Lead story / research

Meta uses AI to help produce 10 health papers in four weeks

Meta disclosed ten health-research manuscripts on October 14 from a four-week, human-led experiment. Medical researchers used AI to generate hypotheses, write and execute code, and collaborate on drafting papers. The system ran on Muse Spark 1.3 and Muse Wonder inside an internal software framework, with access to public data and a licensed database of five million de-identified electronic health records. Those records were retrospective, not collected through a new clinical trial.

Researchers retained responsibility for scientific decisions and conclusions, including choices about patient groups, variables, methods and interpretation. Each manuscript received feedback from peers in the research cohort, analysis review by a biostatistician, and summary reviews from contracted independent experts. Those published reviews were not blanket endorsements: they exposed questions that the AI-assisted workflow and large dataset could not settle.

One paper examined heart-rhythm diagnoses after infective endocarditis, an infection of the heart’s inner lining and valves. Meta’s summary reported a one-year risk of seven percent, versus four percent among patients with bloodstream infections without endocarditis. Penn State cardiologist Michael Farbaniec praised the methods but cautioned that differences in cardiac monitoring could explain the result: more monitoring can uncover more diagnoses. He described the finding as hypothesis-generating, not practice-changing.

Other papers showed similar limits. A behavioral-health emergency study found worse outcomes among patients who left without being seen, but lacked important details about illness severity, diagnoses and reasons for departure. A diabetes study found that documented remission ended more often through restarting treatment than rising blood sugar, without establishing why treatment resumed. Its narrower recommendation was to report those outcomes separately. The experiment produced research leads; it did not establish, on its own, a reason to change patient care.

launch

ArtCraft releases 13 free alternatives to Adobe and Microsoft apps

ArtCraft has released 13 free, open-source apps targeting core creative, document, presentation and spreadsheet functions. Engineer Brandon Thomas says he built them with Claude Opus 5.5 and Rust using public file-format specifications. An intellectual-property scholar questions whether the claimed clean-room process can be verified. Users must weigh subscription savings against setup work and unproven feature parity, security, privacy and maintenance.

earnings

OpenAI UK’s average staff compensation reaches £2.3 million, mostly in shares

OpenAI UK’s 110 staff shared £255 million in compensation in 2025, averaging £2.318 million per person. Shares accounted for £168 million, compared with £87 million in salaries, so the average is not a cash-salary figure. The subsidiary reported £293 million in turnover, more than triple the prior year, and £20.3 million in profit. These are UK figures, not worldwide totals, and the Guardian reported substantial differences between employees’ packages.

launch

Sakana AI lets chatbot users see what it remembers and when it uses it

Sakana AI’s My Fish feature makes Sakana Chat’s saved memories inspectable on web and mobile. Users can trace individual facts to their source conversations, review reads and writes, and see which memories were combined in a reply. Memory remains on by default, with an option to disable it in settings. More accurate selection and updating are future goals; this release exposes the memory already at work.

Also worth knowing

•  Anduril opens employee share sale at roughly $90 billion valuation

•  Google DeepMind’s AI solves nine open math problems in peer-reviewed study

•  Google launches AI document reviews with two free checks a week

•  Block says AI saves 600 security analyst-hours a quarter

•  eBay’s terms require its permission for AI shopping agents

•  California will require labels for prominent AI ad performers in January 2027

•  OpenAI faces lawsuit seeking payouts for children’s ChatGPT data

•  Anthropic adds Claude activity tracing without exporting chats by default

•  OpenAI alleges former employee presented secret product plans to Apple

•  YouTube Music labels AI songs, but AI use alone won’t block monetization

•  Workers fearing AI career loss report more workplace misconduct in study

•  OpenAI will retire five GPT-4 API versions on January 14, 2027

Daily tool drop

6 AI tools worth knowing today

The Internet Had a Point

Meme from the stashAI Needs Adobe to Lose

Reader check-in

Help shape tomorrow's briefing

One click tells us what to keep, improve, or tighten.