ChatGPT Co-Inventor Launches Jev, an AI Built to Make Decisions Instead of Chat

ChatGPT Co-Inventor Launches Jev, an AI Built to Make Decisions Instead of Chat

What if an AI did not need to write a single sentence to be useful?

That is the idea behind Jev, a new AI model launched by Diogo Almeida, a former OpenAI researcher who says he helped develop the instruction-following methods that became part of the research behind ChatGPT.

After two years working quietly on the project, Almeida has launched TypeSafe AI with a very different approach to artificial intelligence.

Instead of building another chatbot that writes paragraphs, explains things, or generates code, Jev is designed to make decisions that software can use directly. And that difference could be a big deal for businesses trying to automate thousands or even millions of small decisions.

Meet Jev: AI That Does Not Need to Chat

GPT, Claude, and other large language models are designed to communicate with humans.

You ask a question.

They generate an answer.

Jev takes a different route.

You give it information and define the decisions you want it to make. Instead of responding with a paragraph, Jev returns structured answers, probabilities, and confidence scores that software can immediately act on.

Think of questions like:

Should this customer be routed to sales?

Is this transaction suspicious?

Which AI model should handle this request?

Does this message require human review?

Which category does this request belong to?

For these kinds of tasks, generating a long answer can actually be unnecessary. Jev is designed to make the decision and get out of the way.

Why Jev Says It Cannot “Hallucinate”

This is where the headline gets interesting.

TypeSafe says Jev “cannot hallucinate” because it does not freely generate text.

Its outputs are constrained to the types and choices defined by the developer.

So instead of an AI inventing an answer that was never requested, the system is designed to return a structured decision that software can understand.

But there is an important catch.

Jev can still be wrong.

If the model makes the wrong classification or chooses the wrong option, that is still an AI error. The “can’t hallucinate” claim is mainly about preventing uncontrolled, off-format generation, not guaranteeing that every decision is correct. That distinction matters.

The Real Goal: Faster AI Automation

TypeSafe is not trying to replace ChatGPT or Claude with another chatbot. The company is targeting the layer underneath AI applications.

Businesses constantly make tiny decisions behind the scenes.

Which customer should see which offer?

Should a request go to a human?

Which model should handle a particular task?

Should a transaction be flagged?

Which information should be included in the next AI prompt?

Traditionally, companies might use complicated rules, traditional machine-learning classifiers, or expensive language models to handle these decisions. Jev wants to become the fast-decision-making layer between those systems.

TypeSafe calls this approach “System One.”

Jev Is Built for Speed

And TypeSafe is making some ambitious performance claims.

The company says Jev can respond in roughly 70 to 500 milliseconds, depending on the task, and claims it can be 20–200 times faster and 40–400 times cheaper than comparable frontier models. The company also says output tokens are free, with input priced at $0.042 per million tokens. Those numbers sound dramatic.

But there is an important caveat: these are TypeSafe's own reported figures, and the model is only just entering early access. Independent testing across a wide range of real-world workloads will be needed to determine how those claims hold up.

Still, the underlying idea is interesting. If a business needs an AI to make millions of small decisions, shaving hundreds of milliseconds and high costs from every decision could add up quickly.

A New Training Method

Jev also comes with a new training approach called Reinforcement Learning for Calibrated Decisions, or RLCD.

Traditional AI systems such as ChatGPT have used reinforcement learning from human feedback, or RLHF, to help models produce responses that people prefer. TypeSafe says RLCD focuses on something different: getting the probabilities behind decisions properly calibrated.

In other words, the system is not just trying to pick an answer. It is also trying to give software a useful indication of how confident that decision is. That could become particularly useful when businesses need to decide when AI should act automatically and when it should hand a case over to a human.

So Where Could Businesses Use Jev?

Quite a lot of places.

Imagine an online store receiving 100,000 customer messages.

Instead of sending every message to a large language model, a smaller decision system could quickly determine:

Sales? Support? Refund? Urgent complaint?

A fraud platform could use a similar system to decide whether a transaction should be:

Approved → Reviewed → Blocked

A company running several AI models could use Jev to determine which model should handle each request. And an AI application could use it before or after a larger language model to classify information, route requests, check conditions, or decide what should happen next. That is the market TypeSafe is chasing.

TypeSafe Is Betting on a Different Future for AI

TypeSafe emerged from stealth with about $40 million in seed funding led by DCVC, according to the company's funding announcement. Almeida founded the company alongside Erik Gafni and Sasha Sheng.

The company's argument is straightforward:

AI does not always need to talk.

Sometimes, AI just needs to decide.

And if that decision can happen quickly, cheaply, and in a format software can immediately understand, businesses may be able to automate far more of their everyday operations. That does not mean Jev replaces ChatGPT, Claude, or other frontier AI models.

In fact, the more interesting possibility may be the opposite. Jev could work alongside them. A large language model could handle the complicated reasoning and communication, while Jev handles the thousands of smaller decisions happening around it.

The Bigger AI Shift

For years, the AI race has largely been about making models better at talking.

Better writing. Better reasoning. Better coding. Better conversations.

Jev is asking a different question: What if the next big AI opportunity is not another chatbot but the decision-making layer that powers everything behind the scenes?

That is the bet TypeSafe is making.

And if the company's early performance claims survive wider testing, businesses may have a new way to automate decisions without paying a frontier language model to generate a paragraph every time.

AI does not always need to say something. Sometimes, it just needs to choose what happens next.