The ChatGPT second in 2022 taught AI to speak to folks. Considered one of its builders now bets the subsequent second is AI that talks to software program, not folks. TypeSafe AI launched Jev. Jev is transformer-based, however it’s not a big language mannequin. It doesn’t generate textual content. You ship a state and typed questions. It returns typed choices with chances that code can department on.
Is it deployable? Sure, as a hosted API in early entry behind a waitlist. TypeSafe has not printed weights, a parameter rely, or a self-hosting possibility.
What’s a System One Mannequin?
The title borrows from Daniel Kahneman’s break up between quick instinct and gradual reasoning. TypeSafe group argues RLHF tuned fashions for human choice. That produced chat, and overconfidence and mode dropping. These flaws maintain a human within the loop.
Jev makes use of a brand new stack: a brand new structure, a parallel sampler, and Reinforcement Studying for Calibrated Choices (RLCD). TypeSafe has not disclosed the structure.
How the Jev API Works
One endpoint handles the whole lot: POST https://api.typesafe.ai/v1/systemone. The physique carries state, mannequin, and a map of questions. The docs outline 3 query varieties.
| Primitive | Asks | Returns |
|---|---|---|
| Selection | Choose 1 possibility from a listing | selection, chances, confidence |
| Rating | Charge towards ordered ranges | rating, chances, confidence |
| Noul | Is that this assertion true? | noul, a chance from 0 to 1 |
Questions run in parallel and in isolation towards the identical state. TypeSafe says including questions barely modifications response time. A Selection helps as much as 255 choices.
from typesafe_sdk import Selection, Noul, TypeSafeClient
consumer = TypeSafeClient() # reads TYPESAFE_API_KEY
r = consumer.system_one(
state=ticket,
questions={
"department": Selection(
directions="Which team should handle this",
standards={"billing": "Payment issues", "technical": "Bugs"},
),
"is_urgent": Noul(directions="The message conveys urgency"),
},
)
print(r.solutions["department"].selection, r.solutions["is_urgent"].noul)Set up with pip set up typesafe-sdk (Python 3.10 or later). A JavaScript SDK ships as @typesafe-ai/sdk. The quickstart additionally covers cURL and an agent ability for Claude Code.
Confidence is the Product
Each Selection and Rating reply carries a confidence worth from 0 to 1. TypeSafe derives it from the form of the chance distribution. Within the docs instance, billing wins at 0.84. Confidence is just 0.596, as a result of technical nonetheless holds 0.159.
The docs counsel 3 paths. Act on excessive confidence. Evaluation the center. Ship low confidence to a human. Thresholds ought to scale with the price of a incorrect motion.
Pricing, Velocity, and the Benchmark Tremendous Print
Jev prices $42 per billion enter tokens. TypeSafe quotes current LLMs at $0.20 to $10 per 1M enter tokens. In its recorded demo, Jev completed in 0.114s for $0.000081. GPT-5.6 Terra took 8.566s for $0.013880.
The TypeSafe group claims it to be 193.6x quicker and 444.6x cheaper. These figures come from TypeSafe’s personal workflow evals. However maintain on listed below are some issues to remember:
- The reference reply is the common of GPT-6 Astra and Fable 5.1.
- TypeSafe’s capabilities group wrote the workflows.
- TypeSafe expects these features to take a seat on the excessive finish of actual use.
- TypeSafe says it can’t show the value is unsubsidized.
‘Zero hallucinations’ means schema matching is assured. The 0% determine is just not empirical. Solutions can nonetheless be incorrect.
What Builders are Constructing with Jev
Group initiatives appeared inside days of launch. Listed below are some examples:
- Command security: Vercel CEO Guillermo Rauch reported Jev as much as 18x quicker at p95 than GPT Luna, and extra correct. His submit stated the fx reviewer nonetheless ran on Luna. Engineer Pranit Sharma shared the benchmark.
- E mail triage:Bryo AI CTO Nikhil Mudholkar discovered Gemini barely extra correct, however 10 to twenty instances dearer.
- Browser brokers: Browser Use’s jev-ultrafast ran a Zürich to London Google Flights search in 7.1 seconds (video).
- Telephone brokers: Droidrun’s mobile-jev drove Uber on an actual Android telephone: 9 actions in about 21 seconds (video). No reserving was accomplished.
- Video scoring: jevmeter scores each sentence of a debate for about $0.05 (demo on X).
- Stay typing: Steve Krouse’s Typewriter updates 16 judgments as you sort (try it).
