The author notes that judgment steps in AI agent development consume too many tokens, and tries connecting the judgment-specific model Jev to Codex so high-performance AI focuses on code and design.
AIエージェント開発では「判断」にトークンを使いすぎる。 コード・設計・デバッグは高性能AIでよいが、テスト成否や分類などの単純判断までCodexに任せる必要はない。 判断専用モデル「Jev」をChatGPTと相談しながらCodexにつないでみた。 詳しくはコメント欄から↓
Gregor Zunic showcases an open-source browser agent combining Browser Use and Jev, finding flights in 7 seconds for $0.0039, using a new action space per step, DOM state space, and small LLM fallback for typing.
Scott Williams predicts custom Jev-style models (parallel constrained decoding) will make existing company agent systems more token efficient, creating millionaires.
The user praises the JEV model for its extremely fast reproduction of each movement and command, believing bots will become smarter as a result.
