The Next 700 Jev-Like Decision Model

At the end of the day it's always data. Our experiments with Kapteeni (and a lot of other people's experiments with models like BERT) has proven that fine-tuning a small open-weight model for classification and other simple tasks is something that can be done relatively easily, at least much easier than collecting enough data to make a difference. We'll see if the data safety sensitive people would even consider Jev for the bootstrapping.

The next 700 decision models (if there would be 700 of them, that is) would be in-house Qwens (or the like) specialized in the one aspect the companies want. I've seen some people on Hacker News trying to make Jev sound superior by calling claiming that it performs "best" while being "general purpose". Unfortunately being general purpose or not is one of the least concerns when the "data → model → decisions → more data → better models &arr; more accurate decisions" pipeline starts working.

This is, of course, not a dig at Typesafe AI per se. Instead, this is obviously a dig at the "Jev-like model craze" in early October 2026.

— Z Lin