Diogo Almeida’s journey from OpenAI’s ChatGPT to his new venture, TypeSafe AI, is a tale of disappointment and innovation. Almeida, who co-invented RLHF, is now focusing on a novel model, Jev, which eschews text outputs for probabilities, promising cheaper and faster automation.

Unlike large language models (LLMs), Jev generates calibrated decisions that developers can act upon, making it particularly useful for software automation. For instance, Vercel, a company building agentic infrastructure, saw a 5 to 18 times increase in speed and accuracy when switching from OpenAI’s Luna to Jev.

The model’s architecture, built on synthetic data using a technique called “reinforcement learning from calibrated decisions,” aims to make intelligent systems more widespread and cost-effective. Almeida envisions this leading to a future where intelligence is seamlessly integrated into everyday software.

While Jev shows promise, its utility could lead to a new wave of competition. Armin Ronacher, CTO of Earendil, predicts that other companies will soon follow suit, capitalizing on Jev’s efficiency.

Almeida remains cautious, emphasizing the model’s practical applications over grandiose visions. For now, Jev stands as a testament to the relentless pursuit of smarter, more cost-efficient AI solutions.