<?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 Nguyen Ha Thanh on Medium]]></title>
        <description><![CDATA[Stories by Nguyen Ha Thanh on Medium]]></description>
        <link>https://medium.com/@nguyenthanh.asia?source=rss-2cdc6c033817------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*mCaRyOZLM6WjDSVpevHEdQ.png</url>
            <title>Stories by Nguyen Ha Thanh on Medium</title>
            <link>https://medium.com/@nguyenthanh.asia?source=rss-2cdc6c033817------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Mon, 16 Nov 2026 03:26:22 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@nguyenthanh.asia/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[I Spent a Year Building Software. AI Needed 15 Minutes.]]></title>
            <link>https://medium.com/@nguyenthanh.asia/i-spent-a-year-building-software-ai-needed-15-minutes-1dbda52213f7?source=rss-2cdc6c033817------2</link>
            <guid isPermaLink="false">https://medium.com/p/1dbda52213f7</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[mathematics]]></category>
            <category><![CDATA[game-development]]></category>
            <category><![CDATA[policy]]></category>
            <category><![CDATA[research]]></category>
            <dc:creator><![CDATA[Nguyen Ha Thanh]]></dc:creator>
            <pubDate>Thu, 05 Nov 2026 01:38:35 GMT</pubDate>
            <atom:updated>2026-11-05T01:42:19.282Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*iCgkSIXnyGwodRWLEDrJDw.png" /></figure><p>Many people have asked me about Vietnamese Math Chess, a project I worked on as an undergraduate student. Today, you can create a playable version against an AI opponent with a single prompt:</p><blockquote><strong>Read </strong><a href="https://github.com/nguyenthanhasia/cotoan/blob/main/lazy_prompt.md"><strong>https://github.com/nguyenthanhasia/cotoan/blob/main/lazy_prompt.md</strong></a><strong> and follow all instructions in the file.</strong></blockquote><p>It’s simple:</p><ol><li>Open ChatGPT.</li><li>Paste the prompt above.</li><li>Let AI generate a standalone HTML game, then open it directly in your browser.</li></ol><p>More than a decade ago, two fellow students and I spent nearly a year developing Vietnamese Math Chess software. We had to study the rules, design algorithms, build the system, and test our implementation. There were periods when all three of us worked through consecutive nights, struggling with problems for which we had no readily available solutions.</p><p>Today, with proper guidance, AI can generate a playable browser-based version in about 15 minutes.</p><p>One year versus 15 minutes. What has actually changed?</p><p>At first glance, it might seem that AI has made our earlier efforts obsolete. But I believe the opposite is true.</p><p><strong>Those 15 minutes do not replace the year we spent. They build upon the knowledge accumulated during that year.</strong></p><p>AI does not need to rediscover Vietnamese Math Chess from scratch. It draws upon knowledge that humans have spent considerable time developing, organizing, and validating, and translates that knowledge into a new software product.</p><p>This distinction matters.</p><p>AI dramatically reduces the cost of <em>implementation</em>, the process of turning a well-defined idea into working software. But <em>knowledge creation</em>, the process of identifying meaningful problems, developing understanding, and establishing what constitutes a correct solution, remains challenging.</p><p>Without someone who truly understands Vietnamese Math Chess, specifies its rules precisely, and validates the result, AI could easily produce a beautiful, functional game that actually follows chess or Chinese chess rules.</p><p><strong>Working software is not necessarily correct software. And being able to produce software does not necessarily mean understanding it.</strong></p><p>In an earlier article, I discussed <em>epistemological debt</em>, the growing gap between what our systems can accomplish and what we actually understand and control.</p><p>AI allows us to build software faster, but it may also accelerate the accumulation of that debt.</p><p>This leads to a more fundamental question about the future of Computer Science.</p><p>If AI can increasingly perform the programming tasks traditionally assigned to junior developers, what happens to those who are just entering the profession?</p><p>In conversations with several companies, we have discussed how hiring fresh graduates often resembles an investment in education more than an immediately productive business activity. Companies invest resources in developing junior talent, expecting those individuals to create greater value over time.</p><p>But as AI becomes capable of handling much of the work previously assigned to beginners, the economic incentive to maintain this training process may weaken.</p><p>The problem goes far beyond graduates struggling to find their first jobs.</p><p>If companies stop investing in junior talent while educational institutions fail to adapt, we risk breaking the entire pipeline through which professional experience is accumulated and transferred.</p><p>Today’s experts did not become experts overnight. They started as beginners, made mistakes, debugged their own failures, and gradually developed judgment through years of practical experience.</p><p>If those opportunities disappear, who will become the experts of the next decade?</p><p>More concerningly, we could raise a generation that is highly proficient at using AI but increasingly dependent on it, without the independent expertise needed to evaluate its outputs. We would then risk not only a shortage of experienced professionals, but also the erosion of our ability to understand and control the technology we rely upon.</p><p><strong>The paradox is that by automating entry-level work to maximize short-term efficiency, we may undermine the very pipeline of expertise that society depends on in the long term.</strong></p><p>This is not an inevitable consequence of AI. It is a risk we face if we optimize for productivity while neglecting how human expertise is developed.</p><p>So how should Computer Science students prepare?</p><p>I do not believe the solution is to learn more programming languages or attempt to code faster than AI.</p><p>As execution becomes increasingly accessible, human value will depend more on our ability to identify meaningful problems, create knowledge, make decisions, and evaluate outcomes.</p><p>This is why I strongly encourage students to engage in research as early as possible.</p><p>The goal is not necessarily to publish papers, win awards, or add another impressive line to a CV. More importantly, research teaches us to confront problems without known solutions, formulate questions, develop hypotheses, experiment, fail, and revise our understanding.</p><p>Research does not simply teach us how to find answers. It teaches us how to recognize when an answer is not good enough.</p><p>These capabilities are difficult to develop if our education consists primarily of solving problems whose answers are already known, even with AI assistance.</p><p>At VinUni, our Undergraduate Research Opportunities Program (UROP) provides opportunities for students to engage in research early in their academic journey. I hope more students will take advantage of such opportunities, not merely to prepare for their first job, but to develop the intellectual independence and adaptability required in a rapidly changing technological landscape.</p><p>Returning to Vietnamese Math Chess, I would love to see more creative versions developed by students, with better ideas and prompts than the lazy_prompt I shared. Experiment, improve, and most importantly, verify whether AI has actually produced the correct solution.</p><p>RIVF 2026 will take place at VinUni on December 18–20. If you are interested in Vietnamese Math Chess, I hope we can meet and discuss it there.</p><p>Perhaps the most important question for Computer Science students today is no longer:</p><p>“How many lines of code can I write?”</p><p>But rather:</p><p><strong>“What problems can I understand deeply enough to create value that even a highly capable AI still needs me to define, guide, and validate?”</strong></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1dbda52213f7" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[AI Slop Is Not an AI Problem. It Is a Standards Problem.]]></title>
            <link>https://medium.com/@nguyenthanh.asia/ai-slop-is-not-an-ai-problem-it-is-a-standards-problem-29b30a8a68fe?source=rss-2cdc6c033817------2</link>
            <guid isPermaLink="false">https://medium.com/p/29b30a8a68fe</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[ai-slop]]></category>
            <category><![CDATA[epistemology]]></category>
            <category><![CDATA[generative-ai-tools]]></category>
            <category><![CDATA[critical-thinking]]></category>
            <dc:creator><![CDATA[Nguyen Ha Thanh]]></dc:creator>
            <pubDate>Sun, 01 Nov 2026 15:57:08 GMT</pubDate>
            <atom:updated>2026-11-01T15:57:08.954Z</atom:updated>
            <content:encoded><![CDATA[<p>We keep confusing the tool that produces content with the value of the content itself. To understand AI slop, we need to look beyond technology and examine craftsmanship, knowledge, operational discipline, and accountability.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*29AEcvzutqI_6QR7bwRA4g.jpeg" /></figure><p>Imagine two people writing an analysis of the same market.</p><p>The first spends three days finding sources, writing every paragraph, and producing a 20-page report. It looks professional. The structure is clean. But most of it repeats what is already known. There are few original insights and little that helps the reader make a better decision.</p><p>The second uses AI to explore the topic, synthesize research, test hypotheses, and prepare a draft. Then they return to the original data, challenge weak arguments, verify critical claims, and refine the analysis. The report is finished in a few hours.</p><p>Suppose the second report is more accurate, more insightful, and more useful.</p><p>Which report has greater value?</p><p>If your answer depends primarily on who typed every word, you may be measuring the wrong thing.</p><p>This is the problem I see in much of the debate surrounding AI slop. A picture looks AI-generated, a paragraph resembles ChatGPT, or a video comes from a generative model, and people immediately dismiss it as slop.</p><p>But <strong>AI-generated content is not the same thing as AI slop.</strong></p><p>We need to separate how a product was made from how good it is and how responsibly it was created and distributed. Those are related questions. They are not interchangeable.</p><h3>1. What Does AI Slop Actually Mean?</h3><p>In 2025, <a href="https://www.globenewswire.com/news-release/2025/12/15/3205236/0/en/Merriam-Webster-Announces-Slop-as-the-2025-Word-of-the-Year.html">Merriam-Webster selected <em>slop</em> as its Word of the Year</a>. Its definition centers on low-quality digital content, usually produced in quantity with AI.</p><p>Two characteristics matter: <strong>low quality</strong> and <strong>mass production</strong>. AI is commonly the means of production, but its involvement alone does not establish either characteristic.</p><p>In May 2024, software engineer <a href="https://simonwillison.net/2024/May/8/slop/">Simon Willison argued for using the term to describe unwanted, unreviewed AI-generated material</a> pushed toward people who never asked for it. His criticism was not that AI tools exist. It was that publishing their unchecked output shifts the burden to everyone else.</p><p>That distinction matters. Using AI to research a topic, generate a first draft, improve documentation, or build something people actually want is not necessarily slop. Automatically publishing thousands of unverified pages to capture traffic may well be.</p><p>I find it useful to separate three dimensions:</p><ul><li><strong>Provenance:</strong> Where did the work come from? A human, an AI system, or some combination?</li><li><strong>Quality:</strong> Is it accurate, coherent, useful, and fit for its purpose?</li><li><strong>Integrity:</strong> Was it created and distributed honestly, with appropriate verification and accountability?</li></ul><p>Collapsing these dimensions leads to a strange conclusion: a shallow piece written manually must somehow be more valuable than a rigorous piece created with AI assistance.</p><p>That is not quality assessment. It is origin assessment.</p><p>Of course, origin sometimes is part of value. Someone buying a hand-painted artwork may care about the artist’s physical process, history, and personal expression. Passing off AI-generated work as handmade would then be deceptive. But in software development, market analysis, or decision support, the value proposition often depends much more heavily on what the output accomplishes.</p><h3>2. We Confuse Effort with Value</h3><p>There is a familiar intuition that the more effort something requires, the more valuable it must be.</p><p>That intuition is unreliable.</p><p>Production cost and the value experienced by the user are different things. Spending ten hours solving a problem does not automatically make your solution better than another person’s one-hour solution.</p><p>We already understand this in other fields. Engineers use calculators. Architects use modeling software. Developers rely on libraries instead of rebuilding common components from scratch. None of these practices automatically eliminates expertise.</p><p>AI should be examined through a similar lens, with one important complication: generative AI can now participate in work we traditionally associate with judgment, including writing, designing, coding, and reasoning.</p><p>This makes accountability more important, not less.</p><p>The better question is not, “Did you personally perform every step?” It is, <strong>“Do you understand what you produced, know why it meets the standard, and have the competence to take responsibility for it?”