This AI can't write a single word. I tested whether that matters.

Narration uses an AI clone of my own voice. The presenter is an animated avatar of me. A new kind of AI model launched on 15 September 2026, and it can't write. TypeSafe AI's Jev returns a choice, a score or a yes/no, each with a probability, and nothing else. Its makers claim it is far faster than a chat model and knows when it might be wrong. I spent a weekend testing both claims with my own evidence: 100 customer emails routed by Jev and by GPT-5.4 mini, three passes each, then 500 real text messages from a public research collection. What held up: about 3.5 times faster than a fair opponent, not the 50 times in the launch material. Same answer on every pass. And when it said it was at least 95% sure, it was right every time. Part 1 of 3. Part 2: I try to break it. Part 3: I put it in my app. CHAPTERS 0:00 Cold open 0:30 How we got here 1:14 What Jev is, and isn't 2:04 How I tested it 2:41 Claim one: speed 3:27 Claim two: is its confidence honest? 4:25 Verdict, and part two ABOUT THIS VIDEO One person, one weekend, a few hundred examples. This is not a benchmark. The test emails were written with Claude's help and the labels are mine; the text messages are the SMS Spam Collection (Almeida and Gómez Hidalgo, UCI Machine Learning Repository), which is old and well known, so both models have probably seen it. Every figure on screen comes from a results file recorded on 21 September 2026 through the Vercel AI Gateway. Prices and access checked before publishing; if Jev changes, a pinned comment will say so.