This weekend X was full of articles on Jev, and the hype had the shape hype always has. 500 articles they don’t say much of anything.

Writing code is the fun part. The boring middle of the loop is Jev promises to save on. Which tool next. Continue, retry, ask me, or stop. Is this output safe enough to act on without me. Every one of those is a yes, a no, or a pick-one, and I’ve been paying a frontier chat model to answer them in prose my code then parses and hopes about.

Jev is a bet that the prose was always the bug. Here’s what’s actually under it.

RLHF is Reinforcement Learning from Human Feedback: you train a model with human preference signals so its answers match what raters like, which is how chat models got good at sounding helpful.

TypeSafe came out of stealth on September 15 with roughly $40M in seed money led by DCVC. Two years dark before that. Diogo Almeida is the co-founder and CEO. He co-authored InstructGPT, he’s acknowledged on ChatGPT, he’s on the GPT-4 technical report. Google Brain before that.

So when he argues that RLHF was a wrong turn, he’s arguing against his own work, which is the only version of that argument worth reading.

His case: RLHF trains a model to please the person in the loop. That’s the correct objective when a person is in the loop. In an agent harness, nobody is. So you get a model tuned to sound right to a human who isn’t there, and every harness any of us has built sits downstream of that mismatch.

They borrow Kahneman for the name. System One is the fast, intuitive call. Their training method is Reinforcement Learning for Calibrated Decisions, RLCD, and the thing it optimizes for is a calibrated probability rather than a sentence you’d enjoy.

The pitch against language models is bold. Language models emit strings humans like, and your code parses, validates, and hopes. Jev only emits values you declared up front. Schema match is the product, so they claim type errors are impossible by construction. Language models also sample one token after another. Jev answers all the declared questions in one parallel shot.

Before the latency claim lands, you need the demo TypeSafe keeps looping on X.

They wired Jev into Doom. Not by watching pixels like a human. The game feeds Jev structured state as text and configs: where you are, what is around you, what moves are legal. Jev returns typed decisions many times a second: turn, shoot, move. No chat reply. No “I think you should strafe left.” Just the decision values the game loop can consume.

That is the Matrix beat. Humans stare at the rendered frame. Jev never sees the frame. It sees the details between the pixels: the state the renderer was built from. Text and configs, not vision. Same world, different interface.

That last part is why the clip keeps circulating. Nobody needs an AI to play Doom. Ten structured decisions a second against live game state is a latency shape chat models hate. You see the difference in the demo more than you feel it reading a pricing table.

Their published numbers: 70 to 500ms end to end, input at $0.042 per million tokens. Output is free. There’s almost nothing to meter.

Two distribution doors opened on top of the launch.

TypeSafe started letting people off the waitlist, with starter credit chatter bouncing around the usual channels.

Vercel put typesafe-ai/jev on AI Gateway and wired it through AI SDK 7’s experimental evaluate API, with Choice, Score, and Boolean as the primitives. They listed it free on September 19 under a promotion that ends September 25, 2026. That’s two days from this posting.

New model class, plus an easy Gateway path, plus a free window with a hard date. That’s why this is eating Sunday timelines instead of sitting in a stealth PDF.

TypeSafe published workflow evals putting Jev on a wild Pareto curve against wrapped LLMs. 193.6x faster and 444.6x cheaper at the top end. Read those footnotes the way we read burn tables.

They score agreement against the average of GPT-6 Astra and Fable 5.1, not against labeled ground truth. Their own team authored the workflows. They say the headline multipliers are the high end and the general range is more like 40x to 200x faster. Independent writeups flagged the same thing.

So run the skepticism the way you’d run cost math. Not back-of-napkin vibes. What can you falsify tonight?

Speed per call, price per call, and “never emits off-schema.” Those you can poke before bed.

“As smart as Astra or Fable on System One tasks” needs your own routing jobs, not their homepage. Early Jev is text and structured state, no images yet. Cardinality caps at 255. This is not a drop-in for Cursor Agent, Grok Bot, or Claude Code. Those still talk. Jev refuses to.

I haven’t dogfooded Jev on my own products. This is a frontier read, not a keep call.

If it holds up in a real harness, the interesting slot isn’t replacing Claude. It’s that boring middle I opened with. Which tool or subagent next. Continue, retry, ask, or stop. Is this output refund-shaped enough that I want a human on it. Can I map-reduce a pile of tickets without burning a frontier chat model on every row.

That’s harness work, and it’s worth being precise about what changes. The model was never the trust boundary. The permission surface and the decision schema are. A fast calibrated router makes the harness cheaper to run. It does not make the IC smarter, and anyone selling it as both is selling you something.

Same lesson as firing the agents I didn’t need: overhead is the tax. Skills on demand beat persistent chatters that only route. Jev is trying to be the router as a primitive.

Chat models got superhuman at talking. Automation still stalls on decisions software can trust at 100ms and a fraction of a cent.

I’m not wiring production on a Sunday night screenshot. I’m putting it on the bench next to the meters I already watch. If the falsifiable parts hold on my own jobs, it earns a seat.

If not, it was a good Doom clip.