TypeSafe's Jev: A Non-LLM Model That Can't Hallucinate — SkimNews
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- TypeSafe AI released Jev, a transformer-based model that outputs "calibrated decisions" (probabilities) rather than text, with input tokens metered by the billion rather than the million and no output token costs.
- Diogo Almeida, a former OpenAI researcher who helped invent RLHF, left to found TypeSafe AI, arguing that optimizing for human language made models "not useful for automation because computers speak a different language."
- Vercel swapped OpenAI's ChatGPT Luna 5.6 safety classifier for Jev and reported results 5 to 18 times faster with greater accuracy, per software engineer Pranit Sharma.
- Bryo AI CTO Nikhil Mudholkar tested Jev against Google's Gemini for email classification, finding Gemini slightly more accurate but 10 to 20 times more expensive, and praised Jev's real probability scores for automating workflows.
- Demand for Jev was so high that TypeSafe briefly lost the ability to serve users from its API, the company told TechCrunch.
- Jev is trained exclusively on synthetic data via a technique Almeida calls "reinforcement learning from calibrated decisions," and the model is named after 19th-century economist William Stanley Jevons, whose paradox describes how falling costs expand consumption.
Why it matters: For developers paying per-token to OpenAI and Google, Jev's probability-output approach undercuts costs by an order of magnitude on classification and routing — Vercel's 5–18x speedup and Bryo's 10–20x cost gap are concrete benchmarks that make non-LLM architectures suddenly competitive. TypeSafe's all-synthetic-data bet also sidesteps the copyright and scraping fights now consuming frontier labs.
Ask SkimNews
