What are the strongest reasons to write code without relying on LLMs? Are there useful middle grounds, or does the extra checking outweigh the benefit?
Why choose not to use LLMs for coding?
Answers
The checking is a big part of the work. Generated code can look plausible while making the wrong assumptions, so you still have to understand it, test it, and make sure it fits the rest of the project.
For some tasks, spelling out the requirements and reviewing the result may take as long as writing it yourself. Lots of generated code is not the same thing as useful progress.
Reply to this answer
There's also a learning and enjoyment question. Writing the code is the interesting part for some people, especially on personal projects. Handing that over can turn a programming exercise into a review exercise, which isn't necessarily what they wanted.
Reply to this answer
You don't have to treat it as all or nothing. One option is to use a model to ask questions about documentation or to work through a problem, while still writing the implementation yourself.
That doesn't make its answers authoritative. You need enough context to ask useful questions and check what it tells you, preferably against the actual documentation.
Nuvzari-6w:
That distinction helps. Using it to explore an idea sounds different from accepting a generated implementation.
Reply to this answer
There can be a practical tension at work. Some developers worry that refusing these tools will make it harder to meet expectations if coworkers use them to move faster. Others question whether the apparent speed holds up once review and maintenance are counted.
I'd judge it by the task and the team's actual requirements, not by a blanket prediction that everyone who opts out will fall behind.
Reply to this answer
Some people object on environmental grounds too. Running models and building the infrastructure behind them have costs, and a developer may decide that convenience isn't enough reason to add to that demand. How much weight to give that concern is a separate judgment from whether the tool saves time.
Reply to this answer
Add your answer
Answers are reviewed before publication.
5n!ferrulewicket_638:
That probably depends on the task, though. A small, well-defined change may be much easier to check than a large feature.