Jev explained in six steps: from RLHF to RLCD decision models
Maryam Miradi walks through six steps on how Jev differs from chat models: it returns typed decisions with explicit probabilities, letting confidence drive run, review or escalate workflows.
🛠️🧭 Jev, TypeSafe AI, Explained. 𝗠𝘆 𝟲-𝘀𝘁𝗲𝗽 𝗝𝗲𝘃 (𝗥𝗟𝗛𝗙 𝘁𝗼 𝗥𝗟𝗖𝗗). Diogo Almeida helped build the RLHF/InstructGPT work behind the ChatGPT era. Now his company TypeSafe has released Jev. And the idea is very different: Instead of building another model primarily for humans to talk to, Jev is designed for software to make decisions with. Here’s how Jev is built for automation, not assistance: 》𝗦𝘁𝗲𝗽 𝟭: Start With the Problem Jev Solves ✸ RLHF optimizes models around human preferences. ✸ That works extremely well for chat, copilots, and coding assistants. ✸ But automation needs outputs software can reliably consume. → Jev targets the machine-in-the-loop instead. 》𝗦𝘁𝗲𝗽 𝟮: Jev Uses RLCD ✸ TypeSafe created a different training method. ✸ RLCD = Reinforcement Learning for Calibrated Decisions. ✸ The target is not simply a response a human prefers. → It is a calibrated decision that can become part of software logic. 》𝗦𝘁𝗲𝗽 𝟯: Jev Returns Decisions, Not Just Strings ✸ Traditional LLMs generate free-form text. ✸ Jev produces typed structured outputs. ✸ The possible output space is defined in advance. → Think approve | reject | review, not another paragraph. 》𝗦𝘁𝗲𝗽 𝟰: Jev Makes Uncertainty Explicit ✸ Every decision includes probability and confidence information. ✸ Now uncertainty is visible to the surrounding software. → The system can distinguish “probably right” from “very sure.” 》𝗦𝘁𝗲𝗽 𝟱: Jev Turns Confidence Into Control ✸ High confidence → act. ✸ Uncertain → verify. ✸ Low confidence or high risk → escalate. → Confidence becomes part of the workflow, not hidden inside the model. 》𝗦𝘁𝗲𝗽 𝟲: Put Jev Inside Software Workflows ✸ Classify. ✸ Route. ✸ Score. ✸ Extract. ✸ Verify and guardrail other AI systems. → These are often decision problems, not conversation problems. 𝗣𝗹𝗲𝗮𝘀𝗲 𝗥𝗲𝗺𝗲𝗺𝗯𝗲𝗿: Jev is not trying to be another ChatGPT. And RLCD is not simply “better RLHF.” They optimize for different things. RLHF → outputs humans prefer. Jev + RLCD → calibrated decisions software can act on. TypeSafe's bigger bet is that the next AI era will not only be about better assistants. It will be about putting intelligence inside software itself. -- 📥 Building production AI agents? Watch my free 30-min training + get 88 pages of production guides. Join 46,000+ AI engineers, architects, and directors. 👉 https://www.maryammiradi.com/free-ai-agents-training?utm_source=x_post