Aaron Steven White

Aaron Steven White

@aaronstevenwhite.io

computational semanticist. into modular synths, film photography, and rum. @anglebas.in is my alter ego.

2.3K followers809 following833 posts

Aaron Steven White reposted

NEW: In March the White House released an FY27 budget request that zeroed out the National Science Foundation's directorate of Social, Behavioral and Economic Sciences (SBE). For months it was unclear what would happen. Now, with FY26 over, the damage to SBE is coming to light. My reporting:

science.org‘Complete devastation’: How NSF gutted its social sciences researchAgency body that funds sociology, linguistics, and archaeology studies in U.S. gave 90% fewer new grants this year

There's also been some improvements to the editor tooling. Treesitter and Pygments can now highlight indexed families, effects, instances, handlers, and resumption grades, and the language server now has checked hovers, completions, rename, symbols, and target-aware diagnostics.

Effect instances are lexical. Two instances of the same interface remain distinct effects, and `with instance` allocates a fresh local identity. A total handler must discharge that identity before it can escape its scope.

Effects can be open, so a computation can say which instances it requires while preserving effects supplied by its caller. `lacks` constraints rule out duplicate lexical instances when unknown.

Effects are now source-level interfaces that can take static parameters. An operation’s result may depend on the parameter, so one interface can distinguish requests whose answers have different indexed types.

Quivers now has a typed, executable IR for QVR that preserves indexed types, effects, handlers, and probabilistic operations after checking and is consumed by the interpreter and transpilers.

QVR now has indexed families, which means a constructor can refine an index, so an invariant such as exact versus loose resolution is represented in the value’s type.

QVR programs can now transpile to BUGS, Church, Edward2, Gen, JAGS, NumPyro, PyMC, Pyro, Stan, Turing, and WebPPL, and the checker will tell you when a transpilation isn't supported (usually because the QVR can't be preserved in the target).

Aaron Steven White reposted

We have a new release (v0.13.0) that introduces a bunch of new community management tooling. There is now one portable path from a governed definition change through consequence analysis, review, verification, signed release, resumable migration, federation, and exit.

github.comRelease v0.13.0 · idiolect-dev/idiolectAdded Community control plane for 0.13.0. The new idiolect-community crate and CLI lifecycle provide versioned idiolect.toml workspaces, role-aware maintainer/consent/vote/steward/hybrid governanc...

Super cool paper by my student @pkuchmiichuk.com that synthesizes Dylan Bumford and @simoncharlow.com's work on effect-driven interpretation with probabilistic dynamic semantics to understand how slack regulators like "exactly" and "approximately" work.

Pavlo Kuchmiichuk@pkuchmiichuk.com

New preprint: In Ukrainian, several approximating devices can stack on a single numeral, and the result conveys the same interval as one device alone. I argue that a numeral introduces the resolutions it can be read at, and that approximators operate on that set the way indefinites and definites do.

View full thread

This week, we got a notification from SBE that the award was in indefinite limbo. Basically, they're going to keep it active, rather than decline it, in case funding becomes available. But they stressed that that is quite unlikely.

I don't know how common this experience is for this seemingly last cycle for NSF Linguistics. But it's probably time to move into the acceptance stage: there's not going to be new funding for linguistics coming from NSF in the near term.

View full thread

For context, we submitted a proposal in January and had a relatively standard pre-recommendation meeting with the then-PO a couple months ago. The one where the PO asks you to get materials ready for them to include with their recommendation.

Stoked for this to finally be up. Think it came out pretty well.

Angle Basin@anglebas.in

Eigengrau is out 6 Nov 2026. The pre-release is up now on @subvert.fm and @bandcamp.com. Subvert: www.subvert.fm/angle-basin/... Bandcamp: anglebasin.bandcamp.com/album/eigeng...

Really excited about this project. You can now access effectively all of UniMorph, Universal Dependencies, Universal Decompositional Semantics, MegaAttitude, and various other linguistic resources in a unified format. And there's a lot more coming down the pike in the near future.

Layers@layers.pub

layers.pub now has an official data PDS at repo.layers.pub as well as a python client and CLI that you can read about at layers.pub/lairs/!

What I find particularly cool about this setup (beyond improved discoverability) is that it makes stand-off annotations a lot easier to work with, especially when multiple annotation layers in distinct datasets point to the same source.

Come work with CM on multilingual and low-resource NLP!!!

C.M. Downey@cmdowney.com

I'm recruiting Linguistics PhD students! Come work with me on all things multilingual and low-resource NLP! NLP background welcome, but commitment to scientific study of language is essential. My recent work has included ASR for field linguistics and compact LLMs for heritage languages

Aaron Steven White reposted

Anyone ever use the geometric Poisson distribution? If I've implemented it correctly in Stan, it's giving me better fits than negative binomial according to PSIS-LOO-CV with a big count occupancy model.

Training a semantic tagger/parser for ~30 datasets right now, and it is truly wild that, even if this ends up SOTA on all of them, this is at best something that gets shoved into an appendix nowadays.

Said sort of neutrally to positively, since SOTA-chasing on these sorts of systems was often quite silly. I generally only want these sorts of systems for downstream analysis anyway, so it's pretty cool how easy it is to build good ones now. But they do still take work.

Hypo is now out! It provides a rich metadata layer on top of @grain.social records, including gear and workflow abstractions as well as tools for photo annotation and search. If you shoot film, it also provides a stockpile tracker, shot logger, and development timer.

hypo.graycard.appHypoHypo is a tool for organizing and sharing your film or digital photography: the gear you shoot, develop, and scan with; the workflows that take a film roll from capture to finished scan; and the prove...