A new kind of AI model from a ChatGPT inventor is thrilling developers (techcrunch.com)

Diogo Almeida, a former OpenAI researcher and co-inventor of reinforcement learning from human feedback (RLHF), has launched a new startup called TypeSafe AI, aimed at improving AI model utility in real-world applications. This week, the company introduced its transformer-based model, Jev, which diverges from traditional large language models by outputting probabilities—termed "calibrated decisions"—instead of text. This approach not only reduces operational costs and speeds up processing but also eliminates issues like hallucinations by allowing users to define expected outputs beforehand. Jev's design is especially appealing for software automation, achieving remarkable performance improvements, such as delivering results five to 18 times faster than existing models like OpenAI’s ChatGPT Luna. Its unique ability to provide confidence scores facilitates intelligent decision-making in workflows and augments the reliability of LLMs by acting as a monitoring tool. Almeida envisions Jev fostering the development of "smart software" across a range of applications, striving for a widespread, emergent deployment reminiscent of the early internet. As interest in Jev surges, Almeida’s commitment to synthetic data also positions TypeSafe AI as a potential leader in this emerging subfield, reflecting a shift towards more practical and efficient AI solutions in the software industry.