Aanchal Mittal

← Architectures

Jev · a decision engine, not a chatbot

TypeSafe's Jev abandons autoregression entirely. Shared state is encoded once, every question runs as an isolated branch in one forward pass, and each branch returns a typed, calibrated decision — not text. Click any node to open it up.

Prefer open source? Laya is the open-weights alternative — same category, published architecture, Apache 2.0. Start there; come back here for the closed-source counterpart.

Why non-autoregressive matters

An LLM used as a decision function is a decoder loop pretending to be a switch statement. You wrap it in JSON mode, hope it picks a real enum value, retry on parse failure, and still get an off-schema string every few hundred calls. Jev's answer is to stop generating text at all: the output space is bounded before the call runs, so a malformed value is not something the architecture can produce.

The other half is speed. Because every question in a Jev call shares the same KV cache and resolves in one parallel forward pass, latency is roughly O(S) instead of O(Q · S). TypeSafe's public claim is 40-200× faster than a frontier LLM on comparable classification and gating workloads.

What "System One" is doing here

Diogo Almeida — co-inventor of RLHF, formerly at OpenAI — frames the automation gap as everyone building System 2 models when most software decisions want a System 1 answer. Fast, intuitive, non-verbal. A router doesn't need to think out loud; it needs to be right and calibrated. Jev is the argument that this class of task deserves its own model shape, not a wrapper over a chat model.

Where the reasoning didn't go

Jev is not a replacement for a reasoning model. It won't write your code, draft your email, or explain its answer. What it does is give the wrapper layer around a reasoning model — the routing, the classification, the gating, the safety checks — a purpose-built substrate. In a production stack you would use both: an LLM for the parts that generate, Jev for the parts that decide.

Sources