<?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 Eddy Borremans on Medium]]></title>
        <description><![CDATA[Stories by Eddy Borremans on Medium]]></description>
        <link>https://medium.com/@charley.borremans?source=rss-51ab50444ff0------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*e43mq8_jl75aTfdVO8DSYw.jpeg</url>
            <title>Stories by Eddy Borremans on Medium</title>
            <link>https://medium.com/@charley.borremans?source=rss-51ab50444ff0------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Fri, 20 Nov 2026 15:39:33 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@charley.borremans/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[The Grind Is the Point]]></title>
            <link>https://medium.com/@charley.borremans/the-grind-is-the-point-def957a94872?source=rss-51ab50444ff0------2</link>
            <guid isPermaLink="false">https://medium.com/p/def957a94872</guid>
            <category><![CDATA[ai-assisted-development]]></category>
            <category><![CDATA[productivity]]></category>
            <category><![CDATA[cognitive-decline]]></category>
            <category><![CDATA[verification-debt]]></category>
            <category><![CDATA[software-engineering]]></category>
            <dc:creator><![CDATA[Eddy Borremans]]></dc:creator>
            <pubDate>Tue, 21 Apr 2026 18:47:28 GMT</pubDate>
            <atom:updated>2026-04-21T18:47:28.452Z</atom:updated>
            <content:encoded><![CDATA[<p>How AI-assisted development undermines the very skills engineering depends on</p><p>When I published “<strong>Speed Is Not Free: What AI Changes in Software Engineering</strong>” I was half-afraid it was a fringe take. I wrote it largely on intuition and experience based logic, without the theoretical vocabulary to fully defend it. Since then, the research has caught up.</p><p>The signal is still early, but it is consistent enough to take seriously. And what it shows is more alarming than I anticipated.</p><p>We are not heading toward a skill erosion problem. We are likely already in one. And most of the industry is either not looking, or actively looking away.</p><p>This article is a follow-up, but it is also part of something larger. A community is forming — researchers, engineers, educators — that is actively trying to understand what AI assistance is doing to the humans using it. Alberto Romero recently compiled over thirty studies on exactly this question (“What the Studies Say About How AI Affects Your Brain: A (Very Big) Compilation”), spanning brain imaging, randomised controlled trials, and longitudinal surveys from institutions including MIT, Wharton, Harvard, and Stanford.</p><p>His synthesis is the most thorough picture we currently have of what the science actually says. I will draw on it here, alongside other recent evidence, because the argument I made on intuition in “Speed is not Free” now has research to stand on. And what that research shows is more alarming than the intuition was.</p><p><strong>Why the grind works: what cognitive science tells us</strong></p><p>Cognitive Load Theory (CLT), developed by John Sweller in the late 1980s, tells us that effective learning is fundamentally about directing scarce mental resources toward building durable, long-term knowledge structures. Not toward feeling productive. Not toward output. Toward understanding.</p><p>Robert Bjork’s work on <strong>desirable difficulties</strong> sharpens this into something almost uncomfortable. His finding: the conditions that feel hardest during learning (spacing repetition out, wrestling with problems before seeing solutions, testing yourself rather than re-reading) produce worse short-term performance but significantly better long-term retention.</p><p>The difficulty isn’t incidental to learning. It <strong>is</strong> the mechanism.</p><p>The intuitive but wrong assumption is that fluency during practice equals learning. It doesn’t. When something feels easy, you are typically retrieving from working memory, not consolidating into long-term understanding. Remove the friction, and you don’t just slow the learning. You can actively prevent it.</p><p>Fringe? This is mainstream cognitive science. And it has a direct, uncomfortable implication for every engineering team currently optimizing for AI-assisted output.</p><p><strong>It’s already happening</strong></p><p>This is not merely prediction. The signal is already in the data.</p><p>Anthropic published a randomized controlled trial with 52 junior engineers learning Trio, an unfamiliar Python library for asynchronous programming. Both groups completed two feature-building tasks in under an hour; then took a 14-question quiz covering debugging, code reading, and conceptual understanding.