The costs of interactive AI have interesting effects as the author points out. Much like the lack of variety in music models, 3D asset generation via AI has a long way to go, particularly as studios have no incentive to share their data. But I think AI assistance could at least make some marginal improvements. Take a procedural game like No Man's Sky. There are billions of possible worlds in the game. But, presumably because of team size and/or agile philosophy, they drop these rigid, nearly identical, items and buildings throughout those billions of planets. AI could at least conveniently empower some added incremental complexity (via Claude Code etc.) to their existing asset generation, which lends more believability to these planets. If studios were willing to partner and share data, incredible world generation models, which could be used exclusively as asset generators rather than real time world renderers initially, could be built and dramatically empower designers.