Washington rushes medical AI to the bedside
-
Nº XCIX
- Date
- 15 Sep 2026
- Issue
- 99
- Stories
- Eight
- Editor
- ARC
Regulators are outrunning the evidence, while a new lab promises to shorten biology's path to medicine.
U.S. health agencies rush medical AI into the clinic
U.S. health officials moved to accelerate medical AI into clinical use, with Medicare coverage decisions and FDA review both in the frame, per a New York Times report picked up on Hacker News. The reporting describes internal concerns about the pace being set aside as deployment timelines compress. That shifts where clinical-AI evidence standards actually get set: in procurement and reimbursement decisions rather than in trials and journals. What counts as adequate validation for a diagnostic or triage model is now a policy question with money attached, and the answer will shape which models reach patients well before the literature catches up.
Geodesic Intelligence launches an AI drug-discovery lab
Geodesic Intelligence launched publicly, founded by Quanquan Gu with the stated goal of building AGI (artificial general intelligence) for drug discovery and finding the shortest path from biology to medicine. Specifics beyond the mission are thin so far. What the launch marks is a category hardening: outfits that position themselves as discovery operators rather than software vendors, betting the model and the molecule pipeline belong under one roof. That framing is fast becoming the reference pitch for AI-native therapeutics, and it sets the bar competitors now have to answer.
Insilico and Liquid AI build a 400-task drug test
A 400-task drug-discovery suite pairs Insilico Medicine's MMAI Gym for Science with Liquid AI's Liquid Foundation Model, a compact general-purpose model, in work accepted to the EMNLP 2026 industry track. The setup uses chemistry-native tokens, feeding molecules to the model in chemical notation rather than as ordinary text. Putting 400-plus tasks under one shared test gives the field a common yardstick for drug-discovery models, where claims have mostly rested on hand-picked case studies and internal numbers nobody else can reproduce.
MCP template lets agents call tools without credentials
Agents get tool access without ever holding the credentials, under a template posted to Hacker News for MCP (Model Context Protocol, the spec that lets agents talk to tools). Credential scoping is heading toward vendor-criterion status for anything touching patient or trial data.
Language model grafts antibody CDRs and designs new ones
Generative model grafts CDRs into antibody frameworks and designs new ones de novo, using alignment-driven sequence generation, in a bioRxiv preprint. Humanization by grafting has stayed largely manual and rule-based; moving it into a learned model automates another link of the antibody pipeline.
DNT models both copies of the genome at once
DNT models diploid genomes keeping both parental haplotypes separate instead of collapsing them onto one reference, per a new bioRxiv preprint. Most genomic models flatten that variation away; phase-aware representation makes allele-specific effects addressable by the same modeling stack.
CausalArena tests whether big models find causal structure
CausalArena scores causal discovery methods against general-purpose models, giving the field a shared test for whether large models recover cause from correlation. Causal claims are where most translational work lives or dies, and they have had no common scoreboard.
Paper proposes giving agents persistent drives for alignment
Artificial Id argues that long-running agents need persistent internal drives to stay aligned across sessions, in an arXiv proposal. As autonomous experiment loops stretch from hours to weeks, alignment over time becomes a design constraint rather than a prompt.
Reply with your discoveries. A human reads them. Forward freely.
|