{
  "id": 8990154,
  "title": "TypeSafe launched Jev because sequential LLMs are “totally useless for computers”",
  "url": "https://urgent.news/2026/09/21/typesafe-launched-jev-because-sequential-llms-are-totally-useless-for",
  "topic": "ai",
  "section": "AI",
  "published": "2026-09-21T19:36:29.000Z",
  "source": {
    "name": "The New Stack",
    "slug": "the-new-stack",
    "url": "https://thenewstack.io/typesafe-jev-system-one/"
  },
  "original_language": "en",
  "account": "TypeSafe, a startup launched two years ago in stealth mode, unveiled its first model, Jev, to the public last week. Backed by $40 million in seed funding led by DCVC, Jev is a text-only model designed to make decisions for software applications, rather than generating text. Co-founded by ex-OpenAI employee Diogo Almeida, Jev is built on a new architecture, sampler, and training algorithm called Reinforcement Learning for Calibrated Decisions (RLCD).\n\nUnlike traditional sequential LLMs, Jev is 20-200x faster and 40-400x cheaper, with free output tokens. It provides decisions alongside calibrated probabilities, addressing issues like mode dropping, overconfidence, and lack of reliability seen in chat LLMs. Almeida, who co-invented RLHF and InstructGPT at OpenAI, believes superhuman chat models haven't led to artificial general intelligence (AGI) because they are \"optimized for human preferences,\" which include problems like hallucinations.\n\nJev is designed to output decisions with probabilities and confidence, making it \"a lot more like code.\" It can't hallucinate, as it's based on code, which is reliable, fast, self-consistent, and type-safe. Developers can use Jev to set thresholds for autonomous actions or human oversight and combine its decisions in code to build larger workflows. Jev is not a chatbot or coding model, but a smart if statement that can be added to an existing application to make judgment calls. Despite being text-only, Jev can handle visual input, albeit with limitations, as demonstrated by a test where it correctly identified an airplane from 400 hand-drawn sketches 35% of the time.",
  "summary": "When TypeSafe emerged last week after two years in stealth, backed by $40 million in seed funding led by DCVC, The post TypeSafe launched Jev because sequential LLMs are “totally useless for computers” appeared first on The New Stack .",
  "key_points": [
    "TypeSafe launches Jev, a text-only model for software applications",
    "Jev 20-200x faster and 40-400x cheaper than traditional LLMs",
    "Jev outputs decisions with calibrated probabilities, avoiding hallucinations"
  ],
  "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."
}