Diogo Almeida, a former OpenAI researcher who contributed to the creation of ChatGPT and the pivotal reinforcement learning from human feedback (RLHF) technique, has launched a new AI startup called TypeSafe AI. Almeida was motivated by his frustration with large language models (LLMs) like ChatGPT, which, despite their linguistic prowess, fall short in automating software tasks effectively because they optimize for human language rather than computer logic. His company recently unveiled Jev, a novel transformer-based model that diverges from typical LLMs by generating “calibrated decisions” — probabilities rather than natural language text outputs.

Jev’s unique design allows it to be faster, cheaper, and more reliable, as it prevents the common problem of hallucination (false or nonsensical outputs) that plagues many LLMs. Users define the output parameters upfront, and the model processes input tokens in the billions rather than millions, enabling highly scalable software automation. This innovation has caught the attention of developers eager to incorporate more efficient AI-driven intelligence into their applications. For instance, at Vercel, Jev replaced OpenAI’s Luna model for command safety classification and achieved performance improvements ranging from five to eighteen times faster with enhanced accuracy.

Industry experts also appreciate Jev’s ability to provide confidence scores as real probabilities, which is invaluable for automating workflows. Unlike some alternatives that might be slightly more accurate but far more costly, Jev offers a balance of affordability and robustness. Beyond replacing LLMs in certain scenarios, Jev can complement them by acting as a monitoring “agent of agents” that detects and prevents errors or inappropriate outputs, such as jailbreak attempts in AI systems. Its speed and low cost also make real-time model routing feasible, which involves quickly determining the best model to handle a particular task.

Almeida envisions Jev driving a wave of distributed, smart software that proliferates broadly and intuitively, akin to the early internet’s decentralized growth. While details of Jev’s architecture remain confidential, its development relies heavily on synthetic data and a new training method called “reinforcement learning from calibrated decisions.” The model is named after economist William Stanley Jevons, symbolizing hopes that decreasing AI costs will spur widespread adoption. Though currently unique, Jev may soon inspire competitors as its advantages become more apparent, with TypeSafe AI planning to expand its model lineup to new applications in the future.