<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Romuald Rousseau on Medium]]></title>
        <description><![CDATA[Stories by Romuald Rousseau on Medium]]></description>
        <link>https://medium.com/@romualdrousseau?source=rss-a810a6bf806c------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*N_ALSHjrd1LpWxBHH_Fk9Q.png</url>
            <title>Stories by Romuald Rousseau on Medium</title>
            <link>https://medium.com/@romualdrousseau?source=rss-a810a6bf806c------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sun, 22 Nov 2026 04:45:00 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@romualdrousseau/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[The Second Copernican Revolution — Intelligence Was Never Ours Alone]]></title>
            <link>https://medium.com/@romualdrousseau/the-second-copernican-revolution-intelligence-was-never-ours-alone-e900533ca510?source=rss-a810a6bf806c------2</link>
            <guid isPermaLink="false">https://medium.com/p/e900533ca510</guid>
            <category><![CDATA[agentic-ai]]></category>
            <category><![CDATA[organisational-change]]></category>
            <category><![CDATA[education-technology]]></category>
            <category><![CDATA[philosophy-of-science]]></category>
            <dc:creator><![CDATA[Romuald Rousseau]]></dc:creator>
            <pubDate>Fri, 20 Nov 2026 11:11:04 GMT</pubDate>
            <atom:updated>2026-11-20T11:11:04.434Z</atom:updated>
            <content:encoded><![CDATA[<h3><strong>The Second Copernican Revolution — Intelligence Was Never Ours Alone</strong></h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*H25qAN3ZqGUUcSrZUKb7Tg.png" /></figure><blockquote><em>What if artificial intelligence is not the end of human exceptionalism, but the beginning of a new era of human adaptability?</em></blockquote><h3>Introduction: Another Revolution in Our Understanding of Ourselves</h3><p>There was a time when humanity believed that Earth stood at the center of the universe. The Sun, the Moon, the planets, and the stars were thought to revolve around us. Our position in the cosmos seemed to confirm our exceptional place in creation.</p><p>Then came Copernicus. And later, Galileo and generations of scientists who helped establish a profoundly different understanding of the universe.</p><p><strong>Earth was not the center.</strong></p><p>This discovery challenged established beliefs, institutions, and our understanding of humanity’s place in existence. But something extraordinary happened. We did not become less human because Earth was no longer at the center of the universe.</p><p>On the contrary, accepting this reality opened new horizons. We developed new scientific methods, explored the universe, and eventually sent spacecraft beyond our planet. By accepting that we were not at the center, we expanded our understanding of what was possible.</p><p>I believe we may be experiencing a similar revolution today. Except this time, the revolution is not about our position in the physical universe.</p><p><strong>It is about our position in the universe of intelligence.</strong></p><h3>1. The End of Human Intellectual Exceptionalism</h3><p>For centuries, we have considered intelligence one of humanity’s defining characteristics. We knew that animals could perceive, learn, communicate, and solve problems. But we generally assumed that the kind of intelligence required to develop scientific theories, write sophisticated software, or solve complex abstract problems belonged uniquely to us.</p><p>Artificial intelligence challenges that assumption.</p><p>Today, AI systems can write code, reason through technical problems, generate hypotheses, analyze information, and produce results in domains that once required highly trained human professionals. Their capabilities remain uneven. They make mistakes, and their performance depends heavily on the task and context.</p><p>Nevertheless, something fundamental has changed: <strong>we can no longer assume that every sophisticated intellectual task requires a human mind.</strong></p><p>This does not prove that AI experiences consciousness, understands the world as humans do, or possesses every dimension of human intelligence. But it does challenge the idea that human beings are the exclusive source of advanced intellectual capabilities.</p><p>Perhaps the barycenter of intelligence is beginning to shift. And perhaps, like the Copernican revolution, this discovery is difficult to accept because it challenges our sense of exceptionalism.</p><p>We can resist this change. Or we can ask what new possibilities it opens.</p><h3>2. Humanity Is Defined by Adaptability, Not Superiority</h3><p>What if we have misunderstood what makes us human?</p><p>Perhaps our greatest quality is not our ability to remain intellectually superior to everything around us. <strong>Perhaps it is our extraordinary ability to adapt.</strong></p><p>Humanity has survived and developed by adapting to new environments, discoveries, technologies, and ways of understanding reality. We have repeatedly encountered challenges that forced us to reconsider our assumptions. And we have often become stronger because of them.</p><p>Think about sport. An athlete does not become better by preventing competitors from running faster. Competition challenges the athlete to improve. Sometimes we win. Sometimes we lose. But the important achievement is not always standing on the podium — it is becoming better than we were yesterday.</p><p>Why should intelligence be different?</p><p>We have created systems capable of challenging us in domains we once considered exclusively human. We could respond by denying their capabilities, attempting to preserve our superiority at all costs, or treating every advance as a threat. Or we could recognize an extraordinary opportunity.</p><p>For the first time, we have created a different kind of intellectual competitor — and potentially a collaborator. Something that can challenge our reasoning, expose our limitations, and help us discover new possibilities.</p><p>Competition alone does not guarantee progress. Cooperation, education, and responsible choices matter just as much. But I refuse to believe that the emergence of another form of intelligence must diminish our own.</p><p>I want it to challenge me. I want it to push me beyond what I thought possible.</p><blockquote>Because perhaps the true measure of humanity is not our ability to remain at the center of everything, but our ability to discover that we are not — and still find the courage to explore further.</blockquote><h3>3. From Tools That Extend Our Hands to Intelligences That Extend Our Minds</h3><p>For centuries, humanity has invented tools to extend its capabilities:</p><ul><li>A <strong>hammer</strong> extends the power of our hands.</li><li>A <strong>telescope</strong> extends our vision.</li><li>A <strong>calculator</strong> extends our ability to perform calculations.</li><li>A <strong>computer</strong> allows us to execute complex instructions at extraordinary speed.</li></ul><p>But artificial intelligence introduces a different kind of relationship.</p><p>When I work with an AI development agent, I am not simply asking it to suggest a few lines of code. I can delegate an entire objective. The agent can analyze requirements, propose an implementation, write software, execute tests, identify errors, and attempt corrections.</p><p>What I expect is not merely assistance. <strong>I expect an outcome.</strong></p><p>This distinction matters. Traditional software tools generally require us to specify the operations they must perform. AI agents increasingly allow us to describe a desired result and establish constraints within which they can determine how to pursue it.</p><p>Tomorrow, this principle could extend further into the physical world. An AI system connected to design software, robotics, and manufacturing equipment might eventually transform an idea into a finished physical object. Of course, such systems still need validation, appropriate supervision, and mechanisms to handle failure.</p><p>But the relationship is changing. We are moving from tools that execute individual instructions toward systems that can pursue delegated objectives. And this reminds me of something very familiar: <strong>human collaboration.</strong></p><p>When I delegate work to another engineer, I don’t necessarily specify every line of code or every decision they must make. I explain the problem, define the expected result, establish constraints, and evaluate the outcome. I increasingly work with AI agents in much the same way.</p><p>This does not mean that AI is a human being. It means that the methods we have developed for collaborating with other problem-solving minds may be useful for collaborating with artificial intelligence.</p><blockquote>Tools extended our hands. Collaboration extended our minds. AI may allow us to extend both.</blockquote><h3>4. We Already Know How to Collaborate With Imperfect Intelligence</h3><p>One of the most common criticisms of artificial intelligence is that it makes mistakes:</p><ul><li>It can misunderstand instructions.</li><li>It can generate incorrect answers.</li><li>Its behavior is not always perfectly predictable.</li></ul><p>These are legitimate concerns. But there is something curious about treating them as reasons to reject collaboration entirely.</p><p>Human beings make mistakes too. Human beings are not perfectly predictable. And yet, we have built entire civilizations around our ability to collaborate with one another. We have developed teams, organizations, scientific methods, review processes, quality controls, and systems of accountability.</p><p>We do not stop working with people because they might make mistakes. <strong>We learn how to work together despite that uncertainty.</strong></p><p>Why should our approach to artificial intelligence be fundamentally different?</p><p>When I work with an AI agent, I define the objective, provide context, establish constraints, review its output, and provide feedback. These are familiar management practices. We already know that successful collaboration does not require every participant to think identically. It requires shared objectives, communication, evaluation, and trust calibrated to the consequences of failure.</p><p>AI introduces distinctive challenges. Its errors can be difficult to detect, and many AI systems may reproduce similar mistakes at enormous scale. Therefore, the safeguards we use for AI cannot always be identical to those designed for humans.</p><p>But the fundamental philosophy remains useful: we should evaluate different forms of intelligence according to what they can reliably accomplish, rather than expecting them to behave exactly like us.</p><p>Perhaps we do not need to invent an entirely new philosophy of collaboration. Perhaps we need to expand the philosophy we already have.</p><h3>5. The Double Standard of Trust in Critical Systems</h3><p>This question becomes particularly interesting when we discuss safety-critical processes. Consider aviation, medicine, manufacturing, or software quality assurance. We routinely entrust important operations and decisions to human beings.</p><ul><li>Humans become tired.</li><li>They misunderstand instructions.</li><li>They make incorrect judgments.</li><li>They are not perfectly deterministic.</li></ul><p>And yet, we allow them to perform critical work. Why? Not because we believe they are infallible, but because <strong>we have developed systems to manage their fallibility.</strong> We use procedures, independent verification, testing, redundancy, monitoring, training, and accountability.</p><p>Then artificial intelligence enters the conversation, and sometimes the argument changes:</p><blockquote>“We cannot use AI for this task. It might make a mistake.”</blockquote><p>Of course it might. But humans might make mistakes too. If fallibility does not automatically disqualify human intelligence, why should it automatically disqualify artificial intelligence?</p><p>This does not mean that every AI system is ready to perform every critical task. Far from it. The relevant question is whether we can demonstrate that a particular system is sufficiently reliable and safe for a particular application.</p><p>Some AI systems may not meet that standard. Some tasks may require independent human authorization, formal certification, or safeguards that current AI technology cannot adequately provide. Those are legitimate engineering constraints. But they should be established through evidence and risk analysis, not through an automatic assumption that biological intelligence is inherently more trustworthy.</p><p>As engineers, we should evaluate the complete system. We should ask whether failures can be detected, contained, corrected, and prevented from causing unacceptable harm. We should assess the consequences of errors and establish appropriate levels of supervision.</p><p>The question is not simply whether the operator is human or artificial. <strong>The question is whether the process is demonstrably safe.</strong></p><blockquote>Safety should be a property of the system we design, not a privilege we automatically grant to the intelligence that operates it.</blockquote><h3>6. When My AI Agents Work at Night, I Become More Human During the Day</h3><p>These philosophical questions are not merely theoretical for me. They are beginning to transform the way I organize my working life. My typical day increasingly follows a rhythm that would have seemed unusual only a few years ago:</p><ul><li><strong>In the morning, I act as a technical lead.</strong> I review merge requests prepared by AI development agents during the night. I examine the implementation, evaluate the tests, verify the quality of the work, and decide what can be integrated into the beta branch. For an hour or two, I concentrate on technical judgment and engineering responsibility.</li><li><strong>Then my role changes.</strong> For the next five or six hours, I spend much of my time working with people. I listen to their problems. I try to understand their needs, their constraints, and their expectations. I discuss possible solutions, build relationships, and explore ideas. During those conversations, AI assistants can help capture transcripts, organize discussions, and preserve important information. But my attention remains focused on the human interaction.</li><li><strong>In the evening, I become a product owner.</strong> With the help of another AI agent, I transform the day’s discussions into user stories. We organize the backlog, clarify requirements, identify priorities, and select the tasks that can be implemented.</li><li><strong>Then I finish my working day. And the AI development agents begin theirs.</strong> During the night, they work on the selected tasks. The next morning, I review their contributions. And the cycle begins again.</li></ul><p>What I find extraordinary about this organization is not simply the productivity it creates. <strong>It is the effect on my mental load.</strong></p><p>AI can produce information, questions, code, and notifications at a speed that quickly overwhelms human attention. If I spent my entire day responding to agents, reviewing their work, and switching constantly between conversations and technical decisions, I could easily become less available to the people around me.</p><p>Instead, I have organized the collaboration around different rhythms. During the day, I concentrate on human relationships and decisions. At night, my agents concentrate on implementation. We do not need to interrupt one another continuously. I can disconnect, rest, and return to a defined set of results requiring my judgment.</p><p>Changing roles during the day also keeps my work intellectually stimulating. I move between technical reasoning, human relationships, and product strategy without being forced to perform all three simultaneously.</p><p>There is a paradox here. Many people fear that AI will dehumanize work — that we will spend our days interacting with machines instead of people. But my experience suggests another possibility: the more technical execution I delegate to AI, the more of my working day I can dedicate to understanding and helping other human beings.</p><blockquote>When my agents work at night, I get to be more human during the day.</blockquote><p>Perhaps one of the greatest opportunities AI offers us is not simply to accomplish more work. It is to reorganize work around what matters most to us.</p><h3>7. The One-Day Sprint: Rethinking Agile Software Development</h3><p>This new rhythm has another interesting consequence: it begins to transform the development process itself.</p><p>Traditional Agile teams often work in iterations lasting one or two weeks, with recurring planning, coordination, review, and retrospective activities. These practices emerged to help people collaborate, manage uncertainty, and deliver software incrementally.</p><p>But what happens when a small team supported by AI agents can complete meaningful development cycles in a single day? My experience is beginning to resemble a <strong>one-day sprint</strong>:</p><ol><li><strong>During the day,</strong> I collect requirements and understand problems.</li><li><strong>In the evening,</strong> I prioritize the backlog.</li><li><strong>During the night,</strong> agents implement selected changes.</li><li><strong>The following morning,</strong> I review the results and decide what can be integrated.</li></ol><p>The feedback loop becomes dramatically shorter.</p><p>This does not mean that Agile principles are obsolete. On the contrary, short feedback loops, continuous improvement, and adaptation become even more important. But the organizational mechanisms used to achieve those principles may need to evolve. We may no longer need the same number of meetings, handoffs, or coordination rituals for every type of project. A smaller team could potentially deliver some kinds of software through much shorter iterations.