Found first: a primary source the press has not covered yet.
Researchers at PHAI Labs, National University of Singapore, Shanghai Jiao Tong University, and the University of Edinburgh have published JEPA-Anything, a single architecture that learns world models across seven domains including molecular dynamics, clinical trajectories, and weather. An intervention nominated by one of the model's learned latent factors was subsequently confirmed in cell co-cultures, patient-derived organoids, tumor fragments, and mice.
What the source says
JEPA-Anything is built on Orthogonal Predictive Factorization (OPF), which extends joint-embedding predictive architectures to decompose dynamics into interpretable latent factors without domain-specific priors. The seven evaluation domains are vision, biology, clinical trajectories, control, molecular dynamics, physical fields, and weather. The method improves on all 10 matched dynamics tasks compared to JEPA baselines, reduces single-intervention prediction error on Interventional Pong by 34.8%, and achieves the lowest one-step and 100-step molecular errors on all four tested molecular systems. On an orbital physics task, recovered latent modes produced a Keplerian scaling slope of -1.4991, against the theoretical value of -1.5. The biological intervention nominated by the model's biology factor received experimental confirmation across four systems including mice.
Why it matters
Building a single predictive architecture that transfers across physics, molecular dynamics, clinical sequences, and vision without domain-specific tuning has been a persistent target; this paper is the first to demonstrate it at the seven-domain scale with competitive numerical results on each. The orbital recovery is notable: a fitted slope of -1.4991 against the theoretical -1.5 suggests the factorization is extracting physically meaningful structure. The wet-lab confirmation is the standout result. AI-nominated biological interventions rarely clear a single experimental system, and confirmation across four, culminating in an in vivo mouse result, is the kind of bridge between computational prediction and bench science that both communities will want to scrutinize. Practitioners building cross-domain predictive models now have a concrete factorized baseline to work from.