TypeSafe AI Releases Jev: A System One Model That Returns Typed, Calibrated Decisions Instead of Text
https://www.marktechpost.com/2026/09/19/typesafe-ai-releases-jev/📌 TypeSafe AI 發布Jev：不寫文字，只回傳型別化決策
TL;DR：Jev是Transformer模型但不生成文字，直接回傳附機率的型別化決策，目前為候補制API。
2022年那場 ChatGPT 時刻，教會了 AI 如何跟人對話。這次 TypeSafe AI 押注的方向完全相反：讓 AI 直接跟程式碼對話。
🤔 RLHF 帶來的對話能力，也帶來了過度自信
TypeSafe 團隊發布 Jev。這是一個以 Transformer 為基礎的模型，但它不是大型語言模型，也不會生成文字。開發者傳入一個狀態（state）與型別化的問題，模型回傳附帶機率的型別化決策，讓程式碼可以直接依此分支處理。TypeSafe 團隊主張，RLHF 把模型調校成迎合人類偏好，因此產生了「聊天」這種輸出形式，也帶來過度自信與模式丟失（mode dropping）等問題，這些缺陷使得系統仍需要人類留在決策迴圈中把關。
Jev 的名字借用了 Daniel Kahneman 對快速直覺與慢速推理的區分。TypeSafe 表示 Jev 使用一套新的技術堆疊：新架構、平行取樣器（parallel sampler），以及「校準決策強化學習」（Reinforcement Learning for Calibrated Decisions，RLCD）。不過具體架構細節，TypeSafe 目前並未公開。
🧩 一個端點處理所有型別化問題
Jev 只用單一端點運作：POST https://api.typesafe.ai/v1/systemone。請求內容帶有 state、model，以及一組問題（questions）的映射。文件定義了3種問題類型，這些問題會針對同一個狀態並行、彼此獨立地執行；TypeSafe 表示，增加問題數量對回應時間幾乎沒有影響。其中 Choice 類型的問題最多可支援255個選項。
開發串接方式包括 pip install typesafe-sdk（需 Python 3.10 以上）、JavaScript SDK @typesafe-ai/sdk，快速入門文件也涵蓋 cURL 呼叫方式，以及供 Claude Code 使用的 Agent Skill。
每一個 Choice 與 Score 類型的答案都會附帶一個介於0到1之間的信心值，這個數值是根據機率分布的形狀推導而來。文件範例中，「billing」選項以0.84的機率勝出，但信心值僅0.596，原因是「technical」選項仍保有0.159的機率。文件建議依此設計3條路徑：高信心值直接執行、中間值交由人工複核、低信心值則轉交人類判斷；判斷門檻應隨錯誤決策的代價高低而調整。
📊 號稱快193.6倍、便宜444.6倍，但數據來自自家評測
Jev 的定價為每10億個輸入 token 收費42美元，TypeSafe 則引用現有 LLM 的報價區間為每百萬輸入 token 0.20美元至10美元。在其公開的錄製示範中，Jev 耗時0.114秒、花費0.000081美元完成任務，而 GPT-5.6 Terra 耗時8.566秒、花費0.013880美元；TypeSafe 據此宣稱快193.6倍、便宜444.6倍。這些數字來自 TypeSafe 自家的 workflow 評測，並非第三方獨立驗證的結果。
⚠️ 「零幻覺」指的是格式，不是答案正確性
文件中所謂的「零幻覺」，指的是結構化輸出的 schema 一定會匹配成功，這個0%的數字並非經驗性的正確率保證——答案本身仍可能是錯的。目前 Jev 僅以早期候補（waitlist）形式提供 API 存取，TypeSafe 尚未公開權重、參數量，也沒有提供自架部署選項。文章也提到，社群專案在發布後幾天內就已經出現。
🎯 實務啟示
如果你的 Agent 系統裡有大量「選A還是選B」「這句話是不是詐騙」這類原子化判斷，Jev 展示的型別化決策介面值得關注：把慢思考（規劃、生成）與快判斷（分類、評分）拆開處理，理論上能同時省下延遲與成本。但在架構未公開、僅有候補制 API 存取的情況下，實際效能與穩定性仍需自行驗證，不宜直接沿用官方評測數字做容量規劃。
🔗 來源
- 標題：TypeSafe AI Releases Jev: A System One Model That Returns Typed, Calibrated Decisions Instead of Text
- 作者／機構：Asif Razzaq, MarkTechPost
- 連結：https://www.marktechpost.com/2026/09/19/typesafe-ai-releases-jev/
#TypeSafeAI #Jev #AgentArchitecture #SystemOneThinking #CalibratedAI #TypedAPI #RLCD #AIInfrastructure #FastInference #AIDecisionMaking
原始資料 MarkTechPost · 收集於 2026-09-20
摘要原文
The ChatGPT moment in 2022 taught AI to talk to people. One of its builders now bets the next moment is AI that talks to software, not people. TypeSafe AI released Jev . Jev is transformer-based, but it is not a large language model. It does not generate text. You send a state and typed questions. It returns typed decisions with probabilities that code can branch on. Is it deployable? Yes, as a hosted API in early access behind a waitlist. TypeSafe has not published weights, a parameter count, or a self-hosting option. The name borrows from Daniel Kahneman’s split between fast intuition and slow reasoning. TypeSafe team argues RLHF tuned models for human preference. That produced chat, and overconfidence and mode dropping. Those flaws keep a human in the loop. Jev uses a new stack: a new architecture, a parallel sampler, and Reinforcement Learning for Calibrated Decisions (RLCD). TypeSafe has not disclosed the architecture. One endpoint handles everything: POST https://api.typesafe.ai/v1/systemone . The body carries state , model , and a map of questions . The docs define 3 question types. Questions run in parallel and in isolation against the same state. TypeSafe says adding questions barely changes response time. A Choice supports up to 255 options. Install with pip install typesafe-sdk (Python 3.10 or later). A JavaScript SDK ships as @typesafe-ai/sdk . The quickstart also covers cURL and an agent skill for Claude Code. Every Choice and Score answer carries a confidence value from 0 to 1. TypeSafe derives it from the shape of the probability distribution. In the docs example, billing wins at 0.84. Confidence is only 0.596, because technical still holds 0.159. The docs suggest 3 paths. Act on high confidence. Review the middle. Send low confidence to a human. Thresholds should scale with the cost of a wrong action. Jev costs $42 per billion input tokens. TypeSafe quotes existing LLMs at $0.20 to $10 per 1M input tokens. In its recorded demo, Jev finished in 0.114s for $0.000081. GPT-5.6 Terra took 8.566s for $0.013880. The TypeSafe team claims it to be 193.6x faster and 444.6x cheaper. Those figures come from TypeSafe’s own workflow evals . But hold on here are some things to keep in mind: ‘Zero hallucinations’ means schema matching is guaranteed. The 0% figure is not empirical. Answers can still be wrong. Community projects appeared within days of launch. Here are some examples: Check out the launch post , docs , and TypeSafe’s GitHub . All credit goes to the researcher of this project. Also, feel free to follow us on Twitter and don’t forget to join our 150k+ML SubReddit and Subscribe to our Newsletter . Wait! are you on telegram? now you can join us on telegram as well. Need to partner with us for promoting your GitHub Repo OR Hugging Face Page OR Product Release OR Webinar etc.? Connect with us The post TypeSafe AI Releases Jev: A System One Model That Returns Typed, Calibrated Decisions Instead of Text appeared first on MarkTechPost .
由 claude-code/sonnet 自動生成