ChatGPT's limitations left its co-creator, Diogo Almeida, unsatisfied despite its groundbreaking capabilities. Almeida, a former OpenAI researcher instrumental in developing the chatbot and the reinforcement learning from human feedback (RLHF) technique, recognized that optimizing for human language presented challenges for automation, as computers operate on a different linguistic framework.
Driven by this realization, Almeida departed from OpenAI two years ago to establish TypeSafe AI. His new venture aims to overcome these limitations, and this week, TypeSafe AI unveiled Jev, a transformer-based model that diverges from the conventional LLM paradigm. Instead of generating text, Jev produces calibrated decisions through probabilities. This innovative approach yields substantial benefits: the model is remarkably cost-effective, processes information at high speeds, and inherently prevents hallucinations because outputs are predefined. Moreover, its output tokens are free, and input tokens are metered at a significantly lower rate.
Jev has already captivated developers, leading to a surge in demand that temporarily overwhelmed the company's API. Its primary utility lies in software automation, offering a more economical and robust method for integrating intelligence into code. For instance, Pranit Sharma, a software engineer at Vercel, noted a significant improvement when replacing OpenAI's ChatGPT Luna 5.6 with Jev for classifier tasks, achieving results 5 to 18 times faster and with enhanced accuracy. Similarly, Nikhil Mudholkar, CTO of Bryo AI, found Jev to be significantly more cost-efficient than Gemini for classifying business emails, despite a slight difference in accuracy. Jev's distinctive confidence scores, which provide real probabilities, were particularly appealing for automating workflows. Beyond replacing LLMs, Jev can also augment them by acting as a 'smart check' to prevent misbehavior, a cost-effective solution for monitoring LLM agent traces and thwarting jailbreaks.
The vision behind Jev, named after the 19th-century economist William Stanley Jevons, is to enable widespread deployment of intelligence by lowering its cost, mirroring Jevons' paradox. Almeida anticipates that this will lead to an emergent and distributed landscape of intelligent software, reminiscent of the early internet. While the model's architecture remains undisclosed, outside observers speculate it builds upon open-weight LLMs. TypeSafe AI refers to Jev as a 'System One model,' emphasizing intuition over reasoning and a focused approach to specific tasks. Almeida credits Jev's success to its exclusive training on synthetic data using 'reinforcement learning from calibrated decisions,' a technique he considers his most significant breakthrough. As Jev currently stands alone in this niche, competitors are expected to emerge as its practical value becomes increasingly evident. Almeida clarifies that TypeSafe AI is not a 'frontier lab' driven by fear or hype, but rather a company focused on delivering tangible intelligence.
