Most of the modern R&D is based on digital models in various representations used in data processing tools. What's becoming more and more challenging nowadays is the consistency of all the models, when failures and errors may occur due to hidden or informal assumptions made by researchers and engineers as well as imperfections of automated converters and generative utilities.

Model representations are not complete and do not fit each other exactly at the same time.

This problem can't be resolved by simply adding more software as the variety of model conversion cases is so wide that no universal algorithm is practically possible to fit them all. Now it's the role of human researchers and developers to bind various representations together and guide the R&D to the goal through the piecewise descriptions made by incomplete and sometimes contradictory models. It's understanding that lives in human intelligence that makes possible solving problems.

Machine understanding is the future basis for automatic binding of different models for solving new problems in various applications.

Despite advances being demonstrated by generative AI on texts and images, the neural network architectures behind the technology still have a number of significant limitations on the way to this purpose. New levels of performance are being achieved by increasing the number of parameters by orders of magnitude, requiring massive volumes of input data, computation resources, and training time.

The development of AI in general goes on in an extensive way not able to breach into the realm of machine understanding.

Machine understanding and abstract thinking need an approach that differs.

The new intensive approach to AI is based on the use of combinatorial agents that learn on limited amounts of data by recognizing patterns and combining them for making decisions on how to solve problems and to learn further. When the agent is trained to express the use of the patterns in an external sign system, it starts exhibiting machine understanding and abstract thinking as the ability to choose right methods for solving problems and explain the choices.

Intensive AI is an approach to solving new complex problems based on the experience of solving other problems.

The process of creating a training program, training, and searching solutions by means of the trained agent in the intensive approach is referred to as train-and-use method. Starting from some high problem complexity, the process may take less time than manual programming the solutions. The way to the acceleration goes through machine understanding, as it allows finding solutions on the level of methods before going to the level of data processing.

The intensive approach comprising combinatorial agents and the train-and-use method leads to machine understanding and opens up new prospects for automation in R&D for

• Solving new problems and tracing the reasons for choosing the methods of solving due to the communication between humans and machines with understanding on both sides
• Keeping models and specifications made in different representations consistent to each other
• Searching new methods of solving problems
• Maintaining R&D workflows elaborated within individual organizations and engineering groups as they correspond to their specific application profiles.

The ultimate goal for the train-and-use is the agent that not only needs less data for further training but learns effectively even when solving individual new problems from the known subject domain, so that additional training takes place at each utile application.

If our approach piques your interest, we invite you to a detailed discussion of the methods and a pathway for integrating the intensive AI in your technologies to face the challenges you encounter.