AI Follows the 80/20 Rule

It is known that 80% of the work requires 20% of the effort, and 20% of the work, requires 80% of the effort. ¹ This is known as the Pareto Principle. Applied to software engineering, this means by volume, the majority of a code base will be the uninteresting plumbing that renders the code functional; boilerplate, error handling, basic CRUD logic, form widgets, etc. all take up a large number of SLOC, but don’t demand the majority of a companies time. By contrast, a small percent of a codebase, nominally 20%, is taken up by the difficult and interesting part. It is the complex business logic, the algorithms, and the edge cases that set ones software product apart from it’s competitors.

Large Language Models excel at the 80%. Anecdotally, AI is best at translation tasks: implement this equation from a paper in rust, create an HTML form to populate the data in this class, write a parser for this JSON. AI is also very good at following tutorials to set up boilerplate, for example, setting up a JavaScript or C++ project. Hence, it does not surprise me that companies are bragging that AI is writing up to 30% of new code ^{2 3 4}, with plans to expand to up 50%.⁵ One suspects that this code is mostly of 80% kind.

This puts a cap on the theoretical efficiency gain from AI. If a Large Language Model only helps with code that previously took 20% of engineering hours, than one can only improve efficiency by 20%. However, AI does not yet work without human guidance. In my experience, AI makes me twice as fast at the tasks it is good at, resulting in projects completed 10% faster.

This is great news for users of AI. Working 10% faster more than justifies the $20 to $200 per month cost of an AI subscription. Furthermore, I personally find the kind of work AI is best at to be tedious, and I am more than happy to delegate it to a machine. This may however not be great news for companies who’s valuations require 90% of code to be written by AI⁶, and who’s inference costs far exceed what they currently charge.