Models and Programming: Understanding the Foundations of Artificial Intelligence

Contemporary artificial intelligence rests on a fundamental interplay between two complementary activities: the construction of models capable of representing a phenomenon, and the programming of mechanisms through which those models can be put to use. The former establishes a way of describing, interpreting or anticipating reality; the latter organises the operations required to transform that representation into calculations, decisions or actions.

This distinction, inherited from applied mathematics, cybernetics and the earliest developments in computing, remains essential to understanding today’s digital systems. From statistical models to large language models, and from simulation environments to autonomous robots developed in Asia, a technology’s performance depends as much on the relevance of its representation as on the way that representation is constructed, trained, executed and evaluated.

The evolution of artificial intelligence therefore renders neither model theory nor the principles of programming obsolete. Rather, it renews their applications by shifting part of the work away from the explicit definition of rules and towards the learning of patterns from data. To understand this transformation, we must return to the foundations of modelling, algorithmic logic and their convergence in automation.

1. What Is a Model? Representing Reality Without Confusing It with Its Representation

A model is a simplified, abstract or constructed representation of an object, system or process. Its purpose is to make certain aspects of a phenomenon sufficiently intelligible for them to be explained, calculated, simulated or used to anticipate how that phenomenon may evolve.

A weather model, for example, does not physically reproduce the atmosphere in its entirety. It represents certain of its properties through variables, mathematical relationships and initial conditions. The resulting calculations make it possible to estimate possible developments, without guaranteeing that every forecast will prove exactly correct.

This process necessarily involves selection. Of all the characteristics of a phenomenon, only some are retained because they appear relevant to the question under investigation. Others are simplified, excluded or represented indirectly.

A model is therefore not reality itself, but a construct designed to make certain aspects of reality accessible to knowledge and action.

This distinction is particularly important in artificial intelligence. A model trained to recognise images, generate text or estimate the trajectory of an object does not automatically possess a comprehensive understanding of the corresponding reality. It operates on a representation constructed from data, assumptions, objectives and technical constraints.

1.1. The History of Modelling: From the Physical Sciences to Digital Systems

Modelling did not originate with computing. It runs throughout the history of science, dating back to the point at which researchers began expressing relationships between observable phenomena through diagrams, proportions, equations and experimental apparatus.

Classical mechanics illustrates this approach: mathematical laws make it possible to describe the motion of bodies in terms of defined quantities. Electromagnetism, thermodynamics and statistical physics subsequently developed other forms of representation, each adapted to phenomena governed by different mechanisms.

During the twentieth century, applied mathematics, systems theory and cybernetics strengthened interest in the relationships between representation, computation and behaviour. Cybernetics, notably associated with the work of Norbert Wiener, helped formalise the study of control, communication and feedback mechanisms in living and artificial systems.

The rise of computing subsequently made it possible to process models whose complexity exceeded the capabilities of manual calculation. Computers enabled the numerical simulation of physical, economic, biological and industrial phenomena by executing large numbers of operations according to defined procedures.

This development paved the way for contemporary artificial intelligence. A model is no longer merely an equation solved by a computer: it can also become a computational structure whose parameters are adjusted from examples, as in machine learning.

1.2. The Main Categories of Models

Models can be classified along several complementary dimensions. These classifications do not necessarily describe mutually exclusive categories: a single model may belong to several of them simultaneously.

• Physical model: represents a system through a material apparatus. A bridge model subjected to mechanical stresses, a test rig or an experimental installation can be used to observe behaviour comparable to that of the system under investigation.
• Mathematical model: expresses relationships between variables through equations, functions, probabilities or other formal structures. It is used, in particular, to calculate results, test hypotheses and analyse developments.
• Conceptual model: describes a phenomenon through concepts, logical relationships and a language, whether natural or specialised. It can be used to organise knowledge, formulate a theory or specify the assumptions underlying a research project.
• Computational model: translates a representation into data structures and executable operations. It makes it possible to calculate results, simulate a system or implement a procedure.
• Learned model: acquires part of its functional structure during a training process. Its parameters are adjusted to reduce error, improve predictions or optimise a defined objective. Neural networks and large language models belong to this category.

These categories overlap. A numerical weather model, for example, combines physical equations, numerical methods, observational data and computer programs. An artificial intelligence system may, in turn, combine a trained neural network, explicit rules and a physical model of the environment in which it operates.

1.3. Explanatory and Predictive Models: Two Distinct Purposes

Not all models pursue the same objective.

An explanatory model seeks to make intelligible the mechanisms or relationships that contribute to the production of a phenomenon. It can be used to test a scientific hypothesis, identify determining variables or compare different interpretations.

A predictive model is primarily intended to estimate an outcome from available information. It may forecast the weather, anticipate commercial demand or determine the probability that an image belongs to a given category.

This distinction is important in artificial intelligence. A system may achieve excellent predictive results without providing a satisfactory explanation of the mechanisms underlying its responses. Conversely, an explanatory scientific model may be of considerable value without being the most effective model for every predictive task.

The two approaches can nevertheless complement one another. A prediction makes it possible to compare a model with observations, while a theoretical explanation can help clarify its limitations and establish the conditions under which its results remain relevant.

1.4. Deterministic and Probabilistic Models

A deterministic model produces a determinate result when its inputs and operating conditions are fixed. An equation describing the idealised fall of an object under specified conditions is a classic example of this approach.

A probabilistic model, by contrast, represents uncertain phenomena through probability distributions. It makes it possible to characterise several possible outcomes, quantify uncertainty or reason about events whose occurrence is not certain.