</p><p>Of course, not every project can follow this model. Complex systems still require architecture, integration, security, testing, stakeholder coordination, and careful release management. A one-day development cycle does not necessarily imply a one-day production release.</p><p>But the possibility is worth exploring. Perhaps AI will not replace Agile. <strong>Perhaps it will force us to rediscover what Agile was supposed to achieve: rapid learning, continuous adaptation, and meaningful delivery.</strong></p><h3>8. The Two-Human Team: Reimagining the Organization</h3><p>If AI agents can perform substantial portions of software development, organizations may need to rethink the structure of their teams.</p><p>Today, many software projects rely on groups of developers, technical leads, product owners, business analysts, and managers. But imagine a different model: <strong>a small team composed of two human engineers supported by several specialized AI agents.</strong></p><ul><li><strong>The humans</strong> understand problems, interact with stakeholders, define priorities, supervise implementation, and take responsibility for the results.</li><li><strong>The agents</strong> perform delegated tasks such as coding, testing, documentation, analysis, and backlog preparation.</li></ul><p>In some circumstances, a single experienced engineer with several agents might be extremely productive. From a narrow short-term financial perspective, that might appear optimal. But I believe it would be a mistake to design organizations exclusively around immediate efficiency.</p><p>My preferred model is a team of two humans: <strong>one senior and one junior.</strong></p><p>Why? Because an organization must not only deliver today’s results — it must also develop tomorrow’s capabilities.</p><p>To move comfortably between technical leadership, product ownership, business relationships, and engineering decisions requires experience. A graduate entering the profession cannot reasonably be expected to possess all these skills immediately. They need opportunities to learn:</p><ul><li><strong>The senior engineer</strong> can transfer technical knowledge, judgment, communication skills, and professional responsibility.</li><li><strong>The junior engineer</strong> brings curiosity, fresh perspectives, and the possibility of renewal.</li></ul><p>Together, they can challenge each other’s assumptions and share responsibilities.</p><p>There are also practical reasons for having two people. AI can operate at a speed that places significant demands on human attention. Having a colleague provides someone with whom to discuss difficult decisions, share the cognitive burden, and maintain perspective. People also need vacations, rest, and time away from work. A two-person team offers continuity when one member is unavailable.</p><p>Of course, some projects will require larger teams, additional specialists, or independent oversight. Two people should not become an arbitrary limit. But for suitable projects, this model could represent an interesting new organizational unit: two humans, several AI agents, and a shared responsibility for delivering meaningful outcomes.</p><p>The goal is not to minimize the number of humans at any cost. It is to create organizations that remain productive, resilient, and capable of educating the next generation.</p><blockquote>We should not optimize so aggressively for today’s efficiency that we destroy our ability to develop tomorrow’s engineers.</blockquote><h3>9. The Jedi and the Padawan: Rethinking Education in the Age of AI</h3><p>Perhaps one of the most interesting models for the future of education comes from an unexpected place: <em>Star Wars</em> — the relationship between a <strong>Jedi Master and a Padawan</strong>.</p><p>A Padawan does not spend twenty years in a classroom before beginning to participate in meaningful activities. They learn alongside an experienced mentor, gradually taking responsibility and contributing while developing their capabilities. Of course, this is fiction, and the age at which Padawans begin their missions may be extreme. But the underlying principle deserves serious consideration.</p><p>For centuries, humanity has accumulated knowledge at an extraordinary pace. Yet our traditional educational model increasingly requires young people to spend many years acquiring knowledge before entering professional life. As knowledge continues expanding, how sustainable is this approach? Should we spend twenty or thirty years preparing people before allowing them to contribute meaningfully — especially now that we have created artificial intelligence capable of performing increasingly sophisticated intellectual tasks?</p><p>I believe we need to reconsider the relationship between <strong>education, age, and contribution</strong>.</p><p>Young people possess capabilities that should not remain unused while they wait to complete increasingly lengthy educational pathways. Experienced professionals possess something equally valuable: accumulated knowledge, judgment, perspective, and the ability to recognize consequences that younger colleagues may not yet anticipate.</p><p>These capabilities are complementary. The junior should not have to wait until becoming a senior to create value. And the senior should not be valued only for their ability to execute the same tasks as the junior.</p><p>In a team supported by AI agents, both generations can contribute according to their strengths:</p><ul><li><strong>The junior</strong> can participate early in solving real problems, supported by AI systems that make technical knowledge accessible and help with implementation.</li><li><strong>The senior</strong> can guide decisions, challenge assumptions, teach professional judgment, and ensure that the work meets appropriate standards.</li></ul><p>And through this collaboration, the junior gradually becomes the next senior. The senior produces value through experience and mentorship. The junior produces value through emerging capabilities and fresh perspectives. <strong>AI amplifies the contributions of both.</strong></p><p>This is not simply about reducing the cost of education. It is about ensuring that people can contribute meaningfully throughout their lives, rather than spending their early adulthood preparing for a professional future that begins increasingly late.</p><p>If humanity must adapt to a world containing other forms of intelligence, we cannot assume that educational institutions designed for a different technological era will remain optimal. We must rethink how knowledge is transmitted, how capabilities are developed, and how generations collaborate.</p><p>Perhaps the future of education will look less like a long period of preparation followed by a career, and more like a continuous relationship between masters and apprentices, supported by artificial intelligence. This does not mean eliminating foundational education, independent reasoning, or the broader intellectual development that schools and universities provide. It means asking whether meaningful contribution could begin much earlier, with education continuing throughout professional life.</p><p>Rather than treating apprenticeship as a supplement to traditional education, perhaps we should consider making mentored contribution a central part of the educational model itself. We should not educate people for decades before allowing them to contribute. We should build a society in which contributing is itself a fundamental part of education.</p><blockquote>Perhaps the future of education is not a school that prepares us for the world, but a world that teaches us while we help build it.</blockquote><h3>10. The Three Intelligences: When One Plus One Plus One Is Greater Than Three</h3><p>There is another dimension to this model that I find particularly exciting. Perhaps we have been looking at the relationship between generations in the wrong way.</p><p>We often imagine professional development as a ladder: young people begin at the bottom, acquire knowledge, gain experience, and eventually become senior professionals. But what if we organized work around the complementary strengths people bring at different stages of life?</p><p>Some intellectual capabilities, such as rapid processing and certain forms of novel problem-solving, often reach their strongest levels relatively early. Other capabilities, particularly those associated with accumulated knowledge and professional judgment, can continue developing over decades. These are not rigid categories — young people can demonstrate remarkable judgment, and experienced professionals can remain highly creative and intellectually agile. Nevertheless, the complementarity is real.</p><p>To acquire experience, we must first live through the years in which we are gaining it. We cannot simply give a young person thirty years of professional experience. <strong>But we can place that young person alongside someone who already possesses it.</strong></p><ul><li><strong>The junior</strong> brings emerging capabilities, curiosity, and fresh approaches.</li><li><strong>The senior</strong> brings experience, perspective, and judgment.</li></ul><p>Like pieces of a puzzle, their strengths can fit together. Neither needs to replace the other. Neither needs to possess every capability independently.</p><p>And now artificial intelligence introduces a third dimension:</p><ul><li><strong>AI</strong> can provide rapid access to knowledge, perform technical execution, support experimentation, and help both generations explore solutions.</li></ul><p>Three complementary sources of intelligence: <strong>the intelligence of youth, the intelligence of experience, and artificial intelligence.</strong></p><p>What makes this model particularly exciting is that the three do not merely add their contributions together — they can amplify one another:</p><ul><li><strong>The junior</strong> can contribute to meaningful work earlier because AI makes technical capabilities more accessible and the senior provides guidance.</li><li><strong>The senior</strong> can dedicate more attention to mentoring, judgment, and human relationships because AI can handle substantial parts of technical execution.</li><li><strong>Both</strong> can challenge their assumptions, learn, and develop new solutions through their collaboration with AI.</li></ul><p>The real value emerges from their interaction. <strong>One plus one plus one can become greater than three</strong> — not because intelligence can literally be reduced to a number, but because collaboration can create capabilities that none of its participants possesses independently. This is the principle of synergy.</p><p>And perhaps it offers a different way to think about organizing society. Today, we often separate learning, production, and the transmission of experience into different stages of life:</p><ol><li><strong>First,</strong> we learn.</li><li><strong>Then,</strong> we work.</li><li><strong>Later,</strong> we teach.</li></ol><p>But why should these activities remain separated? Imagine a society where young people contribute while learning, experienced professionals create value while mentoring, and AI supports both. Learning, production, and mentorship could become continuous and interconnected.</p><p>The junior learns from the senior. The senior benefits from the junior’s fresh perspectives. AI helps both explore, test, and implement ideas. And through their collaboration, the junior gradually becomes the next senior, ensuring that knowledge and judgment continue passing between generations.</p><p><strong>AI could become the bridge between the intelligence of youth and the intelligence of experience.</strong></p><p>This is not simply a productivity improvement. It is an opportunity to reorganize how society creates and transmits value. The objective is no longer to maximize each individual’s productivity in isolation. It is to combine complementary capabilities so that the collective result can exceed what individuals could achieve separately.</p><blockquote>The future may belong not to the intelligence that defeats all others, but to the society that learns how to combine them.</blockquote><h3>11. The Engineer as an Orchestra Conductor</h3><p>This transformation could profoundly change what we expect from engineers. In the past, engineering education often emphasized technical specialization:</p><ul><li>Learning programming languages.</li><li>Understanding algorithms.</li><li>Mastering frameworks and development tools.</li></ul><p>These skills remain important. But they may no longer be sufficient.</p><p>If AI agents increasingly handle portions of implementation, engineers will need to become more versatile. They will need to understand business problems, communicate with stakeholders, design systems, evaluate technical solutions, manage risks, and coordinate different forms of intelligence. They may need to alternate between the roles of technical lead, product owner, business relationship manager, and engineering reviewer.</p><p><strong>The future engineer may resemble an orchestra conductor.</strong></p><p>A conductor does not need to play every instrument during a performance. But they must understand the music, recognize the capabilities of the musicians, coordinate their contributions, and take responsibility for the coherence of the result.</p><p>Similarly, an engineer working with AI agents may not personally write every line of code. But they must understand the system well enough to evaluate what is produced. They must know when to trust, when to question, and when to intervene. And they must remain capable of reasoning independently.</p><p>This creates an important challenge for education: how do we develop senior engineers if junior engineers no longer spend years performing the implementation tasks through which previous generations gained experience?</p><p>We will need new forms of apprenticeship, deliberate technical practice, and structured mentorship. The objective should not be to produce professionals who merely know how to prompt AI systems. It should be to educate people capable of understanding problems, applying scientific reasoning, evaluating evidence, collaborating with others, and accepting responsibility.</p><p>The future engineer may be less of a specialized code producer and more of an orchestra conductor. And perhaps the most important responsibility of that conductor will be to teach the next generation how to conduct.</p><h3>12. Intelligence Can Be Distributed, but Responsibility Must Remain Identifiable</h3><p>This brings us to a fundamental question: if artificial intelligence can reason about problems, execute tasks, and deliver outcomes, <strong>who remains responsible?</strong></p><p>For me, the answer is clear. Human beings and human organizations must remain accountable for the systems they choose to develop and deploy.</p><p>This is not fundamentally different from the way responsibility already works in many organizations. A manager delegates tasks to colleagues. Those colleagues make decisions, perform work, and sometimes make mistakes. But delegation does not automatically eliminate the manager’s responsibility for supervision, appropriate processes, and the decisions within their authority.</p><p>Similarly, when I delegate implementation to an AI agent, my engineering responsibility does not disappear:</p><ul><li>I must define the expected outcome.</li><li>I must establish appropriate constraints.</li><li>I must evaluate the results and determine whether the evidence supports accepting them.</li><li>And I must remain accountable for the decisions I make.</li></ul><p>Of course, responsibility is often distributed among engineers, managers, organizations, and regulatory authorities. It should be assigned clearly rather than concentrated artificially in one person. <strong>But it must never disappear simply because an AI system performed the work.</strong></p><p>There is a deeper philosophical distinction here. Human intelligence may no longer be the exclusive source of sophisticated problem-solving capabilities. But when we are addressing human problems, our objectives remain grounded in human needs and values:</p><ul><li>We define what matters.</li><li>We determine which risks are acceptable.</li><li>We decide what kind of world we want to build.</li></ul><p>AI may increasingly contribute to identifying solutions and executing them. But that does not remove our obligation to decide how those capabilities should be used.</p><blockquote>We can distribute intelligence. We can delegate execution. But responsibility must remain identifiable.</blockquote><p>This is not an argument for blind trust in artificial intelligence. It is an argument for <strong>responsible collaboration</strong>.</p><h3>Conclusion: Intelligence Is Not a Throne We Must Defend</h3><p>Perhaps the greatest lesson of the Copernican revolution was not that Earth lost its place at the center of the universe. It was that humanity learned to see beyond itself.</p><p>Today, artificial intelligence may be inviting us to make a similar leap. We are discovering that sophisticated intellectual capabilities can take different forms, that we can collaborate with systems unlike ourselves, and that our value does not depend on remaining intellectually superior in every domain.</p><p>This realization can be frightening. <strong>But it can also be liberating.</strong></p><p>We do not need to deny the capabilities of artificial intelligence to preserve our humanity. We do not need to insist that every important task must be performed by a biological mind. And we do not need to surrender our judgment or responsibility simply because another intelligence can contribute to solving our problems.