Yeah, some people are is saying "Oh coding is solved, but we still need people to guide the architecture." But coding was "solved" by self-play against "small"

Yeah, some people are is saying "Oh coding is solved, but we still need people to guide the architecture." But coding was "solved" by self-play against "small" problems. Is there any reason training couldn't be scaled up to even longer horizon tasks like evolving an architecture in a growing business? The increasing load from a growing business, both in terms of usage and product accretion, can be entirely simulated in a data-center, and therefore trained against.