AI's hunger for text has a physical cost: rare books are being cut apart and scanned in Amazon warehouses. Meanwhile, Google just paid $10M for Spirit Airlines' entire digital history.
New Zealand startup see1m just put 45 years of global daily rainfall behind a free interactive atlas. The World Rain Player is the firm's first public demo of its persistent data framework: instead of flattening 5 km cells into a single statistic, it preserves every day from 1981 to 2025. That means you can watch seasons move, compare years, and read each cell against its own long-term mean.
FlowingData points to a Washington Post interactive that turns America's march from balanced books to a $40 trillion debt into one long stacked bar chart. Cumulative revenue and spending run side by side, with animated callouts on the spikes that mattered: tax cuts, wars, stimulus, and the pandemic. It is the fiscal cliff, illustrated.
404 Media hid an AirTag inside a rare book it suspected an AI company would buy, then tracked it to an Amazon warehouse in Las Vegas. Workers there say their entire job is cutting bindings off bulk-shipped books and scanning them, destroying each volume in the process. Booksellers believe AI firms are working through ISBN lists to systematically scan every printed book. It is training data with a real, physical cost.
Google Research trained a deep network on 35,000 UK Biobank records plus a 677-adult cohort to estimate body-fat percentage, Apple-to-Pear fat ratio, and visceral fat straight from a standard smartphone photo. On insulin resistance prediction, PhotoScan lands close to DXA scans, the clinical gold standard that's costly and radiation-exposing. That makes early metabolic screening, not just BMI, feasible at population scale.
Prior Labs open-sources three pieces of relational learning software: RelArena-α, a unified framework that standardizes data loading, evaluation and tuning on RelBench v1 so methods can finally be compared fairly. Its TabPFN-Rel harness for the TabPFN-3 foundation model currently ranks first. The takeaway: flattening a relational DB into one table still beats specialized relational architectures on real tasks.
Automating "can you pull this for me" doesn't shrink the queue, it grows it. Incident.io's own agentic "data brain" exposed the real bottleneck: not SQL, but institutional knowledge that never shows up in INFORMATION_SCHEMA or a semantic layer. Text-to-SQL quietly bakes in judgment calls nobody documented, shifting the analyst's job from writing queries to validating them. The takeaway mirrors what a solo analyst's six-week experiment found: easy access begets more requests, so the role becomes framing the right question, not executing queries.
Bankruptcy auctions have a new customer: AI labs. Google outbid Mercor in Spirit Airlines' court-supervised sale, paying $10 million for the carrier's corporate remains, 100M+ emails, 500M Teams chats, 30M lines of code, models, and operational data. The internet was already eaten, so model trainers now harvest corporate graveyards to feed the next generation.
A new CC BY 4.0 panel tracks engineering momentum across 350+ venture-backed GitHub orgs in 15 sectors, refreshed weekly and queryable via JSON/CSV endpoints. The two headline signals, 14-day commit velocity and contributor growth, combine to explain a 3.4× Series-A lift across the 219 startup-period observations that back it. The dataset's thesis: engineering acceleration has historically preceded fundraise announcements by 21-47 days.
An author spent 99 days asking the same 16 questions to OpenAI Search, Gemini, and Claude, scoring every answer under a fixed rubric: 24,882 scored responses in all. The CC BY 4.0 release adds a 6,416-answer no-web control from 10 models, so retrieval effects can be separated from training data, plus a registry of measurement gaps. Rare longitudinal receipts on how consistently answer engines actually find and cite a source.
Historical Swedish handwriting turns OCR output into gibberish, so this release skips text generation entirely: pristine ePub transcriptions are aligned page-by-page with high-res manuscript scans, yielding 278 human-validated image-text rows of Söderberg's Doktor Glas. A 3-page sample is free on Hugging Face, while the full set is locked inside an air-gapped compute-to-data Docker sandbox so enterprises can fine-tune VLMs without ever downloading the images. Small set, clean alignment, weirdly clever security.
You’re caught up.
That was all 10 stories in the 21 August issue.