2025-08-25

links

papers

• Late to the party but Wasserman, Ramdas, Balakrishnan is a really neat, ingenious idea. Given a model with likelihood $L(\theta ;Z^n)$. We split the data into two halves $Z_1^{n/2},Z_2^{n/2}$, and define the split likelihood ratio statistic as

$$T(\theta )=T(\theta ;Z^{(1)},Z^{(2)}):=\frac{L({\hat{\theta }}^{(1)};Z^{(2)})}{L(\theta ;Z^{(2)})}$$

The confidence set for ${\theta }^{\ast }$ is given by $\{\theta \in \Theta \mid T(\theta )\le {\alpha }^{-1}\}$ is a valid $1-\alpha$ level confidence set; $T(\theta )$ is an example of an e-value that has nice properties such as composability [for more, check out the monograph linked on 2025-08-04 ] . This requires us to fit the model once and evaluate the likelihood over a grid of $\theta$ values. This reminds me a bit of weak-identification-robust methods in econometrics such as Anderson-Rubin, which similarly involve inverting a test statistic. Dey et al extend this idea to general ERM problems.

• Hodgson and Lewis build a model of consumer search that accommodates the insight that ‘near things are more related than distant things’, which they call spatial search. Given IO publication timelines I bet the first version was written during SVM times, but I think the paper’s key ideas are especially relevant in the age of LLMs, where chatbots make it easy to traverse latent space quickly; bullishly - perhaps this shows up in less greediness in recommender systems and higher overall welfare, bearishly - this will just descend into (fancy) advertising. I haven’t quite grokked where the identification in their empirical application comes from, which is fairly characteristic for elaborate IO models. Related, see Ursu et al on an empirical instantiation of Weitzman search [featured on 2025-05-09 ].

code, music

• omarchy is a very nice, user-friendly tiling-manager based setup built on arch and hyperland (OMakase ARCh HYprland). I put it on an old thinkpad t490 and it is blazing fast. The need to grow your own rice before ricing your desktop has always put me off arch, but now i can just steal dhh’s homework, which he graciously shared with the world. here are my dotfiles.

• scipp looks like a promising package for fast arrays with labelled dimensions.

• orbital is a funny idea - fit a model once and turn it into sql for deployment. Makes perfect sense for a lot of run-of-the-mill business applications where the model is updated periodically; I can imagine prediction instability across runs setting people’s hair on fire (this is true with vanilla model deployment too, but when you make things sql-able, these things will make it into dashboards, which open you up for a world of pain).

self-promotion

• A simple model of Online Platform Enshittification is a fun little model I sketched out to formalize the problems that plague modern online platforms. Comments/corrections/belligerent flames welcome.

• synthlearners got considerably leaner - the default installation now only uses the standard libraries + adelie for fast penalized regression [check out the benchmarks here] for both unit and time weights. The full installation is now optional - install with uv pip install "synthlearners[full]" @ git+git@github.com/apoorvalal/synthlearners.git.