- Video games: Jev completed StarCraft’s first combat mission (video). It additionally runs the guards in heist-one (video).
- Agent guardrails: jev-guard charges every device name as deny, ask, or enable (78-second video).
- Knowledge and houses: pg-jev provides plain-language filters to Postgres. HA-Jev turns solutions into Dwelling Assistant entities.
Interactive Explainer
“+row(“billing”,0.84)+row(“technical”,0.159)+row(“gross sales”,0.001)+’
confidence 0.596. Billing leads at 0.84, but technical nonetheless holds 0.159. Confidence summarizes the form of the entire distribution, so your code can deal with this as a perhaps.
‘;
if(okay===”score”)h=”
score: 1.035
- Calm, just stating facts
- Frustrated but civil
- Very angry, strong language
confidence 0.842. The score is probability weighted, so it can land between your levels.
“;
if(okay===”noul”)h=”
noul: 0.999
“+row(“P(yes)”,0.999)+’
No separate confidence. A Noul is the chance that the assertion is true. A worth close to 0.5 means sure and no are equally probably.
‘;
$(“#viz”).innerHTML=h;requestAnimationFrame(perform(){requestAnimationFrame(perform(){$$(“#viz .fill”).forEach(perform(e){e.fashion.width=e.dataset.w+”%”});var m=$(“#mk”);if(m)m.fashion.left=(1.035/2*100)+”%”})});ping()}
$$(“#seg button”).forEach(perform(b){b.addEventListener(“click”,perform(){prim(b.dataset.p)})});prim(“choice”);
/* 3. gate: illustrative confidences */
var D=[[“Refund request”,0.97],[“Password reset”,0.95],[“Spam message”,0.93],[“Address change”,0.91],[“Plan upgrade”,0.84],[“Invoice dispute”,0.77],[“Partial outage”,0.66],[“Legal threat?”,0.52],[“Mixed languages”,0.41],[“One-word ticket”,0.28]],prev={};
perform gate(quiet){var lo=+$(“#lo”).worth,hello=+$(“#hi”).worth;
$(“#loV”).textContent=lo.toFixed(2);$(“#hiV”).textContent=hello.toFixed(2);
var b={A:[],R:[],H:[]};D.forEach(perform(d){var okay=d[1]
‘+d[0]+’‘+d[1].toFixed(2)+’
‘);prev[d[0]]=okay});
[“A”,”R”,”H”].forEach(perform(okay){$(“#b”+okay).innerHTML=b[k].be part of(“”);$(“#n”+okay).textContent=b[k].size});
$(“#gateSum”).textContent=b.A.size*10+”% fully automated, “+b.R.size*10+”% reviewed, “+b.H.size*10+”% sent to a person.”+(b.A.size>=8?” Aggressive: fine for low stakes, risky for payments.”:b.A.size+$(“#hi”).worth)$(“#hi”).worth=this.worth;gate()});
$(“#hi”).addEventListener(“input”,perform(){if(+this.worth=1000?”$”+Math.spherical(v).toLocaleString(“en-US”):v>=1?”$”+v.toFixed(2):v>=0.01?”$”+v.toFixed(3):”$”+v.toFixed(5)}
perform value(){var tok=+$(“#tok”).worth,calls=Math.spherical(Math.pow(10,+$(“#call”).worth)),pr=+$(“#price”).worth,out=+$(“#out”).worth;
calls=+calls.toPrecision(2);
$(“#tokV”).textContent=tok.toLocaleString(“en-US”);$(“#callV”).textContent=calls.toLocaleString(“en-US”);$(“#priceV”).textContent=”$”+pr.toFixed(2);$(“#outV”).textContent=out;
var j=tok*calls/1e6*JEV,l=tok*calls/1e6*pr+out*calls/1e6*pr*5,r=l/j;
$(“#dJ”).textContent=cash(j);$(“#dL”).textContent=cash(l);$(“#fL”).fashion.width=”100%”;$(“#fJ”).fashion.width=”max(4px,”+(j/l*100)+”%)”;
$(“#ratio”).innerHTML=Math.spherical(r).toLocaleString(“en-US”)+”x cheaper per dayInput alone is “+(pr/JEV).toFixed(0)+”x. Free output tokens add the rest. $1 buys “+(1/JEV).toFixed(1)+”M Jev input tokens.“}
[“tok”,”call”,”price”,”out”].forEach(perform(i){$(“#”+i).addEventListener(“input”,value)});value();
$$(“details”).forEach(perform(d){d.addEventListener(“toggle”,ping)});
})();