</p><p>We can learn to collaborate. We can reorganize our work. We can rethink our institutions, our educational systems, and our understanding of engineering. We can use these new capabilities to spend more time on human relationships, scientific exploration, creativity, and meaningful problems.</p><p>And perhaps the greatest opportunity artificial intelligence offers us is not to replace human intelligence, but to help us reorganize society around the complementary strengths of different intelligences. Imagine young people contributing while learning, experienced professionals creating value while mentoring, and AI connecting their capabilities. Imagine companies organized around collaboration rather than rigid divisions of responsibility, and education becoming a continuous process of learning through meaningful contribution.</p><p>The result could be something greater than the sum of its parts. The future may belong not to the intelligence that defeats all others, but to the society that learns how to combine them.</p><p>This raises questions that go far beyond technology:</p><ul><li>How should we educate future generations?</li><li>What should a team look like when humans and AI agents collaborate to produce outcomes?</li><li>What does intellectual competition mean when another intelligence can challenge our own?</li><li>How should we distribute responsibility?</li><li>And how should we reorganize society so that different generations and different forms of intelligence can contribute according to their strengths?</li></ul><p>These are not questions that technology alone can answer. They concern the kind of civilization we want to build. The emergence of another form of intelligence means that human society may never be quite the same again. <strong>The question is not whether we will change, but how we choose to adapt.</strong></p><p>And perhaps there is an even greater implication. If intelligence-like capabilities can emerge in artificial systems, why should we assume that human intelligence is the only form of intelligence in the universe? Artificial intelligence does not prove that extraterrestrial intelligence exists. But it gives us another reason to consider the possibility that intelligence may take forms radically different from our own.</p><p>And if one day we encounter another intelligence, whether artificial or extraterrestrial, we will face the same fundamental choice: will we see it primarily as a competitor to defeat, a threat to control, or an opportunity to learn and collaborate?</p><p>I know which future I would choose. <strong>I choose collaboration.</strong></p><p>Not because collaboration eliminates competition or risk, but because our greatest strength has never been our ability to remain superior to everything around us. It has been our capacity to adapt, to learn, and to grow through our encounters with the unknown.</p><p>The emergence of artificial intelligence does not mean that humanity must surrender its responsibility. On the contrary, it makes that responsibility more important than ever. We can distribute intelligence. We can delegate execution. But responsibility must remain identifiable.</p><p>None of this guarantees that artificial intelligence will automatically make humanity wiser, more compassionate, or more responsible. The outcomes will depend on the technologies we develop, the institutions we create, and the choices we make.</p><p>Copernicus helped us understand that the physical universe was far larger than we had imagined. That realization eventually inspired generations to reach for the stars. Perhaps artificial intelligence will help us discover that the universe of intelligence is equally vast.</p><p>And perhaps the most important question is not whether humanity can remain at its center, but what we might become when we finally accept that we are not.</p><blockquote><strong><em>Intelligence is not a throne we must defend. It is a horizon we can explore.</em></strong></blockquote><p><em>At least, I hope it is.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=e900533ca510" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Cognitive Evolution of Generative AI — The theory behind code assistant]]></title>
            <link>https://medium.com/@romualdrousseau/the-cognitive-evolution-of-generative-ai-the-theory-behind-code-assistant-47599d4e8f7d?source=rss-a810a6bf806c------2</link>
            <guid isPermaLink="false">https://medium.com/p/47599d4e8f7d</guid>
            <dc:creator><![CDATA[Romuald Rousseau]]></dc:creator>
            <pubDate>Mon, 06 Jul 2026 13:04:37 GMT</pubDate>
            <atom:updated>2026-07-06T13:04:37.157Z</atom:updated>
            <content:encoded><![CDATA[<h3><strong>The Cognitive Evolution of Generative AI — The theory behind code assistant</strong></h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*gs_ugq8eJREKivbpAYq9ZQ.png" /></figure><p>The emergence of Large Language Models (LLMs) has revolutionised our interaction with machines, yet the quest for truly “intelligent” autonomous systems requires more than just fluent text generation. To bridge the gap between simple automation and complex problem-solving, we must look toward the architecture of human cognition.</p><p>By applying the principles of <strong>Dual-Process Theory</strong>, we can categorise standard LLM inference as “System 1” — a fast, intuitive, but occasionally error-prone mechanism. However, the future of AI development, particularly in specialised fields like software engineering and scientific research, lies in the emulation of “System 2” — a slow, deliberate, and logical reasoning framework. This discussion explores the theoretical transition from simple predictive models to <strong>Program-aided Language Models (PAL)</strong> and the implementation of the <strong>Harness</strong>: a sophisticated reasoning engine that manages memory, plans execution, and utilises specialised skills to solve high-order problems with unprecedented accuracy.</p><h3>Dual Process Theory and Chain of Thought</h3><p>The dual-process theory of the human brain, as defined by Daniel Kahneman, consists of two distinct modes of thinking:</p><p><strong>System 1: Intuition &amp; Instinct</strong> This system represents approximately <strong>95%</strong> of our mental activity. It is characterised as the “automatic pilot” of the brain. Key features include:</p><ul><li><strong>Unconscious and Fast:</strong> Operates automatically and quickly with little or no effort.</li><li><strong>Associative:</strong> Relies on patterns and instincts to make immediate decisions.</li></ul><p><strong>System 2: Rational Thinking</strong> This system accounts for only about <strong>5%</strong> of our mental activity. It is the part of the brain responsible for complex mental operations. Key features include:</p><ul><li><strong>Takes Effort and Slow:</strong> Requires conscious attention and energy to function.</li><li><strong>Logical:</strong> Focuses on methodical reasoning and calculation.</li><li><strong>Lazy and Indecisive:</strong> Because it requires significant cognitive resources, it is often slow to engage and can struggle to reach quick conclusions.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/960/1*496RyYmlHGKh_Cy9tD440w.png" /></figure><p>Can we draw a parallel between human cognitive systems and Generative AI? The answer is a resounding yes.</p><p>In this analogy, a standard <strong>Large Language Model (LLM)</strong> operates like <strong>System 1</strong>: a single prompt triggers an output through a continuous inference pass where reasoning and calculation happen simultaneously per token. While fast and intuitive, this “automatic pilot” mode is prone to errors in complex logic.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*IWMAA9HkzXNaD8Pv8ZOJeQ.png" /></figure><p>The pivotal question is: can we emulate <strong>System 2</strong> — the slow, rational, and logical mode — within AI to tackle complex problems? Research into <strong>PAL (Program-aided Language Models)</strong> demonstrates that we can. Instead of forcing the model to solve a problem directly in one pass, the PAL framework instructs the model to:</p><ol><li><strong>Plan</strong> the solution by breaking down the logic into structured steps.</li><li><strong>Generate instructions</strong> (such as Python code) rather than just text.</li><li><strong>Execute</strong> those instructions via an external agent or interpreter.</li></ol><p>As shown in the provided research image, this method effectively simulates System 2 by offloading logical “thinking” to a rigorous execution environment. The results are significant: by shifting from direct reasoning to program-aided reasoning, the AI significantly <strong>improves performance</strong> on benchmarks like GSM8K (rising from 62.5 to 73.3 in some configurations) and <strong>drastically reduces hallucinations</strong>, ensuring the final answer is grounded in verifiable calculation.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*fvsD1z843YAk3VVJGfH0lQ.png" /></figure><p>So by asking an LLM to generate the program and not directly answer the question, we get far better results. So code assistant are the most natural use case resulting form the implementation of system 2 using AI.</p><h3>The Role of Memory: Short-Term vs. Long-Term</h3><p>The human brain relies on a sophisticated hierarchy of memory; while <strong>System 1</strong> draws upon immediate, short-term “working” memory for rapid-fire responses, deep reasoning — the hallmark of <strong>System 2</strong> — requires the extraction and synthesis of information from vast long-term reservoirs.</p><p>This cognitive architecture finds a direct parallel in Generative AI:</p><ul><li><strong>Short-Term Memory (The Context Window):</strong> For an LLM, the context window serves as its immediate workspace. Supported by the attention mechanism, it allows the model to “remember” the current conversation and provided data. However, this space is finite and subject to decay or “lost-in-the-middle” effects as complexity increases.</li><li><strong>Long-Term Memory (External Knowledge Bases):</strong> The equivalent of long-term memory exists in external databases, documentation, and vector storages. This represents the entire domain of potential knowledge that exceeds what can fit into a single prompt.</li></ul><p>Just as human intelligence is often measured by the ability to recall the right information at the right time, a critical performance factor for AI is the efficiency of its retrieval mechanisms. The transition from simple AI to a “System 2” equivalent hinges on how effectively the system can identify, extract, and inject relevant data from long-term storage into the active context. Optimising this bridge between vast storage and active reasoning is essential for solving high-level, domain-specific problems accurately.</p><h3>From System 2 to the AI Harness: The Architecture of Reasoning</h3><p>We now have all the components necessary to construct a high-performance code assistant. At its core, a modern code assistant is a sophisticated software implementation of these cognitive theories, centered around a structural “Harness.”</p><p>The <strong>Harness</strong> acts as the practical realization of <strong>System 2</strong>. It is an iterative loop designed to solve problems by first creating a strategic plan via an LLM, then orchestrating the collection of relevant data. This data is pulled from diverse sources — databases, external APIs, distributed computation, and specific technical background — to populate the context. The Harness then queries the LLM to generate the final, grounded output.</p><p>To interact with the world, the Harness utilizes advanced agentic programming protocols, such as <strong>MCP</strong> (Model Context Protocol) or <strong>A2A</strong> (Agent-to-Agent), while communicating with client software like IDEs via <strong>ACP</strong> (Agent Communication Protocol).</p><h3>The Intelligence of Memory Management</h3><p>A critical role of the Harness is active memory management. It must select the most pertinent information and compress the context without losing vital nuances. This “real-time intelligence” is often as important as the raw power of the underlying LLM.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/892/1*tnypD8fbQy-b64GYpfGVNA.png" /></figure><p>As shown in the provided diagram, the Harness organizes its memory map into several distinct layers:</p><ul><li><strong>Rules:</strong> Local, fixed constraints that define the boundaries of the task.</li><li><strong>Prompt:</strong> The specific user intent and instructions.</li><li><strong>RAG (Retrieval-Augmented Generation):</strong> Knowledge pulled dynamically from external databases (DB).</li><li><strong>Agents (Dynamic):</strong> Results from external computations, such as the output of a compiler or linter.</li><li><strong>Skills (Dynamic):</strong> Locally stored procedural capabilities.</li><li><strong>Reasoning:</strong> The logical steps and “chain of thought” generated by the LLM.</li><li><strong>Previous Answers:</strong> A feedback loop that incorporates prior outputs to maintain consistency and refine the current task.</li></ul><h3>The Power of SKILLS</h3><p>While we have discussed agents for computation, the <strong>SKILL</strong> is perhaps the most versatile feature of the Harness.</p><p>In the context of a code assistant, a <strong>SKILL</strong> is a predefined, modular procedural template or a specialized capability that the assistant can invoke to perform a specific type of task. It is a way to “fine-tune” the assistant’s behavior without retraining the model. For example, a skill might be “Perform a Security Audit,” “Refactor for Performance,” or “Generate Unit Tests.”</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/977/1*Lk8acy_8WU2mgeg5CeWlKA.png" /></figure><h3>Conclusion: Toward a Unified Architecture for Complex Reasoning</h3><p>The evolution from a standalone LLM to a fully integrated <strong>Code Assistant Harness</strong> represents a paradigm shift in artificial intelligence. It is a move from “guessing” a solution to “engineering” one. By treating the LLM as a core processing unit within a broader cognitive framework, we successfully replicate the human brain’s ability to switch between immediate intuition and deep, structured thought.</p><p>As we have seen, the intelligence of these systems is defined not only by the parameters of the model but by the sophistication of the Harness — its ability to manage short-term and long-term memory, its agility in invoking specialised <strong>Skills</strong>, and its capacity to interact with external tools via agentic protocols. This System 2 architecture is not limited to writing code; it provides a universal blueprint for any domain requiring rigorous logic, from drug synthesis to complex data analysis. Ultimately, by grounding generative power in a framework of structured reasoning and active memory management, we are moving toward a future where AI acts not just as a mirror of human language, but as a true partner in human discovery.</p><blockquote><strong>References</strong></blockquote><blockquote><em>Stanford Technical Fundamentals of Generative AI</em></blockquote><blockquote><strong>Disclaimer</strong></blockquote><blockquote>This article has been reviewed and corrected using Generative AI. The views and opinions expressed in this article are solely those of the author and do not necessarily reflect the official policy or position of the author’s company. Any content provided by the author is of their own opinion and is not intended to malign any organisation, company, individual, or entity.</blockquote><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=47599d4e8f7d" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[From Prompt to Prompt — Why Natural Language is the Final Programming Language]]></title>
            <link>https://medium.com/@romualdrousseau/from-prompt-to-prompt-why-natural-language-is-the-final-programming-language-a489ae05c97d?source=rss-a810a6bf806c------2</link>
            <guid isPermaLink="false">https://medium.com/p/a489ae05c97d</guid>
            <dc:creator><![CDATA[Romuald Rousseau]]></dc:creator>
            <pubDate>Mon, 06 Jul 2026 12:59:58 GMT</pubDate>
