{
  "id": 10912460,
  "title": "Jev: Find Out Why RLCD and System One Models Are Rewriting AI Architecture",
  "url": "https://urgent.news/2026/09/30/jev-find-out-why-rlcd-and-system-one-models-are-rewriting-ai",
  "topic": "ai",
  "section": "AI",
  "published": "2026-09-30T00:58:52.000Z",
  "source": {
    "name": "HackerNoon",
    "slug": "hackernoon",
    "url": "https://hackernoon.com/jev-find-out-why-rlcd-and-system-one-models-are-rewriting-ai-architecture?source=rss"
  },
  "original_language": "en",
  "account": "The software industry has been attempting to make large language models behave like functions for the past three years, but with little success. Developers have faced a multitude of engineering challenges, from parsing layers and retry loops to dealing with hallucination guards and the fear of silent regressions in production systems. The issue lies not in the models' capabilities but in the mismatch in how we perceive their purpose. Large language models are designed to produce human-readable text, while software needs structured values for code. When an LLM is asked to generate values for software, the interface becomes a lossy and statistically opaque pipe.\n\nJev, created by Typesafe.ai, aims to address this impedance mismatch by introducing System One models trained via Reinforcement Learning for Calibrated Decisions (RLCD). This approach marks the end of text generation as the default output of AI inside software. The core issue is the impedance mismatch between software's preference for typed values and LLMs' native output of token sequences.\n\nSystem One models solve this problem by being trained to produce decisions rather than text. They generate well-calibrated probability distributions over a small, well-defined answer space, making them reliable in production. Unlike reasoning models (System Two), which are slow, expensive, and unsuitable for tasks requiring immediate binary judgments, System One models can make decisions in tens of milliseconds at a fraction of a cent. The fast path handles routine decisions, while the slow path deals with more complex tasks requiring deliberation.\n\nThe key to System One models' reliability is their commitment to a closed set of outcomes, inability to invent new categories, and explicit uncertainty expressed through probabilistic outputs. This allows software to consume their decisions directly, without the need for text-parsing layers. By separating fast, quick decisions from slow, deliberate ones, Jev builds a two-speed system that addresses the core issue: the mismatch between text generation and structured value consumption in software.",
  "summary": "By introducing System One models trained via Reinforcement Learning for Calibrated Decisions (RLCD), Jev marks the end of text generation as the default output",
  "key_points": [],
  "editors_take": null,
  "illustration": null,
  "coverage": {
    "outlets": 1,
    "also_reported_by": []
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}