The new architecture answers structured questions in parallel, providing confidence scores and typed choices rather than token-by-token text generation. TypeSafe reports that its Jev model delivers responses in 70 to 500 milliseconds, operating 20 to 200 times faster and 40 to 400 times cheaper than comparable large language models. This System One approach enables software to act directly on AI judgments without the need for a translation layer to parse or validate sequential text.
Training for the model relies on Reinforcement Learning for Calibrated Decisions, or RLCD, to align output probabilities with the accuracy of the model's answers. The system provides output tokens for free and was built by ChatGPT co-inventor Diogo to support AI automation. In a week of testing by developers at Every, the model performed 25 times faster and at a 600 times lower price than a Fable-level model.
Key sources
- SOURCE@stocksavvyshay“first “System One” model built to answer structured questions in parallel instead of generating text token by token”x.com
- SUPPORT@omarsar0“TypeSafe built a new architecture that answers structured questions in parallel, with RLCD training its probabilities”x.com
- SUPPORT@rohanpaul_ai“TypeSafe reports 70-500ms responses and 40-200x faster performance than comparable LLMs”x.com
- SUPPORT@danshipper“in our testing was 25x faster and 600x lower priced”x.com
- SUPPORT@ryan_t_lowe“enable a new paradigm of AI automation (that will grow alongside the current LLM + reasoning paradigm, which is complementary)”x.com
- SOURCEmarketbrief.now