TypeSafe AI debuts Jev, a model using "Reinforcement Learning for Calibrated Decisions" to produce typed probabilistic decisions that software can use directly (Thomas Claburn/The Register)

TypeSafe AI, a startup backed by $40 million in funding, unveiled Jev, a novel AI model tailored for machine-to-machine interaction, on Tuesday. Unlike traditional language models that generate natural language responses, Jev delivers probabilistic decisions suitable for other software or AI systems. Jev's design is based on Type safety, a programming technique that helps catch errors when unexpected data types are processed.

By providing structured values, Jev eliminates the need for text parsing and validation typically required for handling responses from large language models (LLMs). This can prove beneficial in scenarios where AI interactions must adhere to a limited set of outcomes, such as playing video games like Doom when presented with structured game state data.

However, Jev's primary application is likely to be in business workflows, like sorting customer service issues. The model operates by taking a state value, which can be a JSON object or a simple string, and responding with structured answers containing probabilities. For instance, a customer service query might receive a response indicating the department handling the issue, along with a confidence score.

Jev is a System One model that uses Reinforcement Learning for Calibrated Decisions (RLCD) architecture, which allows for faster response times compared to traditional LLMs. The company claims Jev is 40x-200x faster than LLMs and costs significantly less. Jev's output is structured, meaning it doesn't contain natural language and thus avoids the risk of hallucinations associated with LLMs.

Jev is particularly well-suited for AI automation software, real-time applications, and data classification tasks.