Nothing in this engine uses a learned model today. I want to write down where one would go before it goes there, because the word AI has stopped carrying information and I would rather be held to four specific claims than to one vague one.
Start with the strand that is already half built. Modern renderers draw fewer pixels than they show you and rebuild the rest from the frames before, and the upsampler that does it here has no network in it at all. It is a hand-written history rejection rule: take the previous frame, reproject it with the motion vectors, and throw away the parts that disagree with the current one beyond a threshold. At half resolution it measures 34.3 decibels against the full picture, where simply stretching the small one gets 32.0, and it cuts frame-to-frame shimmer on fine detail from 3.4 to 2.1.
A captured scene is the easy case for this, which is the part I find interesting. In a game the world moves: limbs swing, cloth deforms, and a motion vector is an approximation that gets worse the faster things move. A captured street holds still. The camera is the only thing moving, so the reprojection is exact rather than estimated, and the only reason left to throw history away is a disocclusion, where something that was hidden becomes visible. That is a much smaller problem than the one DLSS was built for, and it is why a threshold written by hand gets as far as it does. A network would replace the threshold, not the reprojection.
The second strand is repair, and it is the one I would build first, because I spent today on exactly the defect it would catch. Three of the eleven places here carry Gaussians that are hundreds of metres across. The largest is 974 metres, in a capture whose actual content fits inside a box 280 metres on a side. Seen from inside the street, one of those is a translucent sheet over the whole frame, and it is why two of the places read soft at every distance while the rest read sharp. What shipped is a ceiling: no splat is allowed to be wider than 20 metres by the time it is staged for drawing.
That is a bandage and I would rather say so than let it read as a fix. The ceiling works because the number is not a judgement: the eight captures that read sharp never produce a splat over 13 metres, so 20 metres is a cap on the impossible rather than a tuning knob. The real fix is upstream, in the step that turns a capture into the levels the engine streams, and the reason it is not there yet is that outlier is harder to define than it looks. A 974 metre splat is obviously wrong. A 15 metre one might be a wall, or might be a floater hanging over a courtyard, and the difference is context rather than size. Judging context from examples is the thing a model is actually for.
Repair has a second half that is not classification at all. A capture has holes where the camera never looked: under a parked car, behind a bin, the strip of road nobody walked down. Today those are holes, and the collider patches them so that a walker does not fall through one. Filling them so they look like the rest of the street is generation rather than cleanup, and it is a different problem with a different failure mode.
The third strand is light, which is the biggest thing being worked on here and already written up. A photograph carries the light it was taken in and nothing else. There is no material, no roughness, no measured normal; there is a colour per splat that already includes the sun that happened to be out that day. Asking that surface how it would answer to a lamp it never saw is a question the capture holds no information to answer, which is exactly the shape of problem where a model trained across many surfaces should beat a formula written for one.
The fourth is what was never captured. Around a corner the scan never turned, past the edge of the walk, behind a wall it never passed. I am putting no date on this one and I would not believe a date if somebody gave me one. The test it has to pass is harsh: invented geometry has to stand next to measured geometry without being the thing your eye goes to, and the measured geometry is a photograph, so the bar is a photograph.
And the part I would keep away from a model entirely: the transport. The wire format, the octree, the level selection, the sort. Those are deterministic and measurable, and they already win or lose by numbers I can read off a bench. Putting a model in there would trade something I can debug for something I cannot, and the one lesson that has held through all of this is that the parts I can measure are the parts that actually got fixed.
None of the four is in the build. When one of them is, it will be in the changelog on the day it lands, with the number it beat.