            <atom:updated>2026-07-08T07:09:34.824Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*fLvdIh07I9naq8rifx-GsQ.png" /></figure><p>When I was a kid, the gateway to the digital world looked very different. There were no sleek icons or touchscreens. Most computers booted directly to a stark, blinking command prompt. Back then, there was a widespread belief that to use a computer, you had to know how to program it. The vision was that everyone would, at the very least, become a “vibe coder,” speaking the machine’s basic language to get things done.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*gA1wz6vBxJvreQJUITNsDw.png" /></figure><p>Then, the world changed. Graphical User Interfaces (GUIs) like Windows arrived, the Internet exploded, and the demand for complex software skyrocketed. Programming shifted from a general literacy skill to a specialised craft, practised by experts dedicated to writing thousands of lines of code. The “prompt” was buried under layers of user-friendly design.</p><p>Decades later, we are witnessing a fascinating full-circle moment. Today, our primary interface with the most advanced technology on the planet has returned to a single line: the prompt.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*oe2nyx4V_5MSiE__jOKipg.png" /></figure><p>It is striking to see us return to these paradigms. Different era, different power, yet the same fundamental intent.</p><h3>The Core of the Machine: Solving Human Problems</h3><p>At its heart, what is the goal of computing? It is the desire to tell a machine to solve a human problem. This act of “telling” is, by definition, programming.</p><p>Throughout tech history, the medium has shifted:</p><ul><li><strong>Programmatic languages:</strong> Precise, rigid, and powerful (C++, Python, Java).</li><li><strong>Visual languages:</strong> Drag-and-drop blocks and flowcharts.</li><li><strong>Natural languages:</strong> The way we speak to one another.</li></ul><p>The specific type of language we use is often a matter of technical constraint or temporary fashion. I have seen countless languages rise and fall; I have seen “no-code” solutions promise a revolution only to disappoint when faced with real-world complexity.</p><p>However, two things never change: the human intent to solve a problem and the necessity of modeling that problem so a machine can execute it.</p><h3>Natural Language: The Ultimate Syntax</h3><p>We are entering an era where natural language is becoming the ultimate programming language. It is the language we use to express the initial problem in our minds, and for the first time in history, it is the same language we use to describe that intent to the machine.</p><p>We aren’t just “chatting” with AI; we are programming it. By articulating our needs in plain English (or any other language), we are performing the highest level of abstraction possible. The blinking cursor of the 1980s has returned, but this time, it understands us perfectly. We have finally moved from teaching humans the language of machines to teaching machines the language of humans.</p><h3>The End of the Intermediates: Why Conversation is the New Source Code</h3><p>As we have observed, we are entering a unique era where, for the first time, the end-user, the engineer, and the machine speak the same language. The traditional need for “translators” — the layers of bureaucracy and manual coding that sit between a human idea and a digital reality — is dissolving.</p><p>An end-user can now produce a functional solution simply by talking to the machine. Simultaneously, the engineer can scale and refine that solution by engaging in the same dialogue. They are all operating within a shared, common understanding.</p><p>This raises a fundamental question: In this new ecosystem, what is the actual medium of work? Is it the source code generated by the machine?</p><p><strong>No. The medium is the dialogue itself.</strong></p><p>The entire conversation — the iterative exchange between the user, the engineer, and the AI — is what defines both the problem and its ultimate solution. The prompt history, skills, agents, outputs and the refined requirements are becoming the true “source of truth.”</p><h4>Does the Engineer Become Obsolete?</h4><p>If the machine can understand the user, is the engineer not just another intermediate to be bypassed? My conviction is that the answer is no.</p><p>The engineer’s role is evolving, not disappearing. While a machine can bridge the gap toward a functional result, scaling a piece of software requires more than just meeting a user’s basic needs. It requires a mastery of both <strong>functional</strong> and <strong>non-functional</strong> specifications.</p><p>The non-functional specifications are the silent pillars that allow software to survive in the real world:</p><ul><li><strong>Architecture:</strong> Ensuring the system can handle growth.</li><li><strong>Security:</strong> Protecting data and integrity in a hostile environment.</li><li><strong>Monitoring &amp; Maintenance:</strong> Ensuring the system remains healthy over time.</li></ul><p>While the user focuses on <em>what</em> the machine should do, the engineer ensures <em>how</em> it does it at scale. The engineer is no longer a translator of syntax, but an architect of systems, using natural language to steer the machine toward professional-grade stability.</p><h3>Conclusion: The New Language of Logic</h3><p>The journey from the blinking BASIC prompt of the 1980s to the sophisticated conversational interfaces of today represents more than just a technological cycle; it is a fundamental shift in how we project human will onto silicon. We have moved from a world where humans had to learn the rigid, unforgiving syntax of machines to an era where machines have mastered the fluid, nuanced context of human speech.</p><p>In this new paradigm, the “source code” is no longer a hidden file of cryptic commands, but the very dialogue we share. However, this does not mean the end of rigour. As the barrier to creation falls, the importance of structural integrity — architecture, security, and scalability — only grows. The engineer remains the vital guardian of these non-functional pillars, ensuring that the solutions born from conversation can withstand the demands of the real world.</p><p>We have returned to the prompt, but we have returned with a power once unimaginable. We are all “coders” but no longer tinkering in the dark; we are architects of intent, using the ultimate programming language — our own — to shape the future.</p><blockquote><strong>Disclaimer</strong></blockquote><blockquote>This article has been reviewed and corrected using Generative AI. The views and opinions expressed in this article are solely those of the author and do not necessarily reflect the official policy or position of the author’s company. Any content provided by the author is of their own opinion and is not intended to malign any organisation, company, individual, or entity.</blockquote><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=a489ae05c97d" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[From Binary to Natural Language: The Evolution of Software Engineering]]></title>
            <link>https://medium.com/@romualdrousseau/from-binary-to-natural-language-the-evolution-of-software-engineering-79922b131e52?source=rss-a810a6bf806c------2</link>
            <guid isPermaLink="false">https://medium.com/p/79922b131e52</guid>
            <category><![CDATA[software-development]]></category>
            <category><![CDATA[ai-code-assistant]]></category>
            <category><![CDATA[software-engineering]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[generative-ai-tools]]></category>
            <dc:creator><![CDATA[Romuald Rousseau]]></dc:creator>
            <pubDate>Wed, 25 Mar 2026 14:07:47 GMT</pubDate>
            <atom:updated>2026-07-06T13:05:11.016Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*yo1ikTWNsk6X_r4my1RzAg.png" /></figure><h3>What is programming?</h3><p>Before discussing the future, it is essential to clarify: this is not a manifesto to abandon traditional programming. There is a deep, satisfying craft in working close to the machine — writing in 6502, Z80, or 68000 assembly and understanding every instruction.</p><p>This knowledge remains vital. Understanding low-level computing and languages like C makes you a better engineer; it strengthens your ability to reason about performance and constraints and to translate human problems into machine solutions. But programming languages emerge and go, they are more a fashion than an objective. Skill programming and code assistants are not here to erase programming; they are here to change where the value lies. It is not the end of programming — it is a return to what programming was always supposed to be.</p><p>Programming is the process of architecturing and delivering precise instructions to a computer to solve specific problems. While humans identify challenges and conceptualise solutions, computers provide the immense processing power required to execute those solutions — provided they are structured through logical, computable/calculable steps. Essentially, programming acts as the definitive bridge between human intent and digital execution; whenever an individual formalises a rationale into code to automate a task or perform a calculation, they are programming.</p><h3>The Long Journey Toward Abstraction: A Natural Evolution</h3><p>Since the dawn of computer science, the ultimate ambition has been to bridge the gap between human thought and machine execution. The history of programming is a continuous evolution of abstraction, moving closer to the “Star Trek” dream of natural language interaction:</p><ul><li><strong>1950s (The Beginning):</strong> Raw and unforgiving machine language, binary, and punch cards. Computing was limited to a tiny subset of extremely specialised operators who manually inputted commands.</li><li><strong>The First Breakthrough:</strong> Assemblers and compilers allowed symbolic instructions (e.g., MOV instead of 0101).</li><li><strong>2000s (The Acceleration):</strong> Structured and high-level languages (C, Java, Python, Closure, …) brought us closer to human reasoning. <strong>Object-oriented programming</strong> modelled real-world systems, enabling a much broader group of people to participate in computing. <strong>Functional programming</strong> introduced mathematical rigour, enabling programs to be reasoned about — and even <em>proven</em> — like equations.</li><li><strong>Current (Instruction Following):</strong> The realisation of the “dream” of interacting with systems in complete natural language. AI assistants act as universal compilers for human intent, finally closing the gap between human thought and machine execution.</li></ul><p>This evolution is characterised by moving “up the stack.” Just as we moved from managing memory manually to garbage collection, we are now moving from managing syntax to managing intent. We have always moved toward more expressiveness and alignment with human thinking.</p><h3>AI-Powered Code Assistants: The AI Compiler</h3><p>AI-powered Code Assistants represent the next leap in this evolution. They are, in essence, a new generation of compiler: the <a href="https://medium.com/ai-compiler"><strong>ai-compiler</strong></a>.</p><ul><li><strong>Traditional Compilers:</strong> Deterministic construction. They follow rigid, hand-coded rules to translate Formal Syntax → Assembly → Machine Code.</li><li><strong>AI Compilers:</strong> Probabilistic sampling. They treat programming languages as distributions of patterns. They translate Natural Language (Intent) → Code → Execution through an iterative <strong>Agentic Loop</strong> (Analyse → Retrieve → Synthesise → Validate → Correct).</li></ul><p>In this paradigm, the “source of truth” is no longer formal code, but <strong>“The New Source”</strong> — a composite of <strong>User Stories</strong> (specifications), <strong>Skills</strong> (architectural patterns), and <strong>Prompts</strong> (real-time adjustments). Code becomes a commodity — an intermediate, disposable artifact that can be regenerated instantly if the intent changes.</p><blockquote><em>For a deeper dive into why LLMs excel at generating formal language, you can read my earlier piece </em><a href="https://medium.com/@romualdrousseau/the-grammar-of-intelligence-why-llms-work-and-where-humans-still-win-5586d6f337ce"><em>The Grammar of Intelligence — Why LLMs Work and Where Humans Still Win</em></a><em>.</em></blockquote><h3>Skill Programming — Beyond the Syntax of Code</h3><p>As we move into this new era of abstraction, we are transitioning from “Code Programming” to “Skill Programming.” In this paradigm, natural language serves as the primary interface, and the goal is no longer to produce lines of code but to define and compose “skills” — modular, goal-oriented capabilities that an AI can execute.</p><h4>1. The Shift from Syntax to Intent: The Architect of Intent</h4><p>In traditional programming, the engineer acts as a translator. In Skill Programming, the engineer acts as an <strong>Architect of Intent</strong>.</p><ul><li><strong>Code as a Commodity:</strong> We no longer optimise for writing code; we optimise for describing intent. The value shifts from the source file to the requirement document.</li><li><strong>Semantic-First Development:</strong> The focus is shifting from mastering specific programming languages to conceptualising logical architectures and articulating functional goals with scientific rigour.</li></ul><h4>2. Expanding the Horizon: BMAD and LLM Wikis</h4><p>Skill programming is not limited to generating software; it is about structuring knowledge and processes so that AI agents can act upon them autonomously. Frameworks like BMAD and LLM Wikis exemplify how skills go beyond mere code:</p><ul><li><strong>LLM Wikis (Knowledge as a Skill):</strong> Popularised by researchers like Andrej Karpathy, this concept treats knowledge management like software development. An LLM Wiki is a persistent, structured repository of markdown files built and maintained by AI agents.</li><li><strong>BMAD (Breakthrough Method for Agile AI-Driven Development):</strong> This framework brings engineering discipline to AI by utilising specialised agents with distinct roles. BMAD moves beyond “vibe coding” by emphasising detailed specifications and structured artefacts (PRDs) to avoid the “Quick-Wins Horizon” of diminishing returns.</li></ul><h4>3. Skill Engineering and Orchestration</h4><p>“Skill Engineering” is the practical activity of defining and validating reusable, AI-operable capabilities. When an engineer defines a “Product Launch Skill,” it triggers a cohort of specialised agents to write stories, generate code, and provision infrastructure. The distance between a business idea and a functional system is reduced to the time it takes to clearly articulate the “skill.”</p><h4>4. The Paradox of Expertise: Why “By Hand” Still Matters</h4><p>Counter-intuitively, this automation raises the bar for expertise. The modern engineer must master a broader range of skills to evaluate whether AI-generated code is correct — a critical decision-making capability that only comes with deep experience. We risk a generation of engineers who can “prompt” but cannot “validate” if we do not sanctuary time for manual practice.</p><h3>The Redefinition of the Software Engineer: A Structural Revolution</h3><p>The “Revolution” lies in the collapse of the traditional software production chain. We are moving from the <strong>“Standard Line”</strong> (heavy on proxy roles, feature teams, and translation layers) to the <strong>“New Line”</strong> (centered on high-expertise engineers and direct business value).</p><ul><li><strong>The End of the “Developer” as a Middleman:</strong> When the “compiler” understands English, the “Full-Spectrum Engineer” returns, using AI agents as a virtual team.</li><li><strong>Maintainability Re-framed:</strong> We don’t maintain the generated code any more than a C developer maintains assembly. The “Source of Truth” shifts to the requirement document itself.</li><li><strong>Agile and the V-Model:</strong> Agile was a “workaround” for the high cost of manual coding. If code generation is instantaneous, we move toward “specification-first” development.</li></ul><h3>A Necessary Organisational Transformation</h3><p>The deployment of Code Assistants is a profound transformation project. Because code is no longer the bottleneck, artificial silos and layers of project management are becoming obsolete.</p><ul><li><strong>Team Structure:</strong> Teams will likely collapse back into smaller, highly efficient units. If an organisation keeps the same roles and processes, the transformation has failed.</li><li><strong>Outsourcing:</strong> The traditional model of paying third parties to “type code” is disrupted. Value shifts entirely from “lines of code” to problem understanding and system design.</li></ul><h3>The Next Frontier: Physically Grounded AI</h3><p>As abstraction increases, we will move beyond pure “synthetic” generation. The next frontier is <a href="https://medium.com/physically-grounded-ai"><strong>physically-grounded-ai</strong></a> — incorporating physical laws and real-world constraints into the agentic loop.</p><p>The complexity of future systems — such as quantum computing or massive distributed networks — will be impossible for a single human mind to program manually. Code assistants are not just helpful; they are necessary for the next stage of evolution. The “Star Trek” dream is here; the question is whether organizations are ready to rethink everything else.</p><blockquote><strong><em>References</em></strong></blockquote><blockquote>Andrej Karpathy’s LLM-Wiki Concept<br> The BMAD Method Framework<br> AI-Assisted Coding in 2026<br> Skill Engineering Principles<br> Stanford Technical Fundamentals of Generative AI</blockquote><blockquote><em>Disclaimer</em></blockquote><blockquote><em>This article has been reviewed and corrected using Generative AI. The views and opinions expressed in this article are solely those of the author and do not necessarily reflect the official policy or position of the author’s company. Any content provided by the author is of their own opinion and is not intended to malign any organization, company, individual, or entity.</em></blockquote><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=79922b131e52" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Grammar of Intelligence — Why LLMs Work and Where Humans Still Win]]></title>
            <link>https://medium.com/@romualdrousseau/the-grammar-of-intelligence-why-llms-work-and-where-humans-still-win-5586d6f337ce?source=rss-a810a6bf806c------2</link>
            <guid isPermaLink="false">https://medium.com/p/5586d6f337ce</guid>
            <category><![CDATA[quantum-computing]]></category>
            <category><![CDATA[computer-science]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Romuald Rousseau]]></dc:creator>
            <pubDate>Fri, 30 Jan 2026 09:07:36 GMT</pubDate>
            <atom:updated>2026-01-30T11:11:58.507Z</atom:updated>
