<?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 Grai Rudolf on Medium]]></title>
        <description><![CDATA[Stories by Grai Rudolf on Medium]]></description>
        <link>https://medium.com/@grairudolf?source=rss-0882fd8088fa------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*pNN7j7V7ccGdnyR7sIzJ2g.png</url>
            <title>Stories by Grai Rudolf on Medium</title>
            <link>https://medium.com/@grairudolf?source=rss-0882fd8088fa------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Fri, 20 Nov 2026 20:11:36 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@grairudolf/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[Do LLMs Actually Get Tired?]]></title>
            <link>https://medium.com/@grairudolf/do-llms-actually-get-tired-6e5488bc311d?source=rss-0882fd8088fa------2</link>
            <guid isPermaLink="false">https://medium.com/p/6e5488bc311d</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[ai-experiments]]></category>
            <category><![CDATA[ai-research]]></category>
            <dc:creator><![CDATA[Grai Rudolf]]></dc:creator>
            <pubDate>Wed, 30 Sep 2026 08:22:10 GMT</pubDate>
            <atom:updated>2026-09-30T08:22:10.440Z</atom:updated>
            <content:encoded><![CDATA[<h3>Do LLMs actually get tired?</h3><p><em>A self-directed research experiment into long-context performance degradation in large language models.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ScyYB57K_fgF4xUc1qPbaQ.png" /></figure><p>My most challenging recent project began with a joke.</p><p>A friend was telling me about a feature idea he had for an app he was building. It was a strange suggestion, oddly specific, a little off. I laughed and said, “Are you sure the LLM didn’t just suggest that because it was tired?”</p><p>We both laughed and moved on. Except I didn’t.</p><p>Could that actually be true? Do models get tired? Does performance degrade the longer a session runs? I had noticed it myself in long coding sessions where a model that was sharp at the start would begin giving vague answers, miss context it had already seen, or suggest things that contradicted what it said earlier. I always assumed I was imagining it, or that it was just the nature of working with AI. But what if it wasn’t random? What if it was measurable?</p><p>That question pulled me into a rabbit hole I have not fully climbed out of. 🐇</p><h3>First, I had to understand what was actually happening. Under the hood</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*yOOfMQcMd9plHLep1EEh1g.jpeg" /><figcaption>The Transformer architecture introduced in “Attention Is All You Need” (Vaswani et al., 2017). Understanding this architecture was the starting point for understanding what might happen when context gets long.</figcaption></figure><p>Before I could design any experiment, I needed to understand the mechanics. That meant reading papers. And I want to be honest about what that experience was actually like, because nobody talks about this part.</p><p>Reading ML research papers is not like reading a textbook. No one holds your hand through the notation. You open a PDF and you are immediately inside a world of assumed knowledge: subscripts on subscripts, matrix operations described in a single line, architectural decisions justified by citing three other papers you have not read yet. I would read a paragraph, understand the words individually, and have no idea what the sentence meant. Fun times.</p><p>And this is where I have to be transparent: I used Gemini Notebook to help me understand these papers. I would load a paper into it and ask questions, probe specific equations, ask it to walk me through what was happening mathematically. That became my reading workflow. Not to replace the reading, but to make it possible. The paper gives you the what. Being able to ask questions interactively helped me understand the why. There is no shame in using every tool available to you.</p><p>Here is every paper I read, what it taught me, and why it mattered for this project.</p><p><a href="http://arxiv.org/abs/1706.03762"><strong><em>Attention Is All You Need — Vaswani et al., Google (2017)</em></strong></a></p><p>This is the original transformer paper, the foundation of everything. I had heard about it for months. Actually sitting down to read it properly was a different experience entirely. The scaled dot-product attention equation looks simple until you try to trace what Q, K, and V actually represent in practice. Multi-head attention, positional encodings, residual connections. It took me multiple sessions, drawing diagrams by hand and re-reading the same sections from different angles, before it clicked.</p><p><em>What it gave me: I understood for the first time that attention is not uniform, that every token is competing for the model’s focus, and that distance between tokens is not free.</em></p><p><a href="http://arxiv.org/abs/2307.03172"><strong><em>Lost in the Middle: How Language Models Use Long Contexts — Liu et al., Stanford (2023)</em></strong></a></p><p>This paper felt closest to the question I was actually chasing. The researchers found that models perform best when relevant information appears at the very beginning or very end of a long document. Information buried in the middle gets attended to significantly less, producing a U-shaped performance curve. The phenomenon has a name now. Lost in the Middle.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*HRZ3PqwPkWqbnkPurvIEgg.jpeg" /><figcaption>The “Lost in the Middle” effect: models can perform substantially better when relevant information appears near the beginning or end of a long context than when it is buried in the middle. Source: Liu et al., 2023.</figcaption></figure><p><em>What it gave me: This was the first paper that confirmed my original observation had a name and a measurable shape, and it became the direct theoretical anchor for my experiment design.</em></p><p><a href="http://arxiv.org/abs/2108.12409"><strong><em>Train Short, Test Long: Attention with Linear Biases Enables Input Length Extrapolation — Press, Smith, and Lewis (ICLR 2022)</em></strong></a></p><p>ALiBi replaces traditional positional embeddings with a penalty proportional to the distance between tokens. The further apart two tokens are, the less attention they pay to each other by design. I found this paper easier to follow than the first two, partly because the idea is elegant and partly because my reading stamina was growing. What struck me was the name itself: Train Short, Test Long. The entire premise is that how a model handles positions it was never trained on tells you something deep about how it represents distance.</p><p><em>What it gave me: It showed me that positional encoding is not a solved problem, that different schemes make different assumptions about distance, and that those assumptions directly affect long-context behavior.</em></p><p><a href="http://arxiv.org/abs/2511.17388"><strong><em>Selective Rotary Position Embedding — Movahedi et al. (ICLR 2026)</em></strong></a><br>This one was genuinely hard. RoPE encodes position through rotation in the complex plane rather than by adding a vector. Selective RoPE makes those rotations input-dependent rather than fixed, meaning the model can learn different positional sensitivities for different inputs. Gemini Notebook was essential here. I loaded the paper and worked through the equations section by section until I could explain back what the rotation was doing and why fixed angles were a limitation.