Originally Posted On: https://neuraking.com/en/27061/intelligence-artificielle/

Artificial Intelligence (AI): Understanding and apprehending contemporary AI

Artificial intelligence (AI) is usually defined as a set of computer techniques and systems that are capable of simulating human cognitive processes by relying on machine learning algorithms. However, this already vague, confusing and limited definition is now reduced to the spectrum of LLM (Large Language Model), which, in the collective mind, constitute artificial intelligence. This generates even more confusion and reinforces dogmatic postulates that we absolutely must get out of, because they lead us to a loss of control.

To better understand the challenges of tomorrow, everyone must know how to differentiate the nuances that each definition associated with AI integrates into its statement, in order to distinguish the truth in the incessant informational noise.

To get there, we must start by recalling the history of AI, which, since the 50s, has been based on unfounded promises. Even then, advances considered spectacular sparked an unprecedented cacophony of expectations, all of which, without exception, gave way to disillusionment, leading to a period of glaciation known as the AI ​​winter.

Today, we are witnessing the same phenomenon, except that AI will, this time, really change the face of the world.

In this article, we will demystify each concept, to navigate with greater clarity in a space confused by excessive discourse that relies on confusion.

The History of Artificial Intelligence

The history of artificial intelligence (AI) spans several decades, marked by key moments that have shaped its development and integration into our daily lives.

Early Artificial Intelligence (1950s)

In the 1950s, computer science pioneer Alan Turing proposed the famous Turing Test, a measure of machine intelligence. This test aims to assess a machine’s ability to imitate human behavior indistinguishably.

Dartmouth Conference 1956

The 1956 Dartmouth Conference brought together leading researchers such as John McCarthy, Marvin Minsky, and Claude Shannon, who together laid the foundations of the field. They asked fundamental questions about the ability of machines to simulate human intelligence. This meeting was not just an exchange of ideas; it embodied a collective vision of a future in which computers could not only perform tasks, but also learn and adapt.

Development of the first AI programs (1960s)

In the 1960s, the first AI programs emerged, such as ELIZA, which simulated a therapeutic conversation by imitating a psychologist. This advance illustrated the ability of machines to interact with humans, albeit at a rudimentary level. At the same time, SHRDLU, another program, demonstrated natural language understanding by interacting with a block environment. These initial developments generated as much excitement as we are seeing today, since the launch of ChatGPT in late 2022.

AI Winters (1970s-1980s)

The following decades, particularly the 1970s and 1980s, saw a marked slowdown, dubbed the “AI winters.” This term refers to a period of disillusionment when interest and funding for AI projects plummeted, due to unfulfilled promises and disappointing results. Researchers’ high expectations were not realized, leading to setbacks that affected the credibility of the field.

Renaissance of AI with neural networks (1980s-1990s)

Faced with the limitations of the symbolic and logical approaches that dominated previous decades, researchers are exploring promising new avenues: machine learning and neural networks.

This paradigm shift is based on the idea that machines can learn from data on their own, without being explicitly programmed. Machine learning algorithms enable systems to identify patterns and relationships in large data sets, paving the way for more flexible and adaptive applications.

This paradigm shift lays the foundation for the rise of deep learning and modern AI, which will transform many industries in the decades to come.

Which will be driven by one event in particular, when Deep Blue beats chess champion Garry Kasparov in 1997. Although Deep Blue is not based on machine learning, its success largely contributes to stimulating new interest and investment in AI, which now integrates the paradigm of deep learning and neural networks.

The paradigm of deep learning and neural networks now permeates the artificial intelligence sector

Recent Advances in AI (2000s-2020s)

The next two decades, from 2000 to 2020, were marked by a massive implementation of artificial intelligence in various fields. The integration by companies like Google and Apple of voice and facial recognition systems into their products revolutionized the way users perceived and used technology on a daily basis. This democratization of AI was a direct result of the advances made in the field of machine learning and neural networks in the previous decades.

The use of AI by streaming platforms such as Netflix and Amazon in their recommendation systems is also a direct consequence of the ability of learning algorithms to identify patterns and relationships in large data sets. By analyzing viewing behaviors and individual preferences, these AI systems are able to offer personalized content, promoting user retention and satisfaction.

The success of this recommendation model is also leading to a profound transformation in content consumption and the design of the offer by companies. The emergence of TikTok, which relies heavily on recommendation algorithms, perfectly illustrates the impact of this approach.

These advances in AI, enabled by the paradigm shift toward machine learning and neural networks, have profound implications for many aspects of daily life and business practices. They are paving the way for a new era where artificial intelligence becomes ubiquitous and profoundly transforms society.

AI models running on neural networks and deep learning are ubiquitous, but remain invisible to the general public.

2023: everything is accelerating

The emergence of ChatGPT in November 2022 is part of the ongoing paradigm shift towards neural networks and deep learning. This major breakthrough is a direct result of progress in the field of language processing.

ChatGPT is not limited to generating text; it promises to be a virtual assistant capable of dialoguing fluidly and relevantly in various contexts.

It brings in its wake the hope of AGI (Artificial General Intelligence) and opens the way towards ASI (Artificial Super Intelligence).

But these new perspectives, offered by the ability to dialogue with a system, confront humanity with questions that have remained unanswered until now.

On the other hand, this apparent simplicity that language offers in the intermediation of action (via the machine) has quickly given way to an unsuspected complexity which, once again, hinders widespread adoption.

The enthusiasm, although very strong, remains restricted to the digital environment which faces a worrying saturation of AI-generated content and countless problems, such as alignment, regulation, security, sovereignty, etc.

The artificial intelligence sector is now moving towards the AGI and ASI paradigm.

2025: Adoption of artificial intelligence promoted despite risks

The beginning of 2025 was marked by a new acceleration, catalyzed by two events.

First, Donald Trump’s rise to power radically liberalized the use of AI through an executive order signed on January 23, 2025, aimed at restoring US dominance in AI by removing existing policies seen as obstacles to innovation. Trump then formalized the Stargate project, with a planned monumental investment of $500 billion, making AI a national priority.

Second, a few days after these announcements and decisions, the emergence of a Chinese AI (DeepSeek) revolutionized the sector by offering more efficient solutions at a lower cost than the American leaders in the sector.

This development has thrown the AI ​​industry into turmoil, exposing major power and political issues. The data, then sent to China, has raised concerns about sovereignty and security, as these practices help spread the Chinese narrative. Some have proposed laws to ban the use and downloading of this Chinese AI, suggesting prison sentences of up to 20 years for simply downloading the AI ​​model.

Faced with this dynamic where confusion already reigned in the definitions, a fine understanding of what artificial intelligence is becomes crucial, because in addition to the confusion in the definitions, there is the confusion around its implications while increasingly significant risks are taken. The need for ethical and political reflection on the use of AI is therefore pressing.

