Almeida spent years at OpenAI grappling with the realization that optimizing for human speech is inherently inefficient for machine tasks. While traditional LLMs struggle with hallucinations and high operational costs, Jev functions as a specialized decision engine. Because it produces raw probabilities instead of prose, the model avoids the unpredictability of generative text, allowing developers to integrate intelligence into software with significantly higher speed and lower overhead. Early adopters are already seeing performance gains; Vercel reported that Jev processed safety reviews 5 to 18 times faster than previous OpenAI implementations.
Ex-OpenAI Researcher Ditches Language Models for Pure Probability
Diogo Almeida, a co-inventor of reinforcement learning from human feedback, has launched Jev, a transformer-based model that rejects natural language output entirely. By focusing on calibrated probabilities rather than conversational text, the startup TypeSafe AI aims to solve the automation hurdles that currently plague large language models.
Beyond raw speed, Jev provides confidence scores that allow engineers to set automated thresholds for reliability. This capability enables the model to act as a low-cost supervisor for larger, more expensive systems, effectively monitoring agent behavior or routing tasks without the massive price tag of a full-scale LLM. While Almeida remains protective of the underlying architecture, he emphasizes that the model is trained exclusively on synthetic data. By prioritizing intuition over complex reasoning, TypeSafe AI is betting that the future of software lies in distributed, low-cost intelligence rather than the centralized, high-hype models currently dominating the industry.
Comments (0)
No comments yet. Be the first!