Korva is a scroll of the internet’s best ideas.
Better coders make worse writers
The voice and tone slide. Output settles into stock lines, such as three of the four arguments being right and one being a turning point, followed by a line that opens with here's the rub. It gets to the point where it is sometimes impossible to read what the model is saying.
The problem is fixable. Kimi has always had pretty interesting writing, and its voice and tone are both pretty good.
Kimi is quite good overall, though it is not cyber capable in the way frontier models are and it still trails them on long-horizon tasks. GLM 5.2 was a really big step for open weight models, and Kimi is Moonshot getting up to that step.
Why would coding practice hurt a model's writing?
Coding gives training a clear score, because tests either pass or fail, so labs can reward a model millions of times for getting code right. Prose has no test that passes or fails, so it gets far less of that kind of pressure. A 2023 study by Robert Kirk and colleagues found that reinforcement learning from human feedback reduced the diversity of a model's outputs, a narrowing that can flatten style.
Why do models repeat the same stock phrases?
Training on human or model preferences rewards phrasings that raters liked, and the same few phrasings win again and again. Once a phrase works across millions of outputs, the model leans on it by default. Readers who see many outputs from the same model start to recognize those phrases as that model's habits.
Why is writing quality so hard to measure?
Code can be checked automatically, but good prose depends on taste, audience, and context. Labs often rely on human raters or on other models acting as judges, and both are noisy. Studies of automated judges have found that they tend to favor longer answers.
More from 20VC with Harry Stebbings