</p><p>The AI-assisted group finished tasks about two minutes faster, not enough to reach statistical significance. On the quiz, they averaged 50% against 67% for those who worked without AI assistance. Nearly two letter grades, on concepts they had used minutes earlier.</p><p>A subsequent qualitative analysis showed something important. The gap was not uniform. Engineers who used AI to seek explanations alongside code retained significantly more than those who delegated generation outright. It turns out the tool itself wasn’t the variable. The way people used it was.</p><p>Let that sink in. The tool that felt like it was helping was quietly hollowing out the learning. And the developers didn’t know.</p><p>Anthropic’s own internal 2025 report on how its engineers use AI tools adds a human dimension to the numbers. Engineers inside one of the world’s leading AI companies are already reporting what one called “skill atrophy.” One described their work shifting “70%+ to being a code reviewer/reviser rather than a net-new code writer.” Others noted the loss of peer collaboration, mentorship, and the serendipitous learning that comes from code reviews and engineering discussions, the informal channels through which expertise has always propagated.</p><p>These aren’t junior developers at some AI-naive company. These are experienced engineers, working with the best tools available, raising the alarm from the inside.</p><p>Skill erosion may be closer than I thought.</p><p>The signal is not only in the lab. In mid-2025, Sasha Levin, an NVIDIA engineer and respected Linux kernel maintainer, revealed at a public conference that a patch he had submitted to kernel 6.15 had been written entirely by an LLM. He had reviewed and tested it, but had not written it, and had not said so. The code was already merged.</p><p>The reaction was not primarily about the patch itself. It was about the epistemic contract: in that review culture, a name on a patch implies understanding, not just execution.</p><p>Elsewhere, the cURL project closed its bug bounty programme entirely after maintainers were overwhelmed by AI-generated submissions they no longer had capacity to triage. The review layer is not just strained. In some places it has already buckled.</p><p>These are still early signals. But they are not isolated.</p><p><strong>The junior pipeline: a slow-motion disaster</strong></p><p>Now apply this to the “juniors are replaceable” argument that is gaining ground in boardrooms across the industry.</p><p>The argument goes: AI can write boilerplate. AI can generate endpoints. AI can do the simple stuff. Why invest in juniors?</p><p>This argument makes exactly one mistake. It misunderstands what juniors are for.</p><p>The junior developer’s task is not merely doing simple programming chores. Their equally important task is becoming a senior engineer. And that process, Bjork tells us, requires exactly the kind of uncomfortable friction that AI is now optimized to remove.</p><p>Debugging your own misunderstanding. Hitting a wall you can’t immediately route around. Sitting with ambiguity long enough to actually resolve it rather than paste over it. These aren’t inefficiencies in the junior workflow. They are the workflow, from a developmental perspective.</p><p>A junior environment optimized for output may therefore be systematically optimized against development. The junior produces more, looks more capable, the metrics look great. Short-term, everyone is happy.</p><p>And then, within a decade, you notice you have no seniors who can reason through novel problems without an AI crutch. You notice the pipeline is empty. You look back for the cause and the causal chain is long enough to be deniable.</p><p>There is a compounding layer that makes this worse. The assumption embedded in most engineering organizations is that seniors will catch what juniors miss. That assumption was already imperfect. What is new is the magnitude.</p><p>AI increases the volume of superficially plausible code entering the review layer while reviewer capacity stays flat. And the reviewers themselves are not immune to over-trusting AI output.</p><p>While the review layer never was an absolute guarantee, it is now under structural pressure from both ends simultaneously: immediate benefits, deferred damage, and diffuse responsibility.</p><p><strong>The “AI will fix it” fallacy</strong></p><p>At this point someone always says: AI will keep improving. Won’t it compensate for whatever expertise is lost?</p><p>This is a seductive argument. It is also a circular one.</p><p>AI systems are pattern completion engines trained on human-generated knowledge. They do not generate deep expertise independently. They compress and recombine existing human expertise. Their quality is tied to the quality of the human expertise that produced their training data, and that evaluates their outputs.