Model releases stay quiet while the week's advances come from the lab and the sky
An eclipse-chasing campaign anchors the day, alongside Dyna-1's read on protein motion from missing NMR peaks and Anthropic's new mark on Claude's text output.
Scientists fly jets and balloons into the 12 August total eclipse to read the corona1
On 12 August, NASA-funded teams met a total solar eclipse in the air and the stratosphere. The WB-57 jet flew a four-camera multispectral imager at 50,000 feet, capturing at least 20 corona images per second. That stretched totality to nearly three minutes, against the 2 minutes 18 seconds an observer got on the ground.1 Balloon teams launched about 80 payloads from Iceland and six carrying 360-degree cameras from Spain, running from 18 hours before totality to 8 hours after. Their targets were the corona-heating problem, prominence formation, and the link from the corona to the solar wind.1 European groups observed alongside them: an AFP report names France's CEA, ESA's Solar Orbiter, and the Proba-3 formation-flying mission.2 This is an observing campaign, not a same-day result. The findings arrive once the data is reduced.
Dyna-1 reads protein motion from the NMR peaks that go missing3
Wayment-Steele and colleagues train Dyna-1 on a single assumption. The residues left unassigned in NMR chemical-shift data are the ones broadened out by microsecond-to-millisecond exchange, so absence in the spectrum is itself the label for motion. The training signal comes from roughly 10,000 proteins in the Biological Magnetic Resonance Data Bank, and the best model reads from an intermediate layer of the ESM protein language model.3 The authors curated a test set they call RelaxDB, 133 proteins with relaxation data. On it, they report that Dyna-1 also predicts exchange measured by independent NMR experiments, including the motions that drive enzyme catalysis and ligand binding.3 Independent write-ups from CBIRT and Kiin Bio describe the same result, though both predate the Nature paper: the work circulated as a preprint from early 2025.45
Anthropic starts marking Claude's text output, without saying how6
Anthropic updated its support documentation on 11 August: Claude now weaves an imperceptible watermark into the text it generates. The company says the mark survives copy-paste and may persist through some editing. Generated image files, in SVG, PNG, and JPG, carry signed C2PA provenance metadata, and models launched on or after 2 August mark their output at launch, with older models being retrofitted.6 Anthropic is careful about what a hit means: a detected mark shows only that content may have been processed by Claude, since the model may merely have edited existing material.6 What the company does not disclose is how the text watermark is built, or how much editing removes it. TechCrunch, reporting the change against the EU AI Act's transparency rules, notes that Anthropic had not answered its questions on removability by publication.7
Sources
- NASA Science · NASA science soars during the August total solar eclipse
- phys.org (AFP) · Total eclipse gives scientists a chance to probe the corona
- Nature · Prediction of protein millisecond dynamics from missing NMR assignments (Dyna-1), published 10 August 2026
- CBIRT · Dyna-1: unraveling protein dynamics with deep learning and NMR data (preprint-era, 5 April 2025)
- Kiin Bio Weekly · Dyna-1: learning protein dynamics (preprint-era, 6 November 2025)
- Anthropic Support · How Claude marks AI-generated content (updated 11 August 2026)
- TechCrunch · Anthropic says it will watermark text generated by its AI models