Claude Agents Found a Magnet Candidate Hiding in a 1999 Paper
The headline on Hacker News says agents discovered two room-temperature magnetic semiconductor candidates. The post itself, by Geby Jaff at Vals AI, is more careful than that, and the careful version is more interesting.
The target is a class called Luttinger-compensated magnets. Ordinary ferromagnets sort electrons by spin, which is what spintronic memory needs, but they leak magnetic field and switch slowly. Antiferromagnets have no stray field and switch about a thousand times faster, but their spins are unsorted. A Luttinger-compensated semiconductor would give you both: zero net magnetism with spin-sorted band edges. A 2025 study predicted two such materials, both of which lose magnetic order below room temperature.
Working with a team of Claude Opus agents that ran density functional theory simulations, Jaff reports two candidates. One was designed from scratch: YBaMnFeO5, with a predicted 2.35 eV band gap and magnetic order up to roughly 420 to 490 K. The agents then found the catch themselves. The design needs manganese and iron in a perfect checkerboard, and their simulations show that order falling apart around 950 K, below the temperatures used to make this kind of oxide. It may be very hard to make in useful form.
The second is the better story. KV[Cr(CN)6], a relative of Prussian blue, was synthesized in 1999 and measured as magnetically ordered up to 376 K. A 2008 study even plotted its spin-resolved states without remarking on what they implied. The agents flagged it as a Luttinger-compensated semiconductor, with a predicted gap near 2.1 eV and spin windows of 2.6 and 1.6 eV, against about 26 meV of thermal noise at room temperature.
None of it has been measured. The two simulation methods disagree about how much water in the crystal's pores weakens the effect, and the post says so. Top Hacker News comments bring up LK-99, which is fair. What deserves copying is the packaging: input files, raw outputs, a one-command checker, independent re-runs and a list of known caveats are all in a public repo. The real find is that the agents recognized what a 27-year-old sample was, something the literature had walked past, and then shipped the claim with its own audit kit.
Link: vals.ai/blogs/room-temperature-magnetic-semiconductors
Repo: github.com/spicylemonade/compensated-magnet-ledger
← Back to all articles
The target is a class called Luttinger-compensated magnets. Ordinary ferromagnets sort electrons by spin, which is what spintronic memory needs, but they leak magnetic field and switch slowly. Antiferromagnets have no stray field and switch about a thousand times faster, but their spins are unsorted. A Luttinger-compensated semiconductor would give you both: zero net magnetism with spin-sorted band edges. A 2025 study predicted two such materials, both of which lose magnetic order below room temperature.
Working with a team of Claude Opus agents that ran density functional theory simulations, Jaff reports two candidates. One was designed from scratch: YBaMnFeO5, with a predicted 2.35 eV band gap and magnetic order up to roughly 420 to 490 K. The agents then found the catch themselves. The design needs manganese and iron in a perfect checkerboard, and their simulations show that order falling apart around 950 K, below the temperatures used to make this kind of oxide. It may be very hard to make in useful form.
The second is the better story. KV[Cr(CN)6], a relative of Prussian blue, was synthesized in 1999 and measured as magnetically ordered up to 376 K. A 2008 study even plotted its spin-resolved states without remarking on what they implied. The agents flagged it as a Luttinger-compensated semiconductor, with a predicted gap near 2.1 eV and spin windows of 2.6 and 1.6 eV, against about 26 meV of thermal noise at room temperature.
None of it has been measured. The two simulation methods disagree about how much water in the crystal's pores weakens the effect, and the post says so. Top Hacker News comments bring up LK-99, which is fair. What deserves copying is the packaging: input files, raw outputs, a one-command checker, independent re-runs and a list of known caveats are all in a public repo. The real find is that the agents recognized what a 27-year-old sample was, something the literature had walked past, and then shipped the claim with its own audit kit.
Link: vals.ai/blogs/room-temperature-magnetic-semiconductors
Repo: github.com/spicylemonade/compensated-magnet-ledger
Comments