Artificial intelligence makes extensive use of both families of models. Some control systems rely on deterministic relationships, whereas statistical methods and neural networks exploit probabilistic patterns.

Large language models illustrate the latter approach: they learn statistical relationships between units of text and generate continuations according to context and their generation mechanisms. Their operation, however, cannot be reduced to a simple lookup of isolated probabilities. It depends on an architecture, a training process, learned parameters and, depending on the system, additional mechanisms for retrieval, memory or tool use.

2. The Principles of Model Validity: Usefulness, Empirical Evidence and Limitations

Constructing a model is not enough to establish its validity. A representation may be elegant, coherent and mathematically correct while still providing an imperfect account of the phenomenon it purports to describe.

A model’s quality therefore depends on how its assumptions are formulated, how relevant its selected variables are, and how its results stand up to comparison with independent observations.

2.1. Simplicity as a Principle of Construction

A useful model should avoid introducing complexity that contributes too little to its objective. This requirement is consistent with the principle of parsimony, traditionally associated with Occam’s razor: when several explanations account for the same observations, a simpler explanation may be preferable, provided that it does not sacrifice essential features of the phenomenon.

Simplicity does not mean, however, that the most elementary model is always the best. An overly rudimentary system may overlook determining variables, while an excessively complex representation can become difficult to interpret, verify and maintain.

In machine learning, this balance is particularly evident in the trade-off between learning capacity and generalisation. A model that is too simple may fail to learn the relationships present in the data; one that is too closely fitted to its training examples may reproduce their idiosyncrasies without performing successfully on new observations.

2.2. Universality and the Risk of Overgeneralisation

A model becomes particularly valuable when it can handle several situations through a single set of principles. This capacity for generalisation is one of the principal objectives of both science and machine learning.

However, a pattern observed in one context does not necessarily transfer to another. A model trained on images taken in daylight may behave differently in darkness. A language system that performs well in one language or specialised field may produce less reliable results in another language or in a technical context that is poorly represented in its training data.

A model’s validity must therefore be assessed in relation to an explicitly defined domain of application. Its performance cannot be extrapolated to every conceivable situation simply because it has passed several tests or achieved a high score on a particular evaluation.

2.3. Experimental Validation and the Convergence of Evidence

A model gains credibility when independent observations confirm its results. Such convergence does not guarantee definitive truth, but it helps distinguish a sufficiently substantiated hypothesis from a proposition that remains speculative.

In artificial intelligence, validation generally requires several levels of scrutiny:

• evaluation on data that were not used during training;
• comparison with established baseline methods;
• analysis of errors and edge cases;
• measurement of performance under conditions resembling actual use;
• verification of robustness when the data or environment change;
• monitoring of results after deployment, while the system remains in use.

A result obtained in a laboratory or on a well-defined dataset is not sufficient to establish that a model will be reliable under every operational condition.

2.4. Partial Mismatch: Why Every Model Retains a Margin of Error

A model selects certain properties of the system under investigation and necessarily leaves others aside. This reduction is both the condition that makes the model usable and its principal limitation.

In an economic model, certain human or institutional variables may be difficult to quantify. In a vision system, lighting, camera angle and the diversity of objects may affect the results. In a language model, missing context, ambiguity or the scarcity of certain information may lead to erroneous responses.

It is therefore useful to distinguish three questions:

1. Is the model internally consistent with its own structure?
2. Do its results correspond to the available observations?
3. Are the actual conditions of use sufficiently similar to those under which it was evaluated?

A model may satisfy the first criterion without fully satisfying the other two. Internal consistency, in itself, is not proof of fidelity to reality.

This precaution is crucial in generative artificial intelligence. A confidently worded answer, an apparently coherent demonstration or a visually convincing result does not guarantee that the information produced is accurate.

3. What Is Programming? Turning a Representation into Executable Operations

Programming consists in describing, through a formal language, the operations a computer system must perform to achieve an objective. It stands between the formulation of a problem and its actual processing by a machine.

This activity is not limited to writing lines of code. It includes requirements analysis, data definition, algorithm design, the organisation of processing operations, the choice of technologies, software implementation and the verification of results.

To program is to make a method sufficiently explicit for a machine to execute it under specified conditions.

This definition remains valid in contemporary computing environments, even though development tools, software architectures and machine capabilities have evolved profoundly.

3.1. From Early Calculating Machines to Modern Programming Languages

The history of programming is closely linked to the automation of calculation. Before the advent of electronic computers, several mechanical machines had already embodied arithmetic operations. Charles Babbage’s work on the Analytical Engine, in particular, helped develop the idea of a general-purpose machine whose operations would be governed by instructions.

In the nineteenth century, Ada Lovelace formulated important reflections on the possibilities of the Analytical Engine, particularly its capacity to execute sequences of operations and manipulate symbols beyond straightforward numerical calculation.

During the twentieth century, the development of electronic computers led to the formalisation of programming languages, algorithms and programming methods. Low-level languages gradually came to coexist with higher-level languages, which were better suited to expressing scientific, industrial and administrative problems.

This evolution continued with structured programming, object-oriented programming, functional programming, distributed systems and modern development environments. Software libraries, compilers, interpreters, version-control systems and deployment infrastructures transformed the way programs are designed and maintained.

Generative artificial intelligence adds a further stage to this history: certain systems can now generate code from natural-language instructions. This capability changes the ways in which software is developed without eliminating the need to define objectives, check results and verify that the resulting program actually addresses the problem at hand.

