ChatGPT’s weekly audience has grown about twelvefold in under three years, from 100 million a week in November 2023 to more than 1.2 billion in October 2026. As AI is used more often, the variety of applications also increases, creating a demand for more versatile types of AI.
The AI we are all familiar with today takes a few minutes to process information and provides paragraph-long answers. While writing emails or making study sets requires this kind of precision, many potential applications do not. For example, when determining the risk score of a file, a precise essay down to the nitty gritty details of the risk score is not what the user wants. They just want the risk score to know whether or not they should open the file.
Daniel Kahneman first popularized the terms describing these two types of thinking, System 1 and System 2, through his book Thinking, Fast and Slow. He used the terms to describe the human brain’s two modes of thinking. System 1 thinking operates quickly, automatically, and with little to no conscious effort. However, it trades off detail in favor of speed, so its conclusions are susceptible to error. System 1 is the type of thinking that convinces you that two plus three equals six on a math test. On the other hand, System 2 thinking forces the brain to think analytically. It is slower, but more reliable.
Because of the error component in System 1 thinking, it has largely been ignored when it comes to creating artificial models to simulate the human brain. Up until now, AI has been modeled after System 2 thinking and has gained a reputation as a clunky, time-consuming, yet detail-oriented machine. In the last month, however, System 1 thinking based models were released, the first of which was Jev, created by TypeSafeAI.
When it first launched, Jev generated an enormous amount of buzz in the AI community. It was cheaper, faster, and had new applications—all without the typical AI hallucinations. Ironic, considering it is based on the mode of human thinking which makes the most errors. Unlike large language models (LLMs), which can take any block of text as input, Jev has a specific structure for the types of inputs it can accept, which prevents hallucinations. The model takes in a piece of information, a question about that information, and a list of options from which it must answer. Jev then returns the correct answer, along with a confidence interval for the answer. By design, it will not hallucinate new answer options.
Jev’s specific structure also makes it much cheaper to use. If the same information, question, and answer options are given to both an LLM and to Jev, the LLM will generate at minimum one sentence, whereas Jev will simply pick an answer option. The amount of information that Jev conveys is significantly less relative to the output of LLMs, so it uses fewer tokens (units of information AI models take as input or output, usually around two-thirds of a word). Fewer tokens means a lower cost; Jev averages out to be around 81 times less expensive than LLMs.
Jev is also much faster than an LLM. Given a question, LLMs cross-reference thousands of similar cases in order to predict the most likely answer. The process involves reading thousands of files and is therefore incredibly time consuming. Jev, on the other hand, is able to return an answer in far less time. Compared to the 3 to 329 second response time LLMs have, Jev has a 70 to 500 millisecond response time. Not only is the increase in efficiency convenient, but it also creates a new application for AI. Jev can be written into everyday computer programs without making significant sacrifices in speed.
Ultimately, the emergence of System 1 AI marks a fundamental shift in how artificial intelligence can be deployed. By prioritizing constrained outputs, minimal token usage, and sub-second response times over exhaustive prose, models like Jev unlock new use cases—ranging from immediate file security checks to real-time app interactions—that were previously impractical for traditional LLMs. That being said, LLMs still have various use cases in which models like Jev are infeasible. Models like Jev are not replacing LLMs; instead, they complement them, making AI more versatile than ever before.
