Video Preview
🎧 Podcast episode for paid subscribers only. Also available on Apple Podcasts and Spotify.
To listen to paid episodes in Apple or Spotify, link your Substack subscription via the show settings on those platforms (instructions inside the Substack app under Subscriptions → Podcast).
Table of Contents
Two Science papers and a billing fight, all in one fortnight
What RADAR showed once the telephone game is rewound
The 37,000-agent biotech and the drug Merck already paid for
The production scoreboard reads 67 to 4
Nobody gets paid to find more on a CT
FDA clears findings one at a time and RADAR brought 146
A second reader that leaves a paper trail
Who ends up owning the read
Abstract
RADAR, published in Science on Sept 17, 2026 by Alibaba DAMO Academy and the First Affiliated Hospital of Zhejiang University: 424,911 contrast abdominal CT exams, 146 findings, mean AUC 0.913 vs 0.776 for the next-best vision-language model, better than 23 of 26 radiologists, plus 10% sensitivity and about 30% faster reads when assisting.
Same day in Science: Stanford’s 37,000-agent Virtual Biotech. One week later: BCBSA tied an estimated $942 million of added spending to more-intensive hospital coding between 2023 and 2025, with AI-enabled tools identified as a contributor.
Scoreboard: 67% of provider revenue cycle teams run semi or fully autonomous agents vs 4% in clinical work (Bessemer and Bain, n=226). Imaging AI deployed at 90% of large systems, high success at 19% (JAMIA, n=43).
Blockers covered: 1,164 FDA radiology AI authorizations vs 2 Category 1 CPT codes for newer imaging AI; split technical and professional billing; per-finding FDA validation; a 73% vs 56% juror liability penalty when AI catches what the rad missed; integration and monitoring load.
Thesis: AI reaches production where one P&L owns both the cost and the benefit. For generalist imaging AI that owner is whoever employs radiologists at scale.