</p><p><em>What it gave me: It pushed my understanding of positional encoding from intuitive to mechanical, and it showed me that the field is still actively evolving solutions to the exact problem I was studying.</em></p><p><a href="http://arxiv.org/abs/2603.22608"><strong><em>Understanding LLM Performance Degradation in Multi-Instance Processing: The Roles of Instance Count and Context Length — Chen, Pilehvar, and Camacho-Collados, Cardiff University (ACL 2026)</em></strong></a></p><p>This was the most directly relevant paper to my original observation. The researchers evaluated how models handle being given many instances of a task at once, such as processing dozens of documents in a single prompt. They found a pattern of slight degradation at small instance counts (around 20 to 100), followed by a performance collapse at larger counts. Crucially, they found that the number of instances matters more than raw context length alone.</p><p><em>What it gave me: It reframed my question entirely. The degradation I noticed in long coding sessions was not just about length but about the number of independent things the model was being asked to track simultaneously, which made my multi-hop reasoning experiment feel even more pointed.</em></p><p>By the end of this reading process, something had shifted in how I thought about the problem. Models do not get tired the way humans do. But their ability to connect information across distance is not uniform. It is shaped by the positional encoding scheme, the attention architecture, and by how many separate things are competing for the model’s focus at once. The joke I made to my friend had a real, mechanistic answer buried in these papers. I just had to dig far enough to find it.</p><h3>Then I had to design experiments to test it myself.</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*imNkpw9k-XUmOC39LN1yKg.png" /><figcaption>The first version of my experiment running in Google Colab.</figcaption></figure><p>I wanted to go beyond reading papers. I wanted to reproduce the phenomenon, measure it, and understand its boundaries. So I set up a Python environment in Google Colab and connected to Groq’s API on the free tier to run inference on large models without needing local GPU compute.</p><p>My first approach was what researchers call a “needle in a haystack” test. I generated long documents filled with filler sentences and hid a single labeled fact inside, something like “The secret code is MANGO7742.” I placed the fact at the beginning, middle, or end of documents ranging from 20 to 500 sentences. Then I asked the model to retrieve it.</p><p>The model scored 100% across every condition.</p><p>Every position. Every context length. Perfect retrieval. I was not impressed; I was annoyed. My honest reaction was ‘dammit!!’😅</p><p>My first instinct was that I had done something wrong. But the results were clean. The problem was not my code. The problem was my task. Simple retrieval is too easy for modern frontier models. They are specifically trained to handle it. I was not testing what I thought I was testing.</p><p>So I added distractors. Fake codes scattered throughout the document alongside the real one, up to five of them. The model still got it right every time.</p><p>I had to go back to the drawing board.</p><h3>The Redesign: From Retrieval to Reasoning</h3><p>Here is where the experiment became genuinely interesting. The reason long coding sessions feel like they degrade is not because the model fails to find information. It is because the model fails to connect information. In a real session, the answer to a question rarely lives in a single sentence. It lives across multiple facts, spread over time, that need to be chained together.</p><p>So I redesigned entirely around multi-hop reasoning.</p><p>I constructed scenarios where three separate facts are scattered across a document and the model must connect all three to answer a single question. No individual sentence contains the answer. For example: Dr. Chen leads Project Falcon. Project Falcon has three members. The question is: how many people are on Dr. Chen’s team? The model must find both facts and bridge them.</p><p>I introduced three spatial conditions:</p><ul><li><strong>Clustered</strong> — all three facts sit close together</li><li><strong>Spread</strong> — the facts appear at the beginning, middle, and end</li><li><strong>Reversed</strong> — the facts appear in reverse logical order, forcing the model to reason backward through the chain</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Ptg3yf-gyJUIJ5ridygiGA.png" /><figcaption>The three conditions in my redesigned experiment: clustered, spread, and reversed. The question remains the same while the spatial relationship between the necessary facts changes.</figcaption></figure><p>My hypothesis is that accuracy degrades from clustered to spread to reversed, as the spatial distance between dependent facts increases. That degradation, if it appears, would be a cleaner analogue to what I observed in those long sessions.</p><p>Running this inside the Groq free tier came with hard constraints. I had roughly 7,000 to 8,000 input tokens per request, which meant every experiment had to be tightly scoped. Model selection was also constrained. Some models hit token limits immediately. Others had access issues on my free account. I settled on Qwen 3 8B after testing several alternatives. Not exactly a research lab setup, but you work with what you have.</p><p>I also caught a bug that had been silently corrupting my evaluation for longer than I care to admit. My answer checker was comparing strings like “3” against model responses like “Three.” Exact match was failing on every word-form answer, which meant my accuracy scores were artificially low across all conditions. Once I fixed it, I reran everything from scratch.</p><h3>Where this Is taking me</h3><p>The clean results from the multi-hop phase are still coming in. But even before they do, the process of designing this experiment has already taught me more about how attention mechanisms work in practice than any lecture I have sat through.</p><p>What I plan to do next depends on what the results show. If degradation appears, I want to scale up: longer documents, longer reasoning chains, more conditions. If it does not appear at this scale, I want to push the context further until it does.</p><p>There is also a direction I find genuinely exciting and underexplored. Almost every study on long-context degradation uses English. Very little work exists on how these phenomena behave in low-resource languages, including African languages. That gap matters, because the models being deployed across the continent are being evaluated on benchmarks built for English speakers. I want to eventually test whether the degradation patterns shift when the context language changes.</p><p>The joke about a tired model turned into a real question. The real question turned into an experiment. The experiment broke twice and taught me something both times.</p><p>That is, I think, what research actually feels like.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=6e5488bc311d" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[I tried to replicate a Google DeepMind paper.]]></title>
            <link>https://medium.com/@grairudolf/i-tried-to-replicate-a-google-deepmind-paper-328b4dece48e?source=rss-0882fd8088fa------2</link>
