A compound that researchers now think could be relevant to spintronic memory has been sitting in a 1999 chemistry journal, apparently overlooked by anyone looking for spintronic materials. No published paper connected it to that application, reportedly because no one was reading both bodies of literature at once.
Vals AI reportedly deployed more than 90 Claude agents to comb scientific literature and flag KV[Cr(CN)6] as a candidate Luttinger-compensated semiconductor, a connection the original paper does not appear to discuss. The story is less about the material itself and more about what kind of tool it takes to see what is already there.
The Librarian Who Found Gaps Between Journals
Don Swanson’s 1986 insight was direct: important knowledge can hide in plain sight, split across literatures that never cite each other.
His paper “Undiscovered Public Knowledge,” published in The Library Quarterly, argued that two bodies of research could each be true, publicly available, and not connected by direct citation or discussion. His case study was fish oil and Raynaud’s syndrome. One literature tracked fish oil’s effects on blood viscosity and vascular function; a separate literature described the blood and vascular abnormalities associated with Raynaud’s.
Swanson found no direct link in the literature he examined at the time. He linked them through analysis, and a clinical trial later supported the hypothesis, according to the University of Chicago’s account of his work. That outcome validated the hypothesis-generation method, not literature search as a substitute for clinical evidence.
Swanson and Neil Smalheiser later built Arrowsmith, a system designed to search MEDLINE for shared intermediate concepts between disconnected article sets. Arrowsmith surfaced candidates; humans evaluated them. That division of labor is the through-line connecting 1986 to now.
Ninety Agents, One Previously Reported Compound
In 1999, a paper in the Journal of the American Chemical Society introduced KV[Cr(CN)6] as a Prussian-blue-type molecular magnet and then, largely, moved on.
Stephen M. Holmes and Gregory S. Girolami reported the compound with a magnetic ordering temperature of 376 K, roughly 103 degrees Celsius, notable at the time for retaining magnetic ordering above 100 degrees. Its proposed electronic properties were not part of that original discussion.
What the Vals AI team reportedly found is that density-functional calculations suggest the compound may also exhibit a band gap of approximately 2.1 eV and spin-polarized band edges. Predicted carrier windows run roughly 2.6 eV for holes and 1.6 eV for electrons. These are theoretical predictions, not measured properties.
The next step, according to the available account, is reproducing the material and testing it experimentally. That step matters more than usual here: different computational treatments used in the work do not fully agree on how water in the crystal structure behaves, which introduces uncertainty that only lab work can resolve.
The Swanson parallel is genuine but limited. The proposed electronic significance of this compound emerges from connecting an older material record with concepts from a different research context. Those concepts include compensated magnetism, semiconductor band structure, and spin-selective transport. The original 1999 literature does not appear to emphasize any of that.
What the AI agents did is better described as computational reinterpretation rather than the causal chain Swanson inferred between two biomedical literatures. The accurate label is a reinterpretation of a previously reported material, not a validated technology. No working device has been demonstrated.
What Comes After the Hypothesis
As AI systems generate candidate relationships faster than labs can test them, the bottleneck may shift from finding connections to verifying them.
Open inputs, reproducible calculations, and explicit caveats should be treated as important criteria for judging whether a machine-generated hypothesis is worth a researcher’s time. Swanson demonstrated the method through careful literature analysis and persistence. The Vals AI effort suggests that agent-based search and calculation can be run at a meaningfully larger parallel scale. For KV[Cr(CN)6], the experimental test is still ahead.
