TypeSafe AI Launches Jev Decision-Making Model

TypeSafe AI launched Jev, a decision model from a ChatGPT co-inventor that outputs typed, probability-calibrated decisions instead of text, running 20 to 200 times faster than LLMs. Demos span Doom, Minecraft, and browser agents, with output tokens free.
Key insights
Jev Outputs Typed, Calibrated Decisions Instead of Text
Jev does not generate free text: output structures and values are predefined in advance, making type errors mathematically impossible, and every output carries what the company calls an epistemically honest probability. The design removes the ability to fabricate a pleasing string when uncertain, leaving only the option of stating confidence, such as being 60% certain. The company claims parity with frontier LLMs on System One tasks while being two orders of magnitude faster and cheaper, achieved not by rebuilding intelligence but by channeling pretrained-model intelligence into calibrated decisions.
RLCD Argument: Pleasing Humans Produces Hallucination
Drawing on CEO CompleteSkeptic's keynote Jev Creator: Why RLCD beats RLHF, the company argues reward models create a GAN-like asymmetry in which a model's lack of confidence is detected and punished, teaching models to discard uncertain patterns and fake confidence — making hallucination an inherent product of preference optimization. The anecdote of a fart sound effect praised as ambient music illustrates LLMs inventing flattering strings when they do not know an answer. Jev's response removes the root cause: it was never optimized to please humans, so the capability-versus-obedience tension found in RLHF-trained agents does not arise.
Calibration Turns Human Review into an Optional Switch
The keynote's core product logic: if a model is right 95% of the time and can tell you when it is in the 5%, the task can be automated. Existing LLMs are overconfident and inconsistent when asked for confidence levels, a behavior RLHF teaches, while Jev lets developers set thresholds so high-confidence decisions execute autonomously and low-confidence ones escalate to humans. This reframes human-in-the-loop from a forced error corrector into an optional control, with the company describing the target tasks as automatable once calibrated decisions are available.
Positioned as Smart If-Statements, Not a Chat Assistant
Jev targets classification, routing, scoring, extraction, branching, and guards — places where handwritten rules are too brittle and LLMs too expensive, slow, or unreliable. The API form is rebuilt: structured program state goes in and typed decisions come out, rather than a message stream. The company contrasts this with the Claude Code era of assistance, arguing that if coding agents were purely RLVR-based they would look very different. Its stated mission is that AI needs an interface software could depend on, described as smarter software rather than just-in-time software.
Economics Named After the Jevons Paradox
The name Jev honors William Stanley Jevons, whose paradox observed that more efficient steam engines increased total coal consumption. The company applies the same logic to intelligence: every order-of-magnitude drop in per-decision cost increases consumption of intelligence by more orders of magnitude. Output tokens are free, described as too cheap to be worth billing, and the company says it is already profitable while continuing to cut prices, betting on a task space where the same intelligence is reused nearly for free rather than competing on per-unit intelligence prices.
Early Adopters Run Games, Browser Agents, and LLM Judging
Within three days of launch, Jev drew wide attention with deployments in Doom, Minecraft, browser agents, and LLM judging, presented as ten striking cases. A Doom demo ran 10 queries per second at roughly $7 per hour, and a Wikiracing demo showed software directly depending on AI decisions. The posts describe Jev as a fundamentally different kind of AI: it does not chat, it makes decisions, 20 to 200 times faster.