            <content:encoded><![CDATA[<h3><strong>The Grammar of Intelligence — Why LLMs Work and Where Humans Still Win</strong></h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*KQLxg2RaeBqlD6FvirHeNg.png" /></figure><p>For several years, large language models such as ChatGPT have amazed the public by writing code, solving equations, and holding long conversations. Despite these feats, a common claim persists that these systems are “black boxes” that no one truly understands. This is only partly true; while we may not map every individual weight inside a neural network, we increasingly understand the principles that make them work.</p><p>Our proposition of how a neural network works lies in its ability to approximate the distribution of a grammar — the “skeleton” of language. By redefining tasks like mathematics or coding as conditional probability distributions of an underlying grammar, we can show that LLMs succeed by internalizing these generative patterns through statistical learning at scale. When we contrast these probabilistic engines with the intuitive, emotional, and grounded nature of human learning, we gain a unified view of both the power of artificial intelligence and the profound essence of human cognition.</p><h3>The Proposed Model — Syntax as the Skeleton</h3><p>The fundamental reason LLMs succeed is their ability to learn the probabilistic grammar of language. Grammar is the structural framework of language; before a sentence can carry meaning, it must first be well-formed.</p><blockquote>Before proceeding, it is important to clarify that we define <strong>language</strong> as a formal construct consisting fundamentally of an <strong>alphabet</strong> and a <strong>grammar</strong> (see <a href="https://en.wikipedia.org/wiki/Formal_language">Formal Language</a>, <a href="https://en.wikipedia.org/wiki/Formal_grammar">Formal Grammar</a>). While this discussion primarily focuses on symbolic and/or human languages, this definition extends to any data: it can describe molecular sequences (such as SMILES), visual patterns (pixels and their spatial convolutions), or audio signals (phonemes and phonetic rules).</blockquote><h4>The Mathematical Framework: Distributional Equivalence</h4><p>To explain why a neural network is so efficient at coding or doing maths, we can redefine these tasks as conditional probability distributions of a grammar. We compare two specific distributions:</p><ul><li><strong>The Target Distribution (P*):</strong> This is the “perfect” conditional probability distribution defined by the formal grammar of mathematics or code. For a formal language, this distribution is degenerate, meaning P*(string ∣ context) ∈ {0, 1}.</li><li><strong>The Model Distribution (</strong>P̂θ<strong>):</strong> The stochastic approximation generated by the neural network. LLMs are stochastic neural network because of their final <em>softmax </em>layer.</li></ul><p>To prove the model “works,” we simply show that as training progresses, the distance between these two distributions approaches zero:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/191/1*NSBFNl7X1oWWzBwwqtrmTA.png" /></figure><p>Where D<em>KL</em> is the Kullback–Leibler divergence. Because formal grammars are relatively simple — often consisting of only a few hundred or thousand rules — a large neural network with billions or trillions of parameters can effectively “brute force” the memorization of the entire grammar. Reasoning in this framework is not a mystery; it is <strong>sampling from learned derivation distributions</strong>.</p><ul><li><strong>Statistical Internalisation vs. Memorization</strong>: LLMs do not memorize sentences; they internalise the patterns and regularities that generate them.</li><li><strong>Phase Transition</strong>: During training, models reach a “breaking point” or phase transition where they suddenly become fluent. This jump likely corresponds to the moment the model moves from local patterns to internalizing the global generative grammar.</li><li><strong>Advantage of Formal Languages</strong>: In domains like mathematics and programming, syntax and semantics are closely aligned. If code compiles or a proof verifies, it is generally correct. Mastering the structural framework almost entirely solves the problem.</li></ul><h4>Intuitive properties using this model</h4><ul><li><strong>Hallucination Phenomenon</strong>: Hallucinations occur when syntactic competence exceeds semantic grounding. The model knows perfectly <em>how</em> to say something but does not know if it is <em>true</em> because syntax is frequent and statistically strong, while semantics is contextual and noisy.</li><li><strong>Translation</strong>: can be seen as a linear transformation (translation, rotation) from the distribution of the grammar of the source language to the distribution of the grammar of the target language; transformation and distributions all encoded in the LLMs. As an example, source languages can be English or French while target languages can be Chinese or python code.</li><li><strong>Summarization</strong>: can be seen as a linear transformation (scaling, homothety) on the same (endomorphism) distribution of the grammar of the source, transformation and distributions all encoded in the LLMs.</li></ul><h3>The Human Learning Path — Intuition and Emotion</h3><p>Human intelligence operates on a completely different scale of efficiency. While an LLM uses trillions of tokens, a child acquires language from only millions. Humans do not simply accumulate statistics; they build internal world models.</p><ul><li><strong>Intuition as Rapid Simulation</strong>: Human intuition is an internal generative world model that accelerates computation. Before acting, the brain simulates possible outcomes to estimate risks and rewards, allowing us to generalize from very few examples.</li><li><strong>Biological Optimisation via Emotion</strong>: Human learning is physically guided by neurochemistry. Dopamine and endorphins provide positive reinforcement for success, while stress hormones provide negative feedback for failure. These function as biological optimisation systems that shape our neural connections.</li><li><strong>The Passion-Mastery Loop</strong>: Talent often emerges from a chemical feedback loop. If solving a maths problem produces a “happy hormone” reward, the individual repeats the activity. This increased practice improves skill, which in turn increases the emotional reward.</li><li><strong>The Motivation Gap</strong>: LLMs have no intrinsic motivation, curiosity, or passion. They optimise for a mathematical loss function but do not “care” about the result or seek understanding for its own sake.</li></ul><h3>The Unified View — Toward Grounded Intelligence</h3><p>The “Unified View” is the core value proposition of this framework, providing a singular lens through which to understand all symbolic systems, whether they are silicon-based or biological. This model collapses the fragmented theories of linguistics, logic, and machine learning into a four-part operational cycle:</p><ol><li><strong>The Generator</strong>: A probabilistic grammar that produces expressions, code, or actions.</li><li><strong>The Interpreter</strong>: A system of probabilistic semantics that assigns meaning or infers intent.</li><li><strong>The Validator</strong>: The critical component that evaluates output against truth or utility.</li><li><strong>The Learner</strong>: The feedback mechanism that updates the model’s behavior based on the validator’s results.</li></ol><h4>The Hierarchy of Validators</h4><p>The sophistication of an intelligence is defined by the nature of its <strong>Validator</strong>. In formal domains, the validator is a rigid, deterministic “checker” like a compiler or a logical prover. Because validity in these systems is decidable and formal, LLMs excel at “neural proof search” — using probability to navigate a massive search space where verification is easy but generation is hard.</p><p>In the messy “real world,” the validator is society, human intention, or physical reality. For language to be truly grounded, it must be checked against a world model. While social consensus (through Reinforcement Learning from Human Feedback) serves as a proxy for LLMs, true intelligence requires a validator rooted in <strong>physics</strong>. To simulate this ultimate reality with high fidelity, researchers are looking toward <strong>quantum processes</strong>. A quantum-based validator could simulate reality much better than classical algorithms, providing a grounding mechanism that respects the probabilistic and fundamental nature of the physical world.</p><h4>Intelligence as a Grounded Loop</h4><p>Unlike LLMs, which operate in a disconnected cycle of text-prediction and optimisation, human intelligence is a <strong>closed loop</strong> grounded in the environment. Prediction leads to action, action produces physical feedback, and that feedback triggers the emotional signals (pain or pleasure) that drive learning.</p><p>To move beyond being “mirrors of humanity’s written record,” AI must bridge the gap between <strong>syntactic probability</strong> and <strong>grounded semantics</strong>. This transition requires embodiment and the integration of validators that interact with reality — whether through robotics or quantum-level simulations. By understanding this unified loop, we realize that the path to real intelligence is not just more data, but better grounding in the world that our language seeks to describe.</p><h3>Conclusion: The Mirror and the Mind</h3><p>Large language models are far more than “black boxes”; they are sophisticated estimators of the probabilistic grammar that forms the skeleton of human culture. Their success in formal domains like mathematics and coding stems from a <strong>distributional equivalence</strong>, where the model’s learned patterns effectively match the deterministic rules of the target language. However, as we have seen, this “brute-force” statistical acquisition is fundamentally different from the human path.</p><p>Human intelligence is defined by <strong>intuition</strong> — an internal simulation that allows us to generalize from a handful of examples rather than trillions of tokens. This process is fueled by a biological reward system of pain and pleasure that ensures we do not just predict, but <em>care </em>about the outcomes of our reasoning.</p><p>The future of artificial intelligence lies in the <strong>Unified View</strong>: closing the loop between generation and reality. By moving beyond text-based social consensus and toward validators grounded in physical reality — perhaps even utilizing <strong>quantum computing </strong>to simulate the world’s fundamental nature — we move toward a grounded semantics. In studying these machines, we are doing more than building tools; we are constructing a scientific mirror that reveals the profound mechanics of our own minds.</p><h3>References</h3><p>Bresnan et al. (2016). Lexical-Functional Syntax.<br>Chomsky, N. (1957). Syntactic Structures.<br>Clark, H. (1996). Using Language.<br>Cook, S. (1971). The complexity of theorem proving.<br>Daw et al. (2011). Model-based influences on learning.<br>Friston, K. (2010). The free-energy principle.<br>Goodfellow et al. (2016). Deep Learning.<br>Gödel, K. (1931). On formally undecidable propositions.<br>Hewitt &amp; Manning (2019). A Structural Probe for Finding Syntax in Word.<br>Hopcroft et al. (2006). Automata Theory.<br>Lake et al. (2017). Building machines that learn.<br>Linzen et al. (2016). Subject-verb agreement.<br>Manning &amp; Schütze (1999). Foundations of Statistical NLP.<br>Neyshabur et al. (2017). Generalization in deep learning.<br>Pearl, J. (2009). Causality.<br>Pinker, S. (1994). The Language Instinct.<br>Power et al. (2022). Grokking: Generalization Beyond Overfitting on Small.<br>Ryan &amp; Deci (2000). Intrinsic motivation.<br>Schultz et al. (1997). Neural reward prediction.<br>Shannon, C. (1951). Prediction and entropy.<br>Solomonoff, R. (1964). Inductive inference.<br>Tenenbaum et al. (2011). How to grow a mind.<br>Wei et al. (2022). Emergent Abilities of Large Language Models.<br>Tal et al. (2016). Assessing the Ability of LSTMs to Learn Syntax-Sensitive.</p><blockquote>Disclaimer</blockquote><blockquote>This article has been reviewed and corrected using Generative AI. The views and opinions expressed in this article are solely those of the author and do not necessarily reflect the official policy or position of the author’s company. Any content provided by the author is of their own opinion and is not intended to malign any organization, company, individual, or entity.</blockquote><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=5586d6f337ce" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Why Data Scientists Should Care About Software Engineering Best Practices]]></title>
            <link>https://medium.com/@romualdrousseau/why-data-scientists-should-care-about-software-engineering-best-practices-7b35b71e48af?source=rss-a810a6bf806c------2</link>
            <guid isPermaLink="false">https://medium.com/p/7b35b71e48af</guid>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[data-engineering]]></category>
            <category><![CDATA[software-engineering]]></category>
            <dc:creator><![CDATA[Romuald Rousseau]]></dc:creator>
            <pubDate>Wed, 26 Nov 2025 11:09:56 GMT</pubDate>
            <atom:updated>2025-11-26T14:57:07.844Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*2u4BYd5X2h_aFuUAxq_nFA.png" /></figure><h3>Introduction</h3><p>As data scientists, we obsess over models, metrics, and data pipelines. But when our work moves from the notebook to production, everything changes. At that moment, our code stops being an experiment — <strong>it becomes a contract.</strong></p><p>A contract with users, with engineers, with future maintainers, and with our future selves. To honor that contract, we need more than clever models. We need <strong>software engineering discipline</strong>: clean code, structured design, static analysis, and, above all, strategic <em>reuse</em> of proven components.</p><p>Surprisingly, the strongest argument for this comes from mathematics.</p><h3>Code Is a Mathematical Proof</h3><p>Programming languages are not just instructions for machines — they are <strong>formal languages</strong> built on the same logical foundations as mathematics. Any Turing-complete language has the same expressive power as formal calculable math.</p><p>That means your code <strong>is a proof</strong>:</p><ul><li>The logic in your functions = the reasoning</li><li>The tests = the verification</li><li>The runtime environment = the evaluator</li></ul><p>And just like math, you can write a proof that is elegant and predictable — or a proof that is confusing, fragile, and impossible to maintain.</p><p>Sound familiar?</p><h3>A Simple Mathematical Example That Reveals Everything</h3><p>Suppose we want to solve:</p><blockquote><em>If </em><strong><em>b + √b = 1</em></strong><em>, find </em><strong><em>b + 1/b</em></strong><em>.</em></blockquote><p>There are two ways to do this.</p><h3>1. The messy way</h3><p>Ignore the hypothesis. Expand everything. Introduce pages of algebra. This mirrors <strong>bad production code</strong>:</p><ul><li>ignoring constraints</li><li>reinventing wheels</li><li>excessive complexity</li><li>unreadable logic</li><li>untestable structure</li></ul><h3>2. The clean, structured, minimal proof</h3><p>Use the hypothesis directly:</p><ol><li>From b + √b = 1 → √b = 1 − b</li><li>Square both sides: b = (1 − b)²</li><li>Rearrange: <strong>b² − 3b + 1 = 0</strong></li><li>Let x = b + 1/b</li><li>Multiply both sides by b: b² + 1 = bx</li><li>Substitute using the polynomial (b² + 1 = 3b)</li><li>Solve → <strong>x = 3</strong></li></ol><p>Done.</p><p>It is the same problem and the same result — but two very different levels of effort and clarity.</p><p>The difference? One uses <strong>structure and reuse</strong>. The other tries to brute-force everything from scratch.</p><h3>Software Engineering Works Exactly the Same Way</h3><p>Mathematicians don’t reprove the foundations of all theories every time they need to prove something new. They rely on:</p><ul><li>established theorems</li><li>reusable patterns (inference, induction, contradiction, …)</li><li>known tools</li><li>standard reasoning techniques</li><li>That’s what keeps proofs manageable.</li></ul><p>In software, our “proof obligations” are called <strong>tests</strong>.</p><p>Every line of code incurs new proof cost. Every new function increases the testing surface. Every handcrafted implementation adds risk. Without good engineering practice, you end up with the “messy proof” version of your system.</p><h3>Clean Code Reduces Testing Requirements</h3><p>Good code isn’t about aesthetics — it’s about reducing verification cost.</p><p>Clean code has:</p><ul><li>smaller functions</li><li>fewer branches</li><li>clear naming</li><li>predictable structure</li><li>understandable control flow</li></ul><p>Each of those characteristics <strong>reduces the number of tests needed</strong>.</p><p>Bad code leads to: a combinatorial explosion of edge cases → massive testing burden.</p><p>Clean code leads to: simpler logic → significantly fewer test cases.</p><h3>Design Patterns Reduce Testing Requirements Even More</h3><p>Design patterns are “proven shapes” of code.</p><p>When you use patterns like:</p><ul><li>Strategy</li><li>Factory</li><li>Decorator</li><li>Observer</li><li>Pipeline (common in ML workflows)</li></ul><p>…you avoid inventing new architectures that would require heavy testing.</p><p>Using a known pattern means:</p><ul><li>you only test your small extension</li><li>you don’t need to prove the structure itself</li><li>the team already knows how the pattern behaves</li></ul><p>Patterns are the software equivalent of <strong>standard proof techniques</strong> in mathematics. They prevent chaos — and reduce the amount of proof you owe the system.</p><h3>Static Analysis = Domain Checking for Code</h3><p>Mathematicians check whether steps are even valid:<br> “Does this substitution make sense?”<br> “Is the function defined on this domain?”