LLMs as Language Compilers: Lessons from Fortran for the Future of Coding

3 months ago

Large Language Models (LLMs) have evolved rapidly, now capable of autonomously completing tasks at the scale of full engineering teams.
Stack Overflow's popularity has declined by 77% since 2022 as developers increasingly turn to tools like ChatGPT and coding agents for help.
Coding agents can significantly speed up development, enabling the creation of complex prototypes in hours, though they sometimes struggle with well-defined tasks.
Steven Yegge's 'Gas Town' metaphor illustrates the potential and unpredictability of coding agents in software development.
Historical parallels exist with 'Automatic Programming' in the 1950s, where languages like FORTRAN and COBOL simplified coding but didn't eliminate the need for skilled programmers.
John Backus and Grace Hopper were pioneers in making programming more accessible, despite resistance from the 'Priesthood' of elite programmers.
FORTRAN's optimizing compiler proved highly efficient, nearly matching hand-coded assembly, and helped democratize programming to some extent.
Despite easier programming languages, the number of programmers and the complexity of systems have grown exponentially over the decades.
Jevons Paradox suggests that improvements in efficiency often lead to increased demand rather than reduced resource use, as seen in fields like radiology.
The future of coding may involve higher levels of abstraction, tackling problems we haven't yet named, while essential complexity remains.

Hasty Briefsbeta