New AI Model Focused on Software Automation from Former OpenAI Researcher

Diogo Almeida, who worked on ChatGPT and reinforcement learning at OpenAI, founded TypeSafe AI two years ago as a result of his frustration with language-optimization-focused approaches in artificial intelligence, and has now announced his new model called Jev.

The Emergence of TypeSafe AI and the Jev Model

Former OpenAI researcher Diogo Almeida left OpenAI two years ago and launched TypeSafe AI after realizing that human language optimization is not fully sufficient for automation processes.

This week, the company launched the new transformer-based Jev model, which is not a large language model and produces probabilities and calibrated decisions instead of generating text.

Working Principle and Advantages of the Jev Model

Unlike traditional large language models, Jev does not output text thanks to users pre-defining the outputs, thereby eliminating the possibility of hallucinations.

Designed to produce probabilities and calibrated decisions, this new model has managed to attract great interest from software developers with its low-cost and fast structure.

Real-World Tests in the Software World

Vercel integrated the Jev model instead of the OpenAI model in its security classification processes, achieving higher accuracy rates along with results that are 5 to 18 times faster.

Another company, Bryo AI, compared the Jev model with Gemini in email classification processes, verifying the model's cost advantage and the actual probability scores it provides.

Training Process and Capacity to Support Other Models

Named in memory of economist William Stanley Jevons, Jev was trained entirely on synthetic data using the reinforcement learning technique from calibrated decisions.

This model also has the capability to support large language models by supervising agent traces, preventing security vulnerabilities, and performing model routing at a low cost.