Daily field notes
Ephemeris.
A charred scroll read whole after two thousand years — and the document AI, on-device models, and dev tooling worth your morning.
A scroll sealed since 79 AD, read end to end.
The Vesuvius Challenge unrolled PHerc. 1667 without ever opening it — synchrotron X-ray tomography at the ESRF, then neural nets trained to spot ink one shade off the carbon it sits on. About 1.4 metres of Greek surfaced: a Stoic treatise on ethics, the first Herculaneum papyrus read whole. The scans, models, and code are released openly.
Documents come back as structure, not a wall of text.
Mistral OCR 4 returns each page as blocks — bounding boxes, block types, per-word confidence, markdown — across 170 languages and PDF, DOC, and PPT. Pass a JSON schema and the output is shaped to drop straight into a RAG or agent pipeline. It ships as one container you can run on your own infrastructure, so regulated data never leaves the building. $4 per 1,000 pages, half that batched.
- Languages
- 170
- Per page
- $4 / 1k
- Output
- boxes + md
- Schema
- JSON-typed
- Deploy
- self-host
The on-device model that loads only the weights it needs.
Apple's third-generation Foundation Models put two models on the phone: a 3B dense Core and a 20B sparse Core Advanced that activates just one to four billion parameters per request. Instruction-Following Pruning keeps the full model in flash and pulls only the needed weights into RAM; Quantization-Aware Training compresses the rest without losing accuracy. The recipe is how a 20B model fits where a 3B one used to.
…of which only 1–4B activate on any given request — sparse by design, flash-resident by default.
Errors that arrive with a fix already written.
Next.js 16.3 aims the dev loop at the agent reading your terminal. Build and runtime errors now come with paste-ready fix prompts, first-party Skills teach coding agents the framework's own conventions, and new tooling keeps an agent's view of your project in sync as it edits. The headline feature isn't a page — it's the handoff.
How a team taught its agents taste.
Vercel wrote up how it gets agents to produce product design that survives review. It treats taste as a pipeline — the same scaffolding you would give a junior, written down: Skills that encode the house style, lint rules that fail bad UI the way they fail bad code, evals that score output, and a human feedback loop that keeps tightening the screws.
Skills
House conventions encoded as agent-readable instructions, not tribal memory.
Lint Rules
Bad interface choices fail the build the same way bad code does.
Evals
Output is scored, so "looks right" becomes a number you can move.
Feedback
Human review feeds back in, tightening the loop each round.
A swing at the common cold, fully funded.
Intercept is a $500M nonprofit — backed by Stripe, the OpenAI Foundation, Anthropic, and Bill Gates — taking on colds and flu from two directions. One is broad-spectrum prevention: vaccines, pills, and nasal sprays that hit many respiratory viruses at once. The other is the room itself: filtration, far-UVC, and antimicrobial vapours to cut transmission in shared spaces. It runs on a four-to-seven-year research horizon.
one nonprofit · two fronts · a 4–7 year horizon
Prevent the virus
Broad-spectrum vaccines, pills, and nasal sprays aimed at many respiratory viruses at once.
Clean the room
Filtration, far-UVC light, and antimicrobial vapours to cut transmission in shared spaces.
That's today.
Six picks for 27 June 2026, read off the public feeds so you don't have to.