Shplait-Based RFC Reimagines PL Course for the AI Coding Era

A request for comments published as a Google Doc proposes a wholesale redesign of a programming languages course, built on the argument that a PL course remains essential in the age of AI-generated code, specifically focused on constraining what AI can produce through language design guarantees.

The redesign organizes the course around two axes: whether a language is confined or unconfined, and whether it supports custom property verification or not. These axes produce four quadrants that structure the course. The unconfined-with-custom-properties quadrant covers rich type systems, including refinement types taught through a variant called Liquid-Shplait and information flow control taught through IFC-Shplait. The confined-without-custom-properties quadrant covers languages that provide guarantees by design, including an object capabilities variant called Ocaps-Shplait. The confined-with-custom-properties quadrant covers sub-Turing-complete languages like SQL, Datalog, and regex, plus custom verifiers built with SMT solvers. The fourth quadrant, unconfined with no custom properties, represents conventional programming.

The course runs on a common syntactic foundation called Shrubbery, described as Racket's infix notation. According to the document, using a shared syntax enables focus on conceptual differences rather than syntactic variations. The RFC also states that this approach reportedly helps AI code generators perform better than parenthesized syntax, though the basis for that claim is not spelled out.

The document concludes that AI demonstrates both the necessity and the opportunity for rich programming languages in the future of software construction. The RFC argues that point primarily by design: by showing that the four-quadrant framework gives students tools for specifying and verifying the behavior of AI-generated code. What the source does not give you is any account of whether this course has been piloted, or evidence that students who complete it can apply these frameworks to real AI output.