Diogo Almeida, a former OpenAI researcher instrumental in inventing reinforcement learning from human feedback (RLHF), has launched a new AI startup called TypeSafe AI to address limitations he sees in current language-focused models like ChatGPT. Almeida believes traditional large language models are optimized for human language but fall short in automation because computers operate differently. In response, TypeSafe AI recently introduced Jev, a novel transformer-based AI model that does not generate text but instead produces probabilistic “calibrated decisions.” This design enables Jev to offer faster, cheaper, and more reliable outputs without the hallucination issues common in language models.

Since its release, Jev has attracted significant developer interest, briefly overwhelming the company’s API capacity due to high demand. Developers highlight Jev’s cost-effectiveness and speed, as well as its ability to provide confidence scores, which are highly valuable for automating software workflows. For instance, Vercel experienced much quicker and more accurate classification performance replacing OpenAI’s Luna model with Jev, while Bryo AI found Jev more affordable than competitors like Gemini despite a slight trade-off in accuracy. Jev is also seen as a useful tool to complement traditional LLMs, for example by monitoring outputs to reduce errors or hallucinations while maintaining efficiency.

TypeSafe AI keeps many technical details of Jev's architecture private but notes that it was trained exclusively on synthetic data using a new approach called “reinforcement learning from calibrated decisions.” Unlike typical LLMs that focus on reasoning through text, Jev is described as a “System One” model that relies on intuition and is tailored for specific tasks. The startup aims to foster widespread adoption of AI intelligence by drastically lowering costs and integrating intelligent automation more deeply into software ecosystems, inspired by economic principles behind increased usage when cost falls.

Beyond Jev's standalone role, there is growing discussion among AI leaders about balancing rapid AI development with safety. Anthropic CEO Dario Amodei has called for “pacing the frontier” of innovation through independent safety evaluators and coordination among major labs in democratic countries. However, Nvidia CEO Jensen Huang publicly opposes calls for a slowdown, emphasizing continued rapid progress. Industry insiders debate whether practical consumer choice or government regulation can effectively manage AI risk as the technology advances, especially since major AI companies now primarily serve enterprises rather than individual users.