TypeSafe AI Releases Jev: A Fast Decision Model and Open-Source Variant Kev

TypeSafe AI, founded by ChatGPT co-creator Diogo Almeida, released Jev, a System One model for rapid, type-safe decisions, avoiding text generation. Jev runs 20-200x faster than LLMs, with applications in gaming and browsing. Cognition's Kev, an open-source Jev-like model, was also released, with Kev-8B achieving 79.6% accuracy on out-of-domain data.
Key insights
TypeSafe AI, led by ChatGPT co-creator Diogo Almeida, released Jev, a System One model designed for fast, type-safe decisions, not text generation. It uses RLCD training to provide calibrated probabilities, enabling 20-200x faster decisions than LLMs, with applications in gaming and browsing.
Cognition's open-source Kev, a Jev-like model, offers 0.6B/4B/8B parameters, trainable on a single H100 (8B in 83 minutes), achieving 79.6% accuracy on out-of-domain data, close to Jev's 85.7%.
Jev's design, rooted in the Jevons Paradox, aims to reduce decision costs, potentially increasing AI usage. It enables 'intelligent if statements' for tasks like routing and classification, with demos including real-time DOOM control and Wikipedia racing.