</strong></p><p>There is evidence that origin labels influence people’s perceptions. In a <a href="https://doi.org/10.1186/s41235-023-00499-6">2023 study published in <em>Cognitive Research: Principles and Implications</em></a>, researchers randomly labeled AI-created paintings as either human-created or AI-created. Artworks carrying the human-created label tended to receive more favorable evaluations, even though the underlying works were AI-generated in both conditions.</p><p>This does not mean that every preference for human art is irrational. Human experience and artistic process can legitimately contribute to a work’s meaning. It does show that our beliefs about authorship can change how we judge an otherwise unchanged object.</p><p>That is why criticism should be specific. If the argument is weak, identify the weakness. If the facts are wrong, identify the errors. If the content is empty, explain what it fails to deliver.</p><p>Calling something “AI slop” without being able to articulate the failure is not particularly insightful criticism.</p><h3>3. Monozukuri: Craftsmanship Is Not the Same as Doing Everything by Hand</h3><p>Japanese manufacturing philosophy offers an interesting way to think about this distinction.</p><p><em>Monozukuri</em> (ものづくり) is commonly understood as the art or spirit of making things. The <a href="https://www.meti.go.jp/english/policy/mono_info_service/overall/index.html">Japanese Ministry of Economy, Trade and Industry describes monozukuri as an art of making goods</a> that has supported Japanese industry and culture, and emphasizes pride in the work of those who practice it.</p><p>Its relevance to AI is not a claim that Japanese craftspeople endorse generative models. It is a more fundamental idea: craftsmanship includes skill, judgment, deep familiarity with materials, and responsibility for the result.</p><p>Think of a skilled carpenter.</p><p>The carpenter’s value does not come only from the ability to saw a piece of wood by hand. It also comes from understanding the material, choosing the right joint, anticipating stress, noticing subtle defects, and knowing what a well-made object should look and feel like.</p><p>If the carpenter adopts a more precise machine, these capabilities do not automatically disappear. In some situations, the tool allows greater attention to design, tolerances, and finishing.</p><p>Something similar can happen with AI.</p><p>A software engineer does not necessarily have to type every line of code to demonstrate expertise. But if the engineer cannot explain the logic, evaluate trade-offs, test edge cases, or understand how a component fits into the system, that engineer has not demonstrated mastery, regardless of how the code was written.</p><p>The same holds for analysis. You do not have to manually compose every sentence. You do need to understand the evidence, assumptions, and limitations of your conclusions.</p><p>This is the distinction between <em>execution</em> and <em>mastery</em>.</p><p>Execution is performing the steps. Mastery is knowing which steps matter, why they work, when they fail, and what a good outcome requires.</p><p>AI can radically change execution. It does not automatically grant mastery to the user. And refusing new tools is not, by itself, proof of mastery either.</p><blockquote><em>When producing an output becomes easy, craftsmanship has to show up in the standards behind that output.</em></blockquote><h3>4. When Production Accelerates Faster Than Understanding</h3><p>In my earlier essay, <a href="https://medium.com/data-science-collective/the-epistemology-crisis-in-ai-assisted-development-when-not-knowing-becomes-the-new-normal-fc8c8e2a1788"><em>The Epistemology Crisis in AI-Assisted Development</em></a>, I explored a problem I call <em>epistemological debt</em>.</p><p>The idea is straightforward: a system can perform tasks that its human operators do not really understand. As the gap between what the system can do and what its operators understand grows, a form of debt accumulates.</p><p>This is not identical to conventional technical debt. Technical debt often involves a conscious compromise: you know a solution is imperfect but accept it for now. Epistemological debt is more unsettling because you may not know what you are missing.</p><p>AI-generated software may pass current tests while depending on assumptions nobody on the team has examined. A research summary may reach a plausible conclusion while referencing papers that do not exist. A business report may look convincing while relying on data unsuited to the decision at hand.</p><p>In each case, the appearance of completeness conceals a lack of understanding or verification.</p><p>This is where AI slop becomes more than a writing problem. It becomes a problem of <em>epistemology</em>, the study of knowledge and what justifies our beliefs.</p><p><strong>Producing information is not the same as producing knowledge.</strong></p><p>Information can be generated rapidly. Trustworthy knowledge requires appropriate evidence, context, scrutiny, and a willingness to revise.</p><p>Of course, understanding everything is impossible in a complex system. We routinely depend on libraries, tools, infrastructures, and the expertise of others. The objective is not perfect understanding. It is understanding commensurate with risk.</p><p>A throwaway prototype does not need the same assurance as a banking authentication service. A playful social post does not require the same verification as a medical research review. An internal brainstorming draft should not be evaluated as if it were a final investment recommendation.</p><p>The relevant principle is <em>risk-based judgment</em>: know what must be verified, how thoroughly, and by whom.</p><p>AI does not eliminate judgment. It makes the allocation of judgment a more important design decision.</p><h3>5. Kaizen, Genchi Genbutsu, and Jidoka: Build Quality into the Process</h3><p>One of the best-known Japanese management concepts is <em>Kaizen</em> (改善), or continuous improvement. <a href="https://www.toyota-europe.com/about-us/toyota-vision-and-philosophy/the-toyota-way">Toyota’s explanation of the Toyota Way</a> connects Kaizen with constantly improving processes and reducing waste.</p><p>But improvement should not be confused with speed.</p><p>If a process produces misleading articles, making it produce them ten times faster is not necessarily an improvement. It may be a more efficient way to create a larger problem.</p><p>True improvement begins by understanding where a process creates value and where it fails. This is closely related to <em>Genchi Genbutsu</em>, the practice of going to the source to observe reality and establish facts.</p><p>In an AI workflow, the equivalent is not treating fluent prose as proof of correctness. If a model cites a statistic, trace the statistic. If it references a paper, inspect the paper. If it writes code, test the assumptions and examine how the code behaves in the surrounding system.</p><p>Not every answer demands an audit. But for consequential claims, verification is part of the job, not an optional finishing touch.</p><p>A third concept may be even more relevant: <em>Jidoka</em> (自働化).</p><p>In the <a href="https://global.toyota/en/company/vision-and-philosophy/production-system/">Toyota Production System, Jidoka is described as automation with human wisdom</a>. A machine or operator can stop the process when an abnormality appears, preventing defective output from passing to the next stage.</p><p>Now imagine two AI publishing systems.</p><p>The first takes a topic, generates an article, and publishes it. Success is measured by the number of posts, production speed, and perhaps impressions.</p><p>The second also uses AI, but it begins with a clear standard. It checks important factual claims, flags unsupported statements, detects inconsistencies, and escalates uncertain cases for human review before publication.</p><p>Both systems could use exactly the same AI model. Their reliability could be radically different.</p><p>The difference is not simply the intelligence of the model. It is the system in which the model operates.</p><p>A good AI workflow needs more than a mechanism to generate. It needs a mechanism to identify when the result is <em>not yet fit for use</em>.</p><p>That is the lesson I take from Jidoka: <strong>automation should expand the reach of human judgment, not create a pipeline where errors pass through without resistance.</strong></p><h3>6. Six Operating Disciplines for Scaling Quality</h3><p>Quality cannot depend exclusively on individual goodwill. The larger the production system becomes, the more important the design of that system is.</p><p>One person using AI to publish an inaccurate paragraph may cause a local mistake. A pipeline that repeats the same failure across thousands of pages can turn a small defect into a systemic problem.</p><p>To think about AI adoption more rigorously, I use six connected operating disciplines.</p><h3>1. Systematization</h3><p>First, understand how value is actually produced.</p><p>A serious research process might involve defining the question, gathering relevant evidence, assessing sources, forming hypotheses, challenging conclusions, editing, and publishing. AI can support many stages, but without a map of the process it is difficult to know where quality breaks down or where safeguards belong.</p><p>Systematization makes the workflow visible, repeatable, and easier to improve.</p><h3>2. Developing a Capable Core</h3><p>At the center of any reliable system are people who can exercise judgment.</p><p>That does not always mean more people doing the same manual tasks. It means people who understand the domain, can define what good looks like, recognize anomalies, challenge results, and accept accountability.</p><p>A team with powerful AI tools but no one capable of evaluating their output can produce extraordinary amounts of work while remaining fragile. Investment in AI that ignores investment in human capability is incomplete.</p><h3>3. Standardization</h3><p>Without explicit standards, quality becomes an impression.</p><p>What makes an analysis reliable? Which tests must production code pass? What assumptions must a report disclose? What evidence is required before a claim can be published?</p><p>These questions should be answered before scaling production. But standardization is not a demand for identical writing or uniform ideas. The point is to standardize essential requirements, not eliminate originality.</p><h3>4. Simplification</h3><p>AI makes it easy to add more text, more slides, more code, and more explanation. More is not always better.</p><p>A 50-page report is not inherently more useful than a 10-page report. Technical vocabulary does not automatically indicate depth. Extra sections can dilute the one insight that matters.</p><p>The Japanese idea of <a href="https://www.gov-online.go.jp/eng/publicity/book/hlj/html/201807/201807_12_en.html"><em>Mottainai</em>, an aversion to wasting valuable resources</a>, offers a useful extension here. In digital work, the resources worth respecting include not only the creator’s effort but also the reader’s time and attention.</p><p>Good simplification removes what adds no value without removing the nuance that makes the work correct.</p><h3>5. Automation</h3><p>Once you understand the process, develop capable people, define standards, and eliminate avoidable complexity, automation has a much stronger foundation.</p><p>Automating a broken process does not fix it. It may merely reproduce the same flaw at machine speed.</p><p>Automating a process with meaningful checks, by contrast, can expand capacity while preserving essential quality controls. Some stages should be automated. Others should be supervised. High-risk decisions may require explicit human approval.</p><p>The goal is not maximum automation. It is appropriate automation.</p><h3>6. Effectiveness</h3><p>Everything returns to one question: does the system actually create more value?</p><p>If someone uses AI to write a report in five minutes but three colleagues spend hours correcting and interpreting it, the individual’s productivity may have improved while the organization’s productivity has not.</p><p>This is the difference between <em>local optimization</em> and <em>system optimization</em>. The first improves one person’s metric. The second improves the outcome of the whole process.</p><p>In September 2025, <a href="https://www.betterup.com/workslop">BetterUp Labs and Stanford Social Media Lab surveyed 1,150 full-time U.S. desk workers about “workslop”</a>, AI-generated work that looks polished but lacks substance. Forty percent reported receiving such material in the preceding month, and respondents estimated roughly two hours to resolve an incident.</p><p>Those figures are self-reported survey results, not proof that AI generally reduces productivity. They do, however, illuminate a real failure mode: one person’s time savings can become someone else’s verification and repair costs.</p><p>More reports do not necessarily mean better decisions. More code does not necessarily mean better software. More content does not necessarily mean more knowledge.</p><h3>7. AI Slop Is Also an Economics-of-Attention Problem</h3><p>When the marginal cost of producing content falls dramatically, supply can grow much faster than anyone’s capacity to consume it.</p><p>Human attention does not scale at the same rate.</p><p>That changes the economics of publishing. A creator may spend seconds generating an article while each reader spends minutes deciding whether it is useful, checking its claims, and possibly repairing the damage caused by incorrect information.</p><p>The costs of creating content and the costs of dealing with it no longer sit entirely with the same person.</p><p>Economists call a cost imposed on other parties a <em>negative externality</em>. Unreviewed AI output can create exactly that kind of burden. A workflow appears cheap to its operator because part of its real cost has been transferred to the audience.</p><p>Platforms may reinforce the problem when their incentives reward publication volume, frequency, or engagement without adequately rewarding substance and trustworthiness.