Codegen is not productivity

9 hours ago

Generative AI and LLMs generate a lot of code, but lines of code (LOC) is a poor metric for programmer productivity.
Programming is about managing complexity and representing abstract ideas, not just writing code quickly.
Developers spend most of their time on activities other than coding, making code generation not the primary bottleneck.
LLMs accelerate code writing but may rush implementation, skipping crucial design and planning phases.
High LOC increases maintenance difficulty and reduces understandability, affecting both humans and LLMs.
LLMs often encourage bespoke solutions over using established libraries, leading to unnecessary code.
Collaboration in software development relies on readable and reviewable code, which LLMs may not optimize for.
LLM-generated code often requires significant fixes, leading to frustration and questioning their utility.
The real value of LLMs lies in planning, design, and research, not just code generation.

Hasty Briefsbeta