TypeSafe AI, a startup founded by former OpenAI researcher Diogo Almeida, has released a new transformer-based model named Jev. Unlike large language models (LLMs), Jev does not output text but instead produces probabilities, which the company refers to as "calibrated decisions."
Almeida, who helped build ChatGPT and invent reinforcement learning from human feedback (RLHF), stated that he was disappointed with LLMs' utility for automation due to their optimisation for human language. He left OpenAI two years ago to establish TypeSafe AI to address this issue.
The company claims that eschewing language makes Jev both inexpensive and fast, and because users pre-define outputs, it cannot hallucinate. Its output tokens are free, and input tokens are metered by the billion. Developers have shown significant interest, with demand briefly overwhelming the company's API.
For example, Pranit Sharma, a software engineer at Vercel, reported that replacing OpenAI's ChatGPT Luna 5.6 with Jev for a safety classifier resulted in outcomes 5 to 18 times faster and with improved accuracy. Nikhil Mudholkar, CTO of Bryo AI, found Jev to be 10 to 20 times cheaper than Gemini for classifying business emails, noting its useful confidence scores.
Jev is also seen as a tool to augment LLMs, potentially monitoring agent traces and preventing 'jailbreaks'. Almeida hopes the model's low cost will lead to widespread deployment of 'smart software'.