A day with Jev: a fast classifier, best suited for Computer Use and game AI
jiayuan_jy · x · 2026-09-18
After a day of experimenting with Jev, the author concludes it's essentially a faster general-purpose classifier/decision-maker that LLMs can replicate—its edge is speed, but with unknown (likely small) parameter scale, its world knowledge and complex-scenario accuracy remain questionable.
Good fits: finite-solution-space, low-latency tasks with formal output guarantees; Computer Use (DOM elements form a finite set, looping dom list → jev action → new dom list); and decision-tree-heavy real-time domains like game AI, robotics, and autonomous driving.
Agent integration test: rewriting parts of Pi Agent with Jev—tool selection can't be replaced (each call carries non-finite parameters like line numbers), but Compaction-style classification works fine.
Demos in progress: a Poker AI (long, complex decision trees, battling Jev against other models) and a serious Pokemon VGC doubles AI, intended for actual ranked play.
More from coding & agent
- ty to ship full support for recursive implicit type aliases, built on mu types — charliermarsh · 2026-09-19
- Converting video to 120fps with a single prompt via Runway MCP — tlakomy · 2026-09-19
- Experimental Jev-powered RLS linter catches Supabase policy bugs rule-based lint can't — SeanOliver · 2026-09-19
- Virtual green screen from a RealSense D436: depth-based RGB filtering, no backdrop needed — chrismatthieu · 2026-09-19
- Runway MCP turns video to 120fps with a single prompt — tlakomy · 2026-09-19
- Cost breakdown: 100 LLM calls in 8.5 seconds for roughly a quarter of a cent — narphorium · 2026-09-19