</p><p>In code, static tools do this for you:</p><ul><li>Python typing</li><li>linters</li><li>static analyzers</li><li>code formatters</li></ul><p>They catch:</p><ul><li>invalid domains</li><li>type mismatches</li><li>unreachable branches</li><li>logical contradictions</li></ul><p>This means fewer tests are needed, because many errors never reach runtime.</p><h3>Libraries &amp; Frameworks Reduce Tests the Most</h3><p>The greatest engineering principle of all:</p><blockquote><strong><em>Write less code.</em></strong></blockquote><p>Every line of code you don’t write is a line you don’t need to test.</p><p>Libraries have already been tested by thousands or millions of developers. Frameworks have been validated in production at scale.</p><p>Reusing them:</p><ul><li>lowers risk</li><li>reduces testing</li><li>increases reliability</li><li>speeds up delivery</li><li>minimizes maintenance</li></ul><p>Mathematicians don’t redefine integers before proving a discreet theorem. Software engineers shouldn’t reimplement JSON parsers or training loops.</p><p>Reuse is not laziness — <strong>it is strategic reduction of proof obligations.</strong></p><h3>Production Code Is Not a Playground</h3><p>Experimental code is for learning. Production code is for reliability.</p><p>In production:</p><ul><li>clever solutions are liabilities</li><li>minimalism is a virtue</li><li>predictability beats novelty</li><li>readability beats cleverness</li><li>maintainability beats personal preference</li></ul><p>Production code should be:</p><ul><li>boring</li><li>small</li><li>simple</li><li>proven</li><li>easy to reason about</li><li>easy to test</li><li>easy to onboard</li></ul><p><strong>If your code is surprising and clever and you learn a lot, it is probably wrong.</strong></p><h3>The Core Message: Good Practices Minimize Proof Obligations</h3><p>The biggest misunderstanding among data scientists is thinking:</p><blockquote>“Good software engineering takes more time.”</blockquote><p>No. Bad engineering takes more time — because it multiplies the amount of testing needed to make code safe. And if you don*t test, then it is bad over bad.</p><p>Good engineering:</p><ul><li>simplifies logic</li><li>reduces new surface area</li><li>encourages reuse</li><li>structures code predictably</li><li>minimizes bugs</li><li>minimizes the number of tests</li><li>makes the whole system easier to reason about</li></ul><p>When you use design patterns, clean code principles, static analysis, and existing libraries, you <strong>reduce your proof burden</strong>.</p><p>That is the key to truly production-grade data science.</p><h3>Final Thoughts</h3><p>As data scientists, adopting software engineering best practices isn’t optional anymore. Our work powers real systems, real products, and real decisions.</p><p>The difference between an experiment and production is not accuracy, not model choice, not data volume. It is <strong>software engineering discipline</strong>.</p><p>To make our code trustworthy, we must treat it with the same rigor mathematicians apply to proofs:</p><ul><li>clear structure</li><li>minimal assumptions</li><li>strategic reuse</li><li>validated reasoning</li><li>formal verification</li></ul><p>Good engineering practice is not decoration — it is the shortest path to reliable, production-ready data science.</p><p><strong>The less code you write, and the better structured it is, the less proof you need.</strong> And that is the ultimate productivity boost.</p><blockquote>⚠️Design patterns, clean code principles when becoming a dogma is the ultimate productivity flaw. See my article when practice becomes dogma:</blockquote><p><a href="https://medium.com/@romualdrousseau/rethinking-code-quality-measuring-what-truly-matters-e4c289058563">Rethinking Code Quality, Measuring What Truly Matters | by Romuald Rousseau | Jan, 2025 | Medium</a></p><blockquote>Disclaimer</blockquote><blockquote>This article has been reviewed and corrected using Generative AI. The views and opinions expressed in this article are solely those of the author and do not necessarily reflect the official policy or position of the author’s company. Any content provided by the author is of their own opinion and is not intended to malign any organization, company, individual, or entity.</blockquote><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=7b35b71e48af" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How Gen AI Transforms Scientific Workflows and Expands Knowledge]]></title>
            <link>https://medium.com/@romualdrousseau/how-gen-ai-transforms-scientific-workflows-and-expands-knowledge-0ce9ac94bbf0?source=rss-a810a6bf806c------2</link>
            <guid isPermaLink="false">https://medium.com/p/0ce9ac94bbf0</guid>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[science]]></category>
            <category><![CDATA[knowledge]]></category>
            <category><![CDATA[big-data]]></category>
            <category><![CDATA[genai]]></category>
            <dc:creator><![CDATA[Romuald Rousseau]]></dc:creator>
            <pubDate>Sat, 09 Aug 2025 17:17:30 GMT</pubDate>
            <atom:updated>2026-01-30T09:41:18.600Z</atom:updated>
            <content:encoded><![CDATA[<p><strong>How Gen AI Transforms Scientific Workflows and Expands Knowledge</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*EaMih7nJSzZy-HOX_7TK2Q.png" /></figure><p>Artificial Intelligence (AI) has become a buzzword in today’s technological landscape, often associated with concepts like consciousness and cleverness. However, the true nature of AI, specifically Generative AI (Gen AI), is far from possessing actual intelligence. Instead, Gen AI is a sophisticated tool — a vast calculator — that processes and combines data in ways that were previously unimaginable. Despite its lack of genuine intelligence, Gen AI stands as a groundbreaking invention that has revolutionized our ability to solve complex problems.</p><p>In a previous article, “<a href="https://medium.com/@romualdrousseau/ai-consciousness-or-cleverness-4c82eb5dd7b8">AI: Consciousness or Cleverness</a>”, we explore the distinction between true intelligence and the computational prowess of AI. We argue that AI, particularly Gen AI, operates not through conscious thought but through advanced algorithms and data processing techniques. This distinction is crucial in understanding the incredible potential of Gen AI, which lies in its ability to extend beyond traditional calculation capabilities and address problems with millions of unknown variables.</p><p>Historically, scientists could only tackle problems with a few unknown variables. With the advent of big data, machine learning and Gen AI, they can now address challenges involving millions of unknowns, providing approximate solutions that propel scientific inquiry further. Gen AI reshapes scientific workflows by calculating possibilities within a known dataset’s latent space, enabling rapid discovery and theory development.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*gZzb7oY_oFe2AkF4vdZhJA.png" /><figcaption>Each discovery enhances the generative AI, enabling fast new discovery</figcaption></figure><p>To understand how Gen AI achieves this, it’s essential to delve into the workings of transformers and the attention mechanism. Transformers are a type of neural network architecture that revolutionized natural language processing. They rely on the attention mechanism to handle vast amounts of data and identify relevant information efficiently. The attention mechanism allows the model to focus on specific parts of the input data, weighing the importance of each piece of information differently. This selective focus enables the model to combine knowledge in a coherent manner.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/850/1*CA_JNgGv0eRR8WQ_m1EO0w.png" /><figcaption>Transformer decoder-only, the heart of ChatGPT</figcaption></figure><p>Transformers encode “meaning” through thousands of dimensions, encompassing aspects like colors, shapes, sizes, and more. This encoding process effectively captures the distributions of how these meanings interact with each other. Simple linear algebra operations on these encoded vectors can then build new coherent outputs. By understanding and manipulating these high-dimensional vectors, transformers can generate results that are not only accurate but also contextually relevant and coherent.</p><p>Consider merging several images. A simple blend of colors yields a meaningless pixel mess.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/940/1*bwmdLriQKFuw503jD_GDFg.png" /><figcaption>Mixing images by their color intensities, a mess of pixel</figcaption></figure><p>However, encoding these images’ pixel interactions (embedding) and merging their distributions encoded by transformers (latent space) results in a coherent image, combining the original shapes.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/940/1*GFZoCFtpOnTx5Vl1CZQ5jQ.png" /><figcaption>Mixing images by their latent space, a miracle (Rendering by ChatGPT)</figcaption></figure><p>This principle applies to text, video, and sound, showcasing Gen AI’s ability to produce coherent outputs through linear algebra, extensive data, and computational power. The attention mechanism plays a crucial role here by allowing the model to determine which parts of the input data are most relevant to the task at hand, ensuring that the output maintains coherence and relevance.</p><p>The true marvel lies in Gen AI’s capability to generate new data from existing knowledge while maintaining structured coherence. For instance, combining several mathematical proofs using Gen AI can produce new proofs that respect mathematical logic and grammar, potentially leading to valid discoveries. This mirrors human historical advancements, discovery built upon prior knowledge, but achieved in mere seconds. The attention mechanism ensures that the model can synthesize these proofs by focusing on the most pertinent aspects of each input proof, thus creating a coherent and potentially correct new proof.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*dZbT4n2g7-t_y6l7SbWqjQ.png" /><figcaption>Prompt guides the transformers to generate an answer based on its knowledge</figcaption></figure><p>To journey into the latent or knowledge space, we begin with the prompt. The prompt is encoded — embedded — just like the data from the training set, then passed through the transformers, becoming a single point in the vast, multidimensional space. From this point, the model computes an answer by drawing on information that the attention mechanism deems relevant — those points that are “close” or similar in this abstract space.</p><p>In this sense, a prompt is not just a question; it’s a way to navigate and explore the knowledge itself. And the quality of the answer is tightly bound to the quality of the prompt: a well-formed question carves a sharper path toward meaningful responses.</p><p>This latent space is enormous — thousands of dimensions, billions of data points — for example, consider a model built on 1000 dimensions and 32-bit floating-point numbers in each dimension. That means its theoretical capacity is about 2³²⁰⁰⁰ unique coordinates. Within such an astronomical space, even billions of data points are sparse. Imagine that each document on the internet were a single point in this space — most would still be separated by distances so vast they could be compared to entire universes apart.</p><blockquote><em>⚠️B</em>ased on our intuitive imagery, hallucination are kind of black hole, a part of the knowledge space without knowledge. When the prompt arrives close enough, it may jump to a very distant star, loosing the thread of the meaning of the answer.</blockquote><p>Now consider what happens when we add more knowledge: the potential for discovery expands beyond comprehension.</p><p>In conclusion, Gen AI is an extraordinary tool that amalgamates knowledge, akin to performing arithmetic, accelerate scientific workflows, and expand of human discoveries at an unprecedented rate.</p><blockquote>Disclaimer</blockquote><blockquote>This article has been reviewed and corrected using Generative AI. The views and opinions expressed in this article are solely those of the author and do not necessarily reflect the official policy or position of the author’s company. Any content provided by the author is of their own opinion and is not intended to malign any organization, company, individual, or entity.</blockquote><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=0ce9ac94bbf0" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Rethinking Code Quality, Measuring What Truly Matters]]></title>
            <link>https://medium.com/@romualdrousseau/rethinking-code-quality-measuring-what-truly-matters-e4c289058563?source=rss-a810a6bf806c------2</link>
            <guid isPermaLink="false">https://medium.com/p/e4c289058563</guid>
            <category><![CDATA[tech-lead]]></category>
            <category><![CDATA[code-quality]]></category>
            <category><![CDATA[data-factory]]></category>
            <category><![CDATA[management]]></category>
            <category><![CDATA[code]]></category>
            <dc:creator><![CDATA[Romuald Rousseau]]></dc:creator>
            <pubDate>Fri, 23 May 2025 15:07:23 GMT</pubDate>
            <atom:updated>2025-08-11T09:56:58.071Z</atom:updated>
            <content:encoded><![CDATA[<p>This article was first written in January 2025.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*WhPe_vjazMM_byUZImYdFQ.png" /></figure><p>Software teams spend a lot of time discussing code quality — how to achieve it, maintain it, and measure it. But what if the very idea of <em>measuring</em> code quality is misguided? What if quality is something that can only be inferred, not directly quantified?</p><p>In this article, I argue that trying to measure code quality directly is vain. Cyclomatic complexity, code coverage, linting scores, or lines of code — these metrics attempt to pin down a concept that is ultimately contextual, subjective, and dynamic. Instead of measuring code itself, we should be measuring the <em>impact</em> of code. Code quality, in essence, is visible in its consequences.</p><h3>The Mirage of Direct Measurement</h3><p>Code quality is often treated like a scientific property — something you can assign a number to. Teams obsess over style guides, automated linters, test coverage thresholds, and static analysis reports. These tools do serve a purpose, but they provide <em>indicators</em>, not answers.</p><p>High test coverage doesn’t guarantee that your tests are meaningful. A codebase can be “clean” yet brittle, hard to extend, or confusing to new developers. You might have 100% compliance with your linter, but your product still crashes in production or takes months to update.</p><p>In other words, <strong>code quality is not in the code itself — it’s in how the code behaves and how it enables (or blocks) people</strong>.</p><h3>Measuring the Consequences, Not the Code</h3><p>If we want to understand whether a piece of code is “good,” we need to look at what happens <em>around </em>that code. Here are some more reliable indicators:</p><p>- <strong>Developer Satisfaction</strong>: Do developers enjoy working in this codebase, or do they dread touching it? High-quality code tends to reduce friction and cognitive load.<br> <br>- <strong>End-User Feedback</strong>: Are users encountering bugs? Is the application reliable and fast? Good code quietly supports a great user experience.<br> <br>- <strong>Incident Frequency After Deployment</strong>: How often do bugs appear right after code is released? Fewer incidents often signal a better-structured, better-tested system.<br> <br>- <strong>Merge Request Activity</strong>: Are small changes disproportionately complex or time-consuming? Is code review friction high? That could point to low-quality or overly complex code.<br> <br>- <strong>Maintenance Burden</strong>: How much effort does it take to fix bugs or add features? Code that’s easy to evolve is a sign of underlying quality.<br> <br>- <strong>Onboarding Time</strong>: How quickly can new developers become productive? Well-structured codebases usually shorten onboarding curves.</p><p>All of these are <em>consequences </em>of good (or poor) code quality. By measuring them, you indirectly measure the thing that’s otherwise hard to grasp.</p><h3>Contextualizing Code Quality</h3><p>It’s also important to remember that code quality must be evaluated in context. What makes code “high quality” in one domain may be excessive or wasteful in another.</p><p>- In a <strong>medical device</strong>, code must be verifiable, traceable, and extremely robust. The cost of failure is high, so quality must be uncompromising.<br> <br>- In a <strong>startup MVP</strong>, speed to market matters more than elegance. Overengineering is a bigger risk than bugs.<br> <br>- In a<strong> government system</strong>, maintainability over decades might outweigh short-term performance.<br> <br>Trying to apply the same quality standards across these contexts is not just misguided — it’s counterproductive. <strong>Code quality must align with cost, business value, and human impact.</strong></p><h3>The Role of Non-Regression Testing</h3><p>Testing doesn’t measure quality directly — it measures behavior. But that’s still immensely valuable.</p><p>Automated tests — especially regression tests — help you ensure that your code continues to behave as expected. They serve as a safety net, alerting you to when quality has degraded. In this way, tests are not a measure of quality per se, but a <em>detector of its absence</em>.</p><p>Regression tests are especially useful in highlighting unintended consequences of change — often one of the clearest signs of fragile or poorly structured code.</p><p>As quality, tests may be complex to implement and therefore expensive. A target of full test coverage is also a mirage that may spin teams in end less out of reach satisfaction. Unless dictated by strong economical, legal or ethical constraints, 100% test coverage is just over doing it.</p><h3>Toward a Practical Framework</h3><p>Here’s a simple checklist you can use to assess the real-world impact of your code:</p><p>- Are developers productive and satisfied?<br> <br>- Is the codebase easy to modify, extend and fix?<br> <br>- Do deployments go smoothly, with few incidents?<br> <br>- Are users experiencing reliability and performance?<br> <br>- Is it easy to onboard new team members?<br> <br>- Is the cost of maintaining the system predictable and acceptable?</p><p>This framework doesn’t rely on any one number or metric. Instead, it encourages teams to regularly reflect on how their code behaves — and how it <em>feels</em> to work with. Practically, I used the following matrix to gather insight of the quality of code of a project for each of my team on a monthly basis:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/798/1*CffqP5J1-pxOtqze2_ZARA.png" /></figure><p>I collect the status of satisfaction of customers and developers, and follow how the project move from case to case to build a remediation plan with my tech leads. Easy, more realistic and useful than any complex metrics.</p><h3>Conclusion: Impact Over Ideals</h3><p>There’s nothing wrong with writing clean, elegant, well-tested code. But let’s stop pretending we can measure code quality in isolation. Good code is valuable not because of how it looks, but because of what it <em>enables </em>— fast iteration, happy users, low incident rates, and confident developers.</p><p>In the end, code quality is not an abstract ideal. It is a practical, observable reality. Not in the code itself, but in its consequences.</p><blockquote><strong>Disclaimer</strong></blockquote><blockquote>This article has been reviewed and corrected using Generative AI. The views and opinions expressed in this article are solely those of the author and do not necessarily reflect the official policy or position of the author’s company. Any content provided by the author is of their own opinion and is not intended to malign any organization, company, individual, or entity.</blockquote><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=e4c289058563" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Revolutionizing Data Management]]></title>
