While visiting my cousin, who happens to be a surgeon, we got onto AI.

While visiting my cousin, who happens to be a surgeon, we got onto AI. I always manage to drag AI into the conversation with everyone.

He said today’s chat models feel unreliable and untrustworthy because of hallucinations. Hard to argue with that when the work is a call, not a paragraph.

I told him about Jev from TypeSafe, which just came out. It is a classifier. Super fast. Not a coding model. Not an essay machine. It is built for decisions. At dinner we used examples like cancer or not, polyp or not: the split-second clinical calls where every second can matter. That was our conversation, not a claim that Jev is cleared for clinical care.

What clicked for me is the failure mode. A chat model can invent an answer that was never on the list. A closed-set decision model cannot invent a new label. It can still pick the wrong allowed option. Different failure mode, and one you can design systems around. Confidence helps you act when sure and escalate when not.

There is also a light UI angle: this shape of model can help navigate interfaces quickly by choosing among options it can see. Brief aside. The bigger story for me is medical, risk decisioning, trading, and any low-latency work where inventing a new answer is already a failure. Ultra fast classification with low schema hallucination. That is the bit I wanted him to hear.

TypeSafe and Diogo Almeida are building that System One job.