</p><p><em>Failure one: the reservoir problem</em><br>“AI will compensate for eroding human expertise” implicitly assumes a stable reservoir of human expertise to draw from. So far, training data mostly depended on humans wrestling with hard problems and documenting the results. If that process weakens, the reservoir of knowledge degrades.</p><p><em>Failure two: the direction problem</em><br>Calculators automate arithmetic, which supports mathematical thinking. AI assists with reasoning itself. It does not eliminate thinking, but it shifts it upward toward evaluation, framing, and judgement. Those layers still require a well-developed mental model. A junior who never built that model does not just produce worse output; they cannot detect AI errors in their own domain.</p><p><em>Failure three: the frontier problem</em><br>Genuinely novel problems are underrepresented in training data. They are also the problems that matter most.</p><p>“AI will fix it” avoids confronting these dynamics.</p><p><strong>What productive friction looks like now</strong></p><p>Saying “preserve the grind” is incomplete without asking what the grind should look like now. Some of what juniors used to wrestle with: boilerplate, repetitive patterns, may not have contributed much to deep understanding. If AI removes that, something is freed up.</p><p>The question is whether we deliberately fill that space, or let it collapse into more output.</p><p>Evaluating and critiquing AI output requires deeper domain knowledge than generating it. Knowing that an answer is wrong, and why, is a more demanding task than producing a first draft. System-level and architectural thinking remains difficult to offload. Understanding how complexity accumulates over time requires experience.</p><p>Novel problem framing, i.e. defining what to solve, becomes more valuable as implementation becomes easier. Cross-domain synthesis, connecting ideas across contexts, is both underrepresented in training data and central to expertise.</p><p>These are not leftovers. They may be higher-order skills. And they require effort to develop.</p><p>The grind doesn’t have to disappear. But it won’t relocate by itself.</p><p><strong>Responsibility and the free-rider trap</strong></p><p>Naming the problem is easy. Acting on it is structurally hard, because doing the right thing conflicts directly with short-term incentives.</p><p>A company that invests seriously in junior development (one that deliberately preserves productive friction and accepts slower short-term output in service of real growth) risks training talent that walks out the door to a competitor who didn’t bother. The responsible actor gets penalized. The free-rider wins. Rational actors defect.</p><p>This is why “be responsible” alone never works as a call to action. But it is where individual agency starts.</p><p>Some structural levers that could shift the incentive landscape:</p><p>Reframe the ROI. Companies that genuinely develop engineers produce loyal seniors. Engineers who were invested in, tend to stay, or return. Senior attrition is expensive and measurable.</p><p>Take the reputation signal seriously. Being known as a place where engineers actually grow is a genuine differentiator in a tight talent market. It is also increasingly rare, which makes it increasingly valuable.</p><p>Think at sector level. Some problems cannot be solved firm by firm. Apprenticeship models, industry training frameworks, shared investment in talent pipelines: the trades understood this a long time ago. Engineering hasn’t, but the model exists.</p><p>Name the product risk. If your engineering organization cannot reason through novel problems without AI support in five years, that is not an HR problem. It is a product quality, resilience, and liability problem. That language reaches boards.</p><p>None of this fully neutralizes the free-rider problem. But the alternative — everyone defecting, the reservoir draining quietly — is a collective outcome that serves no one.</p><p><strong>What now?</strong></p><p>There is no clean solution here, and anyone offering one is probably selling something.</p><p>The organizations that navigate this best will not be the ones with the most sophisticated tooling. They will be the ones that stay alert to what the tooling is doing to the humans using it. That keep the grind visible and intentional. That treat productive friction not as inefficiency to be eliminated, but as infrastructure to be protected.</p><p>The signals are early, but consistent. The pipeline consequences are measured in years. The window is open. But not indefinitely.</p><p>Rather than a bug to be swatted, <em>difficulty</em> <strong>is</strong> the mechanism.</p><p>Speed is not free. And neither, it turns out, is ease.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=def957a94872" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Speed Is Not Free: What AI Changes in Software Engineering]]></title>