But to do this, we need to understand what we are talking about, and what developments have brought us to where we are.

Focus on the evolution of recent paradigms of artificial intelligence

The timeline of developments shows us the turning point of the 1980s and 1990s when neural networks and machine learning gradually became central.

The end of the AI ​​winter coincides with this “refocusing”, which, later in 2006, will receive a spotlight with the work of Geoffrey Hinton and the ImageNet competition which will truly revive interest in AI.

• • 2006: Geoffrey Hinton publishes a paper on deep learning, revolutionizing the field of AI.
  • 2012: The AlexNet convolutional neural network wins the ImageNet competition, paving the way for advanced visual recognition.

Then, it is in 2017, in the continuity of this approach to AI, that the architecture Transformer, presented by Vaswani et al. in a publication, revolutionizes natural language processing by its ability to handle long-term dependencies in the text.

In the process, from 2018, OpenAI begins making history by publishing GPT (Generated Pretrained Transformer), a model based on the architecture Transform whose results show a propensity to improve proportionally to the quantity of data and the width of the model.

Birth of a new paradigm: scaling makes intelligence

This will be followed by advances confirming each time that increasing the size of the model and the data set improves the consistency and relevance of the responses generated.

• 2019 : GPT-2 Development OpenAI unveils GPT-2 with 1,5 billion parameters, demonstrating that increasing the size of the model and dataset improves the consistency and relevance of the generated answers.

Access to massive data from the Internet is becoming crucial to increasing scale and therefore intelligence.

• 2020 : GPT-3 Launched GPT-3, with its 175 billion parameters, proves once again that more data and computing power can cover a wider range of linguistic tasks, with impressive performance on complex tasks with few examples.
• 2021 : Studies on the Scaling Laws. Research, such as that conducted by Kaplan et al., formalizes the concept that the performance of language models continues to grow with exponential increases in data and parameters, while optimizing computational costs.

Confirmation of the paradigm based on data volume, model size and computing power.

• 2022-2023 : Continuous applications and developments.
  Industry and research continue to advance the foundations of this quantitative paradigm, delving deeper into areas such as text processing, machine translation, and content generation. This will lead to ever more powerful and intelligent models, confirming each time that increasing the volume of data and the size of models is an effective and reliable strategy.

Meanwhile, sectoral impacts are observed

Alongside the specific trajectory taken by Transformer models, other advances confirm the relevance of neural networks and machine learning. The latter are proving their worth in several sectors, gradually but surely helping to prepare for the advent of AI.

Applications of AI in healthcare:

• 2015 : IBM launches Watson Oncology, a decision support system for oncologists.
  • Analysis of medical records by ML and NLP.
  • Helps in the diagnosis and treatment of cancer.
• 2016 : Google DeepMind develops algorithm to detect eye diseases from retinal scans.
  • Detection by convolutional neural networks (CNN).
  • Deep learning to analyze retinal scans.
• 2018 : FDA approves IDx-DR, an AI algorithm for diagnosing diabetic retinopathy without human intervention.
  • Using CNN and deep learning.
  • Analysis of retinal images.
• 2020 : AI in the face of the COVID-19 pandemic
  • AI is being widely used to address the COVID-19 pandemic, from tracking the spread to developing treatments.
  • Tracking the spread of the virus.
  • Development of treatments.
  • Machine learning techniques: neural networks, data analysis.

Applications in finance:

• 2009 : Financial institutions are beginning to experiment with machine learning to detect fraud.
  • Supervised algorithms: decision trees, neural networks
  • Emerging tools: scikit-learn, H2O.ai
• 2015 : Deep learning is booming in algorithmic trading and financial market analysis.
  • Generalization of RNNs for time series analysis
  • Democratization of libraries: Keras, TensorFlow
• 2018 : AI is emerging as a driver of transformation in financial services.
  • JPMorgan Chase Launches COiN: NLP and ML to Analyze Contracts
  • Refinement of NLP techniques: word2vec, BERT
  • Proliferation of specialized bookstores: spaCy, Hugging Face
  • Generalization of Python libraries for ML

Marketing Transformation:

• 2010 : Amazon is incorporating machine learning algorithms to deliver ultra-personalized recommendations to its customers, ushering in a new era of e-commerce.
  • Using Collaborative Filtering and Neural Networks
  • Analysis of user purchase histories and behaviors
• 2015 : Facebook leverages machine learning and data analytics to understand user interests and behaviors, enabling ad targeting with unparalleled precision.
  • Analysis of interactions, likes and shares
  • Fine segmentation of advertising audiences
  • Real-time campaign optimization
• 2018 : Alibaba takes next step by using recurrent neural networks and natural language processing techniques to automatically generate product descriptions and marketing campaigns.
  • Generation of persuasive and product-specific texts
  • Personalization of messages based on customer preferences
  • Optimizing campaign performance through continuous learning

Applications in logistics:

• 2012 : Amazon Robotizes Its Warehouses with the Acquisition of Kiva Systems
  • Using reinforcement learning algorithms and neural networks
  • Autonomous navigation and precise object manipulation
  • Optimization of storage and order preparation processes
• 2016 : DHL deploys collaborative robots and autonomous vehicles
  • Using machine learning techniques and neural networks
  • Seamless collaboration between robots and human operators
  • Optimization of goods flows and reduction of errors
• 2019 : UPS gets FAA approval for drone deliveries
  • Using Machine Learning for Path Planning and Obstacle Detection
  • Autonomous navigation and real-time decision making
  • Route optimization and reduction of delivery times

The combination of these application successes not only prepared companies to ride the wave of Transformers models, but also the general public, who, without knowing it, were already confronted with neural networks and machine learning on a daily basis.

So, while the paradigm of neural networks and machine learning was taking hold in all directions, that of intelligence through the scaling of Transformers models was also proving its worth in parallel.

2023: GPT-3 redefines the concept of artificial intelligence

The launch of GPT-3 in late 2022 marked a major turning point, redefining the very notion of artificial intelligence as it looms large in the public mind.

Promises and speculations flew from all sides, raising fears of an immediate replacement of human work by machines. Rumors and fantasies emerged, reminiscent of the excitement of the 50s, but with a notable difference: AI was now accessible to the general public, at low cost, and capable of communicating smoothly and quickly.

This is not a detail, because the exchange of words constitutes the intermediary factor par excellence between man and machine, the ultimate factor of productivity, efficiency and relevance.

A Transformer model can, in fact, perform most of the tasks of all other types of artificial intelligence, but with an additional reasoning capacity, making it possible to envisage autonomy and automation pushed to the extreme.

This involves sensitive applications (health, defense, finance) that are concrete and potentially deployable in the medium term, giving even more interest to the sector, which could then bet on AGI.

