Give a very capable model a small problem and plenty of room to think. Surely something excellent comes out the other end.
Sometimes. Sometimes the small problem gets a very large solution.
Testing GPT-6 Astra has produced a slightly annoying discovery: High has been giving better results than Extra High or Ultra for the coding work at hand.
The higher settings kept overthinking and overengineering things. High was more likely to get on with the actual job.
An awkward result for a setting called Ultra. That name sounds like it should at least come with better suspension.
The problem grows legs
Overengineering often arrives looking helpful. A little extra flexibility. An abstraction for later. A more general solution, just in case the small thing eventually becomes a large thing.
Each addition can sound sensible on its own. Together, they can turn a straightforward change into something that needs a guided tour.
The frustrating part is that the result can be clever. The pieces fit. The explanation sounds convincing. There is simply much more of it than the task deserved.
A shelf was requested. The plans include a warehouse.
High was enough
In these sessions, High struck a better balance. The output stayed closer to the request and needed less pulling back towards something simple.
That is what “better” means here. A solution that does the job, is easy to understand, and leaves less unnecessary machinery behind.
For the work in these sessions, turning the setting up kept making the result harder to like. A different problem might benefit from the extra effort. A small change rarely needed to become an architectural event.
There is something quite funny about having access to this much capability and spending part of the conversation saying, “Please make less of a meal of it.”
The clever bit is knowing when to stop
Good software involves a lot of restraint. Solve the problem that exists. Leave the imaginary future requirements somewhere imaginary. Make the next change easy to understand.
AI can produce an impressive amount of code. That makes knowing what to remove even more useful. Every extra layer still becomes somebody’s job to read, maintain, and eventually question.
The same taste that says “that paragraph is unnecessary” can say “that abstraction is unnecessary”. Both are allowed to go, even if they took effort to produce.
For now, High gets the work. A small, readable solution gets the approval.
The warehouse can wait. There is still a shelf to put up.