Skip to content
what-is-jev?-the-ai-that-doesn’t-talk-to-you

What Is Jev? The AI That Doesn’t Talk to You

Written by Coursera Staff • Updated on

Jev AI is a decision model for software. Learn how Jev makes structured decisions and what makes it different from other AI models.

[Featured Image] A smiling professional in a white button-down shirt sits at their office desk in front of a sunny window and types,

Key takeaways

  • Jev is a “System One” artificial intelligence (AI) model designed for classification and decision-making within software systems, with a focus on enabling efficient automation, which distinguishes it from conversational AI models like ChatGPT.

  • The Jev AI model can make thousands of decisions in a single pass, saving significant time compared to traditional AI models.

  • In AI, Jev has the potential to work with traditional models, offering guidance to classify information, label data, and enable better decisions in a fraction of the time, while allowing the traditional AI model to handle reasoning or text generation.

Jev AI is a decision model that can answer three types of questions: Choice, Score, and Noul, providing probability and confidence scores to help you interpret the results. Continue reading to learn more about this different type of AI, including details about the company that created it, its potential uses, and the pros and cons associated with Jev.

To become more familiar with AI and refine your skills, enroll in the Google AI Professional Certificate. This eight-course series takes an average of eight hours and offers hands-on lab assignments that can help you build skills such as adopting a collaborator mindset, data wrangling and visualization, debugging, and application development.

What is Jev in AI?

Jev is TypeSafe AI’s first public System One model, and it works differently from ChatGPT, Claude, Gemini, and other large language models (LLMs) you may have experience with. Instead of chatting with you, this decision model is designed to provide predefined answers rather than generating text or images, and it can classify information and assign it scores for probability and confidence at high speed, which is particularly useful for developers. Potential uses range from routing support tickets to the right department to filtering information to guiding AI agents.

While announcing its release in September 2026, the company’s founder, Diogo Almeida, noted, “Jev achieves similar levels of intelligence on System One tasks compared to existing LLMs, while being two orders of magnitude faster and more efficient. While Jev gives up string generation, it’s optimized for structured outputs and can’t hallucinate,” [1].

This statement refers to the structured output that Jev provides. It cannot deviate from the instructions you give it. User interpretation and judgment remain important, as you will need to ensure that even high-probability and high-confidence results align with your goals and parameters.

It allows you to ask three primary types of questions:

  • Choice: This type of question prompts Jev to select an option from a defined set, making it ideal for model routing and classification. (i.e., Should this ticket go to technical support or billing?) You give Jev the question and a list of up to 255 possible answers, and it will return a probability for each option, along with a confidence score.

  • Score: With this type of question, Jev rates your question as compared to a rubric, which you can define. (i.e., What level of risk would doing XYZ create?) You ask Jev your question and provide an ordered list of two to 10 ratings, with descriptions for each. Jev will score the probabilities for each rating level and provide a confidence score.

  • Noul: This type of question provides binary answers that judge whether a statement is true. (i.e., Does this ticket mention a billing issue?) You provide Jev with a yes-or-no question, and it will deliver results, with answers close to 1 indicating “yes” and close to 0 indicating “no.”

Read more: What Is Probabilistic Programming?

Who is behind Jev AI?

Diogo Almeida worked at OpenAI and co-invented the model training technique, reinforcement learning from human feedback (RLHF), along with InstructGPT, two key methods that power ChatGPT. Despite the success, Almeida felt that optimizing models for human language created a gap in the need for models to communicate effectively with computers and to optimize automation and autonomous decision-making.

Almeida founded TypeSafe AI to focus on machine-native AI research, working to advance the type of AI that can run in the background of computers, quietly guiding intelligent decisions. As part of its manifesto, TypeSafe AI notes, “After years and trillions of dollars invested in the industry, most software still isn’t meaningfully intelligent, and life is mostly the same for the average person. That should tell us something: the bottleneck isn’t raw intelligence. It’s that today’s intelligence is hard to build on,” [2].

The link between “Thinking Fast and Slow” and Jev

TypeSafe AI calls Jev a System One model, a term based on one of two categories of human thinking described in Daniel Kahneman’s book Thinking, Fast and Slow. In the book, Systems One and Two refer to human cognition, which can be either data- or theory-driven. System One is fast and intuitive, taking in data and quickly forming patterns from it. System Two is slower, deliberate, more logical, and capable of creating step-by-step order. Both types of thinking are critical. System One produces an impression. System Two, on the other hand, either approves or denies the System One result. When both work together, they produce efficient, reliable knowledge.

TypeSafe AI classifies Jev as a System One model based on its ability to quickly evaluate the content you provide it and return probabilities and answers that guide decisions.

