The Startup That Ripped the Chatbot Out of the Chatbot
TypeSafe AI's Jev doesn't write a single word. It just answers yes/no and picks-one questions 40 to 200 times faster than an LLM, for a fraction of the price. The catch: it can only ever be as good as the multiple-choice question you hand it.
What was checked, and what could not be.
Verified independently: Almeida's OpenAI background and TypeSafe founding (Forbes, TechCrunch, Wikipedia's Jev (AI model) entry, all consistent). $40M seed / DCVC / $200M valuation, sourced to Forbes (Rashi Shrivastava), Sep 15, 2026 — a single-source figure but from a named reporter at a mainstream outlet, not a press release quote. Launch date (Sep 15, 2026) and current version (jev-1.13.0) per TypeSafe's own docs as reflected in the Wikipedia entry — could not independently verify the version number against a primary TypeSafe page, flagging as company-stated.
The 40-200x speed / 40-400x cost claims, and the specific peak figures (193.6x, 444.6x), are TypeSafe's own benchmarks — the company's own technical notes admit possible best-case bias, and TS2.tech's coverage led with that caveat. Treated in the piece as an estimate/vendor claim, not a fact, per house rule 2.
The Vercel 18x figure and Cua's early-adopter status are third-party reported (The Rundown AI, easternherald.com) rather than TypeSafe's own materials — flagged in the piece as more credible for that reason, but still a single-report figure I could not find corroborated by a second outlet or by Vercel directly.
Could NOT verify: a standalone technical paper from TypeSafe (none exists as of this writing, confirmed by multiple outlets); any on-the-record reaction from Sam Altman, Dario Amodei, or Anthropic specifically about Jev — searched directly and found none. A widely covered Amodei essay ("Pace the Frontier," on coordinating AI safety standards across labs) broke the same week and got covered alongside Jev in some roundups, but the two stories are separate; I did not conflate them and did not manufacture a connection or reaction that doesn't exist.
The attempt to source the original 'Jev is a reward model sold as a product' analysis at startuphub.ai returned a 404 — could not confirm it's still live at that URL. I did not link it in the dossier for that reason and instead described its argument (Plackett-Luce ranking + Brier-score calibration, the reward-model reframing) as reported secondhand via a working fetch of that content earlier in reporting, attributed generally to 'an independent technical breakdown circulated after launch' rather than to a named outlet, since I can no longer confirm the link resolves. This is a case worth double-checking before publish: if you have the working URL, send it and I'll cite it directly.