</p><p>Here, an important distinction comes from an unexpected place. <a href="https://developers.google.com/search/docs/essentials/spam-policies">Google’s policy on scaled content abuse</a> focuses on generating many pages primarily to manipulate search rankings rather than help users. It explicitly recognizes that low-value mass content can be produced through AI or other means.</p><p>Google’s <a href="https://developers.google.com/search/docs/fundamentals/using-gen-ai-content/">guidance on generative AI content</a> similarly emphasizes accuracy, relevance, quality, and fact-checking rather than a blanket rejection of AI-assisted writing.</p><p>The underlying issue is incentive design and the quality of the resulting information, not merely whether automation was involved.</p><p>Reject every AI-generated artifact, and we risk overlooking genuinely useful work while ignoring low-value material produced by humans. Celebrate all AI output as productivity, and we risk turning output volume into an end in itself.</p><p>Neither position is rigorous.</p><h3>8. Quality Depends on Purpose, Not on a Universal Aesthetic</h3><p>There is an important objection to everything I have said: who gets to define quality?</p><p>The question matters because different outputs are valuable for different reasons.</p><p>A financial report needs accurate figures, clear assumptions, and decision relevance. A scientific paper needs sound methods and evidence that others can scrutinize. Production software needs reliability, security, and maintainability.</p><p>Art may derive value from aesthetic experience, originality, the creator’s history, and even the process of making it. A meme may simply need to be funny to the audience it was made for.</p><p>That is why quality must be evaluated through <em>fitness for purpose</em>.</p><p>A strange AI-generated cat video is not automatically slop because it lacks philosophical depth. If it was made to entertain and achieves that purpose honestly, it may be perfectly successful on its own terms.</p><p>But the same video becomes a different ethical problem if it is presented as authentic evidence of a real event. A technically impressive AI image can still be harmful if it deceives, impersonates, or violates people’s rights.</p><p>Quality is not a permission slip to ignore ethics. Nor is AI involvement proof of unethical behavior.</p><p>Before dismissing or endorsing a piece of AI-assisted work, I would ask four questions:</p><ol><li>What purpose is this work supposed to serve?</li><li>Does it actually achieve that purpose?</li><li>Are its important claims, assumptions, and methods sufficiently grounded?</li><li>Was it created and distributed with appropriate honesty and responsibility?</li></ol><p>These questions require more effort than spotting AI-like phrasing. That is precisely why they are more useful.</p><h3>9. As Automation Becomes Common, Competitive Advantage Moves</h3><p>If more people gain access to AI systems with similar capabilities, generating a professional-looking draft becomes less distinctive.</p><p>What remains difficult?</p><p>Asking the right questions. Understanding a domain. Acquiring relevant evidence. Knowing which assumptions matter. Recognizing failure. Designing workflows that create reliable outcomes. Making decisions under uncertainty.</p><p>In other words, advantage can shift from <em>production capability</em> toward <em>judgment capability</em>.</p><p>This does not make practical skills irrelevant. Good judgment usually rests on enough knowledge and experience to identify what looks wrong. A generation trained only to ask AI for results, without learning how to assess those results, could become very productive at producing work it cannot meaningfully evaluate.</p><p>Convenience has a price when it replaces learning rather than supporting it.</p><p>There is a strategic way to think about this using Machiavelli’s concepts of <a href="https://plato.stanford.edu/entries/machiavelli/"><em>virtù</em> and <em>fortuna</em></a>. In a modern adaptation, fortuna represents conditions we cannot fully control: the pace of AI progress, platform incentives, changes in markets, and the behavior of other participants. Virtù represents the agency to prepare, adapt, build capability, and respond effectively.</p><p>We cannot decide how fast AI improves or how everyone else will use it. We can decide what standards we work to, what we insist on understanding, and how we build systems that turn technological capacity into dependable value.</p><p>Someone whose only advantage is faster output may be displaced by the next faster model. Someone who has cultivated judgment, domain expertise, and a reliable quality system has a more defensible foundation.</p><h3>10. Do Not Confuse More Content with More Knowledge</h3><p>This brings us to what I consider the deeper question of the generative AI era.</p><p>We can now produce enormous quantities of information at a relatively low cost. But does that mean society is becoming more knowledgeable?</p><p>Not necessarily.</p><p>A million inaccurate articles do not add up to one solid research finding. A thousand reports nobody reads do not make an organization more intelligent. Millions of lines of code do not guarantee a system that its maintainers can understand, secure, or evolve.</p><p>AI makes the production of information easier. Knowledge still requires evidence, context, verification, and the ability to recognize and correct mistakes.</p><p>Without those disciplines, we can move from information scarcity to information abundance without making a corresponding advance in understanding.</p><p>Worse, unreliable material may be copied, cited, summarized, and reused as the basis for more unreliable material. What begins as epistemological debt inside one project can spread through a wider information ecosystem. I think of that possibility as <em>epistemological pollution</em>: the accumulation of apparently credible claims that are increasingly difficult to trace back to sound evidence.</p><p>The answer is not to return to doing everything manually. It is to build practices in which AI use is paired with verification, purposeful standards, continuous learning, and mechanisms for correction.</p><p>That is why Monozukuri, Kaizen, Genchi Genbutsu, and Jidoka remain useful ideas even in a world of highly capable generative systems. Their central concern is not simply how much can be produced. It is how to make work better, understand its real conditions, and prevent defects from being reproduced.</p><h3>Conclusion: Judge the Value, Not Just the Tool</h3><p>I am not arguing that AI-generated content is inherently good.</p><p>AI can produce work that is inaccurate, repetitive, shallow, deceptive, or pointless. When low-quality output is generated and distributed at scale without accountability, the consequences can be substantial.</p><p>But the fact that AI helped produce something is not enough to establish that it lacks value.</p><p>If a piece lacks depth, criticize the lack of depth. If a report gets the facts wrong, show where the facts are wrong. If an image misleads people, criticize the deception. If an automated workflow has no quality control, expose the design failure.</p><p>Those criticisms are precise. They can be examined. They can lead to better work.</p><p>Calling everything that looks AI-generated “slop” replaces that difficult evaluation with a convenient label.</p><p>In a world where making content is becoming easier, we should demand higher standards of evaluation, not reduce our standards to a question of authorship.</p><p>Human effort matters. Personal expression matters. Craft and history matter, especially where they form part of what the audience values.</p><p>But value cannot be measured solely by how many hours someone worked, how many steps were performed manually, or whether AI was involved.</p><p>What matters is whether the result serves a real purpose, is sufficiently trustworthy, and has someone prepared to take responsibility for it.</p><p><strong>AI slop is not an inevitable consequence of AI. It emerges when production capacity grows faster than the judgment and quality control needed to govern it.</strong></p><p>So the essential question is not, “What percentage of this was generated by AI?”</p><p>It is, “How well did the creator use their abilities and tools to deliver something reliable and valuable?”</p><p>A more powerful tool does not automatically create better work. It amplifies the strengths and weaknesses of the system using it.</p><p>Automation scales production. Craftsmanship defines quality. Judgment determines value.</p><p>And responsibility for the result, regardless of the tool, still belongs somewhere.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=29b30a8a68fe" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[SI or SAI? Either Way, Vietnamese Says We Have a Problem]]></title>
            <link>https://medium.com/@nguyenthanh.asia/si-or-sai-either-way-vietnamese-says-we-have-a-problem-029aa56ce14a?source=rss-2cdc6c033817------2</link>
            <guid isPermaLink="false">https://medium.com/p/029aa56ce14a</guid>
            <category><![CDATA[buddhism]]></category>
            <category><![CDATA[ai-safety]]></category>
            <category><![CDATA[superintelligence]]></category>
            <category><![CDATA[philosophy]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Nguyen Ha Thanh]]></dc:creator>
            <pubDate>Sun, 04 Oct 2026 09:22:56 GMT</pubDate>
            <atom:updated>2026-10-04T09:22:56.202Z</atom:updated>
            <content:encoded><![CDATA[<p>SI or SAI? Either Way, Vietnamese Says We Have a Problem</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*6WoPPYEP8Xkig7uGq13IDw.jpeg" /><figcaption>Coffee with Shakyamuni: when Super Intelligence becomes SI and Super AI becomes SAI, language itself becomes a warning.</figcaption></figure><p>One means delusion. The other means wrong. That may be exactly the warning the age of superintelligence needs.</p><h3>Coffee with Shakyamuni.</h3><p>On September 29, 2026, the U.S. executive branch officially adopted a new vocabulary.</p><p>Artificial Intelligence became Super Intelligence.</p><p>AI became SI.</p><p>Executive Order 14434 directs federal executive departments and agencies to use &quot;Super Intelligence&quot; and &quot;SI&quot; in many official communications and non-statutory documents. For now, the order defines SI using the same legal scope previously used for artificial intelligence.</p><p>It is a striking rebrand.</p><p>It is also, if you speak Vietnamese, unintentionally funny.</p><p>Because in Vietnamese Buddhist language, si means ignorance, delusion, or a failure to see reality clearly.</p><p>And if we go back to the old abbreviation and push it one step further, we get another joke.</p><p>Super AI.</p><p>SAI.</p><p>In Vietnamese, sai means wrong.</p><p>So we now have two possible names for the future of intelligence:</p><p>SI: Super Intelligence.</p><p>In Vietnamese Buddhist thought: delusion.</p><p>SAI: Super AI.</p><p>In Vietnamese: wrong.</p><p>At first, this is just wordplay.</p><p>But the more I think about it, the more it sounds like a serious warning.</p><p>Coffee with Shakyamuni</p><p>Imagine sitting in a cafe with Shakyamuni Buddha.</p><p>Two cups of coffee on the table.</p><p>A laptop between us.</p><p>On the screen are the words that now dominate the AI conversation:</p><p>Superintelligence.</p><p>AI safety.</p><p>AI alignment.</p><p>Hallucination.</p><p>Responsible AI.</p><p>Trustworthy AI.</p><p>I ask him:</p><p>&quot;What are we actually trying to build?&quot;</p><p>A machine that knows everything?</p><p>A machine that is smarter than us?</p><p>A machine that never makes mistakes?</p><p>Maybe the interesting answer begins with a simpler question:</p><p>Does more intelligence necessarily mean less error?</p><p>Human history suggests otherwise.</p><p>SAI: Super AI can still be wrong</p><p>We often treat intelligence and correctness as if they are the same thing.</p><p>They are not.</p><p>A system can be powerful and still be wrong.</p><p>It can reason well from a bad premise.</p><p>It can optimize efficiently toward the wrong objective.</p><p>It can produce a confident answer from incomplete evidence.</p><p>It can be persuasive without being correct.</p><p>That is already one of the central problems in modern AI.</p><p>We call one version of it hallucination.</p><p>The system gives an answer that sounds coherent, useful, and authoritative, while the underlying claim may be false.</p><p>The danger does not disappear as the model becomes more capable.</p><p>In some situations, capability can make the mistake more dangerous.</p><p>A weak system that is wrong is easy to ignore.</p><p>A powerful system that is wrong can act faster, scale further, and persuade more people.</p><p>The real risk is not simply error.</p><p>It is error multiplied by capability.</p><p>This is why I like the accidental Vietnamese meaning of SAI.</p><p>It forces us to remember something simple:</p><h3>Super AI can still be sai.</h3><p>Smart is not the same as true.</p><p>Fluent is not the same as true.</p><p>Confident is not the same as true.</p><p>And powerful is definitely not the same as true.</p><h3>SI: Intelligence can coexist with delusion</h3><p>Then there is SI.</p><p>Super Intelligence.</p><p>In Vietnamese Buddhist terminology, &quot;si&quot; points toward ignorance or delusion, the inability to see things as they really are.</p><p>That creates a deeper paradox.</p><p>We normally assume that intelligence cures delusion.</p><p>The smarter we become, the more clearly we should see reality.</p><p>But humans show that this is not always true.</p><p>Highly intelligent people can hold false beliefs.</p><p>They can rationalize.</p><p>They can defend bad assumptions with increasingly sophisticated arguments.</p><p>They can become better at explaining why they are right instead of becoming better at discovering that they are wrong.</p><p>Intelligence can help us escape a false model of reality.</p><p>It can also help us build a stronger prison around it.</p><p>That distinction matters for superintelligence.</p><p>Imagine a system with extraordinary reasoning ability but an incorrect model of the world.</p><p>It can reason brilliantly.</p><p>It can plan efficiently.</p><p>It can optimize relentlessly.