The Register reports TypeSafe’s new model, Jev, returns typed probabilistic decisions, enabling faster, hallucination‑free AI interactions with software and games like Doom.
TypeSafe AI has unveiled Jev, a new AI model that delivers typed probabilistic decisions, promising faster and hallucination‑free interactions for software and games such as Doom.
What is Jev?
Jev is designed to return decisions that are both typed and probabilistic, meaning each output is accompanied by a confidence score and conforms to a predefined data type. This approach reduces the risk of nonsensical or unsafe responses that have plagued earlier generative models.
Why Typed Decisions Matter
Typed outputs allow developers to integrate the model directly into codebases without extensive validation layers, while the probabilistic component lets systems gauge uncertainty and act accordingly. The combination is especially useful for real‑time environments where speed and reliability are critical.
Jev in Action: Playing Doom
In a live demo, Jev controlled a Doom agent, interpreting game state data and issuing typed commands such as move_forward or fire_weapon. The model’s confidence scores helped the agent decide when to take aggressive actions versus retreating, resulting in smoother gameplay and fewer erratic moves.
- Typed outputs simplify API contracts
- Probabilistic scores enable risk‑aware decision making
- Reduced hallucinations improve safety in critical applications
Implications for Developers
By providing a predictable interface, Jev can be embedded in a range of software—from automated testing tools to interactive game bots—without the overhead of extensive error handling. TypeSafe AI suggests this could accelerate AI adoption in legacy systems that demand strict type safety.
The model also supports fine‑tuning on domain‑specific data, allowing teams to tailor its behavior for particular applications while preserving the core typed‑probabilistic framework.
Comments
No comments yet.