3.2. The Algorithm: The Logical Link Between the Problem and the Program

An algorithm is a procedure consisting of explicit, ordered operations designed to solve a class of problems or perform a task.

It specifies the required data, the transformations to be carried out, the conditions to be examined and the criteria governing whether processing should continue or stop. Depending on the problem, it may involve sequences of instructions, conditional branches, iterations or calls to other procedures.

An algorithm must be distinguished from the program that implements it. The same algorithm can be expressed in several languages and executed on different architectures. Conversely, a single programming language can be used to implement very different algorithms.

Consider the example of a system designed to classify images.

The general algorithm may consist in receiving an image, converting it into a numerical representation, applying a classification method and returning a category together with a confidence score. The program organises these operations in practice, manages the data, calls the necessary components and provides for error handling.

In a machine learning system, the algorithm may also refer to the method used to adjust the model’s parameters during training. In this case, a distinction must be drawn between the learning algorithm, the model obtained after training and the program that uses that model in production.

This distinction is fundamental: an algorithm describes a processing method, a program organises its execution, and a model provides a usable representation of the phenomenon under investigation.

3.3. The Stages of a Programming Process

Designing an effective computer system involves several complementary operations. Their order varies according to the nature of the project, but their functions remain essential.

1. Problem analysis. The requirement is translated into precise objectives. Expected results, constraints, available data and the system’s limitations must be identified.

2. Feasibility analysis. The proposed solution is assessed against available resources, performance requirements and hardware, software or economic constraints.

3. Algorithm design. The necessary operations are broken down into explicit steps. Several methods may be compared in terms of accuracy, complexity and execution cost.

4. Data structuring. Information is organised into formats suited to its intended use. The choice of data structures and access mechanisms affects system performance and maintainability.

5. Choice of programming languages and architectures. Technologies are selected according to project requirements, whether scientific computing, data processing, user interfaces, embedded systems, machine learning or distributed applications.

6. Implementation. The designed methods are translated into executable code, drawing where necessary on existing libraries, services and components.

7. Testing and validation. The program’s behaviour is checked against its specifications. Errors, edge cases, incompatibilities and potential failures are investigated.

8. Deployment and maintenance. The software is integrated into its operating environment, monitored and adapted to changes in data, infrastructure and requirements.

In artificial intelligence systems, these stages are supplemented by operations specific to learning: dataset construction, example preparation, training, parameter validation, independent evaluation and performance monitoring after deployment.

3.4. Programming Languages: Tools Suited to Different Problems

Programming languages offer different ways of expressing operations and organising software systems. No language is universally superior to all others: its value depends on the task, the execution environment, the available expertise and the technical constraints.

Python occupies an important position in artificial intelligence thanks to its ecosystem of libraries for data processing, statistics, machine learning and neural networks. C and C++ remain widely used in embedded systems, performance-critical software and certain computing infrastructures. Java, JavaScript, Rust, Julia and other languages serve specific or complementary purposes.

A further distinction must be made between general-purpose languages, query languages, description languages, configuration tools and languages specialised in knowledge representation or constraint modelling.

Contemporary systems frequently combine several languages. An artificial intelligence application may use Python to orchestrate processing, C++ for certain computationally intensive operations, SQL to query data and JavaScript to present results through a user interface.

This diversity illustrates that programming is less a matter of translation than of architecture: the task is to organise operations, resources and components into a coherent system.

3.5. Why Testing, Documentation and Maintenance Remain Essential

A program may comply with the rules of its syntax and still produce incorrect results. It may also work on a limited dataset before failing when conditions change.

Unit tests verify the behaviour of isolated components. Integration tests examine interactions between components. Functional tests check the software against its intended uses, while performance tests assess response times, memory consumption and the ability to handle a given workload, among other factors.

Documentation makes design decisions, interfaces, dependencies and operating conditions explicit. It facilitates subsequent development and makes it possible to understand why particular decisions were taken.

These requirements become especially important when dealing with AI-generated code. A program produced rapidly may contain logical errors, vulnerabilities, unsuitable dependencies or assumptions that do not match the context. The speed of generation therefore does not dispense with independent verification.

4. Models and Programming: Two Inseparable Dimensions of a Single System

The distinction between models and programming is useful for analytical purposes, but it must not obscure their interdependence. A model without an operational procedure may remain a theoretical construct; a program without a relevant representation of the problem risks executing operations correctly while failing to address the actual requirement.

The relationship between these two fields rests on complementarity: the model defines the properties and relationships the system is intended to represent, while programming determines how those properties will be manipulated, calculated and integrated into an application.

4.1. The Model Defines What the System Represents

A model specifies the variables under consideration, the relationships between them and the assumptions adopted to describe a phenomenon.

In an application designed to forecast energy consumption, it may represent the relationship between temperature, time of day, building type, usage patterns and electricity demand. In a mobile robot, it may represent the vehicle’s position, the geometry of its surroundings, detected obstacles and possible trajectories.

These representations directly influence the results available to the system. If an important variable is missing, the data are biased or an assumption fails to match the actual context, programming alone cannot restore the validity of the whole system.

4.2. Programming Defines How the Model Is Used

The program implements the operations required to use the model. It organises inputs, prepares data, performs calculations, transmits results and manages interactions with other components.

It also determines operational characteristics: calculation frequency, error handling, resource usage, data access and communication with sensors or external services.