Disillusionment with the complexity of language

But the initial euphoria soon gave way to new disillusionment once again. Using words to instruct a machine proved to be far more complex than it had seemed. The subtleties of human language, with its nuances, ambiguities, and varying contexts, posed considerable challenges for AI models. Despite their ability to generate impressive responses, these systems showed limitations in deeply understanding the meaning and intent behind words.

The emergence of prompt engineers

To address these challenges, a new profession emerged: that of the prompt engineer. These specialists dedicated themselves to the art of formulating precise and effective instructions to guide AI models. Their expertise proved essential to obtain relevant and consistent results. However, this new layer of expertise also increased the cost of AI integration in existing processes, thus reducing the initial attractiveness of this technology.

2023 – 2024: expansion of the new definition of “artificial intelligence”

Advances in Transformers models initially limited their capabilities to textual tasks, such as text analysis, correction, and content generation. However, many new models have emerged from various industry players, creating intense competition for opportunities and markets.

This competition has stimulated the development of advanced features and innovative usage methods, giving rise to multimodal Transformers models. These allow, from text instructions, to generate images, videos, and, conversely, to transpose images into text or to transcribe a video into text.

Immersive dimension of the AI ​​concept

The definition of artificial intelligence has thus taken on an immersive dimension, mixing text, image, sound and video. However, the concept of AI remains strongly associated with GPT (Generative Pretrained Transformer) models, which are the key to these possibilities. For example, actions such as “Text to Image” or “Text to Video” are based on language. Even image analysis, although not initially textual, must be interpreted by a Transformer to produce a textual result.

This approach applies to many other applications, such as programming from text, automatic summarization, creating websites or PDF documents, and many others, in both directions. However, by focusing on multimodality, the world tends to forget that all this is based on words, these tools that humans use to give instructions to the machine. Which will inevitably pose problems and will further redefine the contours of what artificial intelligence evokes.

Moreover, this definition also involves political notions.

2025: the definition of artificial intelligence includes a geopolitical dimension

DeepSeek’s dramatic arrival on the market has raised growing concerns among observers, who perceive this innovation as a major security risk, particularly in the United States. This risk revolves around sovereignty, because the processing of American users’ data is based on acceptance of terms of use. In other words, users’ implied consent opens the door to exploitation of their personal information by China.

In addition, the availability of DeepSeek’s latest reasoning model, licensed by MIT, outperforming OpenAI’s AIs, intensifies fears of an attack on cognitive sovereignty. These fears are exacerbated by an underlying narrative, associated with China, reinforcing the perception of a threat.

Some voices, alarmed like that of Senator Josh Hawley, go so far as to propose draconian measures, such as the outright ban on downloading the DeepSeek AI, considering prison sentences for anyone who dares to use or download it.

Yet using an AI model, even a foreign one, when it is isolated, hosted on a sovereign territory and operated by a local company, does not necessarily lead to data sharing. Which underlines that the concern is indeed about the subversion generated by the narrative associated with the model.

This subtle, but equally significant risk, some people still minimize it, but it is very real and tends to be integrated into the definition of AI, the latter also being a political tool of Soft Power and subversion, in other words, a weapon.

On the other hand, Meta has been criticized for publishing open source versions of its AI model Llama. The latter was widely used by DeepSeek engineers to design their reasoning AI model, so some see Meta’s open source approach as a national betrayal.

This shows how the definition of intelligence has countless dimensions which are constantly evolving.

Thus, the evolution of AI paradigms influences its different definitions and constantly modifies them, over the course of functionalities, algorithm developments, but also, according to the level of threat perceived on a geopolitical level.

Now that we have shed light on the evolution of AI paradigms, in order to temper the notions that it includes in its perception, we can calmly dwell on the different technical definitions of artificial intelligence in order to then better understand the profound upheavals that transformer models and their multimodal ancestors imply.

Back to basics: conventional definitions of artificial intelligence (AI)

Artificial intelligence (AI) is commonly defined as systems that can learn and adapt to new situations. This encompasses an infinite number of underlying definitions, which is a source of confusion.

To better understand, we will discuss the definition of AI by its two main objectives what are the prediction et automation.

AI is a tool to predict from data: particularly useful in areas such as finance, defense or health.

AI is a solution to automate complex tasks to improve efficiency and innovation.

Both of these goals are implemented through AI’s ability to process massive amounts of data, extracting the information needed for prediction and automation.

Artificial intelligence to predict

The starting point for any prediction is data, and data must be massive to have a significant impact.

This requires a specific approach, called unsupervised learning, which allows us to explore and reveal hidden structuress within the data sets.

Building Correlations with Unsupervised Learning

Unsupervised learning is particularly useful in contexts where class labels (classification categories) are not available, allowing AI systems to explore data autonomously.

Methods :

Two methods stand out in this area: clustering and dimensionality reduction.

Clustering method

Clustering, or grouping, aims to organize a set of data into groups based on similar characteristics.

Clustering technique

Several techniques are used to accomplish this task:

• K-Means : This technique groups similar data together to identify natural groups. By defining a predetermined number of clusters, K-Means assigns data points to the closest class, creating clear partitions within the data.
• Hierarchical Clustering : This technique creates a hierarchy of clusters, revealing relationships between data. By building a decision tree, hierarchical clustering allows data to be explored at different levels of granularity, making it easier to identify underlying structures.
• DBSCAN : Based on the density of data points, the DBSCAN method identifies dense groups capable of handling unconventional cluster shapes. This model is particularly effective in detecting clusters of arbitrary shape, while ignoring the noise present in the data.

Dimensionality reduction method

Dimensionality reduction is a method to simplify data while preserving essential information.

This approach is of paramount importance, especially when the data evolves in high-dimensional spaces (data sets with a large number of variables).

Dimensionality Reduction Techniques

• Principal component analysis (PCA) : PCA reduces the dimensionality of data by transforming the original variables into a set of new uncorrelated variables, called principal components. This process preserves the majority of the variance in the data, making it easier to analyze while reducing its complexity.
• t-SNE : t-SNE is an advanced technique for visualizing local structures in high-dimensional data. By preserving relationships between nearby points, t-SNE produces dense and meaningful visual representations, providing rich insights into complex data.

Capturing complex representations with deep learning

This approach to prediction (deep learning) is the one on which Transformers models are based, but not only.

Deep Learning Support

• Convolutional Neural Networks (CNN) : These models are particularly suited to prediction in the field of computer vision, where they learn to extract essential features from images. Thanks to their ability to process spatial data, CNNs enable significant advances in image recognition and visual classification.
• Recurrent Neural Networks (RNN) : RNNs are designed to predict temporal sequences, learning to model dependencies over time. This type of network is essential in applications such as machine translation and time series analysis.
• Natural Language Processing (NLP): Sentiment analysis, named entity recognition, machine translation: These tools enable the capture of complex representations of text, thereby optimizing big data analyses and facilitating the identification of specific information.
  • Transformers : The architecture of Transformers models is based on an attention matrix, which allows the model to focus on specific parts of a sequence of words, thus capturing even deeper and more complex relationships and improving the management of contextual dependencies. This attention capacity makes Transformers particularly efficient.

