A new kind of AI model from a ChatGPT inventor is thrilling developers
TypeSafe AI, founded by ex-OpenAI RLHF inventor Diogo Almeida, released Jev, a non-LLM transformer outputting calibrated probabilities for cheap software automation.
TypeSafe AI, a startup founded two years ago by former OpenAI researcher Diogo Almeida, who helped invent RLHF, released a transformer-based model called Jev that is not a large language model. Instead of text, Jev outputs probabilities the company calls 'calibrated decisions,' making it fast, cheap (output tokens free, input metered by the billion), and hallucination-free since users define outputs in advance. Almeida says the 'System One model' is trained exclusively on synthetic data via a technique he calls 'reinforcement learning from calibrated decisions,' and demand was high enough to briefly knock out the company's API. Vercel reported 5-18x faster results than ChatGPT Luna 5.6 for command-safety classification, while Bryo AI found Gemini slightly more accurate but 10-20x more expensive for email classification.
- Jev outputs calibrated probabilities instead of text, making it cheap, fast, and unable to hallucinate by design.
- Vercel replaced ChatGPT Luna 5.6 with Jev for command-safety classification, seeing 5-18x speed gains.
- Proposed uses include monitoring LLM agents to prevent jailbreaks and real-time model routing.
- Almeida says the model is trained only on synthetic data using 'reinforcement learning from calibrated decisions.'
- Architecture is undisclosed; outside observers suspect it builds on an open-weight LLM.
- Products
- JevChatGPT Luna 5.6Gemini
- Organizations
- TypeSafe AIOpenAIVercelBryo AIEarendil
- AI models
- JevChatGPT Luna 5.6Gemini
From the original article
4 min read at techcrunch.com
Jev, a new kind of AI model, is showing developers a cheaper and faster path to software intelligence.
By Tim Fernholz · techcrunch.com
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.