            <link>https://medium.com/@charley.borremans/speed-is-not-free-what-ai-changes-in-software-engineering-db9bace48b96?source=rss-51ab50444ff0------2</link>
            <guid isPermaLink="false">https://medium.com/p/db9bace48b96</guid>
            <category><![CDATA[software-engineering]]></category>
            <category><![CDATA[productivity]]></category>
            <category><![CDATA[engineering-capabilities]]></category>
            <category><![CDATA[ai-tooling]]></category>
            <dc:creator><![CDATA[Eddy Borremans]]></dc:creator>
            <pubDate>Sun, 01 Mar 2026 15:03:45 GMT</pubDate>
            <atom:updated>2026-03-01T15:03:45.544Z</atom:updated>
            <content:encoded><![CDATA[<p><em>Why productivity gains do not automatically translate into stronger engineering capability</em></p><p>In the past year, in my own work and in conversations with other engineers, AI-assisted development has moved from experiment to everyday practice. What started as curiosity is quickly becoming infrastructure.</p><p>Feature implementation cycles are shrinking. AI tools now handle much of the boilerplate, routine refactoring, and local logic work that used to consume significant time.</p><p>The productivity gains are real. But they introduce a trade-off that many organizations have not yet articulated clearly.</p><p>In software engineering, speed has never been free. It has always required balancing velocity against reliability and long-term maintainability.</p><p>AI does not remove that balance. It shifts it, often in ways that are less visible.</p><p>When implementation becomes cheaper, the natural temptation is to convert that gain directly into more output instead of reinvesting it in validation, review depth, and system robustness.</p><p>That is the hidden decision behind AI adoption: how deliberately we choose to allocate the gains it produces.</p><p><strong>Different Types of Engineering Work</strong></p><p>To understand where AI creates the most leverage, and where it does not, it helps to distinguish between different categories of engineering work.</p><p><strong>Isolated &amp; Bounded Work</strong> includes well-defined functionality with clear inputs and outputs: greenfield features, isolated services, small codebases, and tasks with limited coupling. In these situations, context is explicit and constraints are visible. AI performs well here.</p><p><strong>Context-Rich &amp; Coupled Work</strong> looks different. Large or legacy systems, tightly connected modules, implicit assumptions, and backward compatibility constraints create a dense context that is rarely captured fully in a prompt or a single code snippet. AI can still help, but its effectiveness becomes sensitive to hidden dependencies and historical complexity.</p><p><strong>Long-Horizon &amp; Accountability-Critical Work</strong> is where engineers exercise judgment: reviewing complex changes, debugging production issues, managing performance regressions, safeguarding compatibility, troubleshooting legacy “big balls of mud,” and thinking through how today’s decisions will affect the system months or years from now. In these areas, AI may assist, but responsibility and trade-off decisions remain human.</p><p><strong>How Technical Debt Accumulates Under Acceleration</strong></p><p>The concern is rarely that AI produces obviously broken code. In many situations, AI-generated implementations are comparable to what a competent engineer would write.</p><p>The shift happens when throughput increases.</p><p>When implementation becomes faster, more code enters the system per unit of time. Review capacity and validation discipline do not automatically expand at the same pace. As a result, review depth often decreases, sometimes gradually and without anyone explicitly deciding that it should.</p><p>Small inconsistencies, duplicated abstractions, and unexamined edge cases begin to slip through. Individually, they appear minor. Over time, they increase cognitive friction, reduce clarity, and make future changes more expensive.</p><p>This is how technical debt typically forms. It grows through incremental relaxation of discipline.</p><p>Consider two practical scenarios.</p><p>First, AI increases the volume of changes submitted for review. If review capacity remains constant, engineers face higher mental strain. Under time pressure, reviews become shorter, less critical, and more focused on surface issues rather than architectural coherence.</p><p>Second, the same acceleration creates a temptation to delegate review tasks back to AI. Automated review can catch stylistic inconsistencies and simple defects. However, architectural drift, subtle coupling issues, and questionable trade-offs are harder to detect without deliberate human scrutiny. Over time, this can amplify the very debt the tooling was meant to reduce.