            <guid isPermaLink="false">https://medium.com/p/328b4dece48e</guid>
            <category><![CDATA[deep-learning]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[computer-vision]]></category>
            <category><![CDATA[ai-research]]></category>
            <dc:creator><![CDATA[Grai Rudolf]]></dc:creator>
            <pubDate>Tue, 29 Sep 2026 20:32:41 GMT</pubDate>
            <atom:updated>2026-09-29T20:32:43.324Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*kK4fhjFtJFkDGHnv8Jq3mA.jpeg" /><figcaption>D4RT reconstructs dynamic scenes across space and time from video. Source: Google DeepMind.</figcaption></figure><p>I was scrolling LinkedIn on a random afternoon when a post from Google DeepMind showed up on my feed. Normally, I keep scrolling. This time I didn’t.</p><p>It was a link to a research paper. CVPR 2026. I clicked it, started reading, and then I saw it—4D.</p><p>Not 3D. 4D. I actually paused. Space and time, reconstructed together from a regular video. Like the kind anyone shoots on their phone. That word alone made me feel like I was reading something that shouldn’t exist yet. I kept reading.</p><p>By the time I closed the tab, I had already made up my mind. I was going to try to build this.</p><p><strong>Why I Even Tried</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*3OZ0s6a-VHGq_uhONEj2hw.png" /></figure><p>Okay, some context.</p><p>I want to be an AI researcher. Not in a vague “that sounds cool” way—it&#39;s something I think about seriously. But for a long time I genuinely had no idea what research even meant in practice. I’d never written a paper. I didn’t know how to start one. And every time I tried to think of a research direction, I’d hit this wall: everything feels like it’s already been done. Every interesting problem already had five papers on it written by people at Stanford or DeepMind.</p><p>Honestly, I aspire to be an AI researcher in the near future, but without ever writing a paper, I had no clue what research really meant. How to even get started on writing one was completely unknown to me. I always felt like everything had already been done—so I decided to stop overthinking it, get my hands dirty, and work through a real paper from CVPR 2026 myself.</p><p>Replication felt like the right entry point. Not inventing something new—just proving I could take something that existed and actually understand it deeply enough to rebuild it. That felt honest. That felt like a real starting line.</p><p><strong>What the Paper Is Actually About</strong></p><p>The paper is D4RT: Dynamic Dense Depth and Radiance from Time, published at CVPR 2026 by Google DeepMind. You can check it out here.</p><p>The basic idea is this: you give it a regular video of a dynamic scene—people moving, objects shifting—and it reconstructs the full scene in 3D across time. Not just a single frozen moment. The whole thing, moving, frame by frame. That’s the 4D part. Three dimensions of space plus time.</p><p>The architecture is built on top of something called VGGT-1B, a billion-parameter vision transformer that they use as a frozen backbone encoder. Frozen meaning they don’t train it—they just use what it already knows and build everything else on top. Then there’s a query encoding module, a cross-attention transformer that fuses information across frames, a camera geometry module, and a set of loss functions that guide the whole thing during training.</p><p>That’s what I was trying to build.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/954/1*BmDnA8XshHweo5UViqH6kw.jpeg" /></figure><p><strong>What I Was Working With</strong></p><p>Here’s the part I want to be upfront about.</p><p>When I started, I knew Python. I had read Dive into Deep Learning (d2l.ai), which gave me a decent foundation in PyTorch and the basics of deep learning. That was genuinely about it. No computer vision courses. No prior experience implementing papers. No one to check my work or tell me if I was going in the right direction.</p><p>I didn’t know what a frozen backbone meant when I started. I had to look up what VGGT even stood for. Cross-attention transformers were something I had read about but never touched in code. Camera intrinsics and extrinsics—the math that describes how cameras map the 3D world onto a 2D image—were almost completely foreign to me.</p><p>What I had was internet resources, AI guidance through Claude, and the paper itself. That was my starting point.</p><p><strong>The Compute Problem</strong></p><p>The first real wall I hit was compute, and it hit fast.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/980/1*NBEYQ8ux3G-ktR2hQneiYg.jpeg" /></figure><p>VGGT-1B is a billion-parameter model. Even when it’s frozen and you’re just using it as an encoder, it eats GPU memory. I was running out of memory constantly. CUDA out-of-memory errors became such a normal part of my day that I stopped being frustrated by them and just started treating them as information.</p><p>I had to shrink batch sizes way down. I had to think carefully about what I could approximate and what I couldn’t cut corners on. Which parts of the architecture were essential and which were things I could simplify without losing the point of the whole thing?</p><p>Those questions sound annoying. But honestly, they’re what made me understand the paper. When you can’t just run everything and see what happens, you have to actually think about what each piece is doing and why it’s there. The limitations became the lesson.</p><p><strong>Building It</strong></p><p>I planned it out as a 12-week project and set up a proper GitHub repo—structured code, README, and documentation. Not for anyone else. I just wanted to treat it seriously.</p><p>The repo is here if you want to look: github.com/grairudolf/d4rt</p><p>I started with the query encoding module, then moved to the cross-attention transformer. That one took a while. Reading about attention and actually implementing it are two very different experiences. Then the camera geometry module, which is where I spent the most time—the math behind how a camera sees the world and how you recover that from data—is genuinely tricky, and I had to go deep on it. Then the loss functions, then the training data pipeline, which is just as boring as it sounds but completely necessary.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*_ML17MiMilkfu9sp-gVGtw.png" /></figure><p>Every week something new broke. Every week I figured it out and moved forward.</p><p><strong>Claude and the Internet</strong></p><p>I used Claude a lot throughout this. I want to be straightforward about that.</p><p>When something wasn’t working, and I couldn’t figure out why, I’d talk through it. When I needed to understand a concept before I could implement it, I asked. When I was staring at math that wasn’t clicking, having something that could explain it differently and trying a different angle was genuinely useful.</p><p>But Claude can’t run your experiments. It can’t make the judgment calls. It can’t decide what to simplify when you’re out of memory and need to keep moving. All of that was still on me. What it did was help me learn faster and get unstuck when I was stuck—which, at my level, starting from where I started, mattered a lot.</p><p>I don’t think anything is embarrassing about that. I think the people who figure out how to learn well with these tools are going to have a real advantage.