If you picked the first one, you're in good company. It's what most AI apps do, and it looks something like this:

Then you wait for tokens to stream > parse the JSON > validate it > use 3 values out of it.

You just spun up a full generative model to make 3 small decisions, and most of that machinery went to waste. (If you picked a chain of prompts or a trained classifier instead, you're already circling the idea here.)

TypeSafe AI reckons that's the wrong tool for the job, and last week they shipped their answer: Jev, the first of what they're calling "System One" models. 

Instead of writing text, Jev takes some state and answers typed questions with probabilities.

state + questions → probabilities → code

You can think of it as a semantic if-statement. It reads the messy human context, makes the call, and hands your code a clean number to branch on.

You hand it some state:

Then you define the judgments you need. Jev gives you three primitives to work with:

Choice → pick one option
Score  → place something on an ordered scale
Noul   → answer a yes/no question as a probability

For example:

Jev makes the judgment; your code decides what to do with it. 

Now, let's build something.

Install the SDK:

Set up your client and some state:

Spell out the decisions you care about:

Back comes something like department.choice = "technical" with the full probability spread, and urgency.noul = 0.98. 

From there your application logic looks like it always has:

No generated function call, JSON repair loop, or model deciding your business policy. Jev makes the small fuzzy calls and your code handles everything deterministic.

This is where it gets fun. 

Say refundRequested only matters for billing tickets. The instinct with a normal LLM is to stage it: classify the ticket, wait, then ask about the refund only if it turned out to be billing.

Jev nudges you to ask everything at once and ignore the answers you don't need.

TypeSafe calls this speculative fan-out.

Every question runs against the same state in parallel, so you drop the long chains of serial calls, and their docs show batching 13 questions coming out faster and cheaper than firing them one by one. 

Once it clicks, you stop thinking in pipelines and start thinking in one wide question.

Most LLM code throws away uncertainty.

You get department = billing and move on. But billing at 0.97 (technical at 0.02) and billing at 0.38 (technical at 0.35) both give you the same answer, and only one of them is safe to automate.

Jev keeps that uncertainty around, so you can bake it straight into the product:

Your thresholds depend on the stakes: filing an email into a folder and issuing a refund deserve very different levels of nerve.

Jev is barely a week old and the community has completely run with it. The range is wild.

On the serious end, people are using it as a cheap judgment layer inside bigger systems.

Agent routing is the obvious one: before you send every request to an expensive reasoning model, ask a few nouls first, and only spin up the heavy stuff that's needed.

Others use it to check their own LLM output ("did this answer the question? does it make claims it can't back up?"), letting the generative model stay creative while Jev grades its work.

The Browser Use team wired it into an agent that books flights in seconds for fractions of a cent, and Vercel benchmarked it as the safety reviewer inside its fx coding tool and is moving it to the default, up to 18x faster than the model it replaced.

My favorite of the practical bunch is semantic database fields: run your conversations through Jev to get columns like churn_risk = 0.61 and purchase_intent = 0.91, then query them with plain SQL.

I also saw someone ship a Chrome extension that scores your X timeline and blurs the AI slop as you scroll, while someone else built one that skips YouTube sponsor segments with no crowd-sourced database.

TypeSafe themselves have Jev playing Doom at about 10 decisions a second and racing across Wikipedia by choosing between hundreds of links, and people have since had it play Tetris, run NPCs that judge you, and play Smash Bros. against itself.

Personally I want to try it on coding-agent tool routing, browser-agent guardrails, email prioritization, moderation queues, and semantic observability for agents.

We've spent the last couple of years shoving nearly every AI problem through the same shape: messages → model → text, even when all we wanted was a boolean, a category, a score, or a probability.

We paid for a whole essay to get one number. 

Jev makes those outputs first-class without pretending to replace reasoning models.

There's a mountain of production code out there running llm.generate, then JSON.parse, then if, where the model is doing far more work than the app needs.

That last if is where Jev gets interesting, and I think a lot of us are about to swap it in.

Until then,
Vaibhav 🤝🏻

If you read till here, you might find this interesting

#AD 1

Something is being built inside Tesla's facilities that almost no one is talking about — yet.

According to insider sources, Elon Musk has quietly developed a breakthrough product he claims will be "10x bigger than the largest product in history." The target launch date is July 22. And when it drops, the window to position quietly will already be closing.

Most investors will hear about this after the stock has moved. You don't have to be one of them. Our analyst named 3 stocks positioned to ride the launch — with entry guidance, price targets, a bonus 4th supply-chain pick, and a 3-phase playbook for when to buy, add, and take profits.

See the 3 stocks now.

#AD 2

Every day, 4.5 million readers turn to 1440 for their factual news fix. We sift through 100+ sources to bring you a complete summary of politics, global events, business, and culture — all in a brief 5-minute email. No spin. No slant. Just clarity.

Join for free today!