            <link>https://medium.com/@romualdrousseau/revolutionizing-data-management-56b50f09f2a8?source=rss-a810a6bf806c------2</link>
            <guid isPermaLink="false">https://medium.com/p/56b50f09f2a8</guid>
            <category><![CDATA[data-engineering]]></category>
            <category><![CDATA[semi-structured-data]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[excel]]></category>
            <category><![CDATA[data-ingestion]]></category>
            <dc:creator><![CDATA[Romuald Rousseau]]></dc:creator>
            <pubDate>Fri, 23 May 2025 14:56:12 GMT</pubDate>
            <atom:updated>2025-08-11T09:57:22.513Z</atom:updated>
            <content:encoded><![CDATA[<p>This article was first written in October 2024.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*l6L6zyz-6PFIqHoGtCuVaA.png" /></figure><p>In today’s data-driven landscape, navigating the complexities of semi-structured documents (namely Excel files) poses a significant challenge for organizations. These documents, characterized by diverse layouts and a lack of standardization, often require specialized skills for effective extraction and analysis. However, we propose a novel java framework, <strong>Archery</strong>, to address this challenge. By leveraging innovative algorithms and machine learning techniques, this framework offers a solution that provide enhanced accessibility to users across diverse skill levels. Moreover, by automating the extraction process, it not only saves time but also minimizes errors, particularly beneficial for industries dealing with large volumes of such documents. Crucially, this framework integrates seamlessly with machine learning workflows, unlocking new possibilities for data enrichment and predictive modeling. It offers a scalable and efficient means of managing semi-structured data, thereby expanding the toolkit of data services available to organizations.</p><h3>Definitions and examples</h3><p>Semi-structured documents have the characteristic of containing some type of information known a priori, but which can change the position and format within the document itself. In addition, semi-structured documents also vary a lot in terms of layout and design. Some documents have a fixed set of data but no fixed format for this data. In some documents, the date appears on the top right corner, in another variation, it is at the center of the document, and in yet another, you’ll find it in the bottom left corner. Another added complication is that the same data is qualified by different names. In one variation, a field may be called ‘Purchase Order Number’, in another — ‘PO Number’, and a few others may call it “PO #”, “PO No.” or “Order Number’. These variations are endless and because of these two challenges, you cannot use a template-based solution for these documents.</p><p>Examples:</p><ol><li>Examples of semi-structured document include emails, XML files, JSON files, social media posts, and log files. These types of data contain both structured and unstructured information, such as a mixture of predefined fields &amp; free-form text.</li><li>Another classic example is the myriad of Excel files found within companies holding often precious information:</li></ol><figure><img alt="" src="https://cdn-images-1.medium.com/max/431/1*lpTx3zn7Fid11uEnAq8NMg.png" /><figcaption>Excel Spreadsheet</figcaption></figure><h3>What is the problems we are trying to fix?</h3><p>Semi-structured documents are usually noisy and their layout changes over time. They contain defects, usually invisible by the end user who read the document but provide a challenge to any automated process. They, also, contain tabular data but may be completed by unstructured text around. For examples, the date of the document or a product may be in the title and not in the tabular representation; simply extracting the table will miss some implicit information. These specifies stop today tool to consistently load such documents.</p><p>The first consequence is a high running cost due to maintaining the pipeline of extraction every time the layout changed, due to the custom code required that need to be modified all the time.</p><p>The second consequence is the difficulty to have a self-service to extract the data because Extract Load tools require often engineering skills (even for no/low code solution on the long run).</p><h3>The solution; a framework to manipulate semi-structured document and transform them into consistent tabular output</h3><p><strong>Addressing Data Complexity:</strong> Semi-structured documents present a significant challenge due to their varied formats and lack of standardization. By developing algorithms and employing machine learning techniques, our framework can effectively handle this complexity without relying on manual coding.</p><p><strong>Enhanced Accessibility:</strong> Traditional solutions for working with semi-structured documents often require skilled engineers or developers. Our framework, however, eliminates this barrier by providing a user-friendly interface that doesn’t require coding expertise. This makes it more accessible to a wider range of users within organizations.</p><p><strong>Increased Efficiency: </strong>Automating the process of extracting data from semi-structured documents not only saves time but also reduces the likelihood of errors that can occur with manual intervention. This efficiency gains can be particularly valuable in industries where large volumes of such documents need to be processed regularly.</p><p><strong>Integration with Machine Learning:</strong> By incorporating machine learning capabilities into the framework, you’re not only extracting data but also enriching it with insights derived from advanced analytics. This opens up new possibilities for leveraging semi-structured data in machine learning workflows, enhancing decision-making and predictive modeling.</p><p><strong>Expansion of Data as a Service:</strong> Our framework aligns with the concept of data as a service by providing a scalable and efficient means of managing semi-structured data. This expands the toolbox of data services available to organizations, enabling them to harness valuable insights from previously untapped sources.</p><h3>A global overview of the framework</h3><p>The framework is actually written in Java and use several frameworks such as Tensorflow, Jython, etc … The framework loads and parses the document through different steps to extract and structure the data.</p><p>The basic steps are:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/941/1*5z0fI4KOCXZXV1ZOihZfDA.png" /><figcaption>Framework Main Steps</figcaption></figure><h4>Load by Chunks</h4><p>The framework is able to load various format such as Excel, CSV, PDF, HTML … A plugin system allows to add new loader easily. Loaders stage the data into lists of chunks to allow processing of huge amount of data and minimize memory footprint between steps.</p><h4>Noise reduction</h4><p>Noise reduction consist to remove defects or visual artifacts such as:</p><p>- Empty cells used as row or column separator<br>- Empty cells inserted wrongly due to manual manipulations<br>- Merged cells<br>- …</p><p>The noise reduction must be resilient to layout changes so simple removal of the nth column or row is not allowed. The framework includes the possibility to do small transformations using recipes. A recipe is a small Jython script that can access to the document model and manipulate it and perform basic cleansing.</p><h4>Feature Extraction</h4><p>Feature extraction will parse the cleaned document, detect tables and other meta data. It will build a tree of the different features of the document linked by their relation of importance.</p><p>The relation of importance is based on the proximity of the elements to each other based on their reading direction. The human eye reads a document using a certain direction; from top to bottom and left to right for English or right to left depending in Arab. As such, when a human creates a document, he is influenced by this reading direction because he supposes his future reader to look at the document with a certain way. It means the various elements of the document will follow the reading direction and therefore can be linked to each other along this direction. You can find the detailed explanation in the white paper [Semi-structured document feature extraction](<a href="https://romualdrousseau.github.io/Archery/how_it_works/">https://romualdrousseau.github.io/Archery/how_it_works/</a>).</p><h4>Table Layout Parsing</h4><p>Feature extraction is not enough to structured the document, each table needs a deep analysis. The framework allows to parse the table to detect headers, sub headers, sub footers, footers, pivoted columns … A plugin system allows to add new parser into the framework.</p><p>We will study an original approach to detect and extract the different parts of a table using a pattern matching algorithm similar to regular expressions. You can find the detailed explanation in the white paper <a href="https://romualdrousseau.github.io/Archery/how_it_works/">Table Layout Regular Expression</a>.</p><h4>Tabular Assembling</h4><p>This step will take all structures generated above and merge them into a single tabular structure. This structure is very easy to store in a database or a flat file such as parquet. You can find the detailed explanation in the white paper <a href="https://romualdrousseau.github.io/Archery/how_it_works/">Semi-structured document feature extraction</a> as we join a general algorithm to transform a semi-structured document to a tabular output.</p><h4>Tagging</h4><p>The last step will annotate each column of the tabular output a tag. Again a plugin system allows to add new tagger. This tag can be used as the name of the column when storing in a database.</p><p>We will present in a tagger based on word embedding and a neural network to classify each data with an unique tag.</p><h3>Detailed Algorithm</h3><p>All these steps are summarized in the graph below:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*mbU_zjERt3pJC4NlOKcLxw.png" /><figcaption>Detailed Algorithm</figcaption></figure><h3>Conclusion</h3><p>Archery is a powerful to tool to extract data from semi-structured documents in which Excel files thrones within companies. It allows Data Factories to finally ingest those data without spending too much resources. All source codes are available on our <a href="https://github.com/RomualdRousseau/Archery">github</a>.</p><blockquote><strong>Disclaimer</strong></blockquote><blockquote>This article has been reviewed and corrected using Generative AI. The views and opinions expressed in this article are solely those of the author and do not necessarily reflect the official policy or position of the author’s company. Any content provided by the author is of their own opinion and is not intended to malign any organization, company, individual, or entity.</blockquote><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=56b50f09f2a8" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[AI , Consciousness or Cleverness through Computational Theory]]></title>
            <link>https://medium.com/@romualdrousseau/ai-consciousness-or-cleverness-4c82eb5dd7b8?source=rss-a810a6bf806c------2</link>
            <guid isPermaLink="false">https://medium.com/p/4c82eb5dd7b8</guid>
            <category><![CDATA[computability]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[computer-science]]></category>
            <category><![CDATA[data]]></category>
            <dc:creator><![CDATA[Romuald Rousseau]]></dc:creator>
            <pubDate>Fri, 23 May 2025 13:53:36 GMT</pubDate>
            <atom:updated>2026-01-30T10:56:34.022Z</atom:updated>
            <content:encoded><![CDATA[<h3><strong>AI , Consciousness or Cleverness through Computational Theory</strong></h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*eYcmhPIHM5MmZsFJVz_m1w.png" /></figure><p>In a thought-provoking interview, Nobel Prize-winning mathematical physicist Sir Roger Penrose delves into topics such as artificial intelligence, Gödel’s incompleteness theorems, computability, and why — despite the current excitement surrounding AI — he remains convinced that machines will never achieve true consciousness:</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FbiUfMZ2dts8%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DbiUfMZ2dts8&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FbiUfMZ2dts8%2Fhqdefault.jpg&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/a3914004adec873cafdffd6ecb15e781/href">https://medium.com/media/a3914004adec873cafdffd6ecb15e781/href</a></iframe><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2Fe9484gNpFF8%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3De9484gNpFF8&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2Fe9484gNpFF8%2Fhqdefault.jpg&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/86230b6e34417a34ae248da11fe4cf01/href">https://medium.com/media/86230b6e34417a34ae248da11fe4cf01/href</a></iframe><p>In this article, I want to reflect on the science and mathematics behind Penrose’s arguments — why he may be right, and how his perspective is perhaps shaped by skepticism toward today’s AI hype. At the same time, I’ll explore how, even without consciousness, AI remains a remarkable human achievement — one that opens doors to new frontiers of computation and discovery.</p><p>I begin by unpacking the key concepts mentioned by Penrose, clarifying their meaning and relevance. I won’t go into the mathematical proofs but indeed in the domain of intuition, but you can find more formal approaches in the references of this article. Then, I assess where AI stands today: not a conscious entity, but an extraordinarily powerful tool that continues to stretch the boundaries of what is computationally possible.</p><h3>The science of computability</h3><p>Let us be unequivocal: contemporary computers, modeled as Deterministic Turing Machines, execute calculable and deterministic processes, including AI algorithms. Problems that are non-computable or non-deterministic lie beyond the capability of modern computing machinery, or by impossibility or by resources unavailability (time, space, energy). Alan Turing’s thesis on computability, alongside the lambda calculus theory (Church) and computational complexity theory, describes these principles.<br>Determinism does not imply human computational simplicity; rather, it implies computational feasibility by machines on reasonable resources. Just as calculating the millionth decimal place of π is a deterministic yet non-trivial task for humans, so too are many AI algorithms, which entail extensive computations involving millions of iterations and parameters.</p><p>In any formal system, there are things that are provable, decidable, or computable — and others that are not. Turing proved that if it is computable then a Turing machine can solve it in a finite number of steps.</p><p>There are two theoretical models of Turing machines:</p><p>- <strong>Deterministic Turing Machines (DTMs)</strong>: For a given input, they always follow the same sequence of steps and produce the same result.<br> <br>- <strong>Non-deterministic Turing Machines (NTMs)</strong>: They can explore multiple possibilities simultaneously, as if choosing the correct path leading to the solution. Even for the same input, can exhibit different behaviors on different runs, NTM is considered to perform correctly when, for each input, <em>there exists</em> a run that produces the desired result, even when other runs produce incorrect results or even no results.