Beware of Pareto Principle when generating code with LLMs

By Dominik Tarnowski on Mar 2025

did you know:

It’s not always 80/20: sometimes you’ll see this as 90/10 or 70/30. They’re all variations of Zipf’s law, which generally states that outcomes are dominated by a small minority of causes.

LLMs excel at reproducing the frequent patterns in programming (the 80%), but their effectiveness drops significantly for rare, complex, or subtler things (20%).

They are great at quickly spinning up boilerplate, setting up boring test cases, generating common patterns. We should probably utilize that.

But don’t forget about the last 20%. A phenomenon I observed in myself and others when using these AI assisted tools is actually resulting in building poorer quality software due to getting complacent. If 80% of your time you have your code automatically completed for you, you’re going to naturally start thinking about it less, and be less careful scanning the results.