I hear: grinding resin, distorted loons, slide whistles, reversed organs, digital pianos with dying batteries, payphone calls from the Mariana Trench, voices autotuned to raindrops, an unsure violin, dirges, lullabies, french chansons sawed in half.
SO, by SO (Markus Popp + Eriko Toyoda).
Soft melodies, nostalgic and optimistic, high sheen. The music wants to lull you and wake you. Between the two is a beautiful place to be.
Pulled every notice off Grants.gov from 2010 through April 2026 — 76,754 opportunities — and counted how often the phrase "artificial intelligence" shows up in the announcement text. The share is flat through 2019, drifts up through 2024, then jumps sharply after the January 2025 inauguration.
Share mentioning the phrase climbs from ~0% in 2010 to ~10% by early 2026. Total opportunities per year stay between 4k–6k throughout.
The bars are a sanity check: volume of grant notices hasn't collapsed, so the rising share isn't an artifact of a shrinking denominator.
Bartleby is a fallback for Zotero when translators fail, enriching PDFs with structured metadata, tags, and an `[AI Enriched]` audit trail.
Bartleby the Librarian for Zotero 7.
Bartleby sends the first 9,000 characters of a PDF or webpage to OpenAI and attempts to conform the result to Zotero's item schema—distinguishing a Statute from a Book, a Law Review Article from a Report. It also adds tags, restricting the model to tags already in your library.
Bartleby builds a parent item and fills bibliographic fields.
In use, a progress bar appears with the tagline "I would prefer not to..." before Bartleby parses the text, builds a Parent Item, and stamps the record. A recursion guard ensures it doesn't loop on its own output. It does the job—reluctantly, but well.
Update, July 2026 — v2. Auditing 360 enriched items turned up a quiet failure: for statutes, cases, bills, and hearings, Bartleby extracted the title correctly and then dropped it. Zotero stores a statute's title as `nameOfAct` and a case's as `caseName`, and the old field check tested the base name, silently discarding everything but `shortTitle`. Tags had never applied either, in any of those 360 runs. v2 is a rewrite with tests:
Field mapping resolved through Zotero's base-field table, so legal item types keep their titles and dates
A preview of every proposed change before anything is written—apply all, fill empty fields only, or cancel
Paste a front page from the clipboard, or keep it in a `bartleby:` child note, for scans whose OCR is mangled
Bot-check pages are detected and refused instead of being catalogued as the document
OCR for scans with no text layer, via `ocrmypdf`, opt-in because it costs about 40 seconds
Identifiers mined from the filename and URL—`s00253-008-1698-5.pdf` is a Springer DOI, `book_9780262374095.pdf` an ISBN—and resolved against CrossRef, OpenAlex, and OpenLibrary, skipping the model entirely
Files pulled from Zotero storage on demand, which turned out to matter: only 41 of my 5,948 imported PDFs were actually on disk
Duplicate warnings, and stamps written only on success, so a quota error no longer lands in the item's Extra field
