There’s a new way to do AI in town, and it’s not RLHF.
A newer training approach is moving past RLHF. What changes when models learn from outcomes rather than human ratings.
RLHF taught chat models to please people. Fluent answers. Great for writing. Painful when you are paying frontier prices for tiny yes or no choices.
Jev from TypeSafe is a decision model for those tiny forks. Matthew Berman walks through it plainly. You give it the situation. It picks from clear options and tells you how sure it is. No essay. TypeSafe trains that with RLCD, Reinforcement Learning for Calibrated Decisions.
Here is the money story, and it is the point. Berman and TypeSafe say System One work like this can run up to about two hundred times faster and about four hundred times cheaper than burning a frontier chat model on every micro-choice. Soft vendor framing until you measure it yourself. In their pitch, output is free or near free, and input is priced tiny. Browser Use made it tangible: they used Jev to find flights from Zurich to London, and their demo finished in about seven seconds for less than a cent.
So stop paying chat-model prices for routing. Use decision models for the small forks. Keep chat models for writing, coding, and deep thinking.