model-evaluation.

Kimi K3 and the limits of the pelican benchmark

Moonshot AI's 2.8T parameter model is impressive, but my 21-month-old SVG test reveals why we need better ways to evaluate frontier AI models.