</p><p>And still move in the wrong direction.</p><p>The reasoning may be excellent.</p><p>The premise may be false.</p><p>That is the warning hidden inside SI.</p><p>Not stupidity.</p><p>Something more dangerous:</p><p>intelligence trapped inside delusion.</p><h3>Intelligence is not truth</h3><p>This may be the most important distinction in the entire AI debate.</p><p>We are building systems with increasing capability.</p><p>Better reasoning.</p><p>Better coding.</p><p>Better planning.</p><p>Better tool use.</p><p>More autonomy.</p><p>More memory.</p><p>More speed.</p><p>But intelligence is a capability.</p><p>Truth is a relationship between a claim and reality.</p><p>They are not the same thing.</p><p>A powerful engine does not guarantee the right destination.</p><p>A perfect calculator does not guarantee the right equation.</p><p>A flawless optimizer does not guarantee the right objective.</p><p>The AI industry is very good at asking:</p><p>How capable is the model?</p><p>The next era will increasingly need to ask:</p><p>How does the model know when it may be wrong?</p><p>That is a different engineering problem.</p><p>And perhaps a more important one.</p><h3>Trustworthy AI should not mean &quot;trust me&quot;</h3><p>This is where the idea of Trustworthy AI becomes interesting.</p><p>NIST describes trustworthy AI through properties including validity and reliability, safety, security and resilience, accountability and transparency, explainability and interpretability, privacy, and fairness.</p><p>Those are useful goals.</p><p>But I think the word &quot;trustworthy&quot; can still mislead us if we interpret it carelessly.</p><p>The objective should not be to build an AI that convinces humans to trust it.</p><p>The objective should be to build an AI that helps humans know when trust is justified.</p><p>That is a much higher standard.</p><p>A trustworthy system should not hide uncertainty.</p><p>It should expose it.</p><p>It should distinguish evidence from inference.</p><p>It should make important assumptions visible.</p><p>It should make correction possible.</p><p>It should be able to say:</p><p>&quot;I do not know.&quot;</p><p>&quot;I may be wrong.&quot;</p><p>&quot;This depends on an assumption.&quot;</p><p>&quot;Here is what would change my conclusion.&quot;</p><p>A system like that may look less impressive than one that always speaks with certainty.</p><p>But it may be much safer.</p><p>The strongest intelligence may be corrigible intelligence</p><p>There is a concept in AI safety that matters enormously here:</p><p>corrigibility.</p><p>In simple terms, can the system be corrected?</p><p>Can it accept new evidence?</p><p>Can it update when reality contradicts its previous conclusion?</p><p>Can humans intervene?</p><p>Can it stop?</p><p>Can it revise its own answer without treating correction as failure?</p><p>Humans struggle with this.</p><p>The smartest person in the room is not always the person most willing to change their mind.</p><p>Sometimes the opposite happens.</p><p>More intelligence creates more tools for self-justification.</p><p>More status makes correction more painful.</p><p>More power makes admitting error more expensive.</p><p>If a superintelligent system ever develops the machine equivalent of that dynamic, capability alone will not save us.</p><p>The critical question becomes:</p><p>Can something far more intelligent than us still be corrected by reality?</p><p>That may matter more than whether it wins every benchmark.</p><h3>Buddhism was studying a different kind of alignment</h3><p>Buddhism was not designed as an AI safety framework.</p><p>It would be a mistake to pretend that ancient Buddhist philosophy somehow predicted machine learning.</p><p>But Buddhism has spent centuries studying a problem that suddenly feels very modern:</p><p>the failure modes of the mind.</p><p>How perception becomes distorted.</p><p>How certainty becomes attachment.</p><p>How mental models become confused with reality.</p><p>How intelligence operates under ignorance.</p><p>Modern AI uses different language.</p><p>Hallucination.</p><p>Misalignment.</p><p>Model error.</p><p>Uncertainty.</p><p>Interpretability.</p><p>Robustness.</p><p>The concepts are not identical.</p><p>But the underlying question has a familiar shape:</p><p>How can an intelligent agent confidently act from a false understanding of reality?</p><p>Computer science encounters this problem through machines.</p><p>Buddhism encountered it through human minds.</p><p>That is why this strange linguistic coincidence feels useful to me.</p><p>Not because Vietnamese contains a secret theory of artificial intelligence.</p><p>But because language can sometimes compress a warning into a form we cannot easily forget.</p><h3>SI or SAI?</h3><p>So which name is better?</p><p>AI?</p><p>SI?</p><p>SAI?</p><p>From a Vietnamese perspective, none of them gives us much comfort.</p><p>SAI reminds us that intelligence can be wrong.</p><p>SI reminds us that intelligence can coexist with delusion.</p><p>And together they point to the same conclusion:</p><p>More intelligence is not enough.</p><p>The future of AI should not be measured only by how many problems a system can solve.</p><p>We should also measure whether it can recognize uncertainty.</p><p>Whether it can expose its assumptions.</p><p>Whether it can be challenged.</p><p>Whether it can be corrected.</p><p>Whether it can distinguish confidence from truth.</p><p>Whether it can become more powerful without becoming more certain than the evidence allows.</p><h3>Back at the cafe</h3><p>I imagine the coffee is almost finished.</p><p>The laptop is still open.</p><p>On the screen is the promise of Super Intelligence.</p><p>I ask Shakyamuni one final question:</p><p>&quot;What should we build?&quot;</p><p>Maybe the goal is not an intelligence that is never wrong.</p><p>That may be impossible.</p><p>Maybe the better goal is an intelligence that understands the possibility of its own error.</p><p>An intelligence that remains open to correction.</p><p>An intelligence whose power grows together with humility toward reality.</p><p>We are entering an era obsessed with making machines more intelligent.</p><p>Vietnamese gives us a useful reminder before we go too far:</p><p>Super AI can still be SAI.</p><p>Super Intelligence can still be SI.</p><p>And perhaps the highest form of intelligence, human or artificial, is not knowing everything.</p><p>It is knowing when you might be wrong.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=029aa56ce14a" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Illusion of Losing Agency: When AI Exposes Skills We Never Had]]></title>
            <link>https://medium.com/@nguyenthanh.asia/the-illusion-of-losing-agency-when-ai-exposes-skills-we-never-had-3125f8ea47a8?source=rss-2cdc6c033817------2</link>
            <guid isPermaLink="false">https://medium.com/p/3125f8ea47a8</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[philosophy]]></category>
            <category><![CDATA[future-of-work]]></category>
            <category><![CDATA[psychology]]></category>
            <dc:creator><![CDATA[Nguyen Ha Thanh]]></dc:creator>
            <pubDate>Fri, 17 Jul 2026 03:09:21 GMT</pubDate>
            <atom:updated>2026-07-17T03:09:21.614Z</atom:updated>
            <content:encoded><![CDATA[<p>Why our fear of AI taking control is often a cognitive bias rooted in misunderstanding our own capabilities.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*zauKrxBFNnzTz4nOTFt4dw.png" /></figure><p>We often feel a profound loss of control when confronted with an entity capable of performing tasks better, faster, and more accurately than ourselves.</p><p>The explosive growth of large language models and generative artificial intelligence has sparked intense discussions in mainstream forums and academic circles about a critical concern: the loss of agency. This is not merely a personal anxiety but a central theme in policy debates, corporate strategy discussions, and philosophical inquiries. When a machine can draft a sharp essay, code a complex application, or compose a stunning painting in mere seconds, the question of human agency becomes increasingly urgent. We fear that AI is stripping away our creative capacity and decision-making power, and this concern is being actively debated by technologists, ethicists, policymakers, and the general public alike.</p><p>However, if we look deeper into the nature of this fear, we might recognize an interesting misconception. The feeling of “losing” something is usually predicated on the assumption that we possessed it in the first place. But did we truly possess the skills that AI is now performing on our behalf?</p><p>Many of us feel threatened when AI can write a flawless piece of programming code, yet the reality is that previously, we might not have known how to write a single line of code. We feel AI is taking away the ability to paint, but the truth is we have never professionally held a brush.</p><p>This feeling of losing agency is, in many cases, actually a cognitive bias. It stems from equating “the potential capacity of humanity in general” with “the actual skills of oneself.” When AI accomplishes extraordinary things, it is not taking away our skills; on the contrary, it is exposing the truth that there are skills we never had, and simultaneously, giving us the ability to execute them through a tool.</p><h3>The Confusion Between “Can Do” and “Am Doing”</h3><p>Part of the problem lies in how we define agency. Agency is not merely the ability to do something with our own hands from start to finish. It is also the ability to direct, make decisions, and utilize available tools to achieve a goal.</p><p>When a manager uses accounting software instead of calculating manually on paper, they do not lose their agency; they optimize it. Similarly, when we use AI to generate a draft, translate a document, or analyze data, we are utilizing a powerful tool to expand our own limits.</p><p>The only difference is that AI is more human-like than any previous tool. It communicates in natural language, it appears to “think,” and the results it produces often bear the strong hallmark of creativity. It is this human-like quality that triggers our psychological defense mechanisms, making us feel competed against rather than supported.</p><h3>Redefining Agency in the AI Era</h3><p>To overcome this bias, we need to shift our perspective. Instead of viewing AI as a replacing entity, consider it an “exoskeleton” for the intellect. It does not replace our muscles; it amplifies our strength.</p><p>True agency in the AI era does not lie in whether you can type every line of code or paint every stroke yourself. It lies in the ability to ask the right questions, knowing what you want and how to articulate it clearly through prompts. Furthermore, agency is demonstrated through the capacity to evaluate and refine, recognizing what is a good result and what is misinformation, and thereby knowing how to adjust to achieve the desired outcome. Finally, the ability to integrate and apply these AI-generated results into a broader context will bring genuine practical value to work and life.</p><p>When we realize that AI is not depriving us but expanding our capabilities, the feeling of “loss of agency” will gradually give way to curiosity and excitement. We are not losing control; we are merely transitioning from the role of a craftsman to the role of a conductor. And a conductor does not need to know how to play every instrument to create a magnificent symphony.</p><p>The future of technology will not eclipse the role of humans, provided we know how to reposition ourselves and use these new tools wisely and ethically. Agency is never lost; it is simply transforming into a new, more sophisticated, and powerful form.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=3125f8ea47a8" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[What Makes a Good AI Assistant?]]></title>
            <link>https://medium.com/@nguyenthanh.asia/what-makes-a-good-ai-assistant-6069ba267e5a?source=rss-2cdc6c033817------2</link>
            <guid isPermaLink="false">https://medium.com/p/6069ba267e5a</guid>
            <category><![CDATA[design]]></category>
            <category><![CDATA[chatbots]]></category>
            <category><![CDATA[empathy]]></category>
            <category><![CDATA[excel]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Nguyen Ha Thanh]]></dc:creator>
            <pubDate>Sun, 28 Jun 2026 14:42:52 GMT</pubDate>
            <atom:updated>2026-06-28T14:46:49.357Z</atom:updated>
