Constraint decay: why both camps on AI coding are right

In software we have two camps: the "AI can do it all" crowd and the "AI won't replace engineers" crowd.

A new research paper helps explain why both camps are right: "constraint decay" is where an initially capable AI coding tool performs progressively worse the more you constrain it (ie the more it has to fit into an existing ecosystem).

I like that this gives a data-based reason for why AI excels at blue sky development but does less well on large legacy codebases.

The one strange thing is that the paper points to databases as a root cause, whereas in my experience SQL and querying is actually one thing LLMs have done extremely well for a long time now.

Paper in full →