</p><p><strong>Where I Ended Up</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*DeLAEmH5tk00jVFzUaShZQ.jpeg" /></figure><p>I got a partial replication. Not a full one—I&#39;ll be honest about that. The compute constraints were real; some things got approximated, and some things I’m still not sure I got exactly right. But here’s what I actually built:</p><p>The query encoding module<br>The cross-attention transformer<br>The camera geometry module<br>The loss functions<br>The full training data pipeline<br>A 12-week documented research plan</p><p>Starting from Python and a deep learning textbook, I’ll take that.</p><p><strong>What This Did for Me</strong></p><p>I understand now what it means to actually read a paper. Not skim it. Not just grasp the concept and move on. Read it the way you have to when you’re going to build what it describes—where every detail matters and vague understanding isn’t good enough.</p><p>I know what research feels like from the inside now. The confusion is real. The moments where something finally clicks are real. The long stretches where you’re not sure if you’re even doing it right are real.</p><p>This ended up in my CV and fellowship applications. Not because it’s impressive in the way a published paper is impressive, but because it’s real evidence that I actually do the things I say I care about.</p><p>Somewhere between week one and week twelve, I stopped thinking of myself as someone who wants to become a researcher one day.</p><p>I just started being one. Eagerly looking forward to the next thing.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=328b4dece48e" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Computer Engineering at UB]]></title>
            <link>https://medium.com/@grairudolf/computer-engineering-at-ub-1fbc39612464?source=rss-0882fd8088fa------2</link>
            <guid isPermaLink="false">https://medium.com/p/1fbc39612464</guid>
            <category><![CDATA[computer-science]]></category>
            <category><![CDATA[personal-development]]></category>
            <category><![CDATA[education]]></category>
            <category><![CDATA[college]]></category>
            <category><![CDATA[software-engineering]]></category>
            <dc:creator><![CDATA[Grai Rudolf]]></dc:creator>
            <pubDate>Mon, 28 Sep 2026 21:03:34 GMT</pubDate>
            <atom:updated>2026-09-28T21:03:34.525Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/738/1*0SbrhJsYJl58QQK3Qr4Xbw.jpeg" /></figure><p>The Faculty of Engineering and Technology (FET) at the University of Buea is one of Cameroon’s top engineering institutions. Everyone knows that. What no one tells you is what it actually costs to survive there.</p><p>I’m an undergraduate in the computer engineering department, and I think someone needs to provide an honest narrative. Not the polished version you see on LinkedIn, but the real one. So here’s mine. Grab a seat.</p><p><strong>You will learn to manage pressure, or pressure will manage you.</strong></p><p>I was at an event once and met a guy who was one of the organizers. I told him I was a student at FET. His response landed like a badge of honor: <strong>“<em>Oh, you’re the ones who manage pressure.”</em> </strong>That sentence has stayed with me. Because I knew he was right.</p><p>Engineering demands that you know your priorities clearly, manage your time ruthlessly, and make hard calls about what you’re giving up every single day.</p><blockquote><em>As the late Kobe Bryant once said,</em> “Rest at the end, not in the middle.”</blockquote><p>I didn’t fully understand that quote until I lived it at FET. I have rarely had a week where I wasn’t sacrificing one thing to show up for another. That’s just how life is here.</p><p>Some weeks felt like war. My friends jokingly often said, <em>“Let them just kill us at once,”</em> and we laughed it off, but the exhaustion behind that joke was very real. Projects, presentations, research, bulky assignments, submissions, exams, and meetings. Some days, you show up to campus purely because of a team meeting. Not for class. Just because <strong><em>the project has to work.</em></strong></p><p>If you are the kind of person who needs everything to be comfortable before you perform, FET will be a tough teacher. But if you can learn to work under fire, you’ll come out the other side sharper than you went in.</p><p><strong>I refused to be ordinary, and that cost me something.</strong></p><p>Most students wait until their fourth year to think about internships. From my very first year, I had already set a goal: three internships before I graduated. Not because someone told me to. Because I looked at the standard path and decided it wasn’t fast enough for me.</p><p>That decision was strenuous. It was a risk. But I took it anyway.</p><p>On top of coursework, assignments, and exams, I was showing up to internship meetings, delivering tasks for people outside of school, and trying to hold everything together at once. I slept less. I became less free. There were stretches where freedom felt like a luxury I had voluntarily surrendered. But mediocrity felt worse. I was young, I was ambitious, and I wasn’t willing to wait around to become who I wanted to be.</p><blockquote><em>Warren Buffett has a line that stuck with me:</em> “Someone’s sitting in the shade today because someone planted a tree a long time ago.”</blockquote><p>I wanted to plant early. The earlier you plant, the more shade you have later.</p><p>The advice I’d give anyone reading this: <strong>go as hard as you can, as early as you can.</strong> The cost is real, but so is the distance it puts between you and where you started.</p><p><strong>You will learn what teamwork actually means.</strong></p><p>Different courses came with different group projects and different groups of people. One of my teams scored a low mark because not everyone carried their weight. That hurt. But the lesson it taught me was sharper than any grade.</p><p>Leadership is not about doing the work yourself. It’s not about forcing people to do it either. It’s about making people <em>want </em>to do it. John Maxwell puts it well:</p><blockquote><strong><em>“A leader knows the way, goes the way, and shows the way.”</em></strong></blockquote><p><strong><em>Working</em></strong> effectively with people is like solving a jigsaw puzzle. You need the right piece in the right place. The beautiful thing about people is that we have the capacity to become the right piece for one another.</p><p>For me, the key was finding common ground. I didn’t isolate myself from people; I created spaces where I could connect with them naturally around shared interests. That changed everything.</p><p>That low-scoring presentation also hit differently. Feeling like my fate was tied to other people’s commitment shook me. My first instinct was to protect myself by doing everything alone. But I sat with that feeling honestly and realized it leads nowhere. No man is an island. Not in life, and certainly not in engineering.</p><p>The philosophy I eventually landed on was simpler and stronger: <em>as long as my name is on it, it cannot fail. </em>That shift made me stop thinking of projects as things I was doing <em>among</em> others and start thinking of them as wholes I was responsible for. That perspective is a big part of who I am today.</p><p><strong>You will learn how to find the point of what you’re studying.