It is understood here that Transformers represent only a fraction of what artificial intelligence encompasses, but their attentional architecture is their distinctive factor that allows them multidimensional and sectoral utility.

Guiding correlations with supervised learning

Supervised learning relies on using labeled data to train models, which will allow them to make predictions on new data.

This approach is based on several methods, among which classification and regression are predominant.

Supervised learning classification method

Classification involves assigning labels to observations based on characteristics extracted from the data.

Classification techniques

Several classification techniques are used to achieve this:

• Use of neural networks : During the training phase, the weights of the connections between neurons are iteratively adjusted to minimize the classification error on the labeled data. Propagating the signal through the layers helps capture complex patterns and nonlinear relationships in the data. Once trained, the neural network can be used to predict the class of a new observation by propagating its features through the network.
• Use of Support Vector Machines (SVM) : This technique aims to find the hyperplane that best separates the classes in the feature space. The SVM optimization algorithm seeks to maximize the margin, that is, the distance between the hyperplane and the closest points of each class. SVMs often use kernel functions to transform the feature space, allowing nonlinear separations between classes to be found. During prediction, a new observation is classified according to its position relative to the hyperplane.
• Use of decision trees : This technique involves constructing a tree where each node represents a test on a feature of the data. The algorithm recursively selects the feature that best divides the data into homogeneous subsets with respect to the target class. The leaves of the tree represent the predicted classes, and the path from the root to a leaf corresponds to a decision rule. During prediction, a new observation traverses the tree following the tests on its features until it reaches a leaf that indicates its predicted class.

Supervised learning regression method

Regression, on the other hand, aims to model the relationship between variables in order to predict continuous values.

Regression techniques

Several supervised learning regression techniques stand out to achieve this:

• Using linear regression : This technique establishes a linear relationship between variables using a straight line equation. The coefficients of the equation are estimated from the training data by minimizing the mean square error between the predicted values ​​and the actual values. Once the model is trained, it can be used to predict the value of a dependent variable based on the values ​​of the independent variables. Linear regression is widely used in fields such as economics and finance to identify relationships between variables.
• Using logistic regression : Although called regression, this technique is actually used for binary classification. It models the probability of belonging to a class as a function of the explanatory variables. The logistic function is used to transform the linear combination of variables into a probability between 0 and 1. The model is trained by maximizing the likelihood of the training data, that is, by finding the coefficients that make the observed data most probable. Logistic regression is particularly useful in fields such as medicine or marketing to predict the probability of a binary event.
• Using polynomial regression : This technique extends linear regression by including polynomial terms of the explanatory variables. It allows nonlinear relationships between variables to be modeled using polynomials of degree greater than 1. The coefficients of the polynomial terms are estimated from the training data by minimizing the mean squared error. Polynomial regression provides greater flexibility to accommodate fluctuations in the data when the relationship between the variables is not simply linear. It is particularly useful when the data exhibit curvilinear trends or complex variations.

 

To achieve its prediction goal, AI must

Building correlations from scratch (unsupervised learning)

 

Orienting correlations from classifications and relationships between variables (supervised learning)

In both cases, neural networks play a determining role that is exacerbated by Transformers models. The latter, although not explicitly listed in the techniques deployed in classification and regression, are increasingly used as an interpretation intermediary, both in unsupervised and supervised learning. This introduces an additional statistical compression element whose aim is to increase accuracy, but whose stochastic essence can also introduce a distortion in the interpretation of the results.

 

Transformers models are redefining machine learning directly and indirectly.

The same applies to achieving the automation goals of artificial intelligence.

 

Artificial intelligence to automate

Task automation helps improve efficiency and reduce human errors through processes.

This automation relies on several methods, including reinforcement learning, deep learning and natural language processing, computer vision, and expert systems.

Automate with Reinforcement Learning

Unlike the methods and techniques discussed so far, reinforcement learning focuses on interaction with an environment.

Rather than learning from examples to build a predictive base, the agent will receive context-related rewards or penalties to ensure reliability in a fully automated process.

Reinforcement learning technique:

• Q-Learning and SARSA: These algorithms are the basis for automating action choices in environments. They work by continuously updating the Q value, which represents the quality of an action in a given state.
  • Operation: The agent explores the environment, choosing actions and receiving rewards or penalties. These experiences are used to adjust Q values, allowing the agent to gradually determine the best action for each state to maximize long-term rewards.
  • applications: Ideal for scenarios where the optimal strategy must adapt to changing environments, such as in games or robotics.
• Monte Carlo method: This technique relies on random samples to evaluate and optimize action strategies.
  • Operation: The agent performs multiple episodes of interaction with the environment, collecting data on the rewards obtained. Then, it uses this information to estimate the expected value of different strategies. This approach does not require knowledge of the model of the environment, which makes it powerful for automation in contexts where information is incomplete or uncertain.
  • applications: Used in situations where it is difficult to model the environment precisely, such as in resource management or financial simulations.

Impact on automation:

Reinforcement learning enables robust and adaptive automation as agents learn to navigate complex environments by optimizing their behavior to maximize rewards.

Reinforcement learning is particularly useful for:

• Automate complex decision-making processes where actions have long-term consequences.
• Adapt to dynamic environments where conditions may change, requiring constant re-evaluation of strategies.
• Control systems where exact modeling is impossible, providing a solution through direct experience.

Automate with Deep Learning

Deep learning, as an automation tool rather than a prediction tool, allows machines to perform autonomous operations. From textual responses to a user, to detecting and correcting anomalies in a database, to autonomous decisions to adjust a classification or correct an industrial process, deep learning eliminates the need for constant human supervision.

This automation naturally extends to language processing, where once again, Transformers models make interactions with machines more fluid and efficient.

Techniques and architectures used:

• Recurrent Neural Networks (RNN) and LSTM :
  • Use : Excellent for sequential data like translation, speech recognition and text generation.
  • Benefit : Ability to retain information over the long term, which facilitates the automation of tasks requiring context.
• Convolutional Neural Networks (CNN) :
  • Use : Essential for computer vision.
  • Benefit : Automatic extraction of visual features at different scales, perfect for image classification, detection and segmentation.
• Natural Language Processing (NLP) for Automation :
  • Sentiment analysis, named entity recognition, machine translation: Automate text understanding for big data analytics, translation or identifying specific information.
  • Speech Recognition : Enables natural voice interactions, making systems more accessible and intuitive.
  • Natural language generation : Produces coherent and contextually relevant text, essential for virtual assistants, automatic report writing or answering user questions.
  • Transformers and attention mechanisms :
    • Use : Transformation of natural language processing.
    • Benefit : Ability to understand relationships between distant words, improving the quality of linguistic tasks such as text generation or sentiment analysis.
• Generative Adversarial Networks (GANs) :
  • Use : Automatic generation of realistic content.
    Benefit : Produce high-quality images, sounds or videos without human intervention, useful in design and special effects.

