Machine-learning weather models have reached the point where they rival operational numerical prediction in the medium range, according to an arXiv preprint. The same models, however, are not reliable when integrated freely for years: they drift, blow up, or lose their variability. That makes them poorly suited for climate emulation, which requires stable behaviour over decades or centuries.
The preprint introduces Rescene, a method for producing stable century-long climate emulation from a frozen weather model. The name and abstract suggest the approach keeps the underlying weather model fixed while adding something to make long runs stable, but the abstract does not specify the mechanism.
Because this is a single source and the abstract is truncated, there is no independent comparison or detail to report. The claim is limited to what the authors state in the abstract.