From advice to dosing decisions
-
Nº CXXXIV
- Date
- 03 Nov 2026
- Issue
- 134
- Stories
- Six
- Editor
- ARC
Agents stop advising and start acting: glucose control, self-run alignment research, chromatin and transcriptome in one model.
DiabAgent aims AI agents at blood-glucose control
DiabAgent puts a language-model agent (software that plans and acts over several steps on its own) inside a physiological control loop, treating blood-glucose management as a task the agent runs rather than a question it answers. Automated glucose control already exists in closed-loop insulin pumps, but those controllers follow fixed algorithms. The case for agentic control is that a reasoning system can weigh context a fixed controller never sees, including meals, activity, and how a patient responded last time. The 'toward' in the arXiv paper is honest. This is framing and early machinery, not a clinical result.
Research agent tunes alignments for complex prediction
An agent runs its own research on multiple sequence alignments, searching for ways to organize and pair them that improve predicted structures of protein-protein complexes. Alignment construction is one of the few levers left in complex prediction, and groups have tuned it by hand for years. In this arXiv work the tuning is the agent's job: it proposes alignment strategies and evaluates them against prediction accuracy. Agents are starting to propose method changes, not only execute pipelines someone else designed.
Single-cell model links expression and chromatin state
One model reads transcriptome and chromatin together, learning a shared representation across single-cell RNA and accessibility data so one modality can be analyzed in terms of the other. Most single-cell foundation models — general models trained broadly rather than built for one task — handle expression alone, leaving multiome data to separate integration tools. The bioRxiv preprint reports cross-modality analysis coming out of one pretrained model.
Tuned protein models find dual-target peptides
Protein language models retuned cheaply pick out antimicrobial peptides active against two targets at once, with only a small fraction of model weights adjusted. Training and test sets were split under strict homology cutoffs, so near-identical sequences can't inflate the reported performance.
Anima cuts clinic call waits by 75%
Clinic call waits dropped 75% at practices using Anima, which builds patient-facing tools on OpenAI's GPT-6 and GPT-Live, according to OpenAI's customer write-up. What changed is how quickly patients reach a clinic, not what the clinic decides.
OpenAI coordinates fixes as agents find critical bugs
Agents turned up critical vulnerabilities in open-source code, and OpenAI built Patch the Planet to route those findings to maintainers and incident responders ahead of disclosure. Much of the software research depends on is maintained by volunteers.
Reply with your discoveries. A human reads them. Forward freely.
|