Specialized judgment-only AI model Jev gains Silicon Valley attention

Developed by Typesafe AI, the lightweight model Jev offers rapid, intuitive decision-making at a processing cost of approximately 0.3 won, significantly undercutting the higher costs of Claude or GPT models.

The stereotype that artificial intelligence (AI) must be good at everything is being broken. Instead of writing complex text or coding, ultra-lightweight AI models that only make intuitive decisions—such as answering "yes" or "no" or assigning scores—are gaining attention among developers and in Silicon Valley.

According to an analysis by Choi Ji-ung, CEO of YouCanLabs, reported by T Times TV, the recent trend in the AI industry is creating "AI that is good at only one thing." At the center of this movement is Jev, developed by Typesafe AI, which was founded by Diogo Almeida, one of the researchers from ChatGPT. While Jev cannot compose the sophisticated sentences seen in generative AI, it is a model specialized in making rapid judgments for specific situations.

"Hire an elementary student instead of a professor"... 'System 1' AI that only judges

CEO Choi compared the characteristics of Jev to "System 1," a component of the human cognitive system. If System 1 represents answers that pop into the head intuitively without much thought, models that require meticulous reasoning, such as ChatGPT or Claude, correspond to "System 2." While existing Large Language Models (LLMs) have primarily operated in the realm of System 2, Jev targets the System 2 domain exclusively.

The tasks Jev can perform are extremely simple: answering "yes/no," choosing one from given options, or scoring a specific value. Choi explained this by saying, "You know those musical prodigies who are only good at music? They don't know anything about Excel at all." Because it cannot perform other functions, the model is small, making it overwhelmingly fast and very inexpensive.

In a performance comparison test conducted by Choi, Jev demonstrated unrivaled efficiency. The results showed that Jev's processing cost was around 0.3 won, which was overwhelmingly cheaper than Claude or GPT models that incur costs in the tens or hundreds of won range. In terms of speed, it also completed tasks first, capturing both "cost-effectiveness and speed."

An 'AI Judgment Component' that filters 90 out of 100 spam emails

Why, then, are companies paying attention to this small AI that cannot even write? The key lies in "efficient workload processing." Choi suggested a strategy of inserting models like Jev as "judgment components" when building enterprise AI agents. He pointed out that entrusting all tasks to expensive LLMs is a waste of resources.

For example, consider a situation where hundreds of emails pour in every day. Having ChatGPT read every single email would consume massive amounts of cost and time. If Jev is deployed first to intuitively judge "spam status" or "work relevance," it can immediately filter out the 10% of certain advertisements or spam from 100 emails. This method drastically reduces the total system cost by allowing high-performance LLMs to precisely analyze only the remaining 90% of important emails.

This structure is also important in the design of "Agents," a key technology for recent AI performance improvements. The video mentioned cases where generative AI models are used as "evaluators." In the process where an LLM creates a result and then determines if that result was well-made according to the user's intention to trigger a re-work, models like Jev can perform an efficient evaluation role.

From detecting fraudulent texts to game control... the expanding territory of judgment AI

The scope of Jev's application shows potential to expand beyond simple classification into the realm of real-time control. Choi introduced a gameplay demonstration using Jev to analyze real-time situations and decide "whether to dodge or shoot." This structure is similar to the protagonist in the movie Mission: Impossible moving in real-time based on external instructions. Jev makes the judgment, while a programmed program carries out the actual action.

Furthermore, all areas requiring "intuitive judgment"—such as classifying emotions by reading subtle facial expressions or nuances, or finding and classifying appropriate folders among numerous files—are targets for Jev. Choi predicted that "not all AI components need to be luxury items" and that combining appropriate models for specific purposes will be the key to building future AI services.

Currently, Jev is provided in the form of an API (Application Programming Interface) rather than open source, allowing companies to use it by "renting a judging brain" for their services. Big tech companies, including OpenAI, are also releasing similar lightweight judgment models, showing that the center of gravity in the AI market is shifting from "massive intelligence" to "efficient judgment."

Source: original video (YouTube)