In an image-recognition application, the model may consist of a trained neural network. The surrounding program receives the images, preprocesses them, runs inference and passes the predictions to the application.

The model and the program therefore perform different functions, even when they are designed together.

4.3. Numerical Simulation: An Early Form of Convergence

Numerical simulation is a particularly clear historical example of this convergence.

A physical phenomenon is first described through a mathematical model. The resulting equations are then translated into a computational method suited to a computer. The program executes that method, produces results and makes it possible to study how the system evolves under different conditions.

This transformation involves several distinct levels:

• The phenomenon under investigation: the physical or abstract system whose behaviour is to be understood.
• The model: the mathematical or conceptual representation adopted.
• The numerical method: the procedure used to obtain an approximate solution.
• The program: the computational implementation of that procedure.
• The simulation: the execution of the program to investigate the resulting behaviour.
• Validation: the comparison of results with relevant theoretical references or observations.

Each level can introduce its own limitations. A modelling error cannot necessarily be corrected by faster computation; an unsuitable numerical method can produce misleading results even when the model itself is relevant.

This distinction remains essential in artificial intelligence, where increased computing power guarantees neither data quality nor the relevance of the underlying assumptions.

4.4. Feedback Loops: When the Model Influences Action

In automated systems, the relationship between models and programming is not limited to producing a result from an input. The system may also act upon its environment, observe the consequences of its actions and adjust its behaviour accordingly.

This mechanism relies on a feedback loop. A sensor provides information about the system’s state; a processing procedure interprets that information; a command is calculated and executed; new observations then make it possible to assess the result.

A thermostat provides a basic example of this logic. An autonomous robotic system implements it on a much more complex scale, since observations are incomplete, the environment changes and actions may have consequences that are difficult to predict.

The model may then be used to estimate the system’s state, anticipate the consequences of a command or select an action. The program orchestrates observations, calculations, decisions and commands.

This relationship explains why advanced robotic systems require the joint design of models, algorithms, software and physical mechanisms.

5. From Traditional Programming to Artificial Intelligence: What Actually Changes?

Contemporary artificial intelligence does not eliminate the principles of programming. It does, however, change the way certain functions are defined and implemented.

In a traditional program, a substantial part of the system’s behaviour is specified directly through rules written by developers. In a machine learning system, part of that behaviour results from parameters adjusted on the basis of data.

This distinction should not be overstated: artificial intelligence systems remain governed by programs, objectives, architectures, training procedures and execution mechanisms.

5.1. Symbolic Programming: Making Rules Explicit

Symbolic artificial intelligence relies, among other things, on the explicit representation of knowledge, rules and logical relationships.

An expert system, for example, may apply a set of rules to associate symptoms with hypotheses, verify industrial constraints or determine whether a configuration satisfies a set of specifications.

Its operation depends on the quality of the represented knowledge, the precision of its rules and the system’s ability to handle the situations for which it was designed.

This approach remains valuable when rules can be identified, decisions must be traceable or constraints can be formulated explicitly. It can also be combined with machine learning methods.

5.2. Machine Learning: Learning Parameters from Data

Machine learning refers to a set of methods that enable a system to fit a model to examples or interactions.

Rather than manually programming every relationship being sought, developers specify, among other things, an architecture, a learning objective and an optimisation procedure. The system then adjusts its parameters to improve its performance on the training data.

In supervised learning, examples are paired with reference answers. In unsupervised learning, methods seek structures or patterns without necessarily having access to explicit labels. Reinforcement learning relies on interactions between an agent and an environment, using reward signals to guide the learning process.

These approaches share a central principle: the model’s behaviour depends in part on parameters acquired during a learning process, rather than exclusively on rules specified in advance.

Programming nevertheless remains decisive. It defines data preparation, system architecture, the objective function, optimisation mechanisms, evaluation criteria and the conditions under which the model will be used.

5.3. Neural Networks and Deep Models

Artificial neural networks are computational structures consisting of layers and parameters whose values are adjusted during training.

Deep learning uses networks comprising multiple layers of transformations. This approach has enabled major advances in image recognition, language processing, speech synthesis and numerous multimodal tasks.

The performance of these models depends, among other factors, on their architecture, the data used, computing resources, the training procedure and evaluation methods.

A neural network is not an autonomous program in the sense that it can be separated from all supporting infrastructure. Both its training and its deployment require software to organise computations, manage parameters and supply the expected inputs.

The trained model and the inference program thus constitute two distinct elements within a single technical pipeline.

5.4. Large Language Models: A New Relationship Between Code, Data and Behaviour

Large language models (LLMs) are trained on vast collections of textual data and, depending on the system, on other modalities as well. They learn statistical relationships that enable them to process and generate language.

Their behaviour does not result exclusively from manually written linguistic rules. It emerges from the network’s structure, its learned parameters, its training data and subsequent post-training stages.

When a user submits a request, the software passes the context to the model, orchestrates generation and retrieves the result. Additional mechanisms may also come into play, including document retrieval, code execution, tool calls, external memory or the control of certain outputs.

It is therefore useful to distinguish four components:

1. The model architecture, which defines the overall organisation of computations.
2. The learned parameters, which encode some of the patterns acquired during training.
3. The execution software, which organises inputs, inference, tools and outputs.
4. The operating environment, which includes data, interfaces, hardware resources and deployment constraints.

This distinction explains why two applications built around the same model may exhibit different behaviours. They may use different instructions, contextual data, tools, control mechanisms and safety rules.

It also explains why improving a model does not always improve a product: software architecture, the quality of accessible data and the effectiveness of the execution process matter as well.

