Jev, developed by ex-OpenAI researcher Diego Almeida, is a novel AI model that eliminates language generation to function as a fast, cost-effective, and type-safe classifier producing structured outputs like choices, scores, or yes/no answers. This approach addresses common LLM flaws such as over-explanation and hallucinations, making Jev ideal for real-time applications like moderation and gaming, and represents a significant shift towards efficient, reliable AI decision-making.
Last week marked a significant shift in the AI landscape with the introduction of Jev, a novel model developed by ex-OpenAI researcher Diego Almeida and his company, Typesafe AI. Unlike traditional large language models (LLMs) like ChatGPT, Jev cannot generate language, write code, or produce essays. Instead, it functions as a highly efficient classifier that operates without language output, addressing a major flaw in existing LLMs: their tendency to over-explain or generate unnecessary content. Jev is designed to be extremely fast, cost-effective, and free from hallucinations, making it a groundbreaking tool for AI applications requiring quick, reliable decisions.
Jev’s innovation lies in its approach to input and output, which mimics a type-safe programming language such as TypeScript. Users send strongly typed questions that must return one of three specific response shapes: a choice, a score, or a binary yes/no answer. This strict schema matching guarantees type safety and eliminates the possibility of type errors, ensuring that the model’s outputs are predictable and structured. This design philosophy aligns with the concept of a “system one” model, inspired by Daniel Kahneman’s theory of fast, intuitive thinking, contrasting with the slower, more deliberate “system two” models like GPT-6 or Claude.
For developers, Jev offers substantial practical benefits. Its speed and low cost—claimed to be hundreds of times cheaper and faster than major LLMs—make it ideal for real-time applications. For example, it has been used to implement AI moderation in apps, such as identifying and banning non-horse users on a horse-themed dating platform. Additionally, developers are leveraging Jev for NPC behavior in video games and even creating real-time AI calculators. Despite its efficiency, Jev’s outputs are not always deterministic, and it provides a calibrated confidence score with each response, indicating the likelihood of correctness based on reinforcement learning techniques.
The exact architecture of Jev remains undisclosed, with the company keeping details under wraps and hinting at a future research paper. While some critics argue that Jev is similar to earlier zero-shot classifiers developed over the past decade, Typesafe AI does not acknowledge these predecessors. Moreover, open-source efforts like OpenJV have emerged, replicating Jev’s interface and functionality using existing models without additional training, demonstrating the growing interest and experimentation around this new approach.
Overall, Jev represents a paradigm shift in AI by removing language generation from the equation and focusing on fast, type-safe classification. Its potential to revolutionize AI integration in applications is significant, especially for developers seeking cost-effective, real-time solutions. The video also highlights MX, a sponsor offering an easy-to-use video API that complements AI workflows, underscoring the expanding ecosystem of AI tools available to developers today.