The paradigm shift behind Jev

With AI agents, you have a continual cycle: input, LLM decisions, and model evaluation. The process repeats until it fulfills the task at hand. When integrated into software applications, the technology becomes both slow and expensive.

Enter Jev, a decision engine capable of making up to thousands of parallel decisions quickly. Its architecture drives efficiency, and rather than generating tokens as chat models do, Jev produces outputs in a single pass. It relies on reinforcement learning for calibrated decisions (RLCD) instead of RLHF, with calibrating the probability of the returned results as its primary objective. This training method, named as such by TypeSafe AI, rewards the model when the stated probability matches reality, removing human preference ratings from the equation.

When you want to score or classify something or get a yes-or-no decision, Jev can deliver the results quickly, providing the deterministic outcomes that many enterprise AI solutions lack. Now, you can complement the LLM powering AI agents, using LLMs for generation and reasoning, and Jev for structured decisions and model routing.

How is Jev used?

Developers are already using Jev primarily to deliver structured decisions, serve as a classifier, and oversee agents and LLMs. For example, you can deploy Jev to ensure an agent finishes its task or to have agents transfer a request to a human team member for review when the request meets specified criteria. Other examples of how you might use Jev in classification-dominant workflows include:

  • Classifying emails according to set parameters

  • Sorting comments and related community posts on social media

  • Routing support requests to the right queue according to their urgency

  • Ensuring agent requests go to the optimal LLM based on complexity

  • Quickly labeling vast amounts of data according to your criteria

  • Helping simple robots make sound decisions before taking action

What is Jev vs. LLM?

LLMs generate text, while Jev classifies information based on structured outputs that you define. LLMs can also classify information; however, they do so at a higher cost and slower speed. Whereas LLMs can converse with users, TypeSafe AI designed Jev to make structured decisions in AI and software. It can work with LLMs to guide their decision-making abilities.

Pros and cons of Jev

Jev offers a mix of benefits and potential limitations that will likely evolve as it matures. For example, when TypeSafe first released Jev, it could only process text. In October 2026, the company announced it added multimodality for current and future Jev models, image support, and longer context lengths.

Also notable, Jev returns structured outputs based on the options you give it. It cannot generate freeform responses outside of that structure. However, while Jev cannot hallucinate in the traditional sense, it can choose the wrong option. For example, if you ask an LLM to classify emails in four categories, it could invent a fifth. However, with Jev, it will only work within those four categories. Still, it could potentially choose the wrong one, a distinction that’s critical to understand when working with it.

Advantages of working with Jev AI

Jev is a general-purpose AI that you can use across many classification scenarios. Its speed and low cost have many developers considering the costs of using AI.

A few of its advantages include the following [1]:

  • Efficiency: Can take unstructured data and output it using a predetermined structure in a single query

  • Accuracy: Calibrated confidence for higher degrees of consistent accuracy

  • Speed: The end-to-end response time is 40 to 200 times faster, averaging 70 to 500 milliseconds

  • Cost: Output tokens are free, and input tokens are $0.042 per million, making it a fraction of the cost associated with traditional LLMs

Potential drawbacks of working with Jev AI

Additionally, it’s worth noting that lower costs may come with a trade-off: because Jev doesn’t use multi-step reasoning like LLMs do, it can’t explain its answers or debug its performance. For scenarios in which you need to rely on human-in-the-loop workflows, you will still need to use an LLM.

Other limitations to be aware of include the following:

  • Jev can’t generate text or images, or “chat” with you, as it operates strictly as a specialized decision model.

  • Known accuracy issues include potentially unreliable counting, treating dates as text, and an inability to work with open-ended context scenarios.

Explore our free AI resources

Keep pace with industry news, emerging technologies, and professional pathways with a subscription to Career Chat, our LinkedIn newsletter. Then, check out some of our other free resources to continue learning more about AI:

  • Find your path: AI Learning Roadmap: From Beginner to Expert

  • Watch on YouTube: Artificial Intelligence Jobs Unlocked: Pathways to the Future

  • Learn from an expert: The AI Advantage: 9 Questions with UC Davis AI Instructor Sadie St. Lawrence

You can also build your skills, expand your knowledge, and accelerate your growth with Coursera Plus. Your annual or monthly subscription gives you access to more than 10,000 programs from over 350 esteemed partners.

  • cplus logo

    Give yourself skills that last, now 40% off

    40% off [Google, Adobe, IBM, and Microsoft logos]

Written by:

Coursera Staff

Editorial Team

Coursera’s editorial team is comprised of highly experienced professional editors, writers, and fact…

This content has been made available for informational purposes only. Learners are advised to conduct additional research to ensure that courses and other credentials pursued meet their personal, professional, and financial goals.

colind88

Back To Top