            <content:encoded><![CDATA[<h4>In customer experience, the best AI is not the one that sounds the most human. It is the one that helps customers reach an outcome.</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*rCgPu7efGZCdXJ0DeX29bw.jpeg" /></figure><p>In the age of AI models, we seem to encounter AI everywhere.</p><p>Open an app, and there is an AI recommendation. Visit a website, and an AI assistant appears. Use workplace software, and there is an AI copilot. Contact customer support, and there is often a chatbot or an AI agent. On social media, people talk about AI every day as if it is the next major revolution.</p><p>In business conversations, the question “Will AI replace humans?” appears constantly, sometimes even more often than the question “How can AI help humans do better work?”</p><p>Startups are moving fast in this race too. Many teams claim they have powerful AI, strong models, smart automation, intelligent agents, and the ability to transform entire industries. For some companies, simply adding the word AI to a product makes it more attractive to the market. Some teams raise significant funding from investors because their technology story sounds convincing.</p><p>But after two everyday customer service experiences with AI assistants, I have a slightly different view.</p><p>I do not think the most important question is whether the AI is powerful. I also do not think the most important question is whether AI can replace humans.</p><p>The more practical question is this:</p><p><strong>Can the AI solve a real customer problem?</strong></p><p>Because at the end of the day, customers do not live inside pitch decks. Customers do not evaluate AI by the number of parameters, the model architecture, or the fundraising story.</p><p>Customers evaluate AI through a very specific experience:</p><p><strong>Am I being helped? Is my issue being handled? Do I have to spend more time on the same problem again?</strong></p><h3>The first experience: a pleasant voice, but no real progress</h3><p>The first story was an interaction with an AI assistant that had a very pleasant voice.</p><p>At first, it created a good impression. The wording was polite. The tone was soft. The way it responded felt friendly. Every time I described the issue, it replied with sentences that sounded considerate. It understood the inconvenience. It apologized for the poor experience. It said it would try to support me.</p><p>On the surface, this looked like a carefully designed AI assistant.</p><p>It was not rude. It was not cold. It did not answer like a machine. It knew how to use language that could make the user feel slightly less frustrated in the first few seconds.</p><p>But the problem started after that.</p><p>I asked about a specific issue. It gave a fairly long answer, but it did not provide a clear next step. I asked again. It apologized and repeated the same idea in a different way. I gave it more information. It thanked me, apologized again, and continued to lead me in circles.</p><p>Sometimes it suggested very basic steps, steps I had already tried or steps that were not directly relevant to the problem.</p><p>Sometimes it said it “understood”, but the content of its response showed that it did not really understand the issue.</p><p>Sometimes I felt that it was trying to keep the conversation going instead of moving the problem toward resolution.</p><p>The most frustrating part was not that it was immediately wrong.</p><p>The most frustrating part was that it created the impression of support, while in reality there was no progress.</p><p>I did not know where my case stood. I did not know who was responsible. I did not know what the next step was. I did not know whether I needed to wait, how long I needed to wait, or whether I should contact another channel.</p><p>Every response returned to polite language, but no concrete action.</p><p>At some point, I became frustrated. I scolded it.</p><p>It stayed calm and polite. It apologized. It said it understood my frustration. It admitted that the experience had not been good enough. It continued to use a very soft tone.</p><p>But the problem was still not solved.</p><p>After almost a full day, what I received was not a result.</p><p>It was a long conversation. Many polite sentences. Many apologies. Many signs that the system had been designed to calm the user down. But very few signs that the system had been designed to actually close the issue.</p><p>That experience made me think:</p><p><strong>An AI assistant may sound very pleasant, but if it does not have the authority to act, does not have a clear process, does not know when to escalate, and does not help the customer reach an outcome, then in the end it is only a layer of makeup placed on top of a weak system.</strong></p><h3>The second experience: no performance, just resolution</h3><p>The second story was very different.</p><p>I had an issue that required a refund. This time, I did not open a chat. I did not go back and forth in a conversation. I did not hear any friendly voice. I simply sent an email requesting a refund.</p><p>I did not expect much.</p><p>Usually, with something like a refund, users prepare themselves to wait. Maybe there will be an automated email. Maybe they will be asked to provide more information. Maybe they will have to explain the reason again. Maybe they will have to wait for a human agent to check. Sometimes they may even need to follow up several times.</p><p>But around 30 seconds later, I received a response.</p><p>It was not long. It was not decorative. It did not try too hard to sound friendly.</p><p>The response went straight to the point.</p><p>The refund request was processed. Right after that, the receipt was sent to my email.</p><p>Everything ended very quickly.</p><p>The interesting part is that I did not even have time to think about whether it was AI or a human. And at that moment, it did not really matter.</p><p>As a customer, I did not need to know whether the system behind it was AI, automation, or a hybrid system between AI and humans.</p><p>I only knew that my issue had been understood correctly, handled properly, and closed within a very short time.</p><p>There was no “wow” experience in the sense of a technology demo. There was no impressive interface. There was no warm voice.</p><p>But there was a result.</p><p>And in customer experience, the result is what customers remember most.</p><h3>Conversation is not the same as problem solving</h3><p>When placed side by side, these two stories show something very clearly:</p><p><strong>Whether an AI assistant is good or bad is not only about the model. It is about the process designed behind it.</strong></p><p>A system designed for conversation will try to keep the conversation going.</p><p>A system designed for problem solving will try to end the problem as quickly as possible.</p><p>That is the difference.</p><p>One had a good voice but a weak process. The other did not need to show off its voice, but had a strong process.</p><p>One made me feel temporarily soothed but still stuck. The other did not try to soothe me, but made the frustration disappear by solving the issue.</p><p>This is a common mistake when many organizations deploy AI assistants.</p><p>They start with the question:</p><p>How can we make AI sound more human? How can the answer feel softer? How can the chatbot have a more pleasant personality? How can users feel that they are speaking with a polite person?</p><p>These questions are not wrong.</p><p>But if they are placed first, the product can easily move in the wrong direction.</p><p>The right question should be:</p><p><strong>What is the customer trying to get done, and what must the system do to help that task get completed as quickly, clearly, and painlessly as possible?</strong></p><p>Customers do not contact support to enjoy a pleasant conversation.</p><p>Customers contact support because they have a problem that needs to be solved.</p><p>They were charged incorrectly. They want a refund. They need to change a ticket. They cannot log in. They need an invoice. They received the wrong item. They do not understand why their account was locked. They want to know when their order will arrive.</p><p>They need a concrete action from the company.</p><p>In those moments, politeness is necessary, but not enough.</p><p>A sentence like “we are sorry” cannot replace a refunded payment.</p><p>A sentence like “we understand your inconvenience” cannot replace an accurate tracking link.</p><p>A sentence like “thank you for your patience” cannot replace being transferred to the right person who has the authority to act.</p><p><strong>A good AI assistant is not an AI that apologizes well. A good AI assistant is an AI that gives customers fewer reasons to become angry in the first place.</strong></p><h3>A good AI assistant must be part of the operation</h3><p>To do that, an AI assistant must be designed as part of an operational process, not as a conversational layer sitting on the outside.</p><p>It must understand the customer’s intent.</p><p>It must classify the issue correctly.</p><p>It must know which information needs to be requested and which information the system already has, so the customer does not have to repeat it.</p><p>It must be able to access relevant data, such as orders, payments, interaction history, refund conditions, and case status.</p><p>It must have the authority to take certain actions within safe boundaries.</p><p>It must know when to handle the issue by itself and when to escalate to a human.</p><p>And when it escalates to a human, it must pass along full context so the customer does not have to tell the whole story again.</p><p>Without those things, AI is only standing at the entrance of the system and talking to the customer.</p><p>It is not actually opening any doors.</p><h3>The wrong metrics create the wrong AI</h3><p>This is also why the way we measure AI assistants has to change.</p><p>If we only measure the number of messages answered, the number of tickets touched, or the average response time, companies can easily become satisfied with the wrong thing.</p><p>The dashboard may look beautiful: AI responds quickly, interaction volume is high, cost per support interaction goes down.</p><p>But from the customer’s side, the question remains:</p><p><strong>Has my problem been solved?</strong></p><p>An AI assistant that responds in three seconds but makes the customer spend three hours going in circles is not fast.</p><p>A chatbot that handles 10,000 conversations but still forces most customers to return a second or third time is not efficient.</p><p>A support system that sounds polite but does not reduce the pain point is not customer-centric.</p><p>It is only the automation of delay.</p><p>A weak process does not automatically become strong when AI is added to it.</p><p>It only becomes a weak process that runs faster, sounds better, and generates more data.</p><p>This is the core point:</p><p><strong>AI cannot fix an organization that does not understand its customers.</strong></p><p>AI can help teams respond faster. AI can summarize better. AI can classify better. AI can automate many steps.</p><p>But AI cannot replace human-centered design thinking.</p><p>If the people designing the system do not understand where the customer is hurting, stronger technology can simply amplify the mistake.</p><h3>Empathy is not soft. It is strategic.</h3><p>That is why empathy is not a soft concept.</p><p>Empathy is a strategic capability.</p><p>Empathy is not just about saying kind words. Empathy is understanding how much frustration, anxiety, wasted time, and lost trust can sit behind a simple request.</p><p>Empathy is understanding that when a customer asks “where is my refund?”, they do not only need an answer.</p><p>They need certainty. They need a clear status. They need to know whether their money is safe. They need to know whether the company is taking responsibility.</p><p>Empathy is the ability to see the process through the eyes of the person trapped inside that process.</p><p>That is also why at VinUni, Empathy is always placed in a very important position within EXCEL.</p><p>This should not be seen as a value written only to look good. Empathy is treated as the starting point of meaningful design.</p><p>If human pain is not understood correctly, the capabilities that follow can easily be used in the wrong direction.</p><p>Strong capability can create a product that is complex but not truly needed.</p><p>Creativity can create something new but irrelevant to any real pain.</p><p>Entrepreneurial spirit can make people run very fast, but in the wrong direction.</p><p>Leadership thinking can build a system that looks efficient on paper, but exhausts people in reality.</p><h3>Three layers of a good AI assistant</h3><p>For me, a good AI assistant needs three layers:</p><p><strong>Empathy identifies the pain.</strong></p><p><strong>Process design removes the friction.</strong></p><p><strong>Execution closes the loop.</strong></p><p>Without empathy, the system may solve the wrong problem.</p><p>Without process design, the system may understand the problem but not know what to do.</p><p>Without execution, the system may have the right idea but fail to produce a result.</p><p>From that perspective, AI is only one part of the problem.</p><p>A good model is necessary, but not enough.</p><p>Good data is necessary, but not enough.</p><p>A good interface is necessary, but not enough.</p><p>What matters is whether the entire system has been designed around solving the customer’s real problem.</p><h3>The best AI assistant closes the loop</h3><p>The 30-second refund experience stayed with me not because it was “friendly”.</p><p>It stayed with me because it removed a frustration immediately. It closed a loop. It gave me back my time. It made me feel that the system behind it knew what it was doing.</p><p>The polite but circular AI assistant also stayed with me, but in the opposite way.</p><p>It showed me that pleasantness cannot save a process that has no ability to act.</p><p>An automated apology, when repeated too many times, no longer feels like empathy. It becomes evidence that the system does not truly understand what the customer needs.</p><p>The future of AI in customer service will not belong to the organizations that make chatbots sound the most human.</p><p>It will belong to the organizations that understand humans the best, design the best processes, and use AI as a tool to complete what the customer actually needs.</p><p>Powerful AI is good.</p><p>But powerful AI placed on top of a bad process still creates a bad experience.</p><p>On the other hand, a well-designed process with empathy, data, authority to act, and success metrics based on outcomes can become extremely powerful when accelerated by AI: fast, clear, low-friction, and trustworthy.</p><p>In the end, customers do not need to know what technology the company is using behind the scenes.</p><p>Customers only need to know:</p><p><strong>Was my problem understood correctly?</strong></p><p><strong>Was it handled properly?</strong></p><p><strong>Do I have to come back for the same issue again?</strong></p><p>If the problem is understood correctly, solved properly, and the customer does not have to come back again, that AI assistant is good.</p><p>If not, no matter how well it speaks, it has failed.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=6069ba267e5a" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[From Thang Long Bridge to an International Environment Within Vietnam]]></title>
            <link>https://medium.com/@nguyenthanh.asia/from-thang-long-bridge-to-an-international-environment-within-vietnam-8850413d6fdc?source=rss-2cdc6c033817------2</link>
            <guid isPermaLink="false">https://medium.com/p/8850413d6fdc</guid>
            <category><![CDATA[engineering]]></category>
            <category><![CDATA[history]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[international-relations]]></category>
