TypeSafe unveils Jev model for software automation — TechCrunch
Startup TypeSafe AI has unveiled the transformer model Jev, which developers see as a tool for software automation. It is not a large language model and, instead of text, outputs probabilities or, in the company’s definition, “calibrated decisions.” TechCrunch reports on the launch.
The company was founded two years ago by former OpenAI researcher Diogo Almeida, who participated in the creation of ChatGPT and the development of the reinforcement learning from human feedback method, or RLHF. Almeida believes that language models work well with human language but are not always suitable for automation, as computer systems use different interaction formats.
Speed and cost
TypeSafe says users define Jev’s possible outputs in advance, so the model does not generate text and cannot hallucinate in the usual sense associated with language models. The company also claims lower costs and higher operating speed: output tokens are free, while input is measured in billions rather than millions. After its launch, demand for Jev was so high that the company’s API was briefly unable to serve users.
Vercel engineer Pranit Sharma said his company replaced its ChatGPT Luna 5.6-based classifier with Jev to check command security. According to him, after the replacement, results were obtained 5 to 18 times faster and with higher accuracy.
More current news is available on the UA.News Telegram channel Telegram.
Agent monitoring and model routing
Bryo AI CTO Nikhil Mudholkar compared Jev with Gemini in a business email classification task. In his test, Gemini was slightly more accurate but cost 10 to 20 times more. Mudholkar separately noted Jev’s confidence scores, which he believes are suitable for workflow automation.
Jev can also complement language models by tracking AI agents’ actions and helping prevent jailbreak attacks. Earendil CTO Armin Ronacher noted that the model could be useful for routing requests between different models by determining whether a particular task requires a specific model.
TypeSafe calls Jev a “System One” model, focused on intuition rather than reasoning and on performing a specific task. According to Almeida, the model is trained exclusively on synthetic data using an approach the company calls reinforcement learning from calibrated decisions. TypeSafe plans to release new versions of Jev for other modalities.