</strong></p><p>Not every lecture at FET left me inspired. I’ll leave it at that (you know how it is). But here’s unsolicited advice for anyone in the same boat.</p><p>The curriculum is actually well-designed and important.</p><blockquote>The problem isn’t usually <em>what </em>you’re studying. It’s how you’re studying it. Richard Feynman.</blockquote><p>One of the greatest physicists of the 20th century had a learning technique built on one simple principle: if you can’t explain something simply, you don’t understand it yet. That rule changed how I approached every course.</p><p>Look for practical applications of what you learn. Build a roadmap. Connect the dots. Ask yourself where this topic lives in the career you’re building.</p><p>Let me give an example in my context: In my sophomore year, I took a security course. At first it felt abstract, unmoored from anything real. Then I went looking. I found out exactly where and how security operates in the application layer of the TCP/IP model and how I could actually work with it. Suddenly the course had a home. It made sense. That shift in perspective made studying it feel completely different.</p><p>The material isn’t the enemy. Passive learning is.</p><p><strong>Nobody sees what goes on under the hood.</strong></p><p>People, my friends, my coursemates, they think I have it all figured out. I can land internships, I show up, and I keep moving. So they ask me for money sometimes, and more often, they ask me for a roadmap. <em>How do you do it? What should I follow?</em></p><p>But here’s the honest answer: I don’t think we’re walking the same road.</p><p>Most of them want to be software engineers, network engineers, or telecom engineers. Building dashboards, deploying systems, and solving local infrastructure problems. That’s real, important work. But it’s not mine. I’m tired of that lane. My goal is not to solve problems within my community’s current reach. I want to contribute to the global research frontiers of the world. I want to create the biggest impact I possibly can. That’s a different ambition, and it demands a different path. Perhaps become the next Geoffrey Hinton, Fei-Fei Li, Andrej Karpathy, etc.; the list goes on.. Not just because I want to, but because I want to.</p><blockquote>I’m tired of sailing my little boat</blockquote><blockquote>Far inside of the harbor bar;</blockquote><blockquote>I want to be out where the big ships float —</blockquote><blockquote>Out on the deep where the Great Ones are!</blockquote><blockquote>And should my frail craft prove too slight</blockquote><blockquote>For storms that sweep those wide seas o’er,</blockquote><blockquote>Better go down in the stirring fight</blockquote><blockquote>Than drowse to death by the sheltered shore!</blockquote><blockquote>— Daisy Rinehart</blockquote><p>Some people have even told me that’s not what Cameroon needs right now. I understand why they say it. But here’s something I’ve observed: brilliant solutions exist everywhere around us. Genuinely great ideas, great products. And yet people don’t use them. Innovation that should be working just isn’t landing. Why?</p><p>Because building something useful and convincing people they need it are two entirely different skills. As Robert Cialdini writes in <em>Influence</em>, people don’t just act on what is good for them. They act on what <em>feels</em> right to them. You have to build from the perspective of what people will reach for instinctively, not just what solves their problem on paper. Make them feel like they cannot exist without what you’ve built. That’s the gap between a great idea and a product that actually moves. A lesson I’m still learning.</p><p>As for the roadmap requests, I can’t give anyone a blueprint to follow. But I can tell you what I think you should do.</p><p><strong>Stop looking for a map. Start identifying a goal.</strong></p><p>Not a vague wish. A real, specific ambition sitting somewhere in the distant future. Then connect it backwards to today. What does that future require? What does the person who gets there look like, and what are they doing right now?</p><p>Then, and this is the part most people skip, <strong>develop a creed around that goal. </strong>Not just a to-do list, not a vision board. A creed. A set of convictions you return to when things get hard, when people question your path, and when you feel like you’re falling behind. Your creed is what keeps you oriented when the road disappears. Mine has kept me going through nights I don’t write about publicly.</p><p><strong>The goal I have isn’t just an <em>ambition</em> anymore. It has become an <em>apprehension</em></strong>, something I can almost reach out and touch. That shift changes how you operate. You stop asking what’s possible and start asking what’s required.</p><blockquote>I can’t complain about having too much on my plate. My goal was to eat.</blockquote><p><em>This is still a work in progress, much like the person writing it. More to come.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1fbc39612464" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Building Neural Radiance Fields from Scratch]]></title>
            <link>https://medium.com/@grairudolf/building-neural-radiance-fields-from-scratch-54c13d50666a?source=rss-0882fd8088fa------2</link>
            <guid isPermaLink="false">https://medium.com/p/54c13d50666a</guid>
            <category><![CDATA[deep-learning]]></category>
            <category><![CDATA[computer-vision]]></category>
            <category><![CDATA[python]]></category>
            <category><![CDATA[artificalintelligence]]></category>
            <category><![CDATA[machine-learning]]></category>
            <dc:creator><![CDATA[Grai Rudolf]]></dc:creator>
            <pubDate>Mon, 28 Sep 2026 20:40:06 GMT</pubDate>
            <atom:updated>2026-09-28T20:41:01.490Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*wm8QbaWrL_lljoiJDKth8w.jpeg" /></figure><p>A few months before January 2026, I took up the challenge of building a ray tracer. Just me, a blank file, and the stubborn decision to make light behave correctly on a screen.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/960/1*5sb4RJ1cJKSK2rGiM8KDdg.png" /><figcaption>A ray-traced scene. Building a ray tracer was my first step toward understanding how images can emerge from geometry, light, and mathematics.</figcaption></figure><p>It turned out to be a great exercise. My Python and Geometry got sharper, I dusted off geometry I had not touched since high school, and I got a genuine feel for how light bounces, refracts, and falls off across a scene.</p><p>But something about it nagged at me. A ray tracer is <strong>explicit.</strong> Every object, every material, every surface property- you define it yourself. The scene does exactly what you tell it to, nothing more. There is no learning, no inference, no “figuring it out.” It is a calculator with very good taste in physics.</p><p>I wanted something that could <strong>learn</strong> the scene instead of being told about it.</p><p>Then I found a 2020 paper called “<strong><em>NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis.” </em></strong>And it rearranged something in my head. The idea was this: feed a neural network a handful of photographs of an object from different angles, and it learns the object well enough to generate a photo from an angle nobody ever captured.