5.5. Generative AI and Assisted Programming

AI-powered programming assistants can generate functions, explain code, suggest corrections or turn a natural-language description into an initial implementation.

This development sometimes reduces the cost of certain software development tasks and enables less specialised users to explore technical solutions. It also shifts part of the work towards requirements formulation, result verification and the integration of generated code.

Generated code must nevertheless be treated as a proposal that requires scrutiny. It may contain errors, reproduce outdated practices or introduce security risks. It may also appear to satisfy a request while overlooking an implicit constraint.

Developers’ expertise therefore does not disappear; it changes. Analysis, design, architectural understanding, validation and the ability to identify errors remain essential.

6. Embodied Artificial Intelligence: From Digital Representation to Robotic Action

Contemporary robotics makes the interdependence of models and programming particularly clear. Unlike a system that produces only text or images, a robot must connect its digital representations to a physical environment.

It must perceive a situation, estimate its state, choose an action, control its mechanisms and verify the consequences of its behaviour. An error in perception or planning can lead to hardware failure, a collision or an interrupted task.

6.1. How Does a Robot Turn Information into Action?

An autonomous robot generally combines several categories of components.

Sensors collect information about the environment: images, distances, forces, movements or positional data. Algorithms process these signals to estimate the state of the robot and its surroundings. A planner then determines possible actions, while a control system translates those decisions into movements or physical operations.

Models intervene at several levels. A geometric model describes the relationship between joints and the position of a tool. A perception model makes it possible to identify objects. A dynamics model represents the robot’s motion. A learned model may help select actions on the basis of examples or interactions.

Programming ensures that these elements are integrated and executed within the system’s timing constraints.

A robot grasping an object must therefore detect its presence, estimate its position, calculate a trajectory, control its joints and adjust its movement if the object moves. Each stage depends on representations and procedures whose coherence determines the success of the overall task.

6.2. Vision-Language-Action Models

Vision-language-action models, commonly referred to as VLA models (Vision-Language-Action), seek to connect visual information, linguistic instructions and actions executable by a robot.

The aim is to enable a system to relate an instruction such as “pick up the box on the table” to its perception of the environment and the generation of an appropriate sequence of actions.

This approach differs from purely linguistic models because the expected output is not limited to a sentence. It must be translatable into commands compatible with the robot’s physical capabilities.

The transition from representation to action raises several challenges: spatial accuracy, contextual understanding, generalisation to unfamiliar objects, adaptation to environmental changes and compliance with safety constraints.

VLA models therefore do not eliminate the need for conventional control, monitoring and validation mechanisms. They can be integrated into a hybrid architecture in which a general-purpose model contributes to planning while specialised controllers manage short-term movements.

6.3. World Models: Anticipating the Consequences of Actions

A world model is a representation designed to predict certain aspects of how an environment evolves in response to its current state and the actions undertaken.

In robotics, such a model can help anticipate an object’s movement, estimate the consequences of a manipulation or compare several trajectories before executing them.

This approach brings scientific modelling, numerical simulation and machine learning closer together. The system constructs a representation of its environment, uses that representation to evaluate possibilities and then compares its predictions with actual observations.

The effectiveness of this feedback loop nevertheless depends on the model’s fidelity, the relevance of the data and the ability to detect situations in which predictions become uncertain.

A world model is not an exhaustive reproduction of the environment. It remains an approximation whose domain of validity must be assessed.

7. Asia, Robotics and the Evolution of Artificial Intelligence Models

Asia is a major field of observation for understanding the convergence of artificial intelligence, programming and automation. China, Japan, South Korea and Singapore are developing approaches that combine AI models, robotics, industrial infrastructure and applied research in different configurations.

This is not a uniform movement. Research ecosystems, industrial strategies, priorities and economic constraints vary from one country to another. Several trends nevertheless help illuminate current developments.

7.1. China: Bringing AI Models Closer to the Physical World

In China, embodied artificial intelligence is attracting growing interest. This approach seeks to integrate the perception, reasoning and planning capabilities of AI models into physical systems such as robots, drones and autonomous vehicles.

A report published in December 2025 by the German Council on Foreign Relations describes this orientation as a strategic avenue for the development of Chinese AI, complementing approaches centred primarily on large language models. It highlights the role of public policy, innovation centres and collaborations between universities and companies in driving this development.

Source: German Council on Foreign Relations, China’s Embodied AI: A Path to AGI.

The technical significance of this approach lies in exposing models to physical constraints. A representation that appears relevant in a digital environment may fail when a robot encounters unfamiliar objects, variations in lighting, uneven surfaces or unexpected movements.

Embodied robotics therefore requires learning to be connected to sensory data, control mechanisms and actually observed outcomes. It highlights a general limitation of AI systems: the quality of a representation cannot be assessed independently of the conditions under which it is used.

7.2. Japan: Integrating Artificial Intelligence into Robotic Bodies and Systems

Japan has a particularly active research environment in robotics, industrial systems and physical interaction technologies.

Research programmes supported by the Japan Science and Technology Agency (JST) explore embodied AI models, robots capable of contributing to scientific experiments, and architectures incorporating tactile perception, physical adaptation and continual learning.

Source: Japan Science and Technology Agency, projects on embodied AI.

These projects illustrate a fundamental question: a model intended to act in the physical world must account not only for what it perceives, but also for the capabilities and limitations of its own body.

