TypeSafe unveils decision-focused AI model Jev

TypeSafe has introduced Jev, a transformer-based AI model specifically crafted for software-level decision-making rather than generating conversational text. The San Francisco-based company, founded by former OpenAI researcher Diogo Almeida, revealed Jev as its inaugural "System One" model after a two-year development period in stealth mode.

Unlike conventional large language models that generate strings of text, Jev returns typed decisions and probability estimates that software can employ directly for tasks like classification, routing, scoring, verification, and guardrails. These bounded judgments are faster and cheaper compared to traditional language models. TypeSafe claims Jev can perform decisions in 70 to 500 milliseconds and charges $42 per billion input tokens, with no cost for output tokens.

Vercel, a developer platform, integrated Jev into its AI Gateway, describing it as a probabilistic decision model capable of returning Choice, Score, and Boolean answers concurrently. This structure enables software to automate high-confidence cases while directing uncertain ones for review. Initial testing by Vercel's Pranit Sharma showed Jev outperforming OpenAI models in a command reviewer classifier by five to 18 times faster while maintaining higher accuracy.

Bryo AI's CTO, Nikhil Mudholkar, compared Jev with Google's Gemini for classifying business emails, finding Jev to be significantly cheaper despite slightly lower accuracy. The model's training approach, dubbed Reinforcement Learning for Calibrated Decisions (RLCD), relies heavily on synthetic data and is designed for machine-facing automation.

Jev's probability scores, returned with each decision, empower developers to set thresholds for automatic software actions. However, developers must exercise caution, as a probability does not guarantee correctness, and production systems must assess when errors are acceptable. TypeSafe emphasizes that Jev cannot hallucinate, meaning it cannot fabricate open-ended text outside predefined outputs, but it can still make errors.