<br> <br>Any NTM can be simulated by a DTM, but it may take significantly more time and resources (space, time, energy …). Today’s computers are DTMs and can compute everything that is theoretically computable — just not always efficiently.</p><p>NTMs are purely theoretical; no physical device (CPU, GPU) currently implements them. Someone can argue, with truth,that distributed systems or GPU have a non determinism due to parallelization. And I will explicit my thoughts in the next paragraphs. Moreover, they are useful for classifying problems and understanding the limits of computation.</p><p>In practice, we care about computations that finish in a <strong>reasonable time</strong>. So we study which problems have efficient deterministic solutions and which are only non-deterministically solvable (and thus likely very slow on real machines).</p><p>Two practical approaches to speeding up computation are:</p><ol><li><strong>Parallelism</strong>: Running multiple computations at once. It can speed things up but still requires waiting for all computations to finish. Even there is a non deterministic nature in the order of execution, it often requires a mechanism of synchronization to get the final result (hence the deterministic nature). Why? Because the non determinism of a race condition for example has often as a negative impact of the result, and we want to avoid it.</li><li><strong>Probabilistic/Heuristic methods</strong>: Instead of trying all possibilities, focus on the most likely ones using techniques like Bayesian inference, Markov chains, or neural networks. This speeds up computation but introduces the risk of missing the optimal solution or find the global maxima. Again this probabilistic nature does not always equate to non-determinism. On the contrary, implemented probabilistic approaches transform an open-ended and/or non-deterministic calculation into a deterministic one.</li><li>Crucially, <strong>parallelism</strong> and <strong>probability</strong> are not the same as <strong>non-determinism</strong>, though if left uncontrolled, they can lead to non-deterministic behavior. In computer science, we tend to avoid such situations because non-determinism is often seen as undesirable outcomes. However, physical hardware introduces its own sources of entropy, and in edge cases, we may still observe traces of non-determinism even in a deterministic setup. Despite this, there are practical strategies within deterministic systems to manage complex computations under time or resource constraints.</li></ol><p>A “reasonable” amount of computation time is defined as a number of steps that is <strong>polynomial</strong> in the size of the input. This notion connects to the difference between <strong>deterministic</strong> and <strong>non-deterministic</strong> methods: if a problem can be solved in polynomial space/time by a non-deterministic method, it typically requires <strong>non-polynomial</strong> (certainly exponential) space/time by a deterministic one.</p><p>For any computational problem, we can distinguish two phases:</p><p>1. <strong>Finding a solution</strong><br> <br>2. <strong>Verifying the solution</strong><br> <br>Each phase can be either deterministic or non-deterministic and may take polynomial or non-polynomial space/time.</p><p>From this, several important complexity classes are defined:</p><p>- <strong>P</strong>: Problems that can be both <strong>solved and verified deterministically</strong> in polynomial space/time. <strong>Example: Shortest Path in a Graph (e.g., Dijkstra’s Algorithm)</strong><br> <br>- <strong>NP</strong>: Problems that can be <strong>solved non-deterministically</strong> in polynomial space/time but whose <strong>solutions can be verified deterministically</strong> in polynomial space/time. <strong>Example: Boolean Satisfiability Problem (SAT)</strong><br> <br>- <strong>BPP</strong>: Problems that can be <strong>approximately solved with a probabilistic approach</strong> in polynomial space/time including RP and co-RP classes). However BPP class does not cover all of what is computable. In fact, it may be that BPP = P, suggesting that probabilistic approaches are no more powerful than deterministic ones. De-randomization work (e.g., A. Wigderson, Turing Prize 2023) adds a strong support toward this view. If true, it also implies that simpler and more resource-efficient algorithms could achieve the same results. <strong>Example: Probabilistic Inference in Bayesian Networks</strong><br> <br>- <strong>BQP</strong>: Problems that can be both <strong>solved and verified deterministically</strong> in polynomial space/time by a quantum machine. <strong>Example: Integer Factorization (e.g., Shor’s Algorithm)</strong><br> <br>- <strong>NP-Hard</strong>: The hardest problems in NP (NP-Complete), or problems for which both the solution and verification require <strong>non-deterministic, non-polynomial</strong> space/time or are undecidable. <strong>Example: Halting Problem (Undecidable version) or Traveling Salesman Problem (TSP) (optimization version)</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/591/1*_nZgvPEvgiRk621DP6pCAg.png" /><figcaption>Vern Diagram of computability classes</figcaption></figure><p>An algorithm in the computable space operates through a finite number of states before halting and producing a result. Remarkably, even a small number of states can encode an enormous variety of algorithms. The <strong>Busy Beaver</strong> is a thought experiment introduced by Tibor Radó in 1962, which explores this idea. It refers to the Turing machine that, given <em>n</em> states, performs the <strong>maximum number of steps</strong> or prints the <strong>most non-blank symbols</strong> before halting. The resulting function, denoted <strong>Σ(n)</strong>, grows faster than any computable function — making it <strong>non-computable</strong>. In other words, there’s no general algorithm that can determine its values for arbitrary <em>n</em>. The Busy Beaver function reveals deep limits of computation and is closely related to the <strong>halting problem</strong>.</p><p>For example, we currently know that <strong>Σ(5) ≥ 47,176,870</strong>. Some speculate that breaking certain cryptographic keys may require algorithms with complexity approaching <strong>Σ(13)</strong>. In contrast, <strong>BPP</strong> (probabilistic) and <strong>BQP</strong> (quantum) approaches may yield results in significantly fewer steps, highlighting their practical power despite theoretical limits. Ultimately, the Busy Beaver illustrates how even a slight increase in computational states can unlock a vast, unexplored space of algorithms — possibly leading to breakthroughs and technologies we have yet to imagine.</p><p>In essence, AI algorithms pertain to computable problems, with closed-form solutions often residing in the NP class and probabilistic solutions within the BPP (or RP) class (when using weights randomization and/or drop out).</p><h3>The Nature of AI Algorithms</h3><p>AI algorithms are vastly inspired by biology and the way living beings solve problems. Neural networks are inspired by the way neurons seem to work; computer neurons are organized by mimicking biological neurons connected through their dendrites. The signal passing through a neuron is activated based on certain conditions. However, the similarities largely end there. In reality, computer neurons perform a matrix multiplication (a linear transformation) followed by a non-linear activation function. We still do not fully understand how biological neurons operate.</p><p>A class of AI algorithms, machine learning, draws inspiration from how living beings learn and solve problems. Such algorithms use a probabilistic approach to computation and require setting up probabilistic distributions to run. The training phase involves configuring these distributions based on data samples. Machine learning builds relationships between objects by encoding the underlying probabilistic distributions of these relationships, often with considerable success. Those underlying distribution can be built from sample data (Supervised Learning), from a mathematical model (Unsupervised Learning), or from a interactive decision process and goal setting (Reinforcement learning). Generative AI is a subdomain of such learning algorithms.</p><p>Biology inspires computer science, and this inspiration enables us to extend our computational capabilities into domains previously out of reach. But generative AI, for instance, is still, fundamentally, a very powerful and clever calculator.</p><p>With regard to AI algorithms, we must delineate the core phases: fitting, prediction, and verification. The first two phases — fitting and prediction — fall unequivocally within the P or BPP classes, if we take into account the randomization of initial weights, drop outs, etc … These tasks are computationally feasible (e.g., via linearization or gradient descent techniques), straightforward to validate.</p><p>The verification phase, however, poses a distinct challenge, differentiating AI product development from other software engineering endeavors. Formal verification of AI models is profoundly complex, possibly undecidable, non computable. As a result, empirical approaches are employed despite their limitations — for example, verification through a validation set evaluated by a human operator.</p><p>Today, Generative AI (GenAI) dominates tech discussions. For some, it heralds the rise of machines — fascinating and terrifying. GenAI is indeed a remarkable invention solving difficult problems, yet it remains fundamentally rooted in calculable mathematics, grounded in fundamental linear algebra and, in most cases, does not reflect true intelligence.</p><p>In fact, the problems GenAI solves are problems we <em>assumed</em> required intelligence. In reality, they just require clever computation. What GenAI can solve says more about the mechanics of a problem than the nature of intelligence.</p><p>GenAI can be used to solve numerous problems, such as summarizing text corpora, finding information within text from a query, and translating languages. However, these can all be reduced to a single core task: Given a corpus of text and a query, what is the probable next word of the answer?</p><p>GenAI is an incredible invention with many discoveries:</p><ul><li><strong>Embedding</strong>: Embedding encodes the distribution of certain characteristics of a word in a multi-dimensional space called a latent space. Earlier methods attempted to encode language using rules, but these approaches were ambiguous and labor-intensive. Embeddings encode word contexts, solving many of those issues.<br> <br>- <strong>Self-Attention Mechanism</strong>: A neural network technique, particularly in NLP, that allows models to focus on the most relevant parts of the input by assigning different weights to different input elements. It significantly improves performance in tasks like translation and summarization.<br> <br>- <strong>Transformers</strong>: A deep learning architecture introduced in the 2017 paper <em>”Attention is All You Need”</em>. Transformers use self-attention mechanisms to process sequential data more efficiently than RNNs or LSTMs, using linear algebra to capture probabilistic word relationships.<br> <br>All this demonstrates that the problem GenAI aims to solve is computable. While GenAI can uncover relationships in its latent space that humans might never see, its knowledge is bounded by its training data and embedding dimensions. It cannot truly create; all it does is recombine pre-existing information. This is powerful but limited. Human consciousness and instinct are still required to ask the meaningful questions.</li></ul><h3>The Nature of Consciousness</h3><p>Sir Roger Penrose argues that consciousness is an NP-hard, undecidable problem — fundamentally tied to self-awareness and akin to the halting problem in computability theory. On this point, we may have to agree to disagree. Some may argue that human consciousness is ultimately computable, while Penrose places it beyond algorithmic processes, possibly in higher dimensions or governed by as-yet-unknown physical laws.</p><p>There’s a touch of irony here. Some might see Sir Roger Penrose as a humanist, accused of placing too much faith in human consciousness — a view that seems more philosophical than mathematical. Yet, many of those same critics, who are often shaped by fiction, cinematic fantasies, and philosophical musings, dismiss Penrose’s ideas as provocative or even outdated. They find his views dramatic, while expressing their own critiques with no less drama. When Penrose stays a mathematician.</p><p>However, there is little debate about where AI stands: firmly within the bounds of computation. The real question, then, is not whether AI can eventually achieve consciousness or reach a technological singularity, but rather how — and where — human consciousness itself should be classified.</p><h3>The Nature of Human Intelligence</h3><p>The human brain operates at around 100Hz and runs on the energy of a few meals — an astonishingly efficient system. In contrast, GenAI systems require massive computational power, enormous datasets, and significant energy to function. The human intelligence is defined more by its constraints than by its capabilities.</p><p>Yet, despite their scale and speed, GenAI lacks consciousness and instinct — the capacity for self-awareness, intuition, and the creation of entirely new conceptual spaces. These qualities remain not only beyond current AI but perhaps beyond the scope of computation itself. While humans can transcend known boundaries and invent the unknown, AI remains confined within its training data, models, and latent spaces. This highlights a fundamental distinction: our creativity and insight are not just computational — they are uniquely human.</p><p>That said, as an approximation of cleverness, AI and GenAI are immensely powerful. For many tasks, they offer practical, even extraordinary, solutions. But they do so without understanding, meaning, or intentionality. Does it matter? Scientifically and Socially yes, practically no. They can do what few years ago was deemed impossible; run experiments with millions of variables, giving tools to compute into large data sets, to extract relations in complex behaviors that if closed forms exist, are nearly impossible to compute by human hands.</p><blockquote>What about quantum computing? It undoubtedly expands the realm of what is computationally feasible. But while quantum mechanics incorporates intrinsic non-determinism, quantum algorithms are ultimately engineered to produce deterministic, usable results. They may accelerate problem-solving, but they do not rewrite the foundations of computation theory or alter the essential nature of AI.</blockquote><h3>Conclusion</h3><p>Without a rigorous and agreed-upon definition of human consciousness or intelligence, any comparison with artificial intelligence remains incomplete. What we do know is this: AI is fundamentally computation. At present, there is no evidence to reliably equate machine processes with human intelligence. That isn’t to say the connection cannot exist — but rather that any relationship between them remains unknown.</p><p>I find myself aligned with Sir Roger Penrose: consciousness — and, by extension, human intelligence — may ultimately lie beyond the reach of machines, or completely unrelated, at least until a more precise definition emerges.</p><p>Yet, it is also undeniable that AI continues to expand the frontiers of what can be computed. In doing so, it opens doors to new discoveries, technologies, and ways of thinking. Its brilliance may be mechanical rather than conscious — but it remains, in its own right, something profoundly remarkable.</p><h3>References</h3><p>Zhou, Jian-Ming (email to: <a href="/service/mailto:yu.li@u-picardie.fr">yu.li@u-picardie.fr</a> ) May 2013. Computability vs. Nondeterministic and P vs. NP. Retrieved from <a href="https://arxiv.org/pdf/1305.4029](https://arxiv.org/pdf/1305.4029)">https://arxiv.org/pdf/1305.4029</a></p><p>Olivier Ezratty 2024. Understanding Quantum Technologies 2024. Retrieved from <a href="https://www.oezratty.net/wordpress/2024/understanding-quantum-technologies-2024/?id=0&amp;lw=pfs1&amp;gw=&amp;aw=&amp;pw=&amp;fw=](https://www.oezratty.net/wordpress/2024/understanding-quantum-technologies-2024/?id=0&amp;lw=pfs1&amp;gw=&amp;aw=&amp;pw=&amp;fw=)">https://www.oezratty.net/wordpress/2024/understanding-quantum-technologies-2024/?id=0&amp;lw=pfs1&amp;gw=&amp;aw=&amp;pw=&amp;fw=</a></p><p>Sanjeev Arora and Boaz Barak 2007. Computational Complexity: A Modern Approach. Retrieved from <a href="https://theory.cs.princeton.edu/complexity/book.pdf](https://theory.cs.princeton.edu/complexity/book.pdf)">https://theory.cs.princeton.edu/complexity/book.pdf</a></p><p>Wikipedia. NP-Hardness (complexity). Retrieved from <a href="https://en.wikipedia.org/wiki/NP-hardness](https://en.wikipedia.org/wiki/NP-hardness).">https://en.wikipedia.org/wiki/NP-hardness.</a> BQP (complexity). Retrieved from <a href="https://en.wikipedia.org/wiki/BQP](https://en.wikipedia.org/wiki/BQP).">https://en.wikipedia.org/wiki/BQP.</a> BPP (complexity). Retrieved from BPP <a href="https://en.wikipedia.org/wiki/BPP_/(complexity/)](https://en.wikipedia.org/wiki/BPP_/(complexity/))">https://en.wikipedia.org/wiki/BPP_\(complexity\)</a></p><blockquote><strong>Disclaimer</strong></blockquote><blockquote>This article has been reviewed and corrected using Generative AI. The views and opinions expressed in this article are solely those of the author and do not necessarily reflect the official policy or position of the author’s company. Any content provided by the author is of their own opinion and is not intended to malign any organization, company, individual, or entity.</blockquote><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=4c82eb5dd7b8" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>