</p><p><a href="https://github.com/bmild/nerf.git">https://github.com/bmild/nerf</a></p><p>No 3D model. No mesh. No manually defined surfaces. Just pixels in, geometry out, as a side effect of learning to predict color.</p><p>Computer vision had always been my favorite place for deep learning to live, and this felt like the edge of it. I had no course for this, no mentor, no structured path. Just the paper, dense with equations, a lot of broken training runs at 1 am, and an LLM to argue with when I got stuck. I decided to build the entire thing from scratch anyway.</p><p><strong>So What Did I Actually Build?</strong></p><p><em>(Not a technical person? Skip to “What Came Out The Other End” below; you will not miss the plot.)</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Ki-E_t0gleU9j7qz_pHoyA.jpeg" /></figure><p>At a high level, here’s the pipeline I implemented, piece by piece, in PyTorch:</p><p><strong>1. Positional Encoding</strong></p><p>Here’s a fun problem: if you feed a neural network raw (x, y, z) coordinates, it produces blurry, low-detail nonsense. Networks are naturally lazy about high-frequency detail. The fix is to “encode” each coordinate using a series of sine and cosine waves at increasing frequencies, basically translating a plain number into a richer signal the network can actually grab onto. A single 3D point becomes a 63-dimensional vector. It sounds excessive until you try training without it and watch your beautiful object become a gray smudge.</p><p><strong>2. The Network Itself</strong></p><p>An 8-layer MLP (a fairly standard feedforward neural net) takes that encoded position and predicts two things: how dense the “stuff” is at that point in space (is there an object here, or empty air?), and what color it is. Partway through the network, I fed the position encoding back in again as a “skip connection,” basically reminding the network where it is in space, because deep networks tend to forget.</p><p>Viewing direction gets added in near the end, separately, because color can change depending on where you’re looking from (think of how a glossy surface looks different at different angles), but the actual geometry of the object should not change just because you’re looking at it from somewhere else.</p><p><strong>3. Shooting Rays Through the Scene</strong></p><p>For every pixel in an image, imagine a ray shooting out from the camera, through that pixel, into the 3D scene. Along that ray, I sample a bunch of points and ask the network: “what’s here?” First pass, I sample evenly (called stratified sampling). Second pass, I sample more densely wherever the first pass found “stuff” (called hierarchical sampling), so compute gets spent where it matters instead of wasted on empty air.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Fkhi30FXN8xiP6bltaw8dQ.png" /></figure><p><strong>4. Turning Samples Into a Picture</strong></p><p>This is the part that genuinely impressed me. Once you have color and density at a bunch of points along a ray, you combine them using something called volume rendering, basically a weighted sum where points that are “blocked” by denser stuff in front of them contribute less. Do this for every pixel, and you get a full image. The math is the same idea behind how fog or smoke gets rendered in movies, just applied point by point.</p><p><strong>5. Training</strong></p><p>The network never gets told “this is a chair” or “this is a wall.” It only gets shown: “from this camera angle, this pixel should be this color.” That’s it. From thousands of these tiny corrections, the network slowly figures out the entire 3D structure of the object on its own. No 3D supervision, ever.</p><p>I also wrote a full test suite (7 tests) to check that every piece worked: the encoding, the network outputs, the ray math, the sampling, the rendering, and crucially, that gradients flowed correctly through the *entire* pipeline end to end. That last one matters more than it sounds, because if any single step quietly breaks the gradient chain, your network trains on nothing and gives no error message. It just doesn’t learn, and you spend three hours wondering why.</p><p><strong>What Came Out The Other End</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*mURKdfBRMKlA997Bln9Y3g.png" /><figcaption>Output from my portable JBL Clip 3 speaker at home</figcaption></figure><p>After a lot of debugging, broken renders, and staring at gray blobs wondering what I did wrong, it worked. I had a model that could take a handful of photos of a synthetic scene and generate new, convincing views from angles it had never seen. Watching that first coherent render appear, after weeks of equations and failed runs, is genuinely one of the more satisfying moments I have had with a project.</p><p>A few things stuck with me from the process:</p><p>- <strong>Architecture details in papers are not suggestions. </strong>The positional encoding step felt almost decorative until I removed it and watched everything fall apart. Every weird-looking design choice in that paper turned out to be load-bearing.<br>- <strong>Silent failures are the worst failures.</strong> PyTorch will let broken gradient flow pass without complaint. The network just quietly fails to learn. Writing tests for differentiability specifically saved me from a lot of confused late nights.<br>- <strong>Geometry as an emergent property is wild.</strong> The network was never told anything about 3D shape, yet visualizing its internal depth estimates produced a clean, recognizable reconstruction of the object’s geometry. It learned shape as a side effect of learning to match pixel colors. That still feels a little bit like magic, in the best way.</p><p><strong>Where I Want To Take This</strong></p><p>The obvious next step is performance. NeRF in its original form is famously slow (we’re talking around 24 hours of training on a high-end GPU for one scene), and newer techniques like Instant-NGP and Gaussian Splatting get similar results in minutes. Adapting those ideas is on my list.</p><p>But the direction that genuinely excites me is medical imaging. If a neural network can reconstruct accurate 3D structure purely from 2D images taken at different angles, the same underlying idea applied to medical scans raises a real question: could a NeRF-based approach help reconstruct organ or tissue structure from limited-angle imaging and flag structural abnormalities that are hard to catch from individual 2D slices? Could it give medical staff a faster, more reliable spatial picture than current tools allow, especially in settings where high-end imaging equipment is scarce?</p><p>That’s the direction I’m pulling this toward now: not just a clean implementation of a 2020 paper, but a starting point for a research contribution with my own additions, aimed at something that could matter outside of a GitHub repo. Good research only really earns the name once it’s useful to someone. I’d like this to get there.</p><p>GitHub Repo: <a href="https://github.com/grairudolf/nerf">https://github.com/grairudolf/nerf</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=54c13d50666a" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Dare to Be Different: Not for Attention, But for Growth]]></title>
            <link>https://medium.com/@grairudolf/dare-to-be-different-not-for-attention-but-for-growth-7c94495c6a34?source=rss-0882fd8088fa------2</link>
            <guid isPermaLink="false">https://medium.com/p/7c94495c6a34</guid>
            <category><![CDATA[growth]]></category>
