Should developer tooling be reinvented for AI-assisted programming?

This question should have been asked in the 1960s during the first AI summer. We already changed to accommodate machine-generated code and the change was standard by the 1970s:

• There is a distinct phase inserted between authoring and execution: translation
• Instead of manually writing programs, we write high-level specifications which are machine-translated into programs
• The translator performs a search for efficient programs satisfying the specification
• After translation, we must package the resulting program if we want to share it
• Other folks may not be able to use the translated program as-is and they will have to translate it themselves from the specification, driving fragmentation across architectures
• But fortunately the specifications are portable because they do not specify architecture-specific details

Curiously, at the time, we went further:

• Translators are programs too
• Instead of manually writing translators, we write high-level metaspecifications which are translated into translators
• There is a distinct phase inserted between authoring and translation: metatranslation

Later, the biggest innovation was realizing that we can organize specifications into a tree and distribute that as a portable tree.

From this perspective, the main issue today is that folks are not sharing their (hard) prompts, their training data, their model weights, and their coding harnesses. Almost all of the necessary stuff for a reproducible science is kept trade-secret or copyrighted. Incidentally, evidence is that folks' hard prompts are terribly written, their training data is plagiarized, untagged, and dirty in every sense; their model's core weights are fairly similar to everybody else's core weights up to some rotation, and their coding harnesses are also terribly written.