jevseek: wiring a decision model as DeepSeek's sampler to steer every token
cephaloform · x · 2026-09-23
A clever combo hack: the author plugged jev, a decision model that can't generate text, in as DeepSeek's sampler — DeepSeek proposes the top-k next tokens, and jev picks which one to say. The hybrid is dubbed jevseek, and it writes in a distinctive style.
An interesting community experiment in decoupling token selection from token generation by combining two existing models.
More from Fun
- Claude Rebuilds the Iconic Earthrise in 3D, Down to the Second — Simple-Diver-2192 · 2026-09-23
- "Opus 5.5 Dropped and Made Everything Outdated": Community Memes — Able-Line2683 · 2026-09-23
- Instagram co-founder uses Opus 5.5 to build Ciechanowski-style interactive explainers — mikeyk · 2026-09-23
- Dev Laments Learning Vim and US Keyboard Layout for Speed, Now Useless in AI Era — justalexoki · 2026-09-23
- Front-end-only Three.js app simulates acrylic painting, hits top 32 at Hack the North — floguo · 2026-09-23
- Distortion + inpainting + regeneration loop makes trippy AI art — PurzBeats · 2026-09-23