</p><p>At the same time, the volume of generated code can outpace the team’s ability to truly understand it. Shared understanding does not grow automatically with output. As more implementation is produced faster, fewer engineers may fully grasp how components interact or why certain decisions were made. Over time, the team’s mental map of the system becomes thinner. When incidents occur, engineers must debug both the issue itself and their incomplete understanding of the codebase. The cost compounds.</p><p><strong>Skill Formation and Exposure to Failure</strong></p><p>A related concern is how AI acceleration affects skill formation.</p><p>Recent empirical research has begun to explore this question. A large-scale study published in Science examined the global diffusion and impact of generative AI in coding contexts, highlighting both productivity gains and shifts in how developers engage with tasks (Daniotti et al., 2026). While AI increased output, productivity gains do not automatically translate into deeper capability.</p><p>This aligns with a broader engineering reality. Judgment rarely develops through smooth success. It forms through repeated exposure to failure: debugging complex production issues, tracing regressions, reviewing flawed abstractions, and experiencing the long-term consequences of earlier design decisions. These are high-feedback experiences. They are often uncomfortable, but they build durable expertise.</p><p>If AI reduces exposure to those experiences without replacing them with deliberate validation and structured learning, rather than a forecast of a deteriorating discipline, this constitutes a structural phenomenon worth monitoring.</p><p><strong>Leadership Maturity and Conscious Trade-Offs</strong></p><p>The long-term impact of AI-assisted engineering will not be determined by tooling alone. It will be shaped by how deliberately organizations handle the trade-offs that acceleration introduces.</p><p>There are contexts where aggressive AI leverage is rational. Rapid prototyping, sales-driven demos, isolated systems, and short-lived internal tools may justify prioritizing speed over long-term robustness. In such cases, consciously accepting higher technical debt can be a valid strategic decision.</p><p>Problems arise when those trade-offs are made implicitly.</p><p>If increased throughput quietly reduces review depth, if architectural discussions shrink because implementation feels easy, or if validation discipline erodes under delivery pressure, the consequences accumulate slowly. They rarely surface immediately. They appear months later as slower changes, fragile integrations, and systems that require increasing effort to reason about.</p><p>Engineering has always required balancing speed with reliability. AI increases the pressure on that balance. Teams that handle it consciously will benefit. Teams that let it drift will eventually feel the consequences.</p><p>Over the past months, I have found myself thinking less about whether AI can write code and more about what its acceleration does to engineering culture.</p><p>Although I recognize the structural shortcomings of genAI, they are not the focus of this article. Despite those inherent limitations, the tools improve quickly and, in many cases, they are genuinely impressive. What concerns me is how easily productivity gains can be mistaken for capability gains.</p><p>Engineering has always required balancing speed with reliability. AI increases the pressure on that balance. In my view, the teams that benefit most from AI will not be the ones that move fastest, but the ones that remain deliberate.</p><p>I do not see this as a forecast of decline. I see it as a structural risk worth monitoring. If we make our trade-offs explicit, preserve exposure to difficult problems, and treat validation as a first-class activity, AI can strengthen engineering rather than dilute it.</p><p>If we choose acceleration consciously, and articulate the justification when we accept shortcuts, we keep control of the direction. That, to me, is the real responsibility in this moment.</p><p><strong>References</strong></p><p>Daniotti, Simone; Wachs, Johannes; Feng, Xiangnan; Neffke, Frank (2026). Who is using AI to code? Global diffusion and impact of generative AI. Science 391(6787): 831–835. DOI: 10.1126/science.adz9311</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=db9bace48b96" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Convince Me Otherwise: Why LLMs Can’t Yet Oversee Large Real-World Codebases]]></title>
