AI Tech / news
Former OpenAI Researcher's TypeSafe AI Debuts Jev, a Non-Language Model for Software Automation
TypeSafe AI, founded by ex-OpenAI researcher Almeida, has released Jev, a transformer-based model that outputs calibrated decisions instead of text. Developers have shown such strong demand that the company briefly lost API capacity.
Jev is not a large language model. Rather than generating text, it produces probabilities that TypeSafe AI describes as "calibrated decisions," an approach intended to make software automation cheaper, faster and free of hallucination because users define the outputs in advance.
The pricing structure reflects that design. Output tokens are free, while input tokens are metered by the billion rather than the million.
Almeida helped build ChatGPT as an OpenAI researcher and then worked on reinforcement learning from human feedback, the training technique widely credited with shaping the current AI era. He left OpenAI two years ago to start TypeSafe AI, arguing that strong human-language capabilities had not translated into useful automation because computers operate in a different language.
"We have lightning in a bottle, and yet it is not useful," he said, describing the problem he has pursued since then.
Developers responded quickly to the release. Demand was high enough that TypeSafe AI temporarily lost the ability to serve users through its API.
The model's clearest use case so far is software automation, with developers treating it as a cheaper and more robust way to add intelligence to their code.