In Waymo’s implementation of this technology, an extremely large model is trained. This extremely large model is used to create the Teacher models shown in the illustration shown in Figure 1, on page 2. The teacher models, in turn, are used to train their respective small models. Each focuses on a specific domain.

Because of Waymo’s business focus, the three small models are a Driver that is deployed in an autonomous car. A Simulation model that is used to test the Driver. And a Critic model that is used to assess the results of the test in the Simulation model. The Critic can also be deployed in the car to assess the performance of the Driver in actual field operation.

Applications

For mobile chat applications, ways 1.) through 4.) above will be used to access AI's. For the enhancement of other simple mobile applications, way 2), accessing through a cooperative local application communicating with a remote AI, will be used.

For mobile intelligent agent applications, all four ways may be used. The more demanding applications will move to the Waymo distillation approach. Examples of the range of potential mobile applications include those briefly described below.

The technology can be used in other segments of the transportation area, such as drones, railroads, pipelines, etc.

There is a broad range of medical applications, including portable and transportable medical equipment, autonomous crash carts in medical facilities, wearable medical equipment, and medical equipment that moves around in the body (electronics packaged in a swallowable capsule, things that move around in the bloodstream, etc.).

There is a range of mobile applications in field service, construction, and operations. Operations applications can be in electrical, communication, and other kinds of networks. Operations in automated factories also hold promise.

Recent progress in robotics is making humanoid robots seem feasible. These robots would be a particularly fertile application area.

There is interesting work being done with AI applications for the neurodiverse community. One AI application reads human expressions and tells the neurodiverse individual, who has trouble reading them, what the other person is showing on their face. Currently, this is implemented in conjunction with Zoom. Clearly, it would be valuable if implemented in a mobile platform that a neurodiverse person could have with them all the time. One mobile packaging could be in smart glasses.

As can be seen from these examples, the range of applications can be quite broad.

Co-opetition to Move Up the Learning Curve

One thing that is clear is that we are very low on the learning curve. As a society, we are just beginning to learn how best to utilize the power that is being created by GenAI. One way of moving up the learning curve faster is through cooperation. That is, those building the tools and those using the tools come together and share what they have learned. Some fear that this will help their competitors. In past generations of technology, a system of cooperating to build a shared knowledge base, then using that base in a proprietary fashion to compete with each other, has proven effective. That approach has come to be called co-opetition. A contraction of cooperation and competition. An AI group is exploring AI co-opetition. More information about that group can be found at https://www.bacesecurity.org/form/aiwg.

Conclusion

There is a broad range of valuable mobile AI applications. Powerful LLMs are getting very large. There is technical progress being made in moving that power to mobile devices. The challenge that faces us is learning how to convert that power into effective applications. A co-opetition process is the best way to quickly increase the capability to build effective mobile AI applications.