            <category><![CDATA[habit-building]]></category>
            <category><![CDATA[future]]></category>
            <category><![CDATA[tech]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Grai Rudolf]]></dc:creator>
            <pubDate>Mon, 28 Sep 2026 20:06:28 GMT</pubDate>
            <atom:updated>2026-09-28T20:06:28.670Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/600/1*WEEW18q8f1FYYRF3i_la6w.jpeg" /></figure><p>I’ll be honest with you: I’m not the type who naturally loves hard things. I don’t wake up excited to face challenges, and I definitely don’t chase discomfort for fun. If I had it my way, I’d avoid the difficult route most of the time.</p><p>But even with that, I’ve never been someone who follows the norm blindly. Whenever everyone accepts something as “the usual way,” something in me questions it. Why is this the norm? Why can’t we try something else? It’s not rebellion. It’s not an attempt to stand out. It’s simply my instinct to look deeper, because every time I step outside the predictable, I grow.</p><p>I’ve realized that I don’t chase difficulty—I chase growth. I’m tired of being limited by comfort. Playing safe looks attractive, but I know what it leads to: stagnation. And honestly, staying the same scares me more than any challenge ever will.</p><p>While many of my coursemates focus mainly on studying, I often find myself building apps, systems, and small tools. Not because I think I’m smarter or trying to impress anyone, but because my priorities are different. I want what I learn to work in the real world. I want to test it, break it, fix it, and rebuild it. Books alone can’t give me that experience. Projects can.</p><p>That same mindset shows up in other areas of my life. When I teach students, when I apply for programs I’m uncertain about, even when I preach on the streets, I choose the path that stretches me. Not because I enjoy pressure, but because that pressure shapes me. Being different isn’t my personality; it’s my process. It’s how I avoid becoming a spectator in my own life.</p><p>I’m not extraordinary or fearless. I’m not naturally drawn to hard things. Honestly, I thought I was, but it&#39;s really not like that. I’ve learned something important: you don’t need to love the hard path to choose it. You just need to understand what it does to you. And the truth is simple—it transforms you.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/800/1*pBdf8NdTA3HBHZO90ARbuA.jpeg" /></figure><p>As humans, God wired us with almost unlimited potential. He didn’t create us to blend into the background or hide in comfort zones. He blessed us with gifts, strength, creativity, and a mind capable of far more than we often allow ourselves to attempt. If we don’t dare to be different, how will we ever discover what He placed inside us?</p><p>So don’t choose the uncommon path to look special or gain attention. Choose it because it will change you. Choose it because it forces you to grow. Choose it because it unlocks what God has already invested in you.</p><p>Most people run from challenges. Be the one who uses them—not because you enjoy difficulty, but because you value who you become through it.</p><p>Dare to be different. Not for the spotlight, but to make the most of what God has put in you.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=7c94495c6a34" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Why I Started Building]]></title>
            <link>https://medium.com/@grairudolf/why-i-started-building-f542ec739571?source=rss-0882fd8088fa------2</link>
            <guid isPermaLink="false">https://medium.com/p/f542ec739571</guid>
            <category><![CDATA[computer-engineering]]></category>
            <category><![CDATA[cameroon]]></category>
            <category><![CDATA[purpose]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[growth]]></category>
            <dc:creator><![CDATA[Grai Rudolf]]></dc:creator>
            <pubDate>Mon, 28 Sep 2026 19:53:22 GMT</pubDate>
            <atom:updated>2026-09-28T19:53:22.603Z</atom:updated>
            <content:encoded><![CDATA[<p>Back in Form Three (9th grade), I was just a kid fascinated by games, apps, and software. But the big question that kept disturbing me was simple: How do people even create these things?</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*iRvXimslvudRe2pau5gc_Q.jpeg" /></figure><p>One childhood friend mentioned something to me — “programming.” He told me games were programmed. That single word flipped a switch in me. I became curious.</p><p>That curiosity led to what I jokingly call my first internship.</p><p>During the summer holidays of 2019, I got the chance to work as an apprentice phone repairer at a small technician’s shop. I spent months there without owning a phone or a laptop, just learning by watching and touching real devices. By the end of that apprenticeship, I could open up smart devices, troubleshoot problems, and understand how hardware worked.</p><p>But programming? Still a mystery.</p><p>Fast-forward to Form Four: still stubbornly curious, I grabbed my dad’s phone and searched “What is programming?” online.</p><p>Boom — JavaScript.</p><p>I had no roadmap, no mentor, nothing. Just vibes and a long YouTube tutorial. I learned all the concepts blindly, not even understanding what JavaScript was used for. When it came time to build something, I was completely lost.</p><p>After more research, the next year I picked up HTML and CSS. Suddenly, things clicked. I could build basic websites and finally apply the JavaScript I’d learned. That feeling of power — the feeling that I can create something — was addictive.</p><p>By Lower Sixth, I picked up Python and felt unstoppable. My stack was growing, and so was my confidence.</p><p>But everything changed in Upper Sixth.</p><p>I was invited by a classmate to join his team for the TIC Summit. Three of us walked in full of confidence, thinking we’d win. We didn’t. But that loss became one of my biggest turning points.</p><p>I didn’t just learn to code — I learned how to turn code into solutions, how to design ideas that solve real problems, how to pitch, and how to make someone believe in what you’re building. That experience stretched me, shaped me, and humbled me. Even now, I’m ready to talk about any of my projects in front of anyone because of that life-shaping experience.</p><p>I also credit Knowledge Center for planting in me their “Dare to Be Different” ideology. It pushed me to aim higher. By my final year of high school, I became bold enough to apply for jobs and internships abroad — many rejections, yes, but each one made me sharper.</p><p>Eventually, by the grace of God, I landed some internships.</p><p>I met great people. I discovered opportunities I never imagined. I even got to share God’s Word and serve Him through the skills He gave me.</p><p>Today, I don’t only turn ideas into code; I contribute to building a future-ready nexus of people, and, paramount of all, I am serving God in the domain of his call upon my life.</p><p>So here’s the truth:</p><p>Why do I build now? Because I want to make the most of what God has given me. Because building is how I grow, how I serve, and how I give back.</p><p>And honestly, I’m quite ambitious — for the sake of human attention span, I’ll end by saying this: my journey is only just beginning.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f542ec739571" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The One Thing That Truly Defines You]]></title>
