Jev reportedly does tensor logic under the hood: differentiable IF args, no wasted gen tokens

StewartalsopIII · x · 2026-09-22

In a shared thread, notdevin argues Jev did two special things: (1) it effectively practices tensor logic without calling it that—making the args of an IF differentiable, since true/false judgments need gradients to be learnable, and a static context ("a book") is not enough; and (2) it doesn't waste generation tokens on context you already know. The take: don't assume intelligence will come from an LLM's context—it won't, and betting on that only leaves room for competitors.

Original post →

More from Models

Models channel →