See attached

This paper presents an opinion on how climate models can be significantly improved by merging traditional model development (improved parametrizations and higher resolution) with AI techniques. I think it is broadly in line with what many operational centres are already trying to do, i.e. augment models with AI instead of replacing them entirely. It is nonetheless valuable to document this in the literature for those who aren't already involved, and therefore I think the paper is an important contribution. It's also well written and I enjoyed reading it.

I have several comments below, which I think the author should consider in producing a revised manuscript.

• L56 - typo - "climate Turing test"
• L85 - suggest re-wording this slightly - even at kilometer scale, most parametrizations (radiation, cloud macrophysics, cloud microphysics, turbulence, shallow convection, orographic drag) are still required - in fact it's only really deep convection which can plausibly be removed at this scale!
• L105-108 - similar to TOA, the surface precipitation is also required to be accurate to close the water budget, which is critically important for the long-term behaviour of climate models, especially in fully-coupled Earth-systems. Might be worth mentioning this.
• L195 - this is quite a controversial statement that I'm not sure many people would agree with (I certainly don't) - progress undoubtedly is being made, cloud and convection schemes now are measurably better than 10 or 20 years ago when the cited papers were published. I would suggest rephrasing this - what I think is true to say is that progress is not as quick as we would like. The following paragraphs discuss the method by which most groups are making progress in this area (and have been for many years), so it feels slightly disingenuous to present these as new solutions to the problem, when I think they have been known as the solutions by parametrization developers for many years. The issue is that doing as suggested is hard, hence takes a long time.
• L252-256 - whilst I agree with the point being made here about developing scale-aware parametrizations, I'm not really sure the model results presented in Figures 1 & 2 are really a compelling argument for it. All major NWP centres run kilometer scale models without convective parametrizations and see huge improvements in NWP skill from doing so. Super-parametrized climate models have similarly shown measurable improvements in model skill relative to traditional parametrizations. Therefore to state on the basis of one bad model that this "approach has not achieved the anticipated success" seems like cherry picking to support the argument, when actually the weight of evidence shows that turning off the deep convective parametrization is better than including it, although could undoubtedly be improved further by scale-aware approaches. I suggest rephrasing this.
• Sect 4 - whilst I agree with what is being said here, it also neglects that there are important aspects of the climate which are not driven by small scale turbulence, e.g. land-sea contrasts, orography, land-surface type, SST pattern. These aspects become better resolved at higher resolution, and in turn leading to improvements in climate model skill which are not simply governed by the energy spectra. This would be worth mentioning.
• L337-341 - this statement is the one that worries me most in the paper. The whole point of emergent constraints is that they are emergent, i.e. they appear in climate models not because they have been programmed to be there, but because they arise as a function of the underlying model physics leading to their emergence in the same way as it does in reality. As soon as we start to pre-program emergent constraints into the model, they lose all meaning and usefulness. It may give the model a better skill when compared to past observations, but this is no guarantee of future success, since we cannot know how the emergent constraint will evolve in a changing climate.