            <link>https://medium.com/@grairudolf/the-one-thing-that-truly-defines-you-bd28c74ee352?source=rss-0882fd8088fa------2</link>
            <guid isPermaLink="false">https://medium.com/p/bd28c74ee352</guid>
            <category><![CDATA[faith-in-god]]></category>
            <category><![CDATA[purpose]]></category>
            <category><![CDATA[salvation]]></category>
            <category><![CDATA[living-with-purpose]]></category>
            <category><![CDATA[faith]]></category>
            <dc:creator><![CDATA[Grai Rudolf]]></dc:creator>
            <pubDate>Mon, 28 Sep 2026 19:46:02 GMT</pubDate>
            <atom:updated>2026-09-28T19:46:02.641Z</atom:updated>
            <content:encoded><![CDATA[<p>Life, as we all know, is not easy; yes, very cliché of me. But yeah, life tests us, shapes us, and sometimes breaks us. But it becomes even more difficult during life’s transitions — those in-between moments where one chapter closes, and the next hasn’t yet begun.</p><p>Some people believe that life is just a simulation, a system running without a real purpose. But I’ve learned through experience that life isn’t random at all. There’s a thread that ties everything together — and that thread is purpose.</p><p>Growing up, I’ve faced my share of challenges, setbacks, and failures. Many of them came from trying to always be five steps ahead — to stay sharp, ready, and ahead of everyone around me. Some people saw that as natural talent. But it wasn’t. It was my awareness of the kind of person I wanted to become, and more importantly, the kind of person I wanted to be for God.</p><p><strong>What Defines a Man</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*LJN3Ckc-HTIAtB6ynUYfgg.jpeg" /></figure><p>If you strip everything else away — achievements, status, wealth, and recognition — what remains? What defines a man? It’s not talent. It’s not intellect. It’s purpose.</p><blockquote>“For everything God created has a purpose.” — 1 Timothy 4:4 (NIV)</blockquote><p>When something functions according to its purpose, it thrives. The same is true for people. Purpose is the quiet fire that gives direction in chaos. It’s what separates those who wander from those who walk with intention.</p><p>I don’t know what battles you’re facing right now — I have mine too — but the difference between a life that drifts and a life that transforms is purpose. And the truest, deepest kind of purpose begins and ends with God.</p><p><strong>The Model of Purpose: Jesus</strong></p><p>Our perfect example is the Lord Jesus Christ. His entire life on earth was driven by one mission — to fulfill the will of His Father.</p><blockquote>“My food,” said Jesus, “is to do the will of Him who sent me and to finish His work.” — John 4:34 (NIV)</blockquote><p>That’s purpose in its purest form — living not for yourself, but for God’s intention through you.</p><p>I mean, think about it: Jesus Christ, the Son of God, said His purpose was to do the will of He who sent Him, implying that it might not have been His will, but still he forsook Himself to do the will of His Father.</p><p>Before you chase after success, status, or recognition, seek first to do the will of your Heavenly Father.</p><p>Maybe you don’t yet know how. That’s okay. God is patient and kind, and He has a plan for you too. Wherever you are right now, pause — bow your head and speak honestly to Him.</p><p><em>“For all have sinned and fall short of the glory of God.” — Romans 3:23 (NIV)</em></p><p>Sin distances us from the One who created us. It hurts not only us, but the God who loves us beyond measure. Yet in His mercy, He gave us a way back through Jesus Christ.</p><p>Right now, you can make the most important decision of your life — to surrender completely. Confess your sins to God. Tell Him everything. Be honest. Decide to turn away from sin and invite Jesus into your heart.</p><p>If you’ve done that, congratulations — this is the beginning of your new life in Christ. Mark this moment. Cherish it. It’s the day your true purpose begins.</p><p><strong>Walking in Purpose</strong></p><p>On a normal basis, the primary purpose of all creation is to glorify God and this is being made abundantly clear throughout the Bible. If you see your existence, your way of life, your actions portrayed in a manner that does not glorify God, repent immediately! Seek God and live for His exclusive glory.</p><p>When you give your life to Christ, you inherit a divine purpose — the same that all of God’s children share:</p><blockquote>“Therefore go and make disciples of all nations, baptizing them in the name of the Father and of the Son and of the Holy Spirit.” — Matthew 28:19 (NIV)</blockquote><p>But within that great mission, God also gives each of us a unique calling — something special that only you can do. It’s woven into your gifts, passions, and the burdens He places on your heart.</p><p>To uncover it, ask yourself three simple but powerful questions:</p><ul><li>Who am I?</li><li>What do I want to become?</li><li>How will I become that?</li></ul><p>Clarity gives you direction. But direction alone is not enough. You need discipline.</p><p>Purpose without discipline is just a dream — a spark that never becomes a flame.</p><p>Set goals. Think long-term, but start short-term. Ask yourself, “Where do I want to be in two months? In two years? In five?” Then take small, faithful steps every day.</p><p>Because purpose isn’t a single discovery — it’s a daily decision to keep walking with God, no matter how hard the road gets.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*SmXqY5YgxqNckS8E0a6wYA.jpeg" /></figure><p><strong>The One Thing That Truly Defines You</strong></p><p>Life will never be without challenges. There will be seasons of confusion, doubt, and pain. But when your life is anchored in purpose — especially one rooted in God — even your struggles gain meaning.</p><p>You start to realize that every delay was preparation, every disappointment was redirection, and every closed door was grace in disguise.</p><blockquote>“And we know that in all things God works for the good of those who love Him, who have been called according to His purpose.” — Romans 8:28 (NIV)</blockquote><p>Purpose doesn’t make life easier. It makes life meaningful.</p><p>So, if you remember one thing from this: Let purpose define you — not fear, not failure, not the past, but your God-given reason for being.</p><p>Live intentionally. Live purposefully. Live for God.</p><p>Life will never be without challenges. There will be seasons of confusion, doubt, and pain. But when your life is anchored in purpose — especially one rooted in God — even your struggles gain meaning.</p><p>You start to realize that every delay was preparation, every disappointment was redirection, and every closed door was grace in disguise.</p><blockquote>“And we know that in all things God works for the good of those who love Him, who have been called according to His purpose.” — Romans 8:28 (NIV)</blockquote><p>Purpose doesn’t make life easier. It makes life meaningful.</p><p>So, if you remember one thing from this: Let purpose define you — not fear, not failure, not the past, but your God-given reason for being.</p><p>Live intentionally. Live purposefully. Live for God.</p><p>God bless you!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=bd28c74ee352" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>