I just spent 16 hours reviewing changes Astra made to my framework.
Simple nodes had turned into miniature engineering projects. More layers. More abstractions. More work to maintain. None of it made the original problem better defined.
This is the third time I’ve seen this pattern across recent model upgrades from OpenAI and Anthropic.
I didn’t have this problem with earlier models.
The stronger the model becomes, the more possibilities it sees. But seeing a possibility is not the same as knowing it belongs in the solution.
It looks like a machine version of the curse of knowledge.
In software, overengineering is not harmless. It hides intent, increases maintenance cost, and creates more places to fail.
I’d like model makers to treat restraint as a capability worth evaluating.
A stronger model should not only solve harder problems. It should know when the problem is simple.
Have you seen stronger models overengineer work that earlier models kept simple?