Ben Kunkle leads edit predictions at Zed, where he helped develop Zeta2, an open-weight model that anticipates a programmer’s next code change. Its central challenge is interpreting a developer’s intentions accurately enough to offer useful suggestions at the speed of typing.
Kunkle started programming by building Minecraft mods with friends. At California Polytechnic State University, San Luis Obispo, he worked for a web startup and explored systems languages including Rust and Zig. A compilers course introduced him to Zed’s approach to collaborative coding; he joined the company after graduating in fall 2024.
His early projects included redesigning Zed’s settings architecture, replacing scattered configuration registration with strongly typed user and project settings and improving the GPUI framework’s support for focus management and complex forms. He also created the Leptos RSTML extension, which supports Rust-based Leptos templates.
Building predictions developers actually keep
- Production editing traces: Git commits reveal finished code but omit the false starts, typos, and course corrections that shape real programming. Zeta2 instead learns from approximately 100,000 opt-in editing examples collected from open-source repositories, incorporating cursor position, recent changes, diagnostics, and nearby definitions. The collected user data remains private.
- Teacher-student distillation: Kunkle helped train a small, fast model using predictions from a larger model. Because those predictions sometimes delete freshly typed code or exceed the editable region, the pipeline detects unsuitable outputs and repairs them before student training. Reusable intermediate results make it practical to test different context windows and diagnostic inputs.
- Settled-state evaluation: After editing pauses, the eventual code offers a useful but imperfect signal of user intent: developers can change direction, and coding agents can rewrite the same region. Comparing sampled predictions against that settled state separates noisy outcomes and obvious completions from difficult, genuinely instructive examples. Sampling the improving student model reduces reliance on expensive frontier-model queries.
- Real-world acceptance: Kunkle evaluates models through staged production experiments measuring acceptance, latency, retained edits, and diagnostic errors. Zeta2 ultimately used Seed-Coder-8B, and Zed reported a 30 percent acceptance-rate improvement over Zeta1. Its publicly available model weights allow inspection and adaptation without exposing the underlying private editing traces.
