and inference using a pre-trained transformer - nxn.se/p/negative-bin…

The Department of Statistics & Data Science at Carnegie Mellon University invites applications for a tenure track position at the rank of Assistant Professor starting in Fall 2026. cmu.edu/dietrich/stati…

Many of the proofs are similar to those of Vershynin's classic book "High Dimensional Probability", but students may find mine a bit easier to follow since I show all the steps and provide explicit constants.

Main result is that exponential tilting does not affect the tail behavior of the original distribution as long as it is subexponential.

Are you interested in concentration inequalities? Curious about whether there is a middle ground between subexponential and subgaussian bounds? Want to know how this all interacts with exponential tilting? Then check out my new paper recently published in TMLR (link below)!

New #wastewater monitoring dashboard from @ACDHS for illicit substances in Allegheny County, #Pennsylvania glad to see fentanyl on a downward trend. tableau.alleghenycounty.us/t/PublicSite/v…

Polled my PhD stat theory class to see how Bayesian they are, and about a fifth of them say there's a 10% chance that the 10th digit of √2 is 5, if you haven't checked yet. Online radicalization of the youth has gone way too far...

Thoughts on peer review: willtownes.github.io/blog/2025/08/2…

Anyone else having trouble connecting to github due to a "potential security issue"?

@asymmetricinfo Like an economic analogue of second law of thermodynamics

The absolute hardest thing to convince people of is that the optimal amount of fraud in a system is not zero. Obviously it would be ideal if there were no fraud, but at some point the cost of catching it outweighs the benefits