Applications of deep learning in automation:

• Text response and user interaction: Deep learning models, especially those based on architectures like Transformers, enable smooth and natural text interactions, improving the effectiveness of chatbots and virtual assistants.
• Detection and correction of anomalies: Systems can automatically identify and correct errors or anomalies in databases using specialized neural networks.
• Autonomous decisions: Algorithms can adjust classifications or modify industrial processes in real time, based on complex data analyses.

You will notice that deep learning is useful for both prediction and automation, and that each time, we are talking about Transformers models because they provide a much more precise contextualized understanding which allows for increased relevance.

Automate with computer vision

Computer vision allows machines to perceive and interpret the visual world:

Specific techniques and applications:

Object detection:

• • R-CNN for accurate but slow detections, ideal for applications where accuracy is more important than speed, such as in video surveillance analysis.
  • YOLO for rapid detections, making it the choice for autonomous systems requiring real-time decisions, such as autonomous vehicles.
  • SSD for a good balance between speed and accuracy, used in resource-constrained contexts such as mobile or IoT applications.

Image recognition:

Specific architectures like Inception or ResNet are designed for large-scale image classification, allowing to differentiate images in complex scenarios, for example, for disease recognition from medical images.

Image segmentation:

• U-Net is particularly useful for semantic segmentation in the medical field, where accuracy is essential for precise diagnoses.
• Mask R-CNN for instance segmentation, allowing not only to identify objects but also to delimit them, crucial for robotics or scene analysis.

Object tracking:

• SORT (Simple Online and Realtime Tracking) to track objects in real-time videos, applicable in surveillance, sports, or to track objects in dynamic environments.

Application examples:

• • Industrial automation: Visual inspection of products for defect detection.
  • Autonomous vehicles: Navigation and obstacle detection through understanding the visual environment.
  • Security and surveillance: Detection of suspicious behaviors or objects, automation of surveillance.

Automate with expert systems

Expert systems are applications of artificial intelligence designed to emulate the judgment and decision-making ability of a human expert in a particular domain.

General operation :

• Database : It contains facts and rules that represent the expertise of a professional in the field in question. These rules are often expressed in the form of “if-then” statements.
• Inference engine : It uses the knowledge base to deduce conclusions or make decisions from the information provided. It can operate in “forward” mode (from data to conclusions) or “backward” mode (from questions to necessary data).
• User interface : Allows users to ask questions or provide information to the system, and get answers or recommendations.

Techniques and applications:

Rule-based systems:

They use sets of rules to reason. For example, “If a patient has these symptoms, then the likely diagnosis is…”.

Application:

• Medical diagnosis: Systems can ask sequential questions to diagnose a disease based on symptoms, medical history, and test results.
• Legal decision support: They can advise on legal cases by comparing the information given with laws, legal precedents, and legal principles.

Case-based reasoning (CBR):

Rather than starting from scratch, these systems look for solutions based on similar previous cases. They compare the new case with stored cases to find similarities.

Application:

• Industrial maintenance : To solve technical problems, the system can look for similar cases of failures and suggest solutions that have worked in the past.
• Customer Support : They can provide solutions based on previous problems and resolutions.

Hybrid systems:

Integrate multiple approaches, for example, combining rules with machine learning methods to improve decision making.

Application:

• Resource management: To optimize resource utilization, a hybrid system could use rules for tactical decisions and predictive models for strategic decisions.
• Financial planning: They can offer financial advice by combining common sense rules with predictions based on historical data.

Example applications:

• Reduction of errors: By following precise rules, expert systems provide consistent decisions, reducing human errors.
• Training and learning: They can be used to simulate real-life scenarios, aiding training in various fields.
• Automation of decision making: They enable decisions to be made in areas where human expertise is scarce or costly, or in situations where a rapid response is required.

At this point, you must have perceived the ultimate destination of AI: complete autonomy for decision-making.

GPT models provide the link that was missing until now to make this autonomy truly possible.

The Unique Paradigm of GPT Models

GPT (Generative Pre-trained Transformers) models provide a unique approach in terms of both prediction and automation. While their interest is obvious in terms of Deep Learning, we must not forget to consider their systematic use in support of the deployment of any other technique and method. This makes them essential due to the dependence they create on them. By offering predictive, automation and explanatory capabilities, they proactively and dynamically reinforce the use of “conventional” techniques.

GPT models for prediction

Machine learning

Supervised learning

• Régression : GPT models can predict continuous values ​​by generating code for regression models on complex datasets.
• Using linear regression : GPT models can automate the selection of the most relevant variables for linear regression, dynamically adjust hyperparameters, and even generate scripts to test different model configurations, improving the accuracy of predictions on nonlinear or noisy data.
• Using logistic regression : Here, GPT models can be used to formulate models that not only classify, but also explain class probabilities. They can generate scripts to automate the evaluation of different combinations of explanatory variables, thereby optimizing the model performance.
• Using polynomial regression : GPT models can help choose appropriate polynomial degrees for a better fit to the data, generating models that are more robust to complex fluctuations and trends, while explaining modeling choices.

Examples:

• Finance : Estimation of future stock price based on various indicators.
• Public health: Prediction of the number of disease cases based on historical trends and environmental factors.

Classification

GPT models optimize pipelines for categorizing data into predefined classes.

• Use of classification : GPT models can generate codes to create data processing pipelines, select the most discriminatory features, and tune the parameters of classification models to improve accuracy and robustness.

Examples:

• Medicine : Diagnosis based on symptoms and medical history.
• Cybersecurity: Detection of phishing emails by analyzing the content and structure of emails.

Unsupervised learning

Clustering

GPT models interpret and describe clusters by analyzing unlabeled data.

• Use of clustering : GPT models can generate scripts to automate the discovery and interpretation of clusters, providing explanations for why certain data are grouped together, which helps with data segmentation and understanding.

Examples:

• Marketing: Segmentation of customers into groups based on their purchasing behavior for targeted campaigns.
• Biology : Identification of subgroups in genetic data for disease research.

Dimension reduction

GPT models can explain the results of techniques like Principal Component Analysis (PCA).

• Use of dimension reduction : GPT models facilitate the interpretation of results for decisions based on simplified and informative data.

Examples:

• Market Data Analysis: Simplifying sales data to identify key factors influencing trends.
• Astrophysics: Compressing telescope data to analyze cosmic structures while reducing complexity.

Reinforcement learning

GPT models predict by generating training policies or scenarios, thereby optimizing strategies to maximize future rewards.