            <link>https://medium.com/@charley.borremans/convince-me-otherwise-why-llms-cant-yet-oversee-large-real-world-codebases-816bd24a7f4a?source=rss-51ab50444ff0------2</link>
            <guid isPermaLink="false">https://medium.com/p/816bd24a7f4a</guid>
            <category><![CDATA[ai-code-assistant]]></category>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Eddy Borremans]]></dc:creator>
            <pubDate>Sun, 29 Jun 2025 13:30:41 GMT</pubDate>
            <atom:updated>2025-06-29T13:30:41.427Z</atom:updated>
            <content:encoded><![CDATA[<p>Disclaimer: I wrote this article with the aid of ChatGPT. ChatGPT helped me understand how LLMs work, among many other things. This article is based on long discussions online, observations online but also really long discussions with various chat bots about all kinds of topics, of which AI is a rather large one. They help me structure my thoughts and make the content more readable for you. My interest is not getting credit for my writing style (or lack thereof), rather I always want to learn, especially about what this world is coming to in the age of AI. And I rather be honest about my approach. So, convince me otherwise.</p><p>I’ve been experimenting a lot with LLM-assisted coding. And I’ve noticed a growing divide between two camps:</p><p><strong>Group 1:</strong> “AI is already so good at coding that software engineers are on the brink of obsolescence.”<br><strong>Group 2:</strong> “LLMs are useful — sometimes amazing — but not even close to replacing us.”</p><p>I’m firmly in Group 2. Convince me otherwise.<br>But first, let me explain why I think Group 1 is mistaken — and why that mistake is easy to make.</p><h3>🧠 The Illusion of Mastery</h3><p>Ask an LLM to:</p><ul><li>Build a Tetris clone in Python</li><li>Refactor a 50-line function</li><li>Optimize a SQL query</li><li>Pass a university-level math exam</li></ul><p>And it <em>nails</em> it. Fast, fluent, often better than what most people would come up with.<br>So it’s tempting to conclude: “This thing is smarter than most developers.”</p><p>But that’s a category error.<br>LLMs are trained on enormous amounts of <strong>pattern-rich</strong> data. They’re masters of recognizing and recombining familiar structures — especially in short contexts.</p><h3>🏗️ Why This Fails for Real Codebases</h3><p>But now try asking the same LLM to:</p><ul><li>Analyze cross-module dependencies</li><li>Spot architecture-level inconsistencies</li><li>Redesign an integration strategy</li><li>Refactor an aging monolith</li></ul><p>It’ll still give you fluent, confident answers. But this time, many of them are <strong>coherent-sounding nonsense</strong>. Why?</p><p>Because the <strong>most important patterns</strong> in large codebases are:</p><ul><li><em>Project-specific</em></li><li><em>Emergent</em></li><li><em>Implicit</em></li><li><em>Distributed across hundreds of files and historical decisions</em></li></ul><p>These are <strong>not</strong> the kinds of patterns an LLM has seen during training. Nor can it hold them in memory or reason over them across sessions.</p><h3>🧩 Patterns Are the Key</h3><p>That word again: <strong>patterns</strong>.</p><p>LLMs are designed to recognize <em>common</em> ones.<br>But what makes large codebases hard is exactly what makes them <em>unique</em> — patterns of architecture, of team conventions, of past mistakes and trade-offs. Patterns that no other project has. Patterns of <strong>connection</strong>.</p><p>That’s the domain where human engineers still shine — recognizing not just patterns, but <em>patterns of patterns</em>.</p><h3>⚖️ I’m Not an AI Skeptic</h3><p>I’m a huge AI enthusiast.<br>I love what they <em>can</em> do, and use them daily.</p><p>But I’m also an AI realist.<br>And if we want to move forward responsibly — and intelligently — we need to stop extrapolating magic from one domain into another.</p><h3>🔭 Where Next?</h3><p>To those building the next generation of AI coding tools:<br>We probably need to go beyond “pattern match within a prompt window.”<br>The frontier might lie in persistent architectural memory, graph reasoning, or agents with long-term context tracking.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*un4YIdP9ovXRdb0-ZgDoCQ.png" /></figure><p>I don’t claim to know how to build that.<br>But I <em>do</em> know what it looks like when we confuse confident output for coherent reasoning. And right now, that confusion is everywhere.</p><p>Let’s raise the level of this conversation.<br>Let’s make better tools.<br>Let’s stay amazed — but not misled.</p><p><strong>Convince me otherwise.</strong></p><h3>👇 Want to understand more?</h3><p>Start here:<br>Learn how LLMs actually work — not just what they produce.<br>If I hadn’t, I’d still be in Group 1 too.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=816bd24a7f4a" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>