TypeSafe AI 出隐身模式，Jev 做 System 1 决策

Carlos Manuel @Origins_Carlos原推发布于 2026.09.21

在 X 上打开原推

A model just launched that doesn't write a single word, and it might be the most useful thing to happen to AI agents this year TypeSafe AI came out of stealth on Sept 15 with $40M led by DCVC. It's founded by Diogo Almeida, co-inventor of RLHF. Their model, Jev, doesn't chat. It decides You give it a state and a question. It picks an option (Choice), rates something on a scale (Score), or gives you a calibrated probability that something is true (Noul). Latency is 70 to 500 ms, and input costs $0.042 per million tokens with free output (check the diagram 🖼️ below to understand a bit better how it works) Here's why that matters. Most agents today use a giant reasoning model for tiny judgment calls: which tool to use, whether a request is safe, whether the loop should stop. That's 2 to 10 seconds per decision, plus the occasional broken JSON Using a frontier LLM to route a support ticket is like hiring a novelist to sign for your packages So the stack is splitting in two. Fast, cheap "System 1" models make the reflexive calls, and the big models are saved for real thinking. I think that's the right direction But faster decisions create a new problem. An agent making hundreds of micro-decisions per minute is now too fast for any human to watch. One wrong call is a bug, but a fleet repeating the same wrong call in parallel is an incident This is where we spend our time at Origins. Speed tells you nothing about trust, and there are two separate questions to answer: Who deployed this agent? That's identity, and it's why we built a three-tier User / Agent / Session model with single-use session keys that expire when the task ends What did it actually do? That's execution, and it's why Proof of Computation puts verification of the agent's work on-chain Fast brains need auditable receipts If your agent makes 500 decisions a minute, how many of them could you prove? Not sure... Check us out then 🔗: https://www.origins.network/