🏠

🤯 TypeSafe Jev: AI Logic Unlocked 🚀

AI

September 19, 2026 | Author

🎧 Audio Summaries
English flag
French flag
German flag
Japanese flag
Korean flag
Mandarin flag
Spanish flag

🧠Quick Intel


  • TypeSafe launched its Jev model, a System One Model designed for automated, deterministic logic, developed by Diogo Almeida, a ChatGPT co-inventor.
  • Jev utilizes a parallel sampling architecture, taking unstructured state as input and outputting type-safe structured values, avoiding text generation entirely.
  • The Jev architecture is trained via Reinforcement Learning for Calibrated Decisions (RLCD) to ensure correlated confidence scores with output accuracy.
  • Internal evaluations show end-to-end response latencies between 70-500 milliseconds, significantly faster than conversational frontier deployments (3-329 seconds).
  • Jev achieves execution speeds up to 193.6 times faster than consensus baselines (GPT-6 Astra and Fable 5.1) across fixed compute graphs.
  • Input processing costs are $0.042 per million tokens, a substantial reduction compared to standard conversational rates ($0.20 - $10 per million tokens).
  • Real-world testing demonstrated Jev resolving branching rules across Doom game state structures at 10 queries per second, incurring operating costs of approximately $7 per hour.
  • TypeSafe is onboarding engineering teams from Amsterdam, California, and London, beginning early developer access today.

    • 📝Summary


    TypeSafe, founded by a ChatGPT co-inventor, has emerged from stealth to introduce Jev, a system designed for automated decision-making within software. Developed over two years, Jev, created by Diogo Almeida, utilizes a parallel sampling architecture to execute structured probabilistic decisions directly in codebases. Unlike conventional language models, Jev takes unstructured state as input, delivering type-safe, structured values. Internal testing reveals execution speeds up to 193.6 times faster than existing models, with response latencies between 70 and 500 milliseconds. The platform’s design, incorporating Reinforcement Learning for Calibrated Decisions, targets applications like real-time feature extraction and data workflows, demonstrating successful operation across diverse environments including game state and encyclopedia directories. Early developer access commenced today, with initial onboarding underway from Amsterdam, California, and London.

    💡Insights

    ▼


    JEV: A NEW APPROACH TO PROGRAMMATIC DECISION MAKING
    TypeSafe, a newly launched company founded by a ChatGPT co-inventor, is introducing Jev, a specialized System One Model designed to automate deterministic logic within production codebases. Developed by OpenAI veteran Diogo Almeida, Jev represents a significant departure from traditional large language models, focusing on direct, type-safe probabilistic decision execution rather than text generation. This innovative approach addresses the limitations of conversational models in scenarios demanding precise, repeatable outcomes, particularly within software systems. The core of Jev’s design centers around parallel sampling architecture, enabling rapid and reliable decision-making by evaluating multiple options simultaneously.

    THE CORE TECHNOLOGY: RLCD AND PARALLEL SAMPLING
    Jev’s architecture is built upon Reinforcement Learning for Calibrated Decisions (RLCD), a novel training methodology. Unlike conventional frontier models trained via Reinforcement Learning with Human Feedback (RLHF) or Reinforcement Learning with Verifiable Rewards (RLVR), RLCD focuses on generating calibrated probabilities directly tied to execution logic. This ensures that confidence scores accurately reflect output accuracy. The system employs a hardware-aware parallel sampler that evaluates and delivers structured values in parallel, eliminating the need for autoregressive token generation. A key structural constraint restricts outputs to predefined schemas, preventing common issues like syntactic type failures and output hallucinations. The internal sampler efficiently handles high-cardinality selections, up to 255 discrete options, utilizing a two-stage process for independent scoring before selection.

    REAL-WORLD APPLICATIONS AND PERFORMANCE
    Initial testing of Jev demonstrates its potential across a wide range of applications. Production demonstrations showcase its ability to resolve branching rules within dynamic game states, such as those found in Doom, achieving 10 queries per second with operating costs of approximately $7 per hour. Furthermore, Jev has been successfully deployed in evaluating link selection across dense encyclopaedia directories, surpassing external non-reasoning models by avoiding hallucinated dead ends. Key performance metrics reveal execution speeds up to 193.6 times faster compared to conversational frontier deployments, with internal response latencies between 70-500 milliseconds. Input processing costs are significantly lower at $0.042 per million tokens, a substantial reduction from standard conversational rates. This parallel sampling approach generates structured states without the token generation overhead, allowing TypeSafe to provide output tokens without metering charges.

    Related Articles