AI stops assisting, starts discovering
-
Nº CXVII
- Date
- 09 Oct 2026
- Issue
- 117
- Stories
- Five
- Editor
- ARC
An all-sky UV map, trillion-molecule catalogs, and agents that now have to show their sources.
Claude Science helps map the whole sky in UV
Claude Science produced the first complete map of the sky in ultraviolet light, according to a research post from Anthropic describing the project. Claude Science is Anthropic's science-focused offering, and the UV map is the result the post is built around. Astronomy, not biology. But the shape of it travels: a frontier model pointed at a large observational corpus, returning a finished artifact a discipline did not have, rather than a score on a shared test. That changes what a frontier lab has to put on the table when it says its models do science, and biology sits on far more unfinished archival data than the sky does.
Benchmark tests AI agents on molecule prediction tasks
Autonomous LLM agents get benchmarked across molecular prediction and optimization tasks in a new bioRxiv preprint, testing both families together rather than one property set at a time. The pairing matters: prediction asks whether software can read a molecule, optimization asks whether it can improve one, and most prior evaluations pick a lane. A shared test spanning both gives molecular-agent claims a reference point the field has lacked, and turns "our agent does drug design" into a checkable statement instead of a pitch.
New design method searches trillion-molecule chemical spaces
Generative design goes synthesis-aware at trillion-molecule scale in a bioRxiv preprint on automated drug discovery, generating candidates constrained to reactions and building blocks that exist rather than screening for synthesizability after the fact. That distinction decides whether a designed compound can be ordered or only admired. Folding synthetic accessibility into generation at this scale narrows the gap between design output and material in a vial, which is where most generative chemistry still burns weeks.
ARGUS grounds variant interpretation in cited evidence
ARGUS constrains variant calls to cited evidence, an agentic framework (software running multi-step analysis on its own) for interpreting single-nucleotide variants in the regulatory genome. Noncoding interpretation is where clinical sequencing stalls, and evidence-bound output is what makes an automated call auditable rather than merely fast.
OA-MAP forecasts knee osteoarthritis and shows its evidence
OA-MAP forecasts knee osteoarthritis progression from multimodal inputs, with multiple agents attaching supporting evidence to each prediction. Like the variant framework above, it treats grounding as architecture rather than an add-on, which is fast becoming the entry price for anything operating near a clinical decision.
Reply with your discoveries. A human reads them. Forward freely.
|