The trick is in which state does the work. In Hopfield's 1982 model, recall dies at the moment the network slips into a spin glass state, because frustrated spins leave an energy landscape too cluttered to settle into the right valley [7]. Lev's group built its spin glass out of atoms and photons and got retrieval out of the frustrated state itself [8]. The condition that halts recall in Hopfield's model is the substrate this network retrieves from.
The price of a spin is worth stating plainly. Each spin is a Bose-Einstein condensate of 10,000 or more atoms behaving as a single super atom [11], positioned by laser tweezers inside a cavity formed by two curved mirrors [10]. Twenty spins therefore costs at least 200,000 atoms [16], all of them at extremely cold temperatures in a vacuum chamber [14]. The wiring is light: photons bounce thousands of times between the mirrors, making many connections among the atoms and driving each super atom's spin into a low-energy valley [12].
"Greater capacity than a traditional AI network of the same size" [2] is a comparison made at up to 20 spins [13]. At that size the claim is checkable and the physics is genuinely new. What it does not tell you is whether the margin widens, holds, or disappears at a thousand nodes, which is the only version of the question that would matter to anyone building memory hardware. The phys.org account I am working from reports no figure for how many patterns either network held, and none for recall accuracy [18], so the effect size here is a direction, not a magnitude.
The plasticity result deserves the same care its senior author gave it. Lev says the networks "adjust themselves in a way that is somewhat similar to how we believe our brains learn" [4], and the study describes the short-term plasticity as resembling synaptic change during learning [3]. The resemblance described here is behavioral; no shared learning rule with biology is identified in this account.
Forty-three years separate Hopfield's model from the first experimental quantum-optical spin glass, which this same team reported in 2025 [17][9]. Hopfield shared the 2024 Nobel Prize in Physics for the original work [6]. My read: this is a physics-leaderboard result, because storing twenty patterns is not a problem classical computing struggles with. What changed between 2025 and now is that this atomic-scale network, with couplings that adjust themselves, moved from proposal to measurement [1][3].