A robot’s shape, joint mobility, sensors, the forces it can exert and the characteristics of its actuators all influence the actions it can perform. The same abstract representation therefore does not necessarily translate into identical commands on different robots.

Integration between the model, the program and the hardware becomes decisive. Artificial intelligence is not merely a software layer; it must be designed in relation to the physical properties of the system.

7.3. South Korea: Developing Foundation Models for Robotics

South Korea combines its ambitions in artificial intelligence with its expertise in electronics, industry and robotics.

Research at the NAIRL laboratory, for example, focuses on foundation models for robotics, with the aim of integrating perception, decision-making and physical control into systems capable of acting in variable environments. Its publications cover robotic models, object manipulation and navigation, among other topics.

Source: NAIRL, Physical AI and Robotic Foundation Models.

This approach seeks to move beyond architectures in which each robotic task requires an entirely separate model. Foundation models for robotics aim to reuse capabilities learned across multiple tasks while retaining mechanisms suited to the specific constraints of each machine.

The challenge is considerable: a general representation may facilitate the transfer of skills, but it must remain compatible with the requirements of precision, latency and safety inherent in physical action.

7.4. Singapore: Combining High-Level Planning with Motion Control

Singapore is also an important research setting for embodied AI and human-robot collaboration.

The project Embodied AI for Human-Robot Collaborative Task Execution, coordinated through scientific collaborations involving Singapore and France, investigates the integration of models capable of perception, reasoning and planning with learning methods designed for manipulation and short-term control.

Source: CNRS@CREATE, Embodied AI project.

This approach illustrates the value of a multi-level architecture. A general model can interpret a task’s objective and organise its main stages, while specialised mechanisms determine the precise movements required for execution.

The distinction between planning and control makes it possible to reconcile relatively abstract objectives with the immediate constraints of the physical environment.

7.5. Industrial Convergence and Persistent Challenges

These different approaches point to a significant development: artificial intelligence is no longer evolving solely around content generation or data recognition. It is expanding into systems that must perceive, plan and act.

This convergence nevertheless requires several challenges to be addressed:

• collecting data representative of real-world environments;
• generalising to unprecedented situations;
• ensuring accurate actions and controlling errors;
• maintaining compatibility between models and different hardware architectures;
• controlling computing costs and response times;
• ensuring the safety of people, equipment and infrastructure;
• validating behaviour under operational conditions.

Asia is therefore not merely a region in which AI is deployed. It is actively contributing to research into architectures that combine digital models, learning and physical action.

8. From Modelling to Automation: Turning Knowledge into Action

Automation is one of the most direct extensions of the convergence between models and programming. It consists in entrusting a technical system with the execution of tasks according to defined rules, objectives or decision-making mechanisms.

It may rely on entirely deterministic procedures, adaptive control systems or artificial intelligence models. In every case, automation requires a connection between the representation of the problem, executable operations and evaluation criteria.

8.1. Automating a Procedure: Formalising Operations

A task does not automatically become automatable simply because it can be described in natural language. It is necessary to determine whether its operations can be formalised sufficiently, whether the required data are accessible and whether the execution conditions can be defined.

Automated invoice processing provides an example. The system must receive the document, extract information, verify certain fields, apply business rules and transmit the result to an accounting application.

Some stages can be performed using explicit rules. Others, such as document recognition or field extraction from a variable layout, may require a machine learning model.

Programming connects these functions and organises error cases, verification procedures and interactions with other systems.

8.2. Conventional Automation and Intelligent Automation

Conventional automation generally relies on rules and sequences of operations whose behaviour is explicitly defined. It is particularly suitable for repetitive tasks, stable processes and environments in which conditions are predictable.

AI-based automation can process less structured inputs, interpret complex information and adapt certain decisions on the basis of learned patterns.

This distinction does not mean that AI systematically replaces conventional methods. The most robust solutions often combine the two approaches.

An industrial control system, for example, may use a learned model to detect an anomaly, then apply deterministic rules to trigger a safety procedure. A robot may use a general-purpose model to interpret an instruction while relying on a specialised controller to constrain its movements.

This hybrid architecture makes it possible to benefit from the flexibility of learned models without abandoning the explicit constraints that make certain operations verifiable.

8.3. Artificial Intelligence Agents: Models Integrated into Processes

Artificial intelligence agents are systems capable of using a model to select operations, interact with tools and pursue an objective within a given environment.

Depending on their design, they may search for information, call application programming interfaces, execute code, manipulate files or organise several processing stages.

An agent is therefore not reducible to its model. Its behaviour also depends on the tools available to it, its instructions, control mechanisms, available memory and the permissions it has been granted.

An agent tasked with preparing a report, for example, may need to search for documents, extract data, perform calculations and produce a structured result. The model helps interpret the objectives and select certain operations; the program orchestrates tool calls and manages their results.

This arrangement brings the classical principles of programming back into view: task decomposition, state management, execution conditions, error handling and result verification.

The more extensive a system’s capacity for action, the more important it becomes to define its limits and monitor its behaviour.

8.4. Conditions for Reliable Automation

Automation should not be evaluated solely in terms of time saved or the number of operations performed without human intervention. Its quality must also be assessed in terms of output accuracy, process robustness and the ability to detect failures.

Reliable automation requires, in particular:

• a clearly defined objective;
• sufficiently reliable input data;
• rules or models suited to the task;
• validation and error-handling procedures;
• recovery mechanisms in the event of interruption;
• traceability for significant operations;
• supervision proportionate to the consequences of an error.