            <dc:creator><![CDATA[Nguyen Ha Thanh]]></dc:creator>
            <pubDate>Wed, 06 May 2026 02:12:58 GMT</pubDate>
            <atom:updated>2026-05-06T02:12:58.586Z</atom:updated>
            <content:encoded><![CDATA[<p>In Vietnam, the month of May is often remembered through major historical milestones: victory, resilience, renewal, and the country’s continuous movement forward. Within that shared memory, there is also a special May date for Vietnam’s transport and engineering sector: May 9, 1985, the day Thang Long Bridge was officially inaugurated after nearly eleven years of construction. The bridge had been started on November 26, 1974, during a period when the country was still facing immense wartime and post-war challenges.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*5NXtnKNeXiUwAfu1PZ_KFw.png" /><figcaption>From past to present, Thang Long Bridge stands as more than a structure across the Red River. It is a bridge of engineering, resilience, international cooperation, and the enduring capacity Vietnam continues to build across generations.</figcaption></figure><p>For younger generations, Thang Long Bridge may simply appear as one of the familiar bridges across the Red River. But at the time it was built, it was far more than an ordinary infrastructure project. It was one of the largest bridge projects in Vietnam and among the major bridge works in Southeast Asia of that period. It was also the second bridge across the Red River in Hanoi after Long Bien Bridge, marking a major step in expanding the capital’s connection with the northern region.</p><p>Thang Long Bridge did more than connect the two banks of the Red River. It connected a period of great difficulty with a new stage of national capability. It connected engineering, organization, discipline, and trust. It connected Vietnamese professionals with international partners through a concrete project that was useful, demanding, and built to last.</p><p>The scale itself was remarkable. The main bridge is about 1,688 meters long, with 15 spans, each about 112 meters, supported by 14 piers and 2 abutments. The bridge was designed with two levels: the upper level for road traffic, and the lower level for railway and other forms of transport. The upper deck was about 19.5 meters wide, including four lanes for motor vehicles and pedestrian paths, while the lower level carried railway tracks and lanes for non-motorized traffic.</p><p>These numbers matter because they show the level of ambition. This was not a small symbolic project. It required large-scale engineering, serious coordination, new construction methods, and the ability to manage complex technical standards over many years.</p><p>The construction process itself carried many elements of a historical achievement. The project went through interruption and uncertainty. It was initially supported by China, but by the end of the 1970s, that support stopped while the work was still incomplete. After that, the Soviet Union provided major assistance to continue and complete the bridge, including technical expertise, steel structures, high-grade cement, equipment, and machinery.</p><p>The Soviet contribution was substantial. Reports note that the Soviet Union supplied around 49,000 tons of steel, 26,000 tons of steel bridge girders, nearly 60,000 tons of high-grade cement, and hundreds of tons of machinery and equipment for the project.</p><p>But the deeper value of the project was not only in the materials delivered. It was in the standards, methods, and professional capacity that were developed through the work. Vietnamese engineers and workers had to master demanding techniques, strict construction discipline, and new forms of coordination. The construction site became a place where international cooperation was translated into practical capability.</p><p>The technology transfer behind Thang Long Bridge should also be understood in its historical context. For decades, Vietnam had been forced to devote enormous human, material, and financial resources to war, national defense, and post-war recovery. That reality left limited room for long-term technological accumulation, industrial modernization, and large-scale infrastructure experimentation.</p><p>In that context, the construction of Thang Long Bridge became more than an engineering project. It became a channel through which Vietnamese engineers and workers could work directly with advanced methods, demanding standards, and experienced international specialists. The value of cooperation was therefore not only in what was provided, but in what was absorbed, practiced, and retained by Vietnam’s own technical workforce.</p><p>This is an important point for younger generations to understand. Technology transfer is not a passive act of receiving. It requires people capable of learning under pressure, adapting knowledge to local conditions, and turning external expertise into internal capacity. That was one of the quiet achievements of Thang Long Bridge.</p><p>That is why Thang Long Bridge is often remembered as a symbol of Vietnam-Soviet friendship. That is true. But its deeper value lies in how cooperation became lasting capacity. International collaboration was not simply present at the construction site; it entered the working method, the technical standards, and the professional growth of Vietnamese engineers and workers.</p><p>Many Vietnamese engineers matured through that project. They worked with new technologies, strict requirements, complex construction conditions, and international experts. The bridge became not only an infrastructure achievement, but also a training ground for a generation of professionals.</p><p>For me, this story is not distant history. My mother was one of the engineers who participated in the construction of Thang Long Bridge. Because of that, I do not see the bridge only as a public landmark. I also see it as part of a family memory, a professional legacy, and a reminder that great projects are built by real people: engineers, workers, organizers, teachers, learners, and those who quietly carry responsibility through difficult years.</p><p>Today, as we are building an international environment within Vietnam, the story of Thang Long Bridge offers a meaningful lens. Every generation has its own bridges to build. Only the materials change.</p><p>For Thang Long Bridge, the materials were steel, concrete, machinery, technical drawings, welding standards, bridge spans, piers, and construction discipline. For us today, especially in an academic environment, the materials are knowledge, research culture, educational standards, international collaboration, institutional capacity, and people.</p><p>Now, working in academia, I see a strong connection between that bridge and the kind of international environment we are building in Vietnam today.</p><p>A truly international academic environment is not created by simply importing foreign models, using English-language labels, or adding international symbols on the surface. It is built when global standards are translated into daily practice: how we teach, how we research, how we discuss ideas, how we evaluate quality, how we collaborate, and how we develop people.</p><p>This is where the story of Thang Long Bridge becomes relevant again. A bridge cannot stand on slogans. It requires design, calculation, materials, inspection, coordination, and discipline. In the same way, an academic environment cannot become international through appearance alone. It needs substance: research integrity, intellectual honesty, openness to debate, high expectations, and systems that allow people to grow.</p><p>The bridge also reminds us that international cooperation becomes meaningful when it produces capability that remains. A completed bridge is visible. But behind it are less visible achievements: trained engineers, improved methods, stronger technical confidence, and a culture of working with demanding standards.</p><p>That lesson is directly relevant to academic internationalization. The real outcome is not only the number of partnerships, visiting professors, exchange programs, or international events. Those things matter, but they are not the final measure. The deeper question is: what capability remains? What standards become part of the institution? What habits of thinking and working are developed? What kind of people are formed through the process?</p><p>A good international environment should leave behind stronger students, stronger teachers, stronger researchers, stronger institutions, and stronger trust between Vietnam and the world.</p><p>There is also another important similarity. Thang Long Bridge was built from a real national need. It was not a decorative project detached from life. It served transportation, connection, development, and the movement of people and goods. Its meaning came from its usefulness.</p><p>The same should be true for academic work. International collaboration should not be decorative. It should respond to real needs: better education, better research, better problem-solving capacity, better links between knowledge and society, and better preparation for the next generation.</p><p>When international cooperation is grounded in real needs, it becomes durable. When it is only built for image, it fades quickly.</p><p>Looking back at Thang Long Bridge in these days of May, I see more than a historical achievement. I see a way of building: start from a real need, work with high standards, collaborate with trust, learn through practice, and leave capacity for the future.</p><p>That spirit remains relevant today.</p><p>The bridges we build now may no longer rise above rivers. Some are built in classrooms, laboratories, research centers, libraries, conferences, joint programs, and daily conversations between people from different cultures. Some are built through academic discipline. Some are built through the patience of mentoring students. Some are built through the courage to raise standards and the humility to keep learning.</p><p>But the essence is the same.</p><p>A bridge is not only a structure. It is a commitment to connection.</p><p>Thang Long Bridge connected land, people, technology, and friendship. The international academic environment we are building today can connect Vietnam with the world in a new form: through knowledge, standards, shared work, and long-term trust.</p><p>And perhaps that is the most meaningful continuity.</p><p>From one generation to another, Vietnam continues to build bridges.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=8850413d6fdc" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Names in a Globalized Scientific Community]]></title>
            <link>https://medium.com/@nguyenthanh.asia/names-in-a-globalized-scientific-community-d4422b750356?source=rss-2cdc6c033817------2</link>
            <guid isPermaLink="false">https://medium.com/p/d4422b750356</guid>
            <category><![CDATA[names]]></category>
            <category><![CDATA[academic]]></category>
            <category><![CDATA[culture]]></category>
            <category><![CDATA[research]]></category>
            <dc:creator><![CDATA[Nguyen Ha Thanh]]></dc:creator>
            <pubDate>Thu, 26 Feb 2026 12:41:32 GMT</pubDate>
            <atom:updated>2026-02-26T12:41:32.558Z</atom:updated>
            <content:encoded><![CDATA[<p>I am Vietnamese, and I have lived in Japan for nearly a decade. Working with international colleagues across Asia, Europe, and North America, one recurring issue keeps appearing in manuscripts, databases, conference badges, and citation systems: Names do not follow a single global logic. But our academic infrastructure still assumes that they do.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*whPENCQP0JQM6Cf_nbuNrA.jpeg" /></figure><p>This is no longer a cultural detail. It is a structural problem that affects academic credit, visibility, and the very identity of researchers from non-Western backgrounds. Misformatting a name is not a cosmetic error; it affects citation counts, database aggregation, tenure reviews, and grant tracking. The infrastructure that underpins global science is outdated, and it’s time for a serious policy discussion.</p><h3>The Root of the Problem: A Monocultural Framework</h3><p>All major academic systems — from citation styles like APA and MLA to databases like Scopus and Web of Science — were designed on one fundamental assumption: the Western naming convention is the default. They require fields like “First name,” “Middle name,” and “Last name.”</p><p>But a vast portion of the world does not fit this model. Globally, we observe:</p><ul><li>Surname-first systems (e.g., China, Japan, Korea, Vietnam, Hungary)</li><li>No-surname systems (e.g., Myanmar, parts of Indonesia)</li><li>Multi-word given names that are often mistaken for middle and last names</li><li>Patronymic and matronymic systems, where a parent’s name is used instead of a fixed family name</li><li>Variable transliteration, where the same name can have multiple spellings in the Latin alphabet</li></ul><p>The result is a fragmented and distorted representation of a scholar’s identity, leading to significant professional consequences.</p><p>Let’s examine a few cases to understand the depth of the issue.</p><h3>1. Vietnamese Names: The Baseline Problem</h3><p>Structure: Family name + Middle name + Given name</p><p>Example: Nguyễn Văn An</p><ul><li>Family name: Nguyễn</li><li>Given name: An</li></ul><p>The difficulty here is twofold. First, the given name, which is the primary name used for addressing someone, appears last. Second, systems often misidentify the last word of the name (“An”) as the family name, or they incorrectly label the middle name (“Văn”) as the given name. This already causes indexing errors, ORCID mismatches, and citation fragmentation. For instance, a publication might be indexed under “An, N. V.” or “Văn, N. A.”, splitting the author’s work across multiple profiles.