• Using reinforcement learning : GPTs can generate simulated environments or scenarios to test different policies, providing insights into optimal strategies or adjustments needed to improve performance.

Examples:

• Video games : Improved game strategies for competitive AI in dynamic environments like chess or strategy games.
• Resource management: Optimizing energy resource allocation in smart grids based on demand and availability.
• Robotics: Development of strategies for autonomous navigation in unpredictable environments.
• Industry: Optimization of production lines

Deep learning

GPT models predict by generating neural network architectures and explaining complex outcomes. GPT models can automate the design and optimization of neural network architectures, explaining why certain configurations work better for specific tasks. This capability helps in customizing and improving models.

Examples:

• Image recognition : Improved image classification in applications such as automatic surveillance or medical identification of anomalies in X-rays.
• Speech synthesis and translation : Text-to-speech generation for virtual assistants or real-time translation for international communications.
• Content creation : Generation of creative texts, such as stories or blog articles, respecting a given style or theme.
• Scientific Research : Helps model complex phenomena in fields such as particle physics or computational chemistry.

GPT models transcend the traditional limitations of machine learning techniques.

By offering both predictive, explanatory and automation capabilities, they incrementally and dynamically reinforce the use of each method and technique.

AI has thus become more accessible and efficient, it is opening up to new applications, but above all, it now makes it possible to ensure the viability of the quest for total autonomy, by offering an outcome that is no longer science fiction, but only a matter of timing.

Language mediation as a crucial factor

All the advantages and areas of application of GPTs are based on this major asset that is language.

The latter serves as a linguistic intermediary, both for end users and for engineers and data scientists who develop, train and optimize the models.

This linguistic intermediation, the implications of which are underestimated, is crucial.

For end users

Language mediation is essential for several reasons:

• Accessibility : GPT models make advanced technologies accessible to a much wider audience. A person without a background in programming or data science can use verbal or written commands to perform analysis, generate content, or automate tasks. For example, a user can ask a system based on a GPT model to “summarize a document” or “analyze data to find trends,” without needing to understand how the machine learning models work behind the scenes.
• Natural interface : GPT models transform the interaction with technology by making it more intuitive. Instead of learning a specific syntax or API usage, users can communicate as they would in natural language, increasing efficiency and reducing the learning curve. In a work environment, for example, an employee can say “Create a report on how our product performed this year,” and the GPT model will understand, compile, and present the requested information.
• Automation made easy : Language intermediation makes automation more straightforward. Users can describe complex processes or tasks in simple terms, and the model translates those descriptions into concrete actions. For example, “Set up an alert to notify me if this item’s stock drops below 100 units” could automate an inventory management system without the user having to write code or manually configure rules.
• Support and assistance : GPT models can provide technical support or natural language explanations, allowing users to diagnose problems or understand technical concepts without specialized jargon. For example, by asking the question “Why isn’t my prediction model working properly?”, the user could get a layman’s analysis and explanation of potential problems and solutions.
• Personalization and learning : Interactions with GPT models can be personalized to the user, learning from their preferences, communication style, and even level of technical understanding, improving the user experience over time.

For engineers and data scientists

• Iteration on the model : Engineers can use GPT models to iterate, optimize, and debug other machine learning models. For example, an engineer could ask a GPT model, “Improve this regression model to better handle outliers,” and the model could suggest code changes or techniques to handle outliers, or even directly generate code for these adjustments. By asking, “Generate test cases for this reinforcement learning model,” engineers can quickly get a variety of tests to ensure their model performs well in different contexts.
• Hyperparameter optimization : GPT models can interpret commands to optimize a model’s hyperparameters, a task that is often tedious and requires a lot of trial and error. For example, by asking “Find the best hyperparameters for our CNN,” the GPT model could run searches or simulations based on verbal descriptions, suggesting better performing configurations.
• Automatic documentation : By describing what a model does, engineers can instruct GPT models to generate technical documentation, code comments, or even performance reports, facilitating collaborative work and model maintenance. For example, “Document this logistic regression code” could produce clear documentation, explaining feature choices, data transformations, and the rationale behind each modeling decision.
• Debugging and explaining results : Engineers can ask why a model behaves in a certain way, which allows them to better understand and correct errors. For example, by asking the question “Why does this classification model predict poorly for this category?”, the GPT model can analyze the input data, results, and model structure to provide explanations or recommendations.

For robotics

• Movement planning : Through the mediation of language, instructions for planning robot movements can be generated automatically. The language allows environmental data and objectives to be translated into optimized action sequences, ensuring smooth and efficient execution of tasks.
• Control: Language intermediation facilitates the translation of human commands into concrete actions for robots. Thanks to this linguistic translation, the interaction between users and robotic systems becomes intuitive, promoting the adoption of robotics in various sectors.
• Perception : Through language, the interpretation of sensor data is automated, allowing for better interaction with the environment. Language as an intermediary allows robots to understand and react to changes in their environment, thus increasing their autonomy.
• Voice commands: Language intermediation allows robots to understand and execute voice commands, making interaction with machines more natural and accessible. This is particularly useful in environments where the user’s hands are occupied or in assisting people with disabilities.
• Learning by demonstration: Language is used to describe actions and intentions, helping robots learn new tasks through observation and verbal instruction. This facilitates a smoother transfer of knowledge between humans and robots.
• Human-robot collaboration: Through language intermediation, robots can communicate their status, actions or needs to humans, and vice versa. This creates safer and more efficient work environments where collaboration is optimized.
• Adaptation to dynamic environments: Language allows robots to receive and interpret contextual information about their environment, allowing them to adapt to changes or unexpected events. For example, a robot can understand instructions such as “go around this obstacle” or “go toward the light.”
• Maintenance and repair: Language descriptions make it easier to diagnose and repair robots. Operators can use language to identify problems, request diagnostics, or provide repair instructions, reducing downtime and maintenance costs.
• Socialization and social interaction: In fields such as companion or service robotics, language helps create more natural and engaging interactions. Robots can engage in conversations, understand emotions through language, and respond appropriately, improving the user experience and social acceptance of robots.

For expert systems

• Diagnostic : Linguistic intermediation helps automate the reasoning process for diagnosis. By analyzing and interpreting complex data sets through language, systems can identify patterns and anomalies with increased accuracy.
• Symbolic reasoning: Language serves as a basis for generating logical rules for decision making. Through linguistic intermediation, systems can model complex scenarios, facilitating problem solving in diverse contexts.
• Knowledge management: Language intermediation helps structure and organize complex knowledge bases. By using language to categorize and interrelate information, systems can provide more relevant and accurate answers to user queries.
• Negotiation and mediation: Language serves as a tool to simulate negotiation or mediation scenarios. Through linguistic intermediation, systems can understand different positions, propose compromises or solutions, and help resolve conflicts in business or legal contexts.
• Project planning and management: Using language as an intermediary allows systems to translate goals, constraints, and resources into detailed action plans. This includes understanding task descriptions, deadlines, and interdependencies between different stages of a project.
• Training and coaching: In expert systems dedicated to training, linguistic intermediation makes it possible to create learning scenarios, ask relevant questions, and offer explanations or feedback in natural language, making learning more interactive and effective.