The recommendations of the National Institute of Standards and Technology (NIST), notably its Artificial Intelligence Risk Management Framework and its profile devoted to generative AI, provide resources for incorporating risk management into the design, evaluation and operation of artificial intelligence systems.

Source: NIST, AI Risk Management Framework.

These principles extend a requirement already established in traditional computing: a system must be designed not only to perform its task under normal conditions, but also to respond in a controlled manner when its assumptions no longer hold.

9. How Can a System Combining Models, Software and Artificial Intelligence Be Evaluated?

Evaluating an artificial intelligence system requires several levels of performance to be distinguished. A model may produce high-quality predictions without the complete application being reliable; a program may function correctly while relying on an unsuitable model.

A rigorous evaluation method therefore examines the representation, the implementation and the conditions of use together.

9.1. Assessing the Relevance of the Model

The first question is whether the model represents the phenomenon or task under investigation adequately enough.

For a predictive model, evaluation may rely on metrics suited to the objective, such as mean error, precision, recall, false-positive rate or other relevant indicators. The metric selected must reflect the practical consequences of errors.

For a physical model or simulation, comparison with experimental measurements, reference solutions or observed behaviour helps establish its domain of validity.

For a language model, evaluation may examine factual accuracy, instruction following, reasoning quality on defined tasks, robustness and the ability to acknowledge certain limitations.

No single metric is sufficient to characterise every property of a complex system.

9.2. Assessing the Quality of the Program

The program must be examined independently of the model it uses.

It is necessary to verify that data are transmitted correctly, transformations respect the expected formats, errors are handled appropriately and interfaces between components function as intended.

In a generative AI application, tool calls, permission management, information protection and behaviour when the model produces an unexpected response must also be checked.

Automated tests can verify known behaviours, while additional trials make it possible to explore edge cases and situations not fully covered by the specifications.

9.3. Evaluating the System in Its Real Environment

The final stage consists in observing the complete application’s behaviour under conditions resembling those of actual use.

Performance may vary according to users, equipment, data, time constraints or environmental changes. A model that performs well during initial validation may become less effective if the characteristics of the data change.

In robotics, this evaluation must include movement stability, manipulation accuracy, obstacle avoidance and compliance with physical limits.

In a document-processing application, it must examine atypical documents, recognition errors and the consequences of missing data.

In a software agent, it must assess the success of the complete task, the appropriateness of tool calls, failure handling and the ability to regain control when the system encounters an unforeseen situation.

Operational evaluation is thus the point at which all three dimensions — model, program and environment — must be considered together.

10. Epistemological Limitations: When Performance Is Mistaken for Understanding

A system’s technical sophistication is not sufficient to establish the scope of its knowledge. A model may be useful, predictive and effective without faithfully reproducing every mechanism of the phenomenon it represents.

This distinction is particularly important in contemporary debates about artificial intelligence. A model’s ability to produce a relevant answer does not, in itself, establish whether it possesses a reliable representation of every dimension of the problem.

10.1. Correlation Is Not Always Explanation

A learning model may identify highly effective statistical patterns without isolating the causal relationships governing a phenomenon.

If a system frequently associates a visual pattern with a particular category, it may perform well on familiar examples while failing when that pattern appears in a different context.

Similarly, an economic model may correctly forecast certain developments as long as the observed relationships remain relatively stable, only to become less relevant when institutional conditions or behaviour change.

This limitation does not render statistical models useless. It requires us to distinguish predictive accuracy, the interpretation of relationships and an understanding of causal mechanisms.

10.2. Hallucinations in Generative Models

Language models can produce inaccurate information, nonexistent references or incorrect reasoning presented in a convincing form. These errors, commonly referred to as “hallucinations”, arise in part from limitations in the data, the learned representation and the generation mechanism.

Linguistic fluency can lend an answer an appearance of authority that exceeds its actual level of reliability.

Several methods can reduce certain risks: document retrieval, source verification, computational tools, output constraints, testing and human intervention. None, however, guarantees universal accuracy.

A system’s quality must therefore be measured through procedures suited to the task, rather than inferred solely from how convincing its responses appear.

10.3. The Problem of Transfer to Unfamiliar Situations

A model is trained and evaluated on particular data distributions. When it encounters a substantially different situation, its performance may deteriorate.

This problem affects vision models, language systems, scientific models and autonomous robots alike.

Generalisation is a central objective of machine learning, but it does not mean that every future situation has been covered. Real-world environments may contain objects, behaviours or combinations of conditions absent from the training data.

A reliable architecture must therefore provide for the detection of unusual situations, the management of uncertainty and mechanisms that can suspend an action when safety conditions are no longer met.

10.4. The Role of the Observer in Constructing a Model

Modelling involves choices: selecting variables, defining categories, determining objectives and choosing criteria for success.

These choices may introduce biases or privilege certain aspects of a phenomenon at the expense of others. In artificial intelligence systems, the composition of training data, the definition of labels and the metrics selected also help determine the behaviours learned.

It is therefore necessary to document assumptions, specify domains of application and acknowledge the limits of the system’s validity.

This requirement applies not only to scientific models but also to commercial applications, recommendation systems, recruitment tools and automation technologies that influence human decisions.

11. Towards Integrated Design: Thinking About Models, Algorithms and Systems Together

The evolution of artificial intelligence calls for a departure from the view that a model can be designed independently of the program that uses it.

In contemporary systems, model architecture, data, algorithms, hardware resources and the execution environment are closely interconnected. Decisions made at one level can alter the possibilities available at the others.

