In recent days, a new name has frequently appeared on AI developers' timelines: Jev.

Some use it to categorize papers, some use it to select the tools that Agents should call, and others have integrated it into drone simulators to let drones choose actions to bypass obstacles. Platforms like OpenRouter and Cloudflare have also integrated this new model, and application demos around Jev have been appearing one after another.

If JEV makes AI decision-making faster and cheaper, which US stock companies have the opportunity to benefit from it?

Before understanding the investment logic, it is necessary to first clarify what Jev actually is.

Making Multiple-Choice Decisions

Jev is a decision model launched by TypeSafe AI on September 15. The company calls its category System One, or "System 1," borrowing the concept of rapid judgment from the foundational work in behavioral economics, Thinking, Fast and Slow. It can understand text descriptions and mainly outputs options, scores, and probabilities for software to continue processing.

For example. An online store receives a message: "I bought a cup yesterday, and it seems I was charged twice. Can you check?"

The system has several small questions to resolve: Does this matter belong to after-sales or finance? Is the user inquiring about the bill or requesting a refund? Do existing records show a duplicate charge?

Developers can predefine the options and have Jev answer each one separately. Then, the program checks the order and payment records and enters the corresponding processing flow; when an explanation needs to be given to the customer, it is handed over to a model that excels at expression.

This is the basic use of Jev: give it the situation, define the questions, and let it make choices.

General large models can also complete these tasks. Jev's appeal lies in being specifically optimized for classification, scoring, and selection, reducing the time and cost required for a small judgment. It delivers results directly to the program, eliminating the process of generating an explanation.

To extend further, you receive thousands of industry messages every day, and only a few dozen are truly worth reading carefully. You can first have Jev determine whether each message involves companies you care about and whether it contains new information, and then have a general model summarize the filtered content. In this way, the general model only processes the parts worth in-depth reading, while a large number of simple judgments are handed over to the lower-cost Jev.

Similar attempts have already extended to scenarios such as browser operations, search, and gaming. Beating's Jev special topic organizes these projects by use case, introduces which judgments Jev is responsible for in them, and includes source code and some demo entry points. By comparing specific cases, it is easier to understand where this ability to "make multiple-choice decisions" can be applied.

Cutting Costs to the Bone

According to the pricing published by TypeSafe, Jev charges $0.042 per million input tokens, with output free of charge.

The significance of lower costs lies in the ability to conduct item-by-item inspections on work that was previously only subject to spot checks. Customer service records can be classified one by one, retrieved materials can be assessed for relevance individually, and agents can perform an extra round of screening before each tool invocation.

For example, a news analysis program that previously only picked the hottest articles each day for AI processing can now examine all articles first and then categorize them by company, product, order, and personnel changes. Information that is not trending but directly related to a major enterprise can also be surfaced.

Jev is named after the economist Jevons, whose most famous achievement was proposing the "Jevons Paradox." The developers hope that once judgment becomes cheap, software will use it more frequently. Places where it previously wasn't worth making a separate large-model call can now have their cost calculations redone.

Jev and Physical Devices

Developer Roman Slack built a drone project: in the MuJoCo physics simulator, a quadrotor drone tracks a target ahead while navigating through pillars, beams, and moving obstacles.

The program first organizes depth and segmentation information from the camera into data, telling Jev whether the path ahead is blocked, which side has a gap, and how tall the obstacles are. Jev then chooses from actions such as maintaining course, veering left, veering right, climbing, and braking.

For instance, a beam blocks the entire passage, with no way around on either side, but there is space above. Jev can choose to climb, and the control program then translates that choice into motion commands for the drone.

During flight, different modules perform their own roles. Jev decides where to go, while the control program handles balance, motor thrust adjustment, and emergency obstacle avoidance. The camera provides environmental information, the decision model selects actions, and the controller makes the actions happen.

The same division of labor can be understood through a robotic arm grabbing a cup. The camera identifies the cup's position, sensors measure distance and force, the controller adjusts finger movements, and the decision model participates in determining which cup to pick up first and what to do when encountering obstacles.

If this kind of judgment becomes cheap and easy to integrate, the cost of developing automated equipment has the potential to decrease. Devices that were previously not worthwhile due to expensive software development and small use cases will also have new commercial opportunities.

This links Jev to a number of companies. More devices entering factories, warehouses, and homes means more demand for sensing, power, computing, and motion control.