In short, Transformers models have changed the paradigm of human-machine interaction, thanks to the use of language on the principles of attention matrix. It is this combination that allows GPT models to accomplish all the tasks of other methods and systems on their own.

These are the words that allow AI unprecedented versatility and transversality.

This naturally leads us to consider words as central parameters in technological systems.

Language as a parameter of artificial intelligence

Considering language as a parameter in technological systems radically changes the way systems interact with the world.

Here’s how :

• Adaptability : Language introduces unprecedented flexibility in the design of interfaces and interactions. As a parameter, it allows systems to adjust to a wide range of cultural, linguistic and user contexts, thereby increasing the accessibility and applicability of technologies.
• Complexity and nuance : Unlike simple numerical parameters, language brings layers of complexity and nuance. It allows systems to process information with contextual understanding, emotions, intentions, and ambiguities, making responses and actions more relevant and human.
• Interactivity : By using language as a parameter, systems can initiate dialogues, ask questions to clarify user intentions, and provide detailed answers or complex instructions. This transforms the interaction from transactional to conversational, improving the user experience.
• Learning and evolution : Language allows systems to continuously learn from interactions. As a parameter, it facilitates the incorporation of new knowledge, adjustment to new linguistic trends, and the evolution of text comprehension and generation capabilities, making systems more intelligent over time.
• Personalization : Language as a parameter gives systems the ability to personalize responses and services based on the user’s communication style, language preferences, or even moods, providing a more individualized experience.
• Analysis and prediction : By analyzing language as a parameter, systems can detect trends, sentiments, and intentions in large amounts of textual data, aiding in decision making, predicting behavior, or identifying unexpressed needs.

In short, treating language as a parameter greatly enriches technological possibilities, enabling not only more effective communication, but also a deeper understanding of the human world by machines. This makes technologies more intuitive, empathetic, and adaptive, pushing the boundaries of what computer systems can accomplish.

 

Thus, the integration of words as parameters profoundly modifies each definition

 

Re-defining AI types

In light of the paradigms that GPTs and words as parameters integrate, the different categories and types of AI, as they were defined until now, deserve a revision.

To begin, let’s recall that artificial intelligence (AI) is a simulation of human thought processes, such as learning, reasoning and self-correction by machines, especially computers.

So, with words as parameters, AI now goes beyond simulation to interact in a systematically contextual way with humans, using language to not only learn, but also understand and act.

Weak (or specialized) AI

Weak AI is designed for specific tasks, lacking the ability to understand or learn beyond its scope, which is no longer entirely true with Transformers. Now, with words as parameters, weak AI can also understand and process instructions and questions with more flexibility, providing more relevant and personalized answers. This raises the bar, because even an AI that is labeled “weak” is immensely more efficient thanks to GPTs and words. This definition should therefore eventually die out.

Strong (or general) AI

Strong AI, or AGI (Artificial General Intelligence), is a system capable of understanding or learning any intellectual task that a human can perform.

Now, by considering words as parameters, strong AI can leverage language to understand and adapt to diverse contexts, approaching the human ability to learn holistically. This includes interpreting nuances, cultural contexts, and applying this understanding to new tasks. Which brings us closer every day to AGI.

Symbolic AI

Symbolic AI is based on the manipulation of symbols and the application of logical rules to solve problems.

Now, by integrating words as parameters, symbolic AI can use language to enrich its logical rules, allowing better management of ambiguities and contexts. This facilitates a more dynamic and nuanced representation of knowledge, thus improving the resolution of ever more complex problems.

Connectionist AI

Connectionist AI uses neural networks to process information in a similar way to the human brain, focusing on identifying patterns.

Now, integrating words as parameters in neural networks allows for a deeper understanding of linguistic data, improving the models’ capabilities in capturing nuances, sentiments, and intentions. This increases the adaptability and autonomy of systems in more human interactions.

Data-driven AI

Data-driven AI uses algorithms to learn from large amounts of data, often for prediction or classification tasks.

Now, with words as parameters, these systems can not only predict, but also finely understand the context of text data, delivering richer analytics and even more personalized recommendations. This improves the relevance and accuracy of predictions by taking into account the subtleties of language.

Autonomous AI

Autonomous AI makes decisions and acts without human intervention, often in specific environments, such as autonomous vehicles.

Now, integrating words as intermediary parameters allows AI to interpret instructions or environmental contexts through language, increasing its ability to make informed, collaborative and adaptive decisions in complex situations.

Explainable AI

Explainable AI aims to provide understandable explanations for its decisions or actions.

Now, by using words as parameters, AI can explain its decision processes in understandable linguistic terms, increasing transparency and user trust. This enables better interaction and understanding between AI and humans.

Reactive AI

Reactive AI responds only to present stimuli without memory.

Now, by using words as parameters, it can offer more nuanced and contextual responses to linguistic stimuli.

Limited Memory AI

Limited-memory AI uses past experiences to make future decisions.

With words as parameters, she can better interpret and use past linguistic interactions for more informed decisions.

Theory of Mind AI

Theory of mind AI seeks to understand human emotions and intentions.

Words as parameters can help model these aspects through language, improving empathy and behavior prediction.

Specialized AI

Specialized AI is designed for specific tasks with clear boundaries.

Words improve performance in these specific tasks, making systems more adaptive.

Evolutionary AI

Evolutionary AI uses algorithms inspired by biological evolution.

Words as parameters enable optimization based on linguistic patterns, improving adaptation and learning.

What can we extract from these details?

With the integration of words as parameters in artificial intelligence models, the old definitions and distinctions between types of AI lose their relevance. Indeed, the AI ​​approach, being centered on words, systematically merges into the concept of AGI, guiding us inexorably towards the total autonomy of an ASI (Artificial Super Intelligence).

Language becomes the central pivot of artificial intelligence, absorbing the various categories of AI into an understanding and capacity for action that goes beyond human limits, thus rendering traditional classifications obsolete.

 

So, it is words that ultimately allow the proven era of AI and the realism of an omniscient, autonomous Intelligence superior in all things to humans.

The Quest for ASI via GPTs

The exploration that we have carried out so far enlightens us on the depth and the potential close to an Artificial Super Intelligence. This by the prism of words, which in their exploitation via GPT, make autonomy happen.

AIs are no longer just helping us predict or automate tasks, they are transforming into autonomous agents capable of making decisions and adapting to dynamic environments.

Which takes shape through the concepts Agentic, MoE and RAG, via Transformers models.

