A common practice in organic chemistry is to start from the target molecule and work backwards trying to determine simpler precursor molecules that would produce the target. 'Retrosynthesis'

Really really cool! I wonder if we can do something similar with programming. Suppose we have a conceptual framework where each module/system is a widget made up of slightly simpler and well-defined widgets. If we can specify what we want, perhaps a system can be devised that would try to build it from a database of widgets. If some of the constituent widgets are not found, the system could recursively specify each one and find/construct.

In this recursive process, the language used to specify the requirements of a widget can be a changing DSL whose grammar and basic constructs co-evolve with the complexity/abstraction level of the widgets.

This seems feasible because we (software engineers) are one such system. Starting from basic transistors, we build ever higher abstraction layers along with the language used to specify them (circuit diagrams -> microcodes -> assembly -> C -> DSLs). I believe multi-layer neural networks are also a prime example of such a system.

The difference in those targets is that they're uncertain, and that that uncertainty compounds as you reach further into the future. I don't think the nanopass-in-reverse approach lends itself well to resolving uncertainty, because you'd start from the point of most uncertainty, which sort of begs the question a bit.

you're right about that, so I think what it could help us with is to provide a sort of guidance when we have a particular end-goal in mind.