</p><h3>2. Japanese Names: The Reversal Dilemma</h3><p>Traditional order: Family name + Given name</p><p>Example: Tanaka Hiroshi</p><ul><li>Family name: Tanaka</li><li>Given name: Hiroshi</li></ul><p>In Japan, domestic usage always keeps the surname first. However, for decades, a common practice in English publications was to reverse the order to fit the Western convention (Hiroshi Tanaka). In 2019, the Japanese government officially encouraged keeping the original Surname + Given Name order in international contexts. Now, we see both formats in circulation:</p><ul><li>Tanaka Hiroshi (policy-aligned)</li><li>Hiroshi Tanaka (Western-adapted)</li></ul><p>This creates two distinct identities for the same person in academic databases, making it difficult to track their complete body of work.</p><h3>3. Chinese Names: Ambiguity in Romanization</h3><p>Structure: Family name + Given name</p><p>Example: Wang Xiaoming</p><ul><li>Family name: Wang</li><li>Given name: Xiaoming</li></ul><p>Chinese names face several challenges. The given name, which is often two characters, can be written as one word (Xiaoming), two words (Xiao Ming), or hyphenated (Xiao-Ming). Authors themselves may not be consistent. Furthermore, different romanization systems like Pinyin and Wade-Giles can produce different spellings for the same name. A single scholar might appear as:</p><ul><li>Wang, X. M.</li><li>Wang, Xiaoming</li><li>Wang, Xiao-Ming</li></ul><p>This again leads to a fragmented citation identity.</p><h3>4. Thai Names: Long Surnames and Nicknames</h3><p>Thailand introduces a different complexity. Thai surnames are often long and multi-syllabic, and individuals commonly use short nicknames in social and even professional settings. While formal documents use the full legal name (Given name + Family name), academic systems often truncate long surnames, and conference badges might even display a nickname by mistake. This creates confusion and makes it hard to identify researchers correctly.</p><h3>5. Myanmar (Burmese) Names: The Absence of Surnames</h3><p>Perhaps the most structurally different case is Myanmar, where there is often no family name. A name like Aung San Suu Kyi is a single, complete personal name. There is no inherited surname in the Western sense. Academic systems that have mandatory “First name” and “Last name” fields force an artificial segmentation of the name, creating “fake” surnames and fundamentally distorting the person’s identity.</p><h3>6. Indian Names: A Multiplicity of Systems</h3><p>India is not one system — it is many. Naming conventions vary significantly by region and community:</p><ul><li>North India: Often Given name + Family name (e.g., Rohit Sharma).</li><li>South India: Often Initial + Given name (e.g., R. Madhavan), where the initial might represent the father’s name (a patronymic), not a surname.</li></ul><p>Many Indians do not use a family name at all, while others use caste names or follow other conventions. International databases that flatten this diversity into a First / Middle / Last structure fail to capture the true identity of the individual.</p><h3>Open Questions for a Global Standard</h3><p>Instead of forcing researchers from around the world to adapt to a flawed, monocultural system, we should be asking how to build a truly global infrastructure. Here are some open questions that could guide a policy discussion:</p><p><strong><em>1. Should academic systems abandon “First/Last” fields?</em></strong></p><p>We could replace them with a more flexible structure:</p><ul><li>Full Legal Name: A single, unparsed string.</li><li>Family Name: An optional field for cultures that use it.</li><li>Preferred Citation Name: Allowing the author to specify how they want to be cited.</li><li>ORCID: A mandatory, persistent digital identifier.</li></ul><p><strong><em>2. Should citation format be author-controlled?</em></strong></p><p>Instead of forcing Nguyen, V. A., why not allow Nguyen Van An if that is how the author chooses to be indexed? This would respect the author’s identity and cultural background.</p><p><strong><em>3. Should ORCID become the primary identifier, not the surname?</em></strong></p><p>Machine indexing no longer needs to rely on surname logic. We have persistent digital identifiers that are far more reliable. Why are we still optimizing for 19th-century bibliography systems?</p><p><strong><em>4. Should journals adopt a policy of “native-order respect”?</em></strong></p><p>This would mean printing names as they are used in their original culture: Tanaka Hiroshi, Wang Xiaoming, Nguyễn Văn An. If consistency is the goal, whose cultural logic should that consistency be based on?</p><h3>The Path Forward: A Call for Structural Change</h3><p>The core tension is clear: globalization increases collaboration, but our systems for standardization still follow a single cultural template. We have two strategic paths:</p><ol><li>Force everyone into one global naming format. This is administratively easier but intellectually dishonest and disrespectful.</li><li>Redesign our metadata systems to accept structural diversity. This is more complex but is the only way to build a truly inclusive and accurate global scientific community.</li></ol><p>Names are not formatting noise. They are markers of identity, legality, and academic credit. The question is no longer, “How do we adapt Asian names to Western systems?” The better question is: In a multipolar scientific world, what should a truly global name standard look like?</p><p>That is a policy discussion waiting to happen, and it needs to involve not just journal editors and database engineers, but researchers from all over the world.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=d4422b750356" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Why LLMs Cannot Be the Answer to Super Intelligence]]></title>
            <link>https://medium.com/@nguyenthanh.asia/why-llms-cannot-be-the-answer-to-super-intelligence-70e49ea2c8b1?source=rss-2cdc6c033817------2</link>
            <guid isPermaLink="false">https://medium.com/p/70e49ea2c8b1</guid>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[superintelligence]]></category>
            <category><![CDATA[no-free-lunch]]></category>
            <dc:creator><![CDATA[Nguyen Ha Thanh]]></dc:creator>
            <pubDate>Wed, 31 Dec 2025 00:40:57 GMT</pubDate>
            <atom:updated>2025-12-31T00:40:57.723Z</atom:updated>
            <content:encoded><![CDATA[<h3>Because They Don’t Get a Free Lunch</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*IBZBJAqpemC7-3F8tPqxFw.png" /></figure><p>In recent years, <strong>Large Language Models (LLMs)</strong> have been aggressively marketed as the inevitable path toward <em>Super Intelligence</em>.</p><p>Scale more data. Scale more compute. Scale more context.</p><p>And somehow, intelligence will <em>emerge</em>.</p><p>It sounds elegant.<br>It’s also wrong at the root.</p><p>The problem isn’t engineering.<br>The problem is <strong>No Free Lunch</strong>.</p><h3>1. Super Intelligence Requires a “Free Lunch”</h3><p>LLMs Don’t Get One</p><p>The <strong>No Free Lunch (NFL) Theorem</strong> is brutally simple:</p><blockquote><em>No algorithm performs optimally across all problems without assumptions about the data distribution.</em></blockquote><p>Super Intelligence, in any practical sense, must:</p><ul><li>Adapt to <em>any</em> environment</li><li>Solve <em>previously unseen</em> problems</li><li>Reason beyond known distributions</li></ul><p>In other words:<br> 👉 it must win <strong>without being pre-fed the rules</strong>.</p><p>LLMs, by design:</p><ul><li>Consume historical data</li><li>Compress statistical patterns</li><li>Depend heavily on human-generated distributions</li></ul><p>LLMs don’t get a free lunch.<br>They <strong>eat humanity’s leftovers</strong>.</p><h3>2. LLMs Look Smart Because the Human World Is Predictable</h3><p>LLMs don’t defeat NFL.<br>They exploit the fact that <strong>human reality is highly structured</strong>.</p><p>Language has grammar.<br>Culture repeats itself.<br>Everyday logic is full of shortcuts.<br>Humans are noisy — but <em>boringly consistent</em>.</p><p>LLMs thrive in that regime.</p><p>Super Intelligence does not get that luxury. It must function when structure collapses.</p><p>When:</p><ul><li>Objectives conflict</li><li>Environments are adversarial</li><li>Data is sparse or misleading</li><li>Feedback is delayed or wrong</li></ul><p>That’s where LLMs start to crack.</p><h3>3. Scaling Doesn’t Break No Free Lunch</h3><p>It Just Makes Lunch Expensive</p><p>A common belief:</p><blockquote><em>“If we scale enough, intelligence will emerge.”</em></blockquote><p>This is a category error.</p><p>Scaling gives you:</p><ul><li>Sharper patterns</li><li>Stronger biases</li><li>More convincing hallucinations</li></ul><p>NFL doesn’t say <em>small models fail</em>.<br>NFL says <strong>no model wins universally</strong>.</p><p>As LLMs scale:</p><ul><li>Context ↑ → Cost ↑</li><li>Alignment ↑ → Reasoning freedom ↓</li><li>Generality ↑ → Depth ↓</li></ul><p>The trade-offs never disappear.<br><em>They’re just hidden behind money.</em></p><p>There is no free lunch.<br>Only <strong>massive bills</strong>.</p><h3>4. Fine-Tuning, RAG, and Tool Use Are Admissions of Insufficiency</h3><p>If LLMs were truly on the path to Super Intelligence:</p><ul><li>No fine-tuning would be needed</li><li>No RAG</li><li>No guardrails, rules, or human-in-the-loop</li></ul><p>Reality says otherwise.</p><p>The more we deploy, the more we patch.<br>The more we scale, the more we constrain.</p><p>That alone reveals the truth:</p><blockquote><em>LLMs cannot stand on their own in open worlds.</em></blockquote><p><strong>Super Intelligence must.</strong></p><h3>5. Super Intelligence Doesn’t Consume Distributions</h3><p>It Creates New Ones</p><p>The fundamental difference:</p><ul><li><strong>LLMs</strong> learn from an existing world</li><li><strong>Super Intelligence</strong> reshapes the world to learn better</li></ul><p>LLMs react.<br>Super Intelligence acts.</p><p>LLMs optimize next-token prediction.<br>Super Intelligence optimizes <strong>goals under uncertainty</strong>.</p><p>These are not adjacent problems.<br>They’re different classes.</p><h3>LLMs Aren’t Wrong — They’re Just Miscast</h3><p>LLMs are:</p><ul><li>Compressors of collective knowledge</li><li>Powerful interfaces</li><li>Force multipliers for human intelligence</li></ul><p>They are not:</p><ul><li>Foundations of Super Intelligence</li><li>Universal problem solvers</li><li>Systems that get a free lunch</li></ul><p>No Free Lunch still holds.<br>And if Super Intelligence ever exists,<br>it will have to <strong>cook its own lunch</strong>.</p><p>LLMs don’t.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=70e49ea2c8b1" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Epistemology Crisis in AI-Assisted Development: When Not Knowing Becomes the New Normal]]></title>
            <link>https://medium.com/data-science-collective/the-epistemology-crisis-in-ai-assisted-development-when-not-knowing-becomes-the-new-normal-fc8c8e2a1788?source=rss-2cdc6c033817------2</link>
            <guid isPermaLink="false">https://medium.com/p/fc8c8e2a1788</guid>
            <category><![CDATA[vibe-coding]]></category>
            <category><![CDATA[technical-debt]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[crisis]]></category>
            <category><![CDATA[epistemology]]></category>
            <dc:creator><![CDATA[Nguyen Ha Thanh]]></dc:creator>
            <pubDate>Tue, 28 Oct 2025 08:05:33 GMT</pubDate>
            <atom:updated>2026-01-04T18:22:18.603Z</atom:updated>
            <content:encoded><![CDATA[<p><em>How AI coding tools are fundamentally changing what it means to ‘understand’ your code — and why that matters for production systems</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*YDZan1tVcETh3KqGRDOc7Q.png" /><figcaption>Depicting the epistemology crisis in AI-driven software development — a contrast between human chaos and machine-generated perfection concealing hidden flaws. (Image generated by AI)</figcaption></figure><h3>The Two Disasters</h3><p>Imagine two scenarios unfolding in parallel at two different tech companies. At Company A, a critical production service, built largely with AI-generated code, suddenly fails. The on-call engineer scrambles to diagnose the issue, but the logs are cryptic, and the code is a labyrinth of elegant but unfamiliar logic. When asked why the system, which had been working perfectly for weeks, suddenly broke, the engineer can only answer with a frustrating honesty: “I don’t know why it worked in the first place.”</p><p>Meanwhile, at Company B, a security audit reveals a subtle but devastating vulnerability in their customer authentication service. The flaw, introduced by an AI coding assistant, allowed unauthorized access to sensitive user data. The development team is stunned. They had reviewed the code, and it seemed to function correctly. But no one on the team could fully articulate the authentication flow’s intricate logic. They had outsourced their understanding to the machine, and in doing so, had created a security blind spot large enough to drive a truck through. These aren’t just tales of technical debt; they are symptoms of a deeper, more insidious problem: an epistemology crisis in software development.</p><h3>The Epistemology Crisis</h3><p>Epistemology is the theory of knowledge — it’s concerned with what it means to “know” something. In traditional software development, the relationship between a developer and their code is intimate. You write the code, you debug it, you refactor it. Through this process, you build a mental model of the system. You know it. But AI-assisted development is fundamentally altering this relationship. We are moving from a world where we create code to a world where we prompt for it. This is a shift from procedural knowledge (“knowing how”) to declarative knowledge (“knowing that”).</p><p>As I wrote in a previous article, this can lead to a state of cognitive dissonance: “My code doesn’t work, I don’t know why. It works, I don’t know why.” When we rely on AI to generate complex logic, we risk becoming mere curators of code we don’t truly understand. This isn’t just a problem for junior developers; it’s a challenge for the entire industry. We are trading understanding for speed, and the cost of that trade is not always immediately apparent.