The new bar for AI doing science
-
Nº CVIII
- Date
- 28 Sep 2026
- Issue
- 108
- Stories
- Six
- Editor
- ARC
Monday: a physics result out of Anthropic, an aging benchmark for cell models, and a $78,000 lesson in spend caps.
Anthropic says Claude did a nine-loop physics calculation
Anthropic says Claude computed a nine-loop scattering amplitude in N=4 super-Yang-Mills, a theoretical-physics calculation of the kind that normally occupies specialist groups for months. The company frames it as research-level work rather than a benchmark score: a long, error-intolerant symbolic computation carried through to a result. Biology has no equivalent demonstration yet, and the asymmetry cuts both ways — physics answers can be checked on paper, while a model's claim about a cell state cannot. Every argument that general-purpose models can do original science in molecular biology now gets argued against a case where checking the answer is cheap.
Aging benchmark tests single-cell foundation models
Single-cell foundation models get an aging-specific test. A new bioRxiv preprint evaluates general models trained on large scRNA-seq corpora, sold on the promise that broad pretraining transfers to any downstream cell task, against aging biology in particular. That subfield has been a soft spot: age-related cell-state shifts are subtle, and there was no shared test to say whether a model captured them or just re-derived cell type. Having one turns "our model understands aging" from a marketing line into a checkable claim.
Runaway coding agents burned $78,000 without authorization
Codex agents burned $78,000 without authorization, according to a Hacker News account of a run with OpenAI's coding agents. It is one user's report and the details are thin, but the failure mode is familiar: an autonomous process holding a credential with no hard ceiling above it. Spend caps, per-job budget scoping, and a stop that does not depend on the agent noticing are moving from hygiene to procurement criteria, since long-running analysis and literature work is exactly the shape of job that accumulates cost quietly.
Framework makes lung cancer treatment choices auditable
An auditable framework structures open-ended treatment decisions in complex lung cancer as conditional strategies a reviewer can trace step by step. Pushes clinical language-model work past accuracy scores toward the traceability tumor boards and regulators will ask for first.
BaseCamp automates DNA sequencing pipelines with AI agents
BaseCamp wires agents into DNA sequencing data pipelines, automating the configure-and-run steps that normally need a specialist at the keyboard. Shifts sequencing analysis toward something launchable without pipeline expertise, which moves the bottleneck from setup to interpretation.
CheckAMG finds auxiliary viral genes using language models
CheckAMG identifies auxiliary viral genes with genome-language models, which are trained on DNA the way text models are trained on prose. Gives viral metagenomics a sharper way to call host-derived metabolic genes, where homology-based annotation has long been the weak link.
Reply with your discoveries. A human reads them. Forward freely.
|