A larger model may improve certain aspects of performance while increasing computational costs and latency constraints. A simpler representation may be less accurate for some tasks but easier to deploy on an embedded device. A modular software architecture may facilitate future changes while introducing communication overhead between components.

Designing an effective system therefore requires explicit trade-offs between output quality, resource consumption, robustness, safety and ease of maintenance.

11.1. Co-Designing the Model and the Program

Co-design consists in addressing modelling choices and implementation choices together.

In an embedded system, for example, the model must take account of available memory, computing power and the maximum time permitted to produce a response. The program must, in turn, exploit the model’s characteristics without compromising system stability.

Techniques such as quantisation, compression and distillation can help adapt certain models to deployment constraints. However, they may affect performance and must therefore be evaluated in the context in which they will be used.

The same principle applies to robotics: perception and planning models must be compatible with control frequencies, actuator capabilities and safety constraints.

11.2. Hybrid Architectures

Hybrid architectures combine several approaches: explicit rules, learned models, search methods, simulations, knowledge bases and control mechanisms.

This combination makes it possible to allocate functions according to their respective properties. A language model may interpret a request, a rule engine may verify certain constraints, a search algorithm may determine a route and a specialised controller may ensure physical execution.

The value of such an architecture lies in the complementarity of its components. Learned models provide the ability to process complex data, while explicit procedures can constrain certain operations and facilitate verification.

This arrangement does not automatically guarantee reliability: interfaces between components must also be tested, and errors in one module may propagate to others. Nevertheless, it provides a framework for designing systems in which responsibilities are more clearly differentiated.

11.3. Traceability and Technical Governance

As systems become more complex, it becomes necessary to retain information about their design, data, versions and evaluation results.

Traceability makes it possible to establish which model was used, under what conditions it was executed and which transformations were applied to the data. It facilitates the reproducibility of experiments, error analysis and comparisons between versions.

In production AI systems, monitoring can also detect changes in the data or declining performance. Reassessment or updating procedures can then be triggered.

These mechanisms contribute to the responsible design of artificial intelligence. They are not an administrative afterthought: they extend the established requirements of documentation, testing and maintenance that are central to software engineering.

12. Frequently Asked Questions About Models, Programming and Artificial Intelligence

What Is the Difference Between a Model and a Computer Program?

A model represents an object, phenomenon or relationship according to specified assumptions. A computer program organises executable operations that make it possible, among other things, to calculate, simulate or use that representation. A learned model can be used by a program without being identical to it.

What Is the Role of an Algorithm in Artificial Intelligence?

An algorithm defines a processing or learning method. It may organise a search for a solution, the classification of data or the adjustment of a model’s parameters. The program provides an executable implementation of that algorithm within a computing environment.

Is an Artificial Intelligence Model Itself a Program?

Not necessarily, at least not in the strict sense. An AI model generally refers to an architecture and, once trained, a set of parameters that determine part of its behaviour. A program is required to train it, execute it or integrate it into an application.

Why Is Data Essential to Machine Learning?

Data provide the examples from which a model learns patterns. Their quality, representativeness and suitability for the task influence the performance achieved. Biased or insufficient data can limit generalisation and lead to systematic errors.

Will Traditional Programming Disappear Because of Generative AI?

Generative AI can automate certain software development operations, including code generation, function explanation and the search for corrections. It does not eliminate the need for design, validation, security and maintenance. Programming is evolving towards greater collaboration between developers and assistance tools.

What Is the Relationship Between Models, Programming and Robotics?

Robotics combines models representing the environment or the robot’s behaviour, algorithms determining possible actions and programs that execute calculations and control physical mechanisms. Artificial intelligence can enhance perception, planning and adaptation, but its integration must respect timing and safety constraints.

Why Can a High-Performing Model Produce Incorrect Results?

Its performance depends on data, assumptions, learning methods and conditions of use. A model may succeed in familiar situations while failing when faced with unusual data, a change of context or a question outside its domain of validity.

What Is Intelligent Automation?

Intelligent automation combines programs, data and, where appropriate, artificial intelligence models to perform tasks or adapt certain decisions. It can bring together deterministic rules, learned models and control mechanisms to handle more variable situations than conventional automation can accommodate.

Conclusion: Models Represent, Programming Organises, Automation Acts

The history of models and programming reveals a fundamental continuity: transforming knowledge into technical action requires a representation of the problem and a definition of the operations through which that representation can be used.

Models make certain aspects of reality calculable, comparable or predictable. Algorithms organise processing methods. Programs translate these methods into executable operations. Automation finally connects these operations to concrete objectives, whether the task is to process information, simulate a phenomenon or control a physical system.

Contemporary artificial intelligence renews this relationship by enabling models to learn representations from data, generate language, interpret images and contribute to action planning. It nevertheless leaves intact both the need to formalise problems and the need to verify results.

Embodied robotics, particularly as developed in several Asian research ecosystems, makes this interdependence especially visible. When a model must guide a robot through a real-world environment, the quality of the representation, the precision of the program, the properties of the hardware and the safety of the action become inseparable.

The performance of an artificial intelligence system therefore depends not solely on the power of its model, but on the coherence of the entire system formed by its data, algorithms, software, hardware resources and conditions of use.

This perspective makes it possible to understand AI not as an isolated technology, but as a new stage in the history of scientific modelling, computing and automation. The central challenge remains that of every scientific and technical endeavour: to construct useful representations, test them against evidence and transform their results into actions whose effects can be evaluated.