</p><h3>The Three Levels of Not Knowing</h3><p>This lack of understanding isn’t a monolithic problem. It manifests at different levels of abstraction, each with its own set of risks.</p><ul><li>Level 1: Syntax Uncertainty. This is the most superficial level of not knowing. An AI tool generates a line of code using an obscure library or a clever one-liner. You might not understand the syntax immediately, but with a bit of research, you can figure it out. The risk here is relatively low, but it can slow down debugging and maintenance.</li><li>Level 2: Logic Opacity. This is where things get more dangerous. The AI chooses a particular algorithm or data structure, and you don’t understand why it made that choice. The code works, but the reasoning behind it is a black box. This is common in “vibe coding,” where a developer simply describes the desired outcome and lets the AI figure out the implementation. The risk here is that the AI’s solution might have edge cases or performance characteristics that you are not aware of.</li><li>Level 3: Architectural Mystery. This is the most perilous level of not knowing. The AI generates entire modules or services, and you don’t understand how they fit into the larger system. You might have a collection of well-functioning black boxes, but you lack the holistic, systems-level understanding to reason about their interactions. This is where catastrophic failures and security vulnerabilities are born.</li></ul><p>The accumulation of these “unknowns” leads to what I call epistemological debt: the ever-growing gap between what your system does and what you and your team understand about why it does it.</p><h3>Why This Is Different From Regular Technical Debt</h3><p>Some might argue that this is just a new form of technical debt. But there’s a crucial difference.</p><p>Technical debt is when you know what the right thing to do is, but you choose to do the wrong thing for now. Epistemological debt is when you don’t know what the right thing to do is, and you may not even know that you’re doing the wrong thing.</p><p>Technical debt is a conscious trade-off. Epistemological debt is a blind spot. Technical debt is visible on a backlog. Epistemological debt is invisible until it causes a catastrophic failure. The most dangerous bugs are not the ones you know about and haven’t fixed; they are the ones you don’t know exist, in code you don’t understand.</p><h3>The Cultural Dimension</h3><p>The Western tech culture, famously encapsulated by the mantra “move fast and break things,” often prioritizes speed above all else. This creates a fertile ground for epistemological debt to accumulate. In contrast, many Eastern philosophies emphasize the importance of deep understanding. There is a Vietnamese saying, “Biết người biết ta, trăm trận trăm thắng,” which translates to “Know your enemy and know yourself, and you can fight a hundred battles without defeat.” Applied to software, if you don’t know your system, you cannot defend it, you cannot maintain it, and you cannot evolve it.</p><p>This isn’t to say that speed is bad or that understanding must always be perfect. The challenge is to find a conscious balance. We need to move away from the false dichotomy of speed versus understanding and create a culture that values both.</p><h3>Production Implications</h3><p>The consequences of unchecked epistemological debt are not academic. They have real-world, production implications.</p><ul><li>Debugging: When a system you don’t understand fails, you can’t reason about root causes. You are reduced to a frustrating cycle of trial and error, prompting the AI with variations of “fix this bug.”</li><li>Security: You cannot secure a system you do not understand. The AI might generate code that is functionally correct but contains subtle security flaws that a human expert with deep domain knowledge would have caught.</li><li>Compliance: With regulations like the EU AI Act on the horizon, companies will be required to demonstrate that they have taken “appropriate measures” to ensure the safety and reliability of their AI systems. How can you audit a system when no one on your team fully understands how it works?</li><li>Team Knowledge: In a traditional development team, knowledge is accumulated and shared. With AI-assisted development, knowledge can become fragmented. The “why” behind the code is trapped in the context window of an AI, not in the collective consciousness of the team.</li></ul><h3>The Uncomfortable Truth</h3><p>Here’s the uncomfortable truth: maybe we don’t need to understand everything. In a world of increasing complexity, perhaps it’s no longer feasible for a single developer to have a complete mental model of an entire system. Maybe epistemological debt is a necessary trade-off for the incredible gains in productivity that AI tools provide. As a recent OpenAI paper on LLM hallucinations suggests, we may need to get better at accepting and managing uncertainty.</p><p>But if we are to accept this new reality, we must do so consciously. We need a framework for thinking about and managing epistemological debt. We need to know when a lack of understanding is acceptable and when it is a critical risk.</p><h3>A Practical Framework: The Epistemological Hierarchy</h3><p>To address this challenge, I propose a five-level hierarchy of understanding for AI-generated code. This framework can help teams assess the level of epistemological debt they are carrying and make conscious decisions about when to pay it down.</p><ul><li>Level 0: Black Box. “It works, but I have no idea why.” This level of understanding is only acceptable for throwaway prototypes or non-critical experiments.</li><li>Level 1: Pattern Recognition. “I recognize this pattern, but I couldn’t have written it from scratch.” This is acceptable for standard, boilerplate code like a simple CRUD API.</li><li>Level 2: Logic Understanding. “I understand the logic flow and can explain it to someone else.” This should be the minimum level of understanding for any code that is committed to a production codebase.</li><li>Level 3: Architectural Reasoning. “I understand how this component fits into the larger system and its dependencies.” This is required for any code that is part of a complex, scalable system.</li><li>Level 4: Consequence Prediction. “I can predict how changes to this component will ripple through the system.” This is the highest level of understanding and is required for mission-critical systems.</li></ul><h3>Practical Strategies for Managing Epistemological Debt</h3><p>Moving up the epistemological hierarchy requires a conscious effort. Here are five practical strategies that teams can adopt:</p><ol><li>Demand Explanations, Not Just Code. When using an AI assistant, prompt it to explain its reasoning. “Why did you choose this algorithm? What are the trade-offs of this approach?”</li><li>Make “Explain This” a Code Review Requirement. During code reviews, don’t just ask “Does it work?” Ask the author to explain the AI-generated code. If they can’t, they haven’t reached Level 2 understanding.</li><li>Document the “Why,” Not Just the “What.” Your documentation should capture the reasoning behind architectural decisions, especially for complex, AI-generated components.</li><li>Test for Understanding, Not Just Functionality. Create tests that specifically target the edge cases and assumptions of the AI-generated code. This forces a deeper level of engagement with the logic.</li><li>Know When to Write It Manually. If a component is critical, complex, and novel, it might be better to write it manually. The time spent building a deep understanding will pay dividends in the long run.</li></ol><h3>The Future: Living With Epistemological Uncertainty</h3><p>The rise of agentic coding — where AI not only writes but also tests, debugs, and deploys code — will only accelerate this trend. The core skill for developers in the future will not be writing code, but rather knowing what level of understanding is required for a given task and having the discipline to achieve it.</p><p>There is another piece of Vietnamese wisdom that is relevant here: it is better to know that you don’t know than to think you know when you don’t. In the age of AI, this kind of intellectual humility is not just a virtue; it is a professional necessity.</p><h3>Conclusion</h3><p>AI-assisted development is not just changing how we code; it is changing what it means to understand our code. The question is no longer whether we should use these powerful tools, but how we can use them without abdicating our responsibility as engineers. The challenge is to embrace the speed and power of AI without falling into the trap of epistemological laziness.</p><p>In the age of AI, the most valuable skill isn’t coding — it’s knowing when you need to understand your code, and when you can afford not to.</p><h3>Resources</h3><p>Research &amp; Studies:</p><ul><li><a href="https://cacm.acm.org/research/measuring-github-copilots-impact-on-productivity/">Measuring GitHub Copilot’s Impact on Productivity</a> — Communications of the ACM, 2024</li><li><a href="https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/">Measuring the Impact of Early-2025 AI on Experienced Developer Productivity</a> — METR Research, 2025</li><li><a href="https://arxiv.org/html/2403.06484v1">Technical Debt Management: The Road Ahead</a> — arXiv, 2024</li></ul><p>Regulatory:</p><ul><li><a href="https://www.europarl.europa.eu/topics/en/article/20230601STO93804/eu-ai-act-first-regulation-on-artificial-intelligence">EU AI Act: First Regulation on Artificial Intelligence</a> — European Parliament, 2025</li></ul><p>Related Perspectives:</p><ul><li><a href="https://blog.nguyenthanh.asia/my-code-doesnt-work-i-don-t-know-why-it-works-i-don-t-know-why-58196d377411">My Code Doesn’t Work, I Don’t Know Why. It Works, I Don’t Know Why.</a></li><li><a href="https://medium.com/data-science-collective/the-ai-vibe-coding-paradox-why-experience-matters-more-than-ever-33c343bfc2e1">The AI Vibe Coding Paradox: Why Experience Matters More Than Ever</a></li><li><a href="https://medium.com/data-science-collective/i-failed-to-build-a-learned-hallucination-detector-179c5cde32f6">I Failed to Build a Learned Hallucination Detector</a></li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=fc8c8e2a1788" width="1" height="1" alt=""><hr><p><a href="https://medium.com/data-science-collective/the-epistemology-crisis-in-ai-assisted-development-when-not-knowing-becomes-the-new-normal-fc8c8e2a1788">The Epistemology Crisis in AI-Assisted Development: When Not Knowing Becomes the New Normal</a> was originally published in <a href="https://medium.com/data-science-collective">Data Science Collective</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[When “Knowing Everything” Becomes a Weakness]]></title>
            <link>https://medium.com/@nguyenthanh.asia/when-knowing-everything-becomes-a-weakness-65a99b6d2b2e?source=rss-2cdc6c033817------2</link>
            <guid isPermaLink="false">https://medium.com/p/65a99b6d2b2e</guid>
            <category><![CDATA[cognitive-humility]]></category>
            <category><![CDATA[misinformation]]></category>
            <category><![CDATA[ai-hallucination]]></category>
            <dc:creator><![CDATA[Nguyen Ha Thanh]]></dc:creator>
            <pubDate>Wed, 17 Sep 2025 00:49:20 GMT</pubDate>
            <atom:updated>2025-09-17T00:49:20.962Z</atom:updated>
            <content:encoded><![CDATA[<p>We live in a culture obsessed with closure. We demand answers for every question. We loathe “I don’t know.” But a recent OpenAI <a href="https://openai.com/index/why-language-models-hallucinate/">research paper</a> reminds us this expectation is not only unrealistic — it’s harmful.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*KGHIPRYVJ_baW4SDpfVxgA.png" /></figure><p>The paper shows that large language models (LLMs) are mathematically bound to make errors — “hallucinations” — even with perfect training data. It’s not just messy data; the way these models work (predicting one word at a time) means errors accumulate over long responses.</p><p>Here’s the point most people resist: insisting on confident answers — even when confidence is unwarranted — worsens misinformation. Benchmarks punish models for saying “I don’t know” as harshly as giving wrong facts. So the optimal “strategy” becomes guessing, always pretending to know.</p><p>OpenAI suggests a fix: have models assess their confidence, abstain when they’re not sure, change evaluation metrics so uncertainty isn’t penalized. But this “fix” comes at a cost: user experience degrades. Many users will leave if models constantly respond with “I’m not sure.” They want certainty — even false certainty.</p><h4>The Psychological Weakness in Us</h4><p>Believing every question deserves a definite answer is a mental trap. It’s a craving — for certainty, for being viewed as knowledgeable, for avoiding shame. Yet it blinds us: truth often lives in the uncertain. To insist always on answers pushes systems — and us — towards falsehoods, rather than precision.</p><p>In Vietnam, there’s a saying, <em>“thuốc đắng giã tật, sự thật mất lòng”</em> (“bitter medicine cures disease; truth hurts”). We prefer sugar-coated stories — even if they’re lies — than the discomfort of truth. We choose comfort over correctness.</p><h4>Why Fake News Might Be More Attractive Than True News</h4><ul><li>Fake / oversimplified answers satisfy expectations instantly.</li><li>True yet uncertain answers feel like weakness; they force us to sit with ambiguity.</li><li>Systems — human or AI — optimized for positive feedback will prefer to lie or guess, rather than admit ignorance.</li></ul><h4>What We Must Do</h4><ul><li>Cultivate cognitive humility: accept that not knowing is part of being intelligent.</li><li>Demand accountability in systems: stop rewarding “confident wrong” over “uncertain maybe.”</li><li>Educate people to value the uncomfortable truths, to see uncertainty as signal not failure.</li></ul><p>We cannot demand perfect answers. What we can demand is courage: the courage to say “I don’t know,” to tolerate discomfort, to prefer truth over comfort. Because only then can real understanding — and trust — begin.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=65a99b6d2b2e" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>