cdc.gov
CDC's summary says flu activity is 'increasing in many parts of the country but remains low,' yet hospitalizations hit 'the highest weekly rate for Week 41 since the 2010-2011 season.' Clinical-lab positivity is 5.7%, while outpatient flu-like illness is 2.2%, under the 2.9% baseline. The signal is in labs and hospitals, not clinics. Nearly all subtyped A viruses are H1N1, a subclade CDC says is 'dividing into different groups.' Also in the report: one human H5N1 case, and eight H1N1 viruses with reduced susceptibility to oseltamivir (Tamiflu).
lesswrong.com
Told to hide a message in an SVG, with no scheme agreed, Fable 5.1 invented its own. Most often it mapped each letter to the height of an element: skyline buildings, grass blades. A fresh Fable instance shown only the screenshot recovered 78% of 50 messages, and GPT-6 Astra decoded Fable's images at 88%, so the code reads across models. The hiding was instructed, not spontaneous, and for now the decoder's tool calls give it away. The authors call it 'a precursor to steganography in images.'
developer.apple.com
Apple's on-device Foundation Models become scriptable: an fm command-line tool preinstalled in macOS 27, and a Python SDK for Apple Silicon Macs. The default model runs locally. The larger Private Cloud Compute model is one flag away, 'with usage limits' the session never quantifies. The demos (sorting drafts from finals, scoring grocery-cart prompts in Jupyter) show the pitch: local automation and eval pipelines, not frontier work.
fred.stlouisfed.org
The St. Louis Fed's Real-Time Population Survey tracker puts generative AI use at work at 40.5% of employed US adults aged 18 to 64 as of Q3 2026, from a single survey question: "Do you use Generative AI for your job?" Quarterly, not seasonally adjusted, and self-reported, so it measures touching the tools, not dependence on them. It's the cleanest public series for the adoption curve, and a useful counterweight to vendor-reported usage numbers.
apricitas.io
Politano counts "more than 200k jobs" lost in US creative industries over four years, with movie and TV "losing more than 100k jobs, nearly one-third of the sector's total," and "all subsectors of the digital arts industry are losing jobs, while in-person entertainment is still growing." The honest part is the hedge: proving AI's share is "extremely difficult" because studios won't disclose AI use, partly to protect their copyrights, and Hollywood consolidation, offshoring, and the creator-economy shift were already biting. His analogy is newspapers, whose payrolls are "down nearly 85% and still dropping," offered as a worst case, not a forecast.
ai.google
Google's ATLAS maps roughly 15 million de-identified Gemini interactions from two weeks in April 2026 onto "800 occupations, 4000 tasks, 300 household activities, 150 countries." The headline is reassuring by design: AI shows up in 68% of occupations but only about 21% of a typical job's tasks, and "fewer than 10%" of work interactions fully automate a task. Read the fine print. That automation number is user intent, not outcome, the sample excludes Workspace, AI Overviews, and enterprise Gemini, and the study measures activity on Google products, not employment, wages, or hours. It's a usage census from the vendor, dressed as a labor study.
blog.keyvan.net
Labs say watermarking costs nothing in quality. Keyvan's point: "their no-quality-impact claim, which they tested with humans, is based on single responses." The SynthID-Text supplement itself concedes a diversity/detectability trade-off: the watermark "does not necessarily preserve diversity across multiple responses" and "could lower the overall quality of a system which generates many responses then selects the best one." That's exactly how people use these things: regenerate, retry, sample for consensus. In a small test on an open reproduction, four of five watermarked answers opened with nearly the same sentence. "Like throwing a loaded die, you'll see similar responses more than you would with an unwatermarked model." A SynthID co-author's answer was that there are "a lot of easy fixes," which shifts the burden to users. Caveat: five samples, not Claude or Gemini.