ChatGPT Visionary Unleashes 'Jev' Model for Programmatic Logic Breakthrough

TypeSafe, a new artificial intelligence company founded by OpenAI veteran and ChatGPT co-inventor Diogo Almeida, has emerged from two years in stealth mode to launch its groundbreaking Jev model. This innovative system is designed to automate programmatic decisions, offering a specialized System One Model that directly executes structured probabilistic decisions within production codebases, thereby providing an alternative to conventional conversational language models for specific software systems requiring automated deterministic logic.

A core differentiator of the Jev model lies in its complete abandonment of text and string generation. Unlike traditional large language models that autoregressively generate sequential text tokens, Jev takes an unstructured state as input and swiftly outputs type-safe structured values in a single parallel query. This architectural approach specifically targets deterministic code integration, inherently preventing syntactic type failures and output hallucinations, which are common challenges with more generalized AI models. As Diogo Almeida, Founder at TypeSafe, succinctly puts it: “Think of Jev as a frontier-intelligence function call: unstructured state in, typed probabilistic decisions out.”

The engineering team behind Jev developed the platform around an alternative training methodology known as Reinforcement Learning for Calibrated Decisions (RLCD). This approach stands in contrast to methods like Reinforcement Learning with Human Feedback (RLHF) or Reinforcement Learning with Verifiable Rewards (RLVR), commonly used by other frontier models to satisfy human conversational preferences or isolated programmatic checks. RLCD specifically trains Jev to produce calibrated probabilities on execution logic, ensuring that the returned confidence scores directly correlate with the accuracy of the output. This focus on calibration is critical for applications demanding high reliability and precision.

Further distinguishing its architecture, Jev utilizes a hardware-aware parallel sampler. This mechanism evaluates and delivers all structured values simultaneously, moving away from the sequential nature of autoregressive token generation. The inherent structural constraint of this design restricts outputs to predefined schemas, effectively eliminating the need for complex parsing pipelines and extensive guardrail wrappers that are often indispensable when deploying traditional large language models in business automation workflows. The internal sampler is robust, capable of handling high-cardinality selections of up to 255 discrete options through a sophisticated two-stage scoring process conducted independently prior to the final selection.

The performance metrics of the Jev model are particularly striking. Internal evaluations conducted on TypeSafe's West Coast servers recorded end-to-end response latencies ranging between 70 and 500 milliseconds. This starkly contrasts with the 3 to 329 seconds often observed across conversational frontier deployments. Furthermore, published workflow evaluations, which measured multi-step decision branching against consensus baselines from leading models like GPT-6 Astra and Fable 5.1 across fixed compute graphs, demonstrated execution speeds up to an astounding 193.6 times faster. Beyond speed, Jev also offers significant cost efficiencies; input processing is priced at a mere $0.042 per million tokens, a substantial reduction from standard conversational rates that typically span $0.20 to $10 per million tokens. Crucially, because its parallel sampling architecture generates structured states without requiring autoregressive token generation passes, TypeSafe provides output tokens without any metering charges.

Real-world deployment testing has showcased Jev's capabilities across diverse and demanding scenarios. Production demonstrations reveal its efficacy in resolving complex branching rules within high-speed game states and intricate web traversal trees. In dynamic stress testing, engineers successfully operated a real-time reactive bot navigating Doom game state structures at a rate of 10 queries per second, with operating costs approximately $7 per hour. A secondary Wikiracing test evaluated Jev's link selection across dense encyclopaedia directories, demonstrating its ability to complete traversals in fewer steps than external non-reasoning models by intelligently evaluating options without generating hallucinated dead ends. These field tests confirm Jev's primary target applications include real-time feature extraction, petabyte-scale data workflows, output verification layers, and automated branching logic where traditional hand-written rules often prove brittle and difficult to maintain.

TypeSafe has officially opened early developer access for its Jev model and has begun onboarding engineering teams from its deployment waitlist, marking a new chapter in the automation of programmatic decisions with precise, high-performance AI.