Jev is the rediscovery of BERT — $40M, 17M views, a 2018 architecture

Jev by TypeSafe AI is the most hyped AI launch of 2026 — and the CEO agreed on day one that it is, in his words, "exactly right" to call it basically a zero-shot classifier. I got early access, tried it, and walked through what it actually is under the hood: an encoder with classification heads scoring probabilities in a single pass. Sound familiar? That's BERT. From 2018. Recorded live, no script — screen share, receipts, and the receipts are all cited. Chapters: 0:00 The most hyped AI launch of 2026 0:45 Am I missing something? My first Jev tests 1:31 The launch post: RLCD and "calibrated decisions" 2:27 Structured program states in, typed answers out 3:13 The sampling trick: all probabilities in one pass 4:05 Predicting options instead of generating text 4:54 The eval problem: comparing against other models' average 5:42 What the 70–500ms latency actually buys 6:29 Where Jev fits: smart if-statements 8:12 AI-powered workflows and map-reduce over data 9:07 What LLMs were actually doing all along 10:44 The 255-option ceiling 11:31 The encoder, drawn out 12:19 What models like BERT already did in 2018 13:07 The architecture: what Jev almost certainly is 14:08 More than 255 options? Two-stage scoring 14:59 If this sounds strangely familiar — BERT, masked outputs 15:48 Why the industry forgot bidirectional encoders 16:37 Schemas on top: Pydantic and structured outputs 18:01 We haven't saturated what encoders can do 19:08 "Maybe you are missing something" — the fair reading 20:12 The hype machine and the VC cycle 21:10 We are in an era of deception 22:12 How to read launches like this 23:09 Stay skeptical and keep on learning Sources cited in the video: TypeSafe's launch post, the Hacker News thread, independent benchmarks (Novel Cognition's Jev File, webofmike's jev-benchmark), the BERT paper (arXiv:1810.04805), ModernBERT (arXiv:2412.13663), and OpenJev. https://toothfairyai.com #AI #Jev #BERT #MachineLearning #ToothFairyAI