This week, TypeSafe AI, founded by former OpenAI researcher Diogo Almeida, released Jev, a new transformer-based model designed for automation tasks. Unlike traditional large language models, Jev does not output text but instead generates probabilities or 'calibrated decisions', aiming to provide fast and cost-effective solutions for software automation.
Jev is engineered to be highly efficient, with free output tokens and input tokens metered by the billion. This design choice eliminates the risk of hallucinations common in language models and significantly reduces operational costs. The model's speed and efficiency have attracted significant interest, temporarily overwhelming the company's API capacity due to high demand from developers.
Technically, Jev stands apart by not being a large language model but rather a 'System One model', which focuses on intuition rather than reasoning. It is trained exclusively on synthetic data, employing a method called 'reinforcement learning from calibrated decisions'. Almeida emphasizes that the model's architecture is built to optimize for these specific tasks, though details remain closely held.
The primary audience for Jev includes software developers and companies looking to integrate intelligent automation into their systems. Use cases highlighted include replacing traditional LLMs for tasks like command classification and business email sorting, where it offers increased speed and cost savings. Additionally, Jev can augment existing LLMs by acting as a low-cost, efficient check on their outputs, preventing errors like jailbreaks.
Work implications: Jev could enhance workflows in software development and automation by providing a more reliable and affordable alternative to LLMs, particularly for tasks requiring real-time decision-making and classification.
Originally reported by TechCrunch