Korva is a scroll of the internet’s best ideas.
Kimi K3 beat Fable at front-end
It was the first time an open-source model had beaten the leading American closed models on a subset of tasks that important. A large share of developers work on the web, and the result showed up on Arena, which ranks models by what real people prefer.
For years a persistent American narrative has held that Chinese labs only keep up by distilling American models, meaning they train on the outputs of those models. Kimi's result cuts against that. Distillation may still be one step in their training, but it is only part of the story, and something else those labs do is lifting performance above what the American labs currently reach.
The result shifted how people view the whole field, from American scientific dominance to the economics of selling AI.
How does Arena turn votes into rankings?
Arena began as Chatbot Arena, a project from researchers at UC Berkeley. A visitor asks a question, receives answers from two anonymous models, and votes for the better one. The votes feed a statistical model in the Elo family, the kind of system used to rate chess players, which turns millions of head-to-head results into a leaderboard. Arpad Elo designed his rating system for the United States Chess Federation, which adopted it in 1960.
How does distillation train a smaller model?
A small student model is trained to imitate the outputs of a larger teacher. Geoffrey Hinton, Oriol Vinyals, and Jeff Dean described the method in a 2015 paper, showing that the teacher's full probability distribution over answers carries more information than a single correct label. With language models, the teacher's generated text can serve as training data for the student. The student often gets close to the teacher on narrow tasks at a fraction of the size.
Why does front-end coding matter so much?
Front-end code builds what users see in a browser: layouts, buttons, forms, and animations. The output is easy for people to judge, since a page looks right or it does not. That makes web tasks a natural fit for a leaderboard built on human preference. Stack Overflow's annual developer surveys have consistently listed JavaScript and HTML/CSS among the most commonly used languages.
More from 20VC with Harry Stebbings