Which names could benefit

Along these lines, what is worth watching is the components devices need to go from sensing their environment to executing actions. No matter what model the software chooses, a robot needs to know where it is, what is around it, and convert electricity into precise motion.

Texas Instruments (TXN), covering the supporting hardware for sensing and execution

Analog chips handle continuously varying signals such as voltage and current, and also take on power supply and drive tasks. TXN's robotics products cover signal acquisition, motor control, power, and embedded processing.

Once a robot decides to raise its arm, how current is delivered to the motor and how force is controlled all depend on these components. If cheaper automation brings more robots, robotic arms, and smart devices, TXN has an opportunity to sell more supporting chips into these devices.

Analog Devices (ADI), corresponding to more precise sensing and motion.

ADI's robotics solutions involve inertial sensing, camera connectivity, industrial communications, and motion control interfaces. For example, when a warehouse robot enters a poorly lit aisle, it needs to combine inertial information to judge its own movement; when multiple cameras and sensors work at the same time, data also needs to be synchronized and sent to the processor.

The more complex the work a device takes on, the higher the requirements for sensing precision, connection stability, and motion control. ADI's benefit logic is providing these supporting components in more autonomous devices.

STMicroelectronics (STM), corresponding to AI on low-power devices.

So-called edge AI means letting devices process part of the data locally. STM's STM32N6 microcontroller already integrates an accelerator for neural network computing, aimed at cost- and power-sensitive applications such as vision and sound.

For example, a small camera device can recognize objects locally, and an industrial device can determine from sound whether operation is abnormal. The discussion around Jev has drawn investors' attention to this type of use case: a large number of simple, repetitive intelligent tasks have the opportunity to move into lower-priced, smaller devices. What STM provides is exactly the processing capability these devices need.

ON Semiconductor (ON), corresponding to vision and power devices.

ON's products for robots include image sensors, depth sensors, as well as power transistors and motor driver chips. The first two categories handle capturing visual and distance data, while the latter two work together to control the current required by the motors.

Take warehouse transport robots as an example: the image sensor in the camera captures the view of the shelves ahead, the depth sensor measures the distance to obstacles, and the motor driver chip works with the power transistors to control the current and make the wheels turn. As more robots and smart camera devices are put into use, there is an opportunity to increase demand for these sensors and chips.

Beyond these four companies, there are three more that can be added to the watchlist.

NXP Semiconductors (NXPI)

Related business: i.MX processors, microcontrollers, connectivity and control products

Beneficiary logic: providing computing, communication, and control capabilities for more robots and industrial equipment

Qualcomm (QCOM)

Related business: the Dragonwing platform for robots, drones, and on-device AI

Beneficiary logic: capturing demand for on-device processing of vision, computing, and connectivity tasks

Cloudflare (NET)

Related business: providing access to and invocation of Jev third-party models

Beneficiary logic: when developers use Jev for customer service triage, material screening, and agent scheduling, they still need to deploy applications, connect data, and manage calls. Cloudflare's opportunity is to keep these applications running on its own platform.

Still in the early stages

Jev still has issues that need to be resolved, including judgment accuracy, question design, and whether the service can support long-running business operations.

First, the claim that Jev has "zero hallucinations" is easily misunderstood. Restricting it to choosing only among A, B, and C can prevent it from returning a D that the program does not recognize, but it can still choose wrong. Going back to the customer service example, the program being able to smoothly read out "route to finance" does not mean that this ticket should really be handled by finance.

A developer tested eight types of tasks with Jev and completed 16,000 judgments. He acknowledged Jev's speed and cost advantages, while also finding that the questions and options still need further adjustment to cover real-world situations. He also hopes to integrate it into production operations only after suppliers accumulate more operational experience and alternatives emerge in the market.

How much work Jev can handle is currently unknown, but a trend has already emerged: delegate a large number of repetitive judgments to JEV, and leave complex tasks to general-purpose models.

As the cost of developing and using automation decreases, will analog chips, edge AI, and autonomous devices usher in their own ChatGPT moment?

Welcome to join the official BlockBeats community:

Telegram Subscription Group: https://t.me/theblockbeats

Telegram Discussion Group: https://t.me/BlockBeats_App

Official Twitter Account: https://twitter.com/BlockBeatsAsia