Jev picks an option and reports confidence instead of generating text
Jev, built by RLHF co-inventor Diogo Almeida, chooses among given options and returns a confidence score in 70-500ms with no output tokens, though the claimed 193x speedup and benchmark methodology are questioned.
THE GUY WHO HELPED BUILD CHATGPT JUST BUILT AN AI THAT CAN'T WRITE A SINGLE WORD. And developers can't stop talking about it. It's called Jev. Diogo Almeida co-invented RLHF, the training behind ChatGPT. Then he disappeared into stealth for 2 years asking one question... Why are we using chat models to make yes or no decisions? Is this spam. Approve or reject. Refund or not. You ask an LLM and it writes you a paragraph. Just to say "yes." Jev skips the writing completely. You give it the options → it picks one → it tells you how sure it is. → 70 to 500 ms per decision → $0.042 per million input tokens → Output tokens cost nothing. There aren't any. But the speed isn't the interesting part. The confidence score is. 95% accurate is useless if you don't know which 5% is wrong. Jev tells you every time. Sure? Let it run. Unsure? Send it to a human. That's what makes it usable in production. Now the catch nobody's reposting. The 193x faster number is from TypeSafe's own tests. Independent measurements so far land between 1.16x and 6x. Still faster. Just not 193x. And their benchmark scores "correct" as agreeing with two frontier LLMs. So it measures agreement, not truth. They say so themselves. The launch still hit 1,500+ points on Hacker News in a day and briefly knocked their API over from demand. Because the idea is right even if the headline is loud. Most of what we pay LLMs to do was never a conversation. It was a checkbox.