A hypothesis on why people say models get dumber
Published: Mar 19, 2026
People mostly use models to write code to write new features fast, adding lots of code to the codebase. You can make progress on individual pieces of work quite fast, the models are good at finding their way around a codebase and implementing a feature in a total-session-token efficient way.
So naturally each new feature feels good to implement, and the model feels smart.
But as the project grows, and you uncover bugs, it becomes harder to do things in it, the model spends more time looking at different parts of the repo, gets distracted. And so you are disappointed in the model's intellect, and you complain about it on twitter dot com.
The next step change model drops, you try it, it is very good because it is a step change, and the cycle repeats again.