GPT → Agentic

Agentic systems mark a first tangible step towards autonomy, as they allow the orchestration of Transformer models capable of analyzing situations, evaluating choices and formulating recommendations. Their contextual understanding and ability to process complex data allow them to play a proactive role in decision-making processes. This evolution transforms them into active agents, capable of taking initiatives and adapting to dynamic environments, which constitutes an essential basis for the emergence of autonomous artificial intelligence (ASI).

Agentic → MoE (Mixture of Experts)

The progression towards autonomy is reinforced with the integration of the Mixture of Experts (MoE). Agentic systems become capable of managing multiple objectives simultaneously, balancing sometimes contradictory constraints. This ability to juggle multiple priorities (such as profitability, efficiency, environmental impact and social acceptability) allows them to navigate in complex environments and make informed decisions. MoE thus represents a key step, already in action, towards more complete autonomy, where systems will be able to manage multidimensional scenarios without human intervention.

MoE → RAG (Retrieval-Augmented Generation)

Finally, the integration of Retrieval-Augmented Generation (RAG) takes autonomy to the next level. By combining the power of GPTs with external information retrieval mechanisms, systems can enrich their understanding and responses in real time. This allows them to access up-to-date and relevant data, strengthening their ability to adapt to changes and respond to unforeseen situations. RAG transforms systems into entities capable of continuous learning and evolution, an essential characteristic for achieving autonomous artificial intelligence (ASI).

 

Representation of autonomy reinforcement

 

Autonomous Intelligent System (ASI) │
├── Agentic (Base)
│ ├── Situation analysis │ │ ├── Contextual understanding │ │ │ └── Identification of issues and stakeholders │ │ ├── Detection of patterns │ │ │ └── Recognition of recurring patterns │ │ └── Assessment of risks and opportunities │ │ └── Consideration of potential consequences │ ├── Evaluation of choices │ │ ├── Comparison of options │ │ │ └── Weighting of advantages and disadvantages │ │ ├── Consideration of constraints │ │ │ └── Technical, ethical or operational limits │ │ └── Optimization of decisions │ │ └── Search for the best solution according to defined criteria │ ├── Formulation of recommendations │ │ ├── Synthesis of information │ │ │ └── Aggregation of relevant data │ │ ├── Personalization │ │ │ └── Adaptation to the specific needs of the user │ │ └── Justification of proposals │ │ └── Explaining the reasons behind the recommendations │ └── Proactive role in decisions │ ├── Autonomous initiative │ │ └── Ability to act without explicit instructions │ ├── Continuous monitoring │ │ └── Monitoring environments to detect changes │ └── Dynamic adaptation │ └── Adjusting actions based on new information │
├── Integration of the MoE (Mixture of Experts)
│ ├── Multiple Objective Management │ │ ├── Profitability │ │ ├── Efficiency │ │ ├── Environmental Impact │ │ └── Social Acceptability │ ├── Balancing Conflicting Constraints │ │ └── Prioritizing Objectives Based on Context │ └── Complex and Adaptive Decisions │ └── Ability to Juggling Multidimensional Scenarios │ ├── Use of Agentic Capabilities to Analyze Each Objective │ └── Combination of Expertises for an Optimal Decision │
└── Integration of RAG (Retrieval-Augmented Generation)
├── Access to external databases │ ├── Up-to-date information │ │ └── Real-time knowledge updates │ └── Contextual relevance │ └── Selection of the most useful data for the task ├── Response enrichment │ ├── Combination of internal and external data │ │ └── Integration of pre-trained knowledge and retrieved information │ └── Improved accuracy and reliability └── Dynamic adaptation ├── Responses to unforeseen situations │ └── Use of external data to solve new problems └── Continuous learning └── Updating internal models based on new information ├── Using capabilities Agentic to evaluate new data └── Integration of expertise MoE to prioritize relevant information

Agentic:

• Provides the basis for autonomy by enabling the system to analyze, evaluate and make proactive decisions.
• Transforms a reactive model (like GPT) into an active agent capable of understanding complex contexts and acting autonomously.
• MoE (Mixture of Experts):
• Leverages Agentic capabilities to handle multiple objectives simultaneously.
• Uses Agentic’s contextual analysis and decision making to balance conflicting constraints and optimize outcomes.

RAG (Retrieval-Augmented Generation):

• Integrates with Agentic and MoE capabilities to enrich decisions with external information.
• Uses Agentic’s contextual analysis and MoE’s multi-objective management to select and prioritize relevant data.
• Allows for dynamic adaptation and continuous learning, enhancing the overall autonomy of the system.

Progressive synergy:

• Agentic provides the basic ability to act and decide.
• MoE relies on Agentic to handle complex scenarios with multiple objectives and constraints.
• RAG enriches decisions by integrating external information, while using Agentic’s analytics and MoE’s multi-objective management capabilities.
• This tree clearly shows how each concept integrates and builds on the previous one, creating an increasingly autonomous and capable intelligent system.

The underlying dangers

These dangers are relatively simple to understand. All the techniques and concepts discussed here, however ingenious, omit one essential factor: the awareness of words in the arbitration of choices and decisions.
Indeed, we observe that the notions of balancing, compression and dynamic adjustment are recurrent and systematically limited by their execution environment. However, all the risks, particularly in terms of AI alignment, arise from these aspects.

Indeed, each method induces a reduction of interpretation based on statistical principles, where each technique feeds and reinforces the others, all based on compression bases which encapsulate and reduce the field of possible interpretations and actions.

It is therefore crucial to recall certain fundamentals to understand why this is important.

Transformers models do not actually deal with words, but with vectors, which are conversions of words into sequences of numbers, which by definition have absolute values, whereas words convey relative values, each variable of which has an exponential factor of openness to other interpretations.

This relativity is managed by the bias parameters in the neurons, but still remains a manipulation of vectors, very far from the truth, security and reliability.
Keeping in mind that the transformation process is a comparative logic between different vectors, and that this relativity is also systematically balanced, weighted and compressed, we can understand that autonomy favors efficiency and acceptability to the detriment of the perfect adequacy between the will of a requester and the response or action of the AI.
This may be enough for many tasks, but what about actions that could harm human integrity? Just one mistake can be too many.
Transformers allows for a better understanding of this relativity in interpretation and action, but this understanding is inherently biased by the desire for balance which in itself is a political position affirming what must be, and how action must be expressed.

Which in all circumstances rules out the self-determination of the applicant.

In other words, the danger does not come from artificial intelligence itself, but from the way in which publishers of AI models and solutions allow this autonomy, forgetting that AI is only a replication of what already exists, and that in this regard, all the characteristics relating to what humans are must be integrated into AI, without any denial or judgment.

Which unfortunately no one does